xref: /openbmc/qemu/tcg/tcg-op-gvec.c (revision d525f73f)
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-op.h"
23 #include "tcg/tcg-op-gvec.h"
24 #include "qemu/main-loop.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_new_ptr();
121     a1 = tcg_temp_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_new_ptr();
142     a1 = tcg_temp_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_new_ptr();
162     a1 = tcg_temp_new_ptr();
163     a2 = tcg_temp_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_new_ptr();
185     a1 = tcg_temp_new_ptr();
186     a2 = tcg_temp_new_ptr();
187     a3 = tcg_temp_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_new_ptr();
211     a1 = tcg_temp_new_ptr();
212     a2 = tcg_temp_new_ptr();
213     a3 = tcg_temp_new_ptr();
214     a4 = tcg_temp_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_new_ptr();
241     a1 = tcg_temp_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_new_ptr();
262     a1 = tcg_temp_new_ptr();
263     a2 = tcg_temp_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_new_ptr();
287     a1 = tcg_temp_new_ptr();
288     a2 = tcg_temp_new_ptr();
289     a3 = tcg_temp_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_new_ptr();
315     a1 = tcg_temp_new_ptr();
316     a2 = tcg_temp_new_ptr();
317     a3 = tcg_temp_new_ptr();
318     a4 = tcg_temp_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_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_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_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_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_const_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_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_size);
650         tcg_temp_free_ptr(t_ptr);
651         return;
652     }
653 
654     t_desc = tcg_constant_i32(simd_desc(oprsz, maxsz, 0));
655 
656     if (vece == MO_64) {
657         if (in_64) {
658             gen_helper_gvec_dup64(t_ptr, t_desc, in_64);
659         } else {
660             t_64 = tcg_constant_i64(in_c);
661             gen_helper_gvec_dup64(t_ptr, t_desc, t_64);
662         }
663     } else {
664         typedef void dup_fn(TCGv_ptr, TCGv_i32, TCGv_i32);
665         static dup_fn * const fns[3] = {
666             gen_helper_gvec_dup8,
667             gen_helper_gvec_dup16,
668             gen_helper_gvec_dup32
669         };
670 
671         if (in_32) {
672             fns[vece](t_ptr, t_desc, in_32);
673         } else if (in_64) {
674             t_32 = tcg_temp_new_i32();
675             tcg_gen_extrl_i64_i32(t_32, in_64);
676             fns[vece](t_ptr, t_desc, t_32);
677             tcg_temp_free_i32(t_32);
678         } else {
679             if (vece == MO_8) {
680                 in_c &= 0xff;
681             } else if (vece == MO_16) {
682                 in_c &= 0xffff;
683             }
684             t_32 = tcg_constant_i32(in_c);
685             fns[vece](t_ptr, t_desc, t_32);
686         }
687     }
688 
689     tcg_temp_free_ptr(t_ptr);
690     return;
691 
692  done:
693     if (oprsz < maxsz) {
694         expand_clr(dofs + oprsz, maxsz - oprsz);
695     }
696 }
697 
698 /* Likewise, but with zero.  */
699 static void expand_clr(uint32_t dofs, uint32_t maxsz)
700 {
701     do_dup(MO_8, dofs, maxsz, maxsz, NULL, NULL, 0);
702 }
703 
704 /* Expand OPSZ bytes worth of two-operand operations using i32 elements.  */
705 static void expand_2_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
706                          bool load_dest, void (*fni)(TCGv_i32, TCGv_i32))
707 {
708     TCGv_i32 t0 = tcg_temp_new_i32();
709     TCGv_i32 t1 = tcg_temp_new_i32();
710     uint32_t i;
711 
712     for (i = 0; i < oprsz; i += 4) {
713         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
714         if (load_dest) {
715             tcg_gen_ld_i32(t1, cpu_env, dofs + i);
716         }
717         fni(t1, t0);
718         tcg_gen_st_i32(t1, cpu_env, dofs + i);
719     }
720     tcg_temp_free_i32(t0);
721     tcg_temp_free_i32(t1);
722 }
723 
724 static void expand_2i_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
725                           int32_t c, bool load_dest,
726                           void (*fni)(TCGv_i32, TCGv_i32, int32_t))
727 {
728     TCGv_i32 t0 = tcg_temp_new_i32();
729     TCGv_i32 t1 = tcg_temp_new_i32();
730     uint32_t i;
731 
732     for (i = 0; i < oprsz; i += 4) {
733         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
734         if (load_dest) {
735             tcg_gen_ld_i32(t1, cpu_env, dofs + i);
736         }
737         fni(t1, t0, c);
738         tcg_gen_st_i32(t1, cpu_env, dofs + i);
739     }
740     tcg_temp_free_i32(t0);
741     tcg_temp_free_i32(t1);
742 }
743 
744 static void expand_2s_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
745                           TCGv_i32 c, bool scalar_first,
746                           void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32))
747 {
748     TCGv_i32 t0 = tcg_temp_new_i32();
749     TCGv_i32 t1 = tcg_temp_new_i32();
750     uint32_t i;
751 
752     for (i = 0; i < oprsz; i += 4) {
753         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
754         if (scalar_first) {
755             fni(t1, c, t0);
756         } else {
757             fni(t1, t0, c);
758         }
759         tcg_gen_st_i32(t1, cpu_env, dofs + i);
760     }
761     tcg_temp_free_i32(t0);
762     tcg_temp_free_i32(t1);
763 }
764 
765 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
766 static void expand_3_i32(uint32_t dofs, uint32_t aofs,
767                          uint32_t bofs, uint32_t oprsz, bool load_dest,
768                          void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32))
769 {
770     TCGv_i32 t0 = tcg_temp_new_i32();
771     TCGv_i32 t1 = tcg_temp_new_i32();
772     TCGv_i32 t2 = tcg_temp_new_i32();
773     uint32_t i;
774 
775     for (i = 0; i < oprsz; i += 4) {
776         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
777         tcg_gen_ld_i32(t1, cpu_env, bofs + i);
778         if (load_dest) {
779             tcg_gen_ld_i32(t2, cpu_env, dofs + i);
780         }
781         fni(t2, t0, t1);
782         tcg_gen_st_i32(t2, cpu_env, dofs + i);
783     }
784     tcg_temp_free_i32(t2);
785     tcg_temp_free_i32(t1);
786     tcg_temp_free_i32(t0);
787 }
788 
789 static void expand_3i_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
790                           uint32_t oprsz, int32_t c, bool load_dest,
791                           void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32, int32_t))
792 {
793     TCGv_i32 t0 = tcg_temp_new_i32();
794     TCGv_i32 t1 = tcg_temp_new_i32();
795     TCGv_i32 t2 = tcg_temp_new_i32();
796     uint32_t i;
797 
798     for (i = 0; i < oprsz; i += 4) {
799         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
800         tcg_gen_ld_i32(t1, cpu_env, bofs + i);
801         if (load_dest) {
802             tcg_gen_ld_i32(t2, cpu_env, dofs + i);
803         }
804         fni(t2, t0, t1, c);
805         tcg_gen_st_i32(t2, cpu_env, dofs + i);
806     }
807     tcg_temp_free_i32(t0);
808     tcg_temp_free_i32(t1);
809     tcg_temp_free_i32(t2);
810 }
811 
812 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
813 static void expand_4_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
814                          uint32_t cofs, uint32_t oprsz, bool write_aofs,
815                          void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32))
816 {
817     TCGv_i32 t0 = tcg_temp_new_i32();
818     TCGv_i32 t1 = tcg_temp_new_i32();
819     TCGv_i32 t2 = tcg_temp_new_i32();
820     TCGv_i32 t3 = tcg_temp_new_i32();
821     uint32_t i;
822 
823     for (i = 0; i < oprsz; i += 4) {
824         tcg_gen_ld_i32(t1, cpu_env, aofs + i);
825         tcg_gen_ld_i32(t2, cpu_env, bofs + i);
826         tcg_gen_ld_i32(t3, cpu_env, cofs + i);
827         fni(t0, t1, t2, t3);
828         tcg_gen_st_i32(t0, cpu_env, dofs + i);
829         if (write_aofs) {
830             tcg_gen_st_i32(t1, cpu_env, aofs + i);
831         }
832     }
833     tcg_temp_free_i32(t3);
834     tcg_temp_free_i32(t2);
835     tcg_temp_free_i32(t1);
836     tcg_temp_free_i32(t0);
837 }
838 
839 static void expand_4i_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
840                           uint32_t cofs, uint32_t oprsz, int32_t c,
841                           void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32,
842                                       int32_t))
843 {
844     TCGv_i32 t0 = tcg_temp_new_i32();
845     TCGv_i32 t1 = tcg_temp_new_i32();
846     TCGv_i32 t2 = tcg_temp_new_i32();
847     TCGv_i32 t3 = tcg_temp_new_i32();
848     uint32_t i;
849 
850     for (i = 0; i < oprsz; i += 4) {
851         tcg_gen_ld_i32(t1, cpu_env, aofs + i);
852         tcg_gen_ld_i32(t2, cpu_env, bofs + i);
853         tcg_gen_ld_i32(t3, cpu_env, cofs + i);
854         fni(t0, t1, t2, t3, c);
855         tcg_gen_st_i32(t0, cpu_env, dofs + i);
856     }
857     tcg_temp_free_i32(t3);
858     tcg_temp_free_i32(t2);
859     tcg_temp_free_i32(t1);
860     tcg_temp_free_i32(t0);
861 }
862 
863 /* Expand OPSZ bytes worth of two-operand operations using i64 elements.  */
864 static void expand_2_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
865                          bool load_dest, void (*fni)(TCGv_i64, TCGv_i64))
866 {
867     TCGv_i64 t0 = tcg_temp_new_i64();
868     TCGv_i64 t1 = tcg_temp_new_i64();
869     uint32_t i;
870 
871     for (i = 0; i < oprsz; i += 8) {
872         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
873         if (load_dest) {
874             tcg_gen_ld_i64(t1, cpu_env, dofs + i);
875         }
876         fni(t1, t0);
877         tcg_gen_st_i64(t1, cpu_env, dofs + i);
878     }
879     tcg_temp_free_i64(t0);
880     tcg_temp_free_i64(t1);
881 }
882 
883 static void expand_2i_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
884                           int64_t c, bool load_dest,
885                           void (*fni)(TCGv_i64, TCGv_i64, int64_t))
886 {
887     TCGv_i64 t0 = tcg_temp_new_i64();
888     TCGv_i64 t1 = tcg_temp_new_i64();
889     uint32_t i;
890 
891     for (i = 0; i < oprsz; i += 8) {
892         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
893         if (load_dest) {
894             tcg_gen_ld_i64(t1, cpu_env, dofs + i);
895         }
896         fni(t1, t0, c);
897         tcg_gen_st_i64(t1, cpu_env, dofs + i);
898     }
899     tcg_temp_free_i64(t0);
900     tcg_temp_free_i64(t1);
901 }
902 
903 static void expand_2s_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
904                           TCGv_i64 c, bool scalar_first,
905                           void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64))
906 {
907     TCGv_i64 t0 = tcg_temp_new_i64();
908     TCGv_i64 t1 = tcg_temp_new_i64();
909     uint32_t i;
910 
911     for (i = 0; i < oprsz; i += 8) {
912         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
913         if (scalar_first) {
914             fni(t1, c, t0);
915         } else {
916             fni(t1, t0, c);
917         }
918         tcg_gen_st_i64(t1, cpu_env, dofs + i);
919     }
920     tcg_temp_free_i64(t0);
921     tcg_temp_free_i64(t1);
922 }
923 
924 /* Expand OPSZ bytes worth of three-operand operations using i64 elements.  */
925 static void expand_3_i64(uint32_t dofs, uint32_t aofs,
926                          uint32_t bofs, uint32_t oprsz, bool load_dest,
927                          void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64))
928 {
929     TCGv_i64 t0 = tcg_temp_new_i64();
930     TCGv_i64 t1 = tcg_temp_new_i64();
931     TCGv_i64 t2 = tcg_temp_new_i64();
932     uint32_t i;
933 
934     for (i = 0; i < oprsz; i += 8) {
935         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
936         tcg_gen_ld_i64(t1, cpu_env, bofs + i);
937         if (load_dest) {
938             tcg_gen_ld_i64(t2, cpu_env, dofs + i);
939         }
940         fni(t2, t0, t1);
941         tcg_gen_st_i64(t2, cpu_env, dofs + i);
942     }
943     tcg_temp_free_i64(t2);
944     tcg_temp_free_i64(t1);
945     tcg_temp_free_i64(t0);
946 }
947 
948 static void expand_3i_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
949                           uint32_t oprsz, int64_t c, bool load_dest,
950                           void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64, int64_t))
951 {
952     TCGv_i64 t0 = tcg_temp_new_i64();
953     TCGv_i64 t1 = tcg_temp_new_i64();
954     TCGv_i64 t2 = tcg_temp_new_i64();
955     uint32_t i;
956 
957     for (i = 0; i < oprsz; i += 8) {
958         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
959         tcg_gen_ld_i64(t1, cpu_env, bofs + i);
960         if (load_dest) {
961             tcg_gen_ld_i64(t2, cpu_env, dofs + i);
962         }
963         fni(t2, t0, t1, c);
964         tcg_gen_st_i64(t2, cpu_env, dofs + i);
965     }
966     tcg_temp_free_i64(t0);
967     tcg_temp_free_i64(t1);
968     tcg_temp_free_i64(t2);
969 }
970 
971 /* Expand OPSZ bytes worth of three-operand operations using i64 elements.  */
972 static void expand_4_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
973                          uint32_t cofs, uint32_t oprsz, bool write_aofs,
974                          void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64, TCGv_i64))
975 {
976     TCGv_i64 t0 = tcg_temp_new_i64();
977     TCGv_i64 t1 = tcg_temp_new_i64();
978     TCGv_i64 t2 = tcg_temp_new_i64();
979     TCGv_i64 t3 = tcg_temp_new_i64();
980     uint32_t i;
981 
982     for (i = 0; i < oprsz; i += 8) {
983         tcg_gen_ld_i64(t1, cpu_env, aofs + i);
984         tcg_gen_ld_i64(t2, cpu_env, bofs + i);
985         tcg_gen_ld_i64(t3, cpu_env, cofs + i);
986         fni(t0, t1, t2, t3);
987         tcg_gen_st_i64(t0, cpu_env, dofs + i);
988         if (write_aofs) {
989             tcg_gen_st_i64(t1, cpu_env, aofs + i);
990         }
991     }
992     tcg_temp_free_i64(t3);
993     tcg_temp_free_i64(t2);
994     tcg_temp_free_i64(t1);
995     tcg_temp_free_i64(t0);
996 }
997 
998 static void expand_4i_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
999                           uint32_t cofs, uint32_t oprsz, int64_t c,
1000                           void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64, TCGv_i64,
1001                                       int64_t))
1002 {
1003     TCGv_i64 t0 = tcg_temp_new_i64();
1004     TCGv_i64 t1 = tcg_temp_new_i64();
1005     TCGv_i64 t2 = tcg_temp_new_i64();
1006     TCGv_i64 t3 = tcg_temp_new_i64();
1007     uint32_t i;
1008 
1009     for (i = 0; i < oprsz; i += 8) {
1010         tcg_gen_ld_i64(t1, cpu_env, aofs + i);
1011         tcg_gen_ld_i64(t2, cpu_env, bofs + i);
1012         tcg_gen_ld_i64(t3, cpu_env, cofs + i);
1013         fni(t0, t1, t2, t3, c);
1014         tcg_gen_st_i64(t0, cpu_env, dofs + i);
1015     }
1016     tcg_temp_free_i64(t3);
1017     tcg_temp_free_i64(t2);
1018     tcg_temp_free_i64(t1);
1019     tcg_temp_free_i64(t0);
1020 }
1021 
1022 /* Expand OPSZ bytes worth of two-operand operations using host vectors.  */
1023 static void expand_2_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1024                          uint32_t oprsz, uint32_t tysz, TCGType type,
1025                          bool load_dest,
1026                          void (*fni)(unsigned, TCGv_vec, TCGv_vec))
1027 {
1028     TCGv_vec t0 = tcg_temp_new_vec(type);
1029     TCGv_vec t1 = tcg_temp_new_vec(type);
1030     uint32_t i;
1031 
1032     for (i = 0; i < oprsz; i += tysz) {
1033         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
1034         if (load_dest) {
1035             tcg_gen_ld_vec(t1, cpu_env, dofs + i);
1036         }
1037         fni(vece, t1, t0);
1038         tcg_gen_st_vec(t1, cpu_env, dofs + i);
1039     }
1040     tcg_temp_free_vec(t0);
1041     tcg_temp_free_vec(t1);
1042 }
1043 
1044 /* Expand OPSZ bytes worth of two-vector operands and an immediate operand
1045    using host vectors.  */
1046 static void expand_2i_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1047                           uint32_t oprsz, uint32_t tysz, TCGType type,
1048                           int64_t c, bool load_dest,
1049                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, int64_t))
1050 {
1051     TCGv_vec t0 = tcg_temp_new_vec(type);
1052     TCGv_vec t1 = tcg_temp_new_vec(type);
1053     uint32_t i;
1054 
1055     for (i = 0; i < oprsz; i += tysz) {
1056         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
1057         if (load_dest) {
1058             tcg_gen_ld_vec(t1, cpu_env, dofs + i);
1059         }
1060         fni(vece, t1, t0, c);
1061         tcg_gen_st_vec(t1, cpu_env, dofs + i);
1062     }
1063     tcg_temp_free_vec(t0);
1064     tcg_temp_free_vec(t1);
1065 }
1066 
1067 static void expand_2s_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1068                           uint32_t oprsz, uint32_t tysz, TCGType type,
1069                           TCGv_vec c, bool scalar_first,
1070                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec))
1071 {
1072     TCGv_vec t0 = tcg_temp_new_vec(type);
1073     TCGv_vec t1 = tcg_temp_new_vec(type);
1074     uint32_t i;
1075 
1076     for (i = 0; i < oprsz; i += tysz) {
1077         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
1078         if (scalar_first) {
1079             fni(vece, t1, c, t0);
1080         } else {
1081             fni(vece, t1, t0, c);
1082         }
1083         tcg_gen_st_vec(t1, cpu_env, dofs + i);
1084     }
1085     tcg_temp_free_vec(t0);
1086     tcg_temp_free_vec(t1);
1087 }
1088 
1089 /* Expand OPSZ bytes worth of three-operand operations using host vectors.  */
1090 static void expand_3_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1091                          uint32_t bofs, uint32_t oprsz,
1092                          uint32_t tysz, TCGType type, bool load_dest,
1093                          void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec))
1094 {
1095     TCGv_vec t0 = tcg_temp_new_vec(type);
1096     TCGv_vec t1 = tcg_temp_new_vec(type);
1097     TCGv_vec t2 = tcg_temp_new_vec(type);
1098     uint32_t i;
1099 
1100     for (i = 0; i < oprsz; i += tysz) {
1101         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
1102         tcg_gen_ld_vec(t1, cpu_env, bofs + i);
1103         if (load_dest) {
1104             tcg_gen_ld_vec(t2, cpu_env, dofs + i);
1105         }
1106         fni(vece, t2, t0, t1);
1107         tcg_gen_st_vec(t2, cpu_env, dofs + i);
1108     }
1109     tcg_temp_free_vec(t2);
1110     tcg_temp_free_vec(t1);
1111     tcg_temp_free_vec(t0);
1112 }
1113 
1114 /*
1115  * Expand OPSZ bytes worth of three-vector operands and an immediate operand
1116  * using host vectors.
1117  */
1118 static void expand_3i_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1119                           uint32_t bofs, uint32_t oprsz, uint32_t tysz,
1120                           TCGType type, int64_t c, bool load_dest,
1121                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec,
1122                                       int64_t))
1123 {
1124     TCGv_vec t0 = tcg_temp_new_vec(type);
1125     TCGv_vec t1 = tcg_temp_new_vec(type);
1126     TCGv_vec t2 = tcg_temp_new_vec(type);
1127     uint32_t i;
1128 
1129     for (i = 0; i < oprsz; i += tysz) {
1130         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
1131         tcg_gen_ld_vec(t1, cpu_env, bofs + i);
1132         if (load_dest) {
1133             tcg_gen_ld_vec(t2, cpu_env, dofs + i);
1134         }
1135         fni(vece, t2, t0, t1, c);
1136         tcg_gen_st_vec(t2, cpu_env, dofs + i);
1137     }
1138     tcg_temp_free_vec(t0);
1139     tcg_temp_free_vec(t1);
1140     tcg_temp_free_vec(t2);
1141 }
1142 
1143 /* Expand OPSZ bytes worth of four-operand operations using host vectors.  */
1144 static void expand_4_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1145                          uint32_t bofs, uint32_t cofs, uint32_t oprsz,
1146                          uint32_t tysz, TCGType type, bool write_aofs,
1147                          void (*fni)(unsigned, TCGv_vec, TCGv_vec,
1148                                      TCGv_vec, TCGv_vec))
1149 {
1150     TCGv_vec t0 = tcg_temp_new_vec(type);
1151     TCGv_vec t1 = tcg_temp_new_vec(type);
1152     TCGv_vec t2 = tcg_temp_new_vec(type);
1153     TCGv_vec t3 = tcg_temp_new_vec(type);
1154     uint32_t i;
1155 
1156     for (i = 0; i < oprsz; i += tysz) {
1157         tcg_gen_ld_vec(t1, cpu_env, aofs + i);
1158         tcg_gen_ld_vec(t2, cpu_env, bofs + i);
1159         tcg_gen_ld_vec(t3, cpu_env, cofs + i);
1160         fni(vece, t0, t1, t2, t3);
1161         tcg_gen_st_vec(t0, cpu_env, dofs + i);
1162         if (write_aofs) {
1163             tcg_gen_st_vec(t1, cpu_env, aofs + i);
1164         }
1165     }
1166     tcg_temp_free_vec(t3);
1167     tcg_temp_free_vec(t2);
1168     tcg_temp_free_vec(t1);
1169     tcg_temp_free_vec(t0);
1170 }
1171 
1172 /*
1173  * Expand OPSZ bytes worth of four-vector operands and an immediate operand
1174  * using host vectors.
1175  */
1176 static void expand_4i_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1177                           uint32_t bofs, uint32_t cofs, uint32_t oprsz,
1178                           uint32_t tysz, TCGType type, int64_t c,
1179                           void (*fni)(unsigned, TCGv_vec, TCGv_vec,
1180                                      TCGv_vec, TCGv_vec, int64_t))
1181 {
1182     TCGv_vec t0 = tcg_temp_new_vec(type);
1183     TCGv_vec t1 = tcg_temp_new_vec(type);
1184     TCGv_vec t2 = tcg_temp_new_vec(type);
1185     TCGv_vec t3 = tcg_temp_new_vec(type);
1186     uint32_t i;
1187 
1188     for (i = 0; i < oprsz; i += tysz) {
1189         tcg_gen_ld_vec(t1, cpu_env, aofs + i);
1190         tcg_gen_ld_vec(t2, cpu_env, bofs + i);
1191         tcg_gen_ld_vec(t3, cpu_env, cofs + i);
1192         fni(vece, t0, t1, t2, t3, c);
1193         tcg_gen_st_vec(t0, cpu_env, dofs + i);
1194     }
1195     tcg_temp_free_vec(t3);
1196     tcg_temp_free_vec(t2);
1197     tcg_temp_free_vec(t1);
1198     tcg_temp_free_vec(t0);
1199 }
1200 
1201 /* Expand a vector two-operand operation.  */
1202 void tcg_gen_gvec_2(uint32_t dofs, uint32_t aofs,
1203                     uint32_t oprsz, uint32_t maxsz, const GVecGen2 *g)
1204 {
1205     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1206     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1207     TCGType type;
1208     uint32_t some;
1209 
1210     check_size_align(oprsz, maxsz, dofs | aofs);
1211     check_overlap_2(dofs, aofs, maxsz);
1212 
1213     type = 0;
1214     if (g->fniv) {
1215         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1216     }
1217     switch (type) {
1218     case TCG_TYPE_V256:
1219         /* Recall that ARM SVE allows vector sizes that are not a
1220          * power of 2, but always a multiple of 16.  The intent is
1221          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1222          */
1223         some = QEMU_ALIGN_DOWN(oprsz, 32);
1224         expand_2_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
1225                      g->load_dest, g->fniv);
1226         if (some == oprsz) {
1227             break;
1228         }
1229         dofs += some;
1230         aofs += some;
1231         oprsz -= some;
1232         maxsz -= some;
1233         /* fallthru */
1234     case TCG_TYPE_V128:
1235         expand_2_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
1236                      g->load_dest, g->fniv);
1237         break;
1238     case TCG_TYPE_V64:
1239         expand_2_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
1240                      g->load_dest, g->fniv);
1241         break;
1242 
1243     case 0:
1244         if (g->fni8 && check_size_impl(oprsz, 8)) {
1245             expand_2_i64(dofs, aofs, oprsz, g->load_dest, g->fni8);
1246         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1247             expand_2_i32(dofs, aofs, oprsz, g->load_dest, g->fni4);
1248         } else {
1249             assert(g->fno != NULL);
1250             tcg_gen_gvec_2_ool(dofs, aofs, oprsz, maxsz, g->data, g->fno);
1251             oprsz = maxsz;
1252         }
1253         break;
1254 
1255     default:
1256         g_assert_not_reached();
1257     }
1258     tcg_swap_vecop_list(hold_list);
1259 
1260     if (oprsz < maxsz) {
1261         expand_clr(dofs + oprsz, maxsz - oprsz);
1262     }
1263 }
1264 
1265 /* Expand a vector operation with two vectors and an immediate.  */
1266 void tcg_gen_gvec_2i(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
1267                      uint32_t maxsz, int64_t c, const GVecGen2i *g)
1268 {
1269     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1270     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1271     TCGType type;
1272     uint32_t some;
1273 
1274     check_size_align(oprsz, maxsz, dofs | aofs);
1275     check_overlap_2(dofs, aofs, maxsz);
1276 
1277     type = 0;
1278     if (g->fniv) {
1279         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1280     }
1281     switch (type) {
1282     case TCG_TYPE_V256:
1283         /* Recall that ARM SVE allows vector sizes that are not a
1284          * power of 2, but always a multiple of 16.  The intent is
1285          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1286          */
1287         some = QEMU_ALIGN_DOWN(oprsz, 32);
1288         expand_2i_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
1289                       c, g->load_dest, g->fniv);
1290         if (some == oprsz) {
1291             break;
1292         }
1293         dofs += some;
1294         aofs += some;
1295         oprsz -= some;
1296         maxsz -= some;
1297         /* fallthru */
1298     case TCG_TYPE_V128:
1299         expand_2i_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
1300                       c, g->load_dest, g->fniv);
1301         break;
1302     case TCG_TYPE_V64:
1303         expand_2i_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
1304                       c, g->load_dest, g->fniv);
1305         break;
1306 
1307     case 0:
1308         if (g->fni8 && check_size_impl(oprsz, 8)) {
1309             expand_2i_i64(dofs, aofs, oprsz, c, g->load_dest, g->fni8);
1310         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1311             expand_2i_i32(dofs, aofs, oprsz, c, g->load_dest, g->fni4);
1312         } else {
1313             if (g->fno) {
1314                 tcg_gen_gvec_2_ool(dofs, aofs, oprsz, maxsz, c, g->fno);
1315             } else {
1316                 TCGv_i64 tcg_c = tcg_constant_i64(c);
1317                 tcg_gen_gvec_2i_ool(dofs, aofs, tcg_c, oprsz,
1318                                     maxsz, c, g->fnoi);
1319             }
1320             oprsz = maxsz;
1321         }
1322         break;
1323 
1324     default:
1325         g_assert_not_reached();
1326     }
1327     tcg_swap_vecop_list(hold_list);
1328 
1329     if (oprsz < maxsz) {
1330         expand_clr(dofs + oprsz, maxsz - oprsz);
1331     }
1332 }
1333 
1334 /* Expand a vector operation with two vectors and a scalar.  */
1335 void tcg_gen_gvec_2s(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
1336                      uint32_t maxsz, TCGv_i64 c, const GVecGen2s *g)
1337 {
1338     TCGType type;
1339 
1340     check_size_align(oprsz, maxsz, dofs | aofs);
1341     check_overlap_2(dofs, aofs, maxsz);
1342 
1343     type = 0;
1344     if (g->fniv) {
1345         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1346     }
1347     if (type != 0) {
1348         const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1349         const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1350         TCGv_vec t_vec = tcg_temp_new_vec(type);
1351         uint32_t some;
1352 
1353         tcg_gen_dup_i64_vec(g->vece, t_vec, c);
1354 
1355         switch (type) {
1356         case TCG_TYPE_V256:
1357             /* Recall that ARM SVE allows vector sizes that are not a
1358              * power of 2, but always a multiple of 16.  The intent is
1359              * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1360              */
1361             some = QEMU_ALIGN_DOWN(oprsz, 32);
1362             expand_2s_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
1363                           t_vec, g->scalar_first, g->fniv);
1364             if (some == oprsz) {
1365                 break;
1366             }
1367             dofs += some;
1368             aofs += some;
1369             oprsz -= some;
1370             maxsz -= some;
1371             /* fallthru */
1372 
1373         case TCG_TYPE_V128:
1374             expand_2s_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
1375                           t_vec, g->scalar_first, g->fniv);
1376             break;
1377 
1378         case TCG_TYPE_V64:
1379             expand_2s_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
1380                           t_vec, g->scalar_first, g->fniv);
1381             break;
1382 
1383         default:
1384             g_assert_not_reached();
1385         }
1386         tcg_temp_free_vec(t_vec);
1387         tcg_swap_vecop_list(hold_list);
1388     } else if (g->fni8 && check_size_impl(oprsz, 8)) {
1389         TCGv_i64 t64 = tcg_temp_new_i64();
1390 
1391         tcg_gen_dup_i64(g->vece, t64, c);
1392         expand_2s_i64(dofs, aofs, oprsz, t64, g->scalar_first, g->fni8);
1393         tcg_temp_free_i64(t64);
1394     } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1395         TCGv_i32 t32 = tcg_temp_new_i32();
1396 
1397         tcg_gen_extrl_i64_i32(t32, c);
1398         tcg_gen_dup_i32(g->vece, t32, t32);
1399         expand_2s_i32(dofs, aofs, oprsz, t32, g->scalar_first, g->fni4);
1400         tcg_temp_free_i32(t32);
1401     } else {
1402         tcg_gen_gvec_2i_ool(dofs, aofs, c, oprsz, maxsz, 0, g->fno);
1403         return;
1404     }
1405 
1406     if (oprsz < maxsz) {
1407         expand_clr(dofs + oprsz, maxsz - oprsz);
1408     }
1409 }
1410 
1411 /* Expand a vector three-operand operation.  */
1412 void tcg_gen_gvec_3(uint32_t dofs, uint32_t aofs, uint32_t bofs,
1413                     uint32_t oprsz, uint32_t maxsz, const GVecGen3 *g)
1414 {
1415     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1416     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1417     TCGType type;
1418     uint32_t some;
1419 
1420     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
1421     check_overlap_3(dofs, aofs, bofs, maxsz);
1422 
1423     type = 0;
1424     if (g->fniv) {
1425         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1426     }
1427     switch (type) {
1428     case TCG_TYPE_V256:
1429         /* Recall that ARM SVE allows vector sizes that are not a
1430          * power of 2, but always a multiple of 16.  The intent is
1431          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1432          */
1433         some = QEMU_ALIGN_DOWN(oprsz, 32);
1434         expand_3_vec(g->vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256,
1435                      g->load_dest, g->fniv);
1436         if (some == oprsz) {
1437             break;
1438         }
1439         dofs += some;
1440         aofs += some;
1441         bofs += some;
1442         oprsz -= some;
1443         maxsz -= some;
1444         /* fallthru */
1445     case TCG_TYPE_V128:
1446         expand_3_vec(g->vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128,
1447                      g->load_dest, g->fniv);
1448         break;
1449     case TCG_TYPE_V64:
1450         expand_3_vec(g->vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64,
1451                      g->load_dest, g->fniv);
1452         break;
1453 
1454     case 0:
1455         if (g->fni8 && check_size_impl(oprsz, 8)) {
1456             expand_3_i64(dofs, aofs, bofs, oprsz, g->load_dest, g->fni8);
1457         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1458             expand_3_i32(dofs, aofs, bofs, oprsz, g->load_dest, g->fni4);
1459         } else {
1460             assert(g->fno != NULL);
1461             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz,
1462                                maxsz, g->data, g->fno);
1463             oprsz = maxsz;
1464         }
1465         break;
1466 
1467     default:
1468         g_assert_not_reached();
1469     }
1470     tcg_swap_vecop_list(hold_list);
1471 
1472     if (oprsz < maxsz) {
1473         expand_clr(dofs + oprsz, maxsz - oprsz);
1474     }
1475 }
1476 
1477 /* Expand a vector operation with three vectors and an immediate.  */
1478 void tcg_gen_gvec_3i(uint32_t dofs, uint32_t aofs, uint32_t bofs,
1479                      uint32_t oprsz, uint32_t maxsz, int64_t c,
1480                      const GVecGen3i *g)
1481 {
1482     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1483     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1484     TCGType type;
1485     uint32_t some;
1486 
1487     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
1488     check_overlap_3(dofs, aofs, bofs, maxsz);
1489 
1490     type = 0;
1491     if (g->fniv) {
1492         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1493     }
1494     switch (type) {
1495     case TCG_TYPE_V256:
1496         /*
1497          * Recall that ARM SVE allows vector sizes that are not a
1498          * power of 2, but always a multiple of 16.  The intent is
1499          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1500          */
1501         some = QEMU_ALIGN_DOWN(oprsz, 32);
1502         expand_3i_vec(g->vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256,
1503                       c, g->load_dest, g->fniv);
1504         if (some == oprsz) {
1505             break;
1506         }
1507         dofs += some;
1508         aofs += some;
1509         bofs += some;
1510         oprsz -= some;
1511         maxsz -= some;
1512         /* fallthru */
1513     case TCG_TYPE_V128:
1514         expand_3i_vec(g->vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128,
1515                       c, g->load_dest, g->fniv);
1516         break;
1517     case TCG_TYPE_V64:
1518         expand_3i_vec(g->vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64,
1519                       c, g->load_dest, g->fniv);
1520         break;
1521 
1522     case 0:
1523         if (g->fni8 && check_size_impl(oprsz, 8)) {
1524             expand_3i_i64(dofs, aofs, bofs, oprsz, c, g->load_dest, g->fni8);
1525         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1526             expand_3i_i32(dofs, aofs, bofs, oprsz, c, g->load_dest, g->fni4);
1527         } else {
1528             assert(g->fno != NULL);
1529             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz, maxsz, c, g->fno);
1530             oprsz = maxsz;
1531         }
1532         break;
1533 
1534     default:
1535         g_assert_not_reached();
1536     }
1537     tcg_swap_vecop_list(hold_list);
1538 
1539     if (oprsz < maxsz) {
1540         expand_clr(dofs + oprsz, maxsz - oprsz);
1541     }
1542 }
1543 
1544 /* Expand a vector four-operand operation.  */
1545 void tcg_gen_gvec_4(uint32_t dofs, uint32_t aofs, uint32_t bofs, uint32_t cofs,
1546                     uint32_t oprsz, uint32_t maxsz, const GVecGen4 *g)
1547 {
1548     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1549     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1550     TCGType type;
1551     uint32_t some;
1552 
1553     check_size_align(oprsz, maxsz, dofs | aofs | bofs | cofs);
1554     check_overlap_4(dofs, aofs, bofs, cofs, maxsz);
1555 
1556     type = 0;
1557     if (g->fniv) {
1558         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1559     }
1560     switch (type) {
1561     case TCG_TYPE_V256:
1562         /* Recall that ARM SVE allows vector sizes that are not a
1563          * power of 2, but always a multiple of 16.  The intent is
1564          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1565          */
1566         some = QEMU_ALIGN_DOWN(oprsz, 32);
1567         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, some,
1568                      32, TCG_TYPE_V256, g->write_aofs, g->fniv);
1569         if (some == oprsz) {
1570             break;
1571         }
1572         dofs += some;
1573         aofs += some;
1574         bofs += some;
1575         cofs += some;
1576         oprsz -= some;
1577         maxsz -= some;
1578         /* fallthru */
1579     case TCG_TYPE_V128:
1580         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1581                      16, TCG_TYPE_V128, g->write_aofs, g->fniv);
1582         break;
1583     case TCG_TYPE_V64:
1584         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1585                      8, TCG_TYPE_V64, g->write_aofs, g->fniv);
1586         break;
1587 
1588     case 0:
1589         if (g->fni8 && check_size_impl(oprsz, 8)) {
1590             expand_4_i64(dofs, aofs, bofs, cofs, oprsz,
1591                          g->write_aofs, g->fni8);
1592         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1593             expand_4_i32(dofs, aofs, bofs, cofs, oprsz,
1594                          g->write_aofs, g->fni4);
1595         } else {
1596             assert(g->fno != NULL);
1597             tcg_gen_gvec_4_ool(dofs, aofs, bofs, cofs,
1598                                oprsz, maxsz, g->data, g->fno);
1599             oprsz = maxsz;
1600         }
1601         break;
1602 
1603     default:
1604         g_assert_not_reached();
1605     }
1606     tcg_swap_vecop_list(hold_list);
1607 
1608     if (oprsz < maxsz) {
1609         expand_clr(dofs + oprsz, maxsz - oprsz);
1610     }
1611 }
1612 
1613 /* Expand a vector four-operand operation.  */
1614 void tcg_gen_gvec_4i(uint32_t dofs, uint32_t aofs, uint32_t bofs, uint32_t cofs,
1615                      uint32_t oprsz, uint32_t maxsz, int64_t c,
1616                      const GVecGen4i *g)
1617 {
1618     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1619     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1620     TCGType type;
1621     uint32_t some;
1622 
1623     check_size_align(oprsz, maxsz, dofs | aofs | bofs | cofs);
1624     check_overlap_4(dofs, aofs, bofs, cofs, maxsz);
1625 
1626     type = 0;
1627     if (g->fniv) {
1628         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1629     }
1630     switch (type) {
1631     case TCG_TYPE_V256:
1632         /*
1633          * Recall that ARM SVE allows vector sizes that are not a
1634          * power of 2, but always a multiple of 16.  The intent is
1635          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1636          */
1637         some = QEMU_ALIGN_DOWN(oprsz, 32);
1638         expand_4i_vec(g->vece, dofs, aofs, bofs, cofs, some,
1639                       32, TCG_TYPE_V256, c, g->fniv);
1640         if (some == oprsz) {
1641             break;
1642         }
1643         dofs += some;
1644         aofs += some;
1645         bofs += some;
1646         cofs += some;
1647         oprsz -= some;
1648         maxsz -= some;
1649         /* fallthru */
1650     case TCG_TYPE_V128:
1651         expand_4i_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1652                        16, TCG_TYPE_V128, c, g->fniv);
1653         break;
1654     case TCG_TYPE_V64:
1655         expand_4i_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1656                       8, TCG_TYPE_V64, c, g->fniv);
1657         break;
1658 
1659     case 0:
1660         if (g->fni8 && check_size_impl(oprsz, 8)) {
1661             expand_4i_i64(dofs, aofs, bofs, cofs, oprsz, c, g->fni8);
1662         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1663             expand_4i_i32(dofs, aofs, bofs, cofs, oprsz, c, g->fni4);
1664         } else {
1665             assert(g->fno != NULL);
1666             tcg_gen_gvec_4_ool(dofs, aofs, bofs, cofs,
1667                                oprsz, maxsz, c, g->fno);
1668             oprsz = maxsz;
1669         }
1670         break;
1671 
1672     default:
1673         g_assert_not_reached();
1674     }
1675     tcg_swap_vecop_list(hold_list);
1676 
1677     if (oprsz < maxsz) {
1678         expand_clr(dofs + oprsz, maxsz - oprsz);
1679     }
1680 }
1681 
1682 /*
1683  * Expand specific vector operations.
1684  */
1685 
1686 static void vec_mov2(unsigned vece, TCGv_vec a, TCGv_vec b)
1687 {
1688     tcg_gen_mov_vec(a, b);
1689 }
1690 
1691 void tcg_gen_gvec_mov(unsigned vece, uint32_t dofs, uint32_t aofs,
1692                       uint32_t oprsz, uint32_t maxsz)
1693 {
1694     static const GVecGen2 g = {
1695         .fni8 = tcg_gen_mov_i64,
1696         .fniv = vec_mov2,
1697         .fno = gen_helper_gvec_mov,
1698         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1699     };
1700     if (dofs != aofs) {
1701         tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g);
1702     } else {
1703         check_size_align(oprsz, maxsz, dofs);
1704         if (oprsz < maxsz) {
1705             expand_clr(dofs + oprsz, maxsz - oprsz);
1706         }
1707     }
1708 }
1709 
1710 void tcg_gen_gvec_dup_i32(unsigned vece, uint32_t dofs, uint32_t oprsz,
1711                           uint32_t maxsz, TCGv_i32 in)
1712 {
1713     check_size_align(oprsz, maxsz, dofs);
1714     tcg_debug_assert(vece <= MO_32);
1715     do_dup(vece, dofs, oprsz, maxsz, in, NULL, 0);
1716 }
1717 
1718 void tcg_gen_gvec_dup_i64(unsigned vece, uint32_t dofs, uint32_t oprsz,
1719                           uint32_t maxsz, TCGv_i64 in)
1720 {
1721     check_size_align(oprsz, maxsz, dofs);
1722     tcg_debug_assert(vece <= MO_64);
1723     do_dup(vece, dofs, oprsz, maxsz, NULL, in, 0);
1724 }
1725 
1726 void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t dofs, uint32_t aofs,
1727                           uint32_t oprsz, uint32_t maxsz)
1728 {
1729     check_size_align(oprsz, maxsz, dofs);
1730     if (vece <= MO_64) {
1731         TCGType type = choose_vector_type(NULL, vece, oprsz, 0);
1732         if (type != 0) {
1733             TCGv_vec t_vec = tcg_temp_new_vec(type);
1734             tcg_gen_dup_mem_vec(vece, t_vec, cpu_env, aofs);
1735             do_dup_store(type, dofs, oprsz, maxsz, t_vec);
1736             tcg_temp_free_vec(t_vec);
1737         } else if (vece <= MO_32) {
1738             TCGv_i32 in = tcg_temp_new_i32();
1739             switch (vece) {
1740             case MO_8:
1741                 tcg_gen_ld8u_i32(in, cpu_env, aofs);
1742                 break;
1743             case MO_16:
1744                 tcg_gen_ld16u_i32(in, cpu_env, aofs);
1745                 break;
1746             default:
1747                 tcg_gen_ld_i32(in, cpu_env, aofs);
1748                 break;
1749             }
1750             do_dup(vece, dofs, oprsz, maxsz, in, NULL, 0);
1751             tcg_temp_free_i32(in);
1752         } else {
1753             TCGv_i64 in = tcg_temp_new_i64();
1754             tcg_gen_ld_i64(in, cpu_env, aofs);
1755             do_dup(vece, dofs, oprsz, maxsz, NULL, in, 0);
1756             tcg_temp_free_i64(in);
1757         }
1758     } else if (vece == 4) {
1759         /* 128-bit duplicate.  */
1760         int i;
1761 
1762         tcg_debug_assert(oprsz >= 16);
1763         if (TCG_TARGET_HAS_v128) {
1764             TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V128);
1765 
1766             tcg_gen_ld_vec(in, cpu_env, aofs);
1767             for (i = (aofs == dofs) * 16; i < oprsz; i += 16) {
1768                 tcg_gen_st_vec(in, cpu_env, dofs + i);
1769             }
1770             tcg_temp_free_vec(in);
1771         } else {
1772             TCGv_i64 in0 = tcg_temp_new_i64();
1773             TCGv_i64 in1 = tcg_temp_new_i64();
1774 
1775             tcg_gen_ld_i64(in0, cpu_env, aofs);
1776             tcg_gen_ld_i64(in1, cpu_env, aofs + 8);
1777             for (i = (aofs == dofs) * 16; i < oprsz; i += 16) {
1778                 tcg_gen_st_i64(in0, cpu_env, dofs + i);
1779                 tcg_gen_st_i64(in1, cpu_env, dofs + i + 8);
1780             }
1781             tcg_temp_free_i64(in0);
1782             tcg_temp_free_i64(in1);
1783         }
1784         if (oprsz < maxsz) {
1785             expand_clr(dofs + oprsz, maxsz - oprsz);
1786         }
1787     } else if (vece == 5) {
1788         /* 256-bit duplicate.  */
1789         int i;
1790 
1791         tcg_debug_assert(oprsz >= 32);
1792         tcg_debug_assert(oprsz % 32 == 0);
1793         if (TCG_TARGET_HAS_v256) {
1794             TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V256);
1795 
1796             tcg_gen_ld_vec(in, cpu_env, aofs);
1797             for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
1798                 tcg_gen_st_vec(in, cpu_env, dofs + i);
1799             }
1800             tcg_temp_free_vec(in);
1801         } else if (TCG_TARGET_HAS_v128) {
1802             TCGv_vec in0 = tcg_temp_new_vec(TCG_TYPE_V128);
1803             TCGv_vec in1 = tcg_temp_new_vec(TCG_TYPE_V128);
1804 
1805             tcg_gen_ld_vec(in0, cpu_env, aofs);
1806             tcg_gen_ld_vec(in1, cpu_env, aofs + 16);
1807             for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
1808                 tcg_gen_st_vec(in0, cpu_env, dofs + i);
1809                 tcg_gen_st_vec(in1, cpu_env, dofs + i + 16);
1810             }
1811             tcg_temp_free_vec(in0);
1812             tcg_temp_free_vec(in1);
1813         } else {
1814             TCGv_i64 in[4];
1815             int j;
1816 
1817             for (j = 0; j < 4; ++j) {
1818                 in[j] = tcg_temp_new_i64();
1819                 tcg_gen_ld_i64(in[j], cpu_env, aofs + j * 8);
1820             }
1821             for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
1822                 for (j = 0; j < 4; ++j) {
1823                     tcg_gen_st_i64(in[j], cpu_env, dofs + i + j * 8);
1824                 }
1825             }
1826             for (j = 0; j < 4; ++j) {
1827                 tcg_temp_free_i64(in[j]);
1828             }
1829         }
1830         if (oprsz < maxsz) {
1831             expand_clr(dofs + oprsz, maxsz - oprsz);
1832         }
1833     } else {
1834         g_assert_not_reached();
1835     }
1836 }
1837 
1838 void tcg_gen_gvec_dup_imm(unsigned vece, uint32_t dofs, uint32_t oprsz,
1839                           uint32_t maxsz, uint64_t x)
1840 {
1841     check_size_align(oprsz, maxsz, dofs);
1842     do_dup(vece, dofs, oprsz, maxsz, NULL, NULL, x);
1843 }
1844 
1845 void tcg_gen_gvec_not(unsigned vece, uint32_t dofs, uint32_t aofs,
1846                       uint32_t oprsz, uint32_t maxsz)
1847 {
1848     static const GVecGen2 g = {
1849         .fni8 = tcg_gen_not_i64,
1850         .fniv = tcg_gen_not_vec,
1851         .fno = gen_helper_gvec_not,
1852         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1853     };
1854     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g);
1855 }
1856 
1857 /* Perform a vector addition using normal addition and a mask.  The mask
1858    should be the sign bit of each lane.  This 6-operation form is more
1859    efficient than separate additions when there are 4 or more lanes in
1860    the 64-bit operation.  */
1861 static void gen_addv_mask(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 m)
1862 {
1863     TCGv_i64 t1 = tcg_temp_new_i64();
1864     TCGv_i64 t2 = tcg_temp_new_i64();
1865     TCGv_i64 t3 = tcg_temp_new_i64();
1866 
1867     tcg_gen_andc_i64(t1, a, m);
1868     tcg_gen_andc_i64(t2, b, m);
1869     tcg_gen_xor_i64(t3, a, b);
1870     tcg_gen_add_i64(d, t1, t2);
1871     tcg_gen_and_i64(t3, t3, m);
1872     tcg_gen_xor_i64(d, d, t3);
1873 
1874     tcg_temp_free_i64(t1);
1875     tcg_temp_free_i64(t2);
1876     tcg_temp_free_i64(t3);
1877 }
1878 
1879 void tcg_gen_vec_add8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1880 {
1881     TCGv_i64 m = tcg_constant_i64(dup_const(MO_8, 0x80));
1882     gen_addv_mask(d, a, b, m);
1883 }
1884 
1885 void tcg_gen_vec_add8_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
1886 {
1887     TCGv_i32 m = tcg_constant_i32((int32_t)dup_const(MO_8, 0x80));
1888     TCGv_i32 t1 = tcg_temp_new_i32();
1889     TCGv_i32 t2 = tcg_temp_new_i32();
1890     TCGv_i32 t3 = tcg_temp_new_i32();
1891 
1892     tcg_gen_andc_i32(t1, a, m);
1893     tcg_gen_andc_i32(t2, b, m);
1894     tcg_gen_xor_i32(t3, a, b);
1895     tcg_gen_add_i32(d, t1, t2);
1896     tcg_gen_and_i32(t3, t3, m);
1897     tcg_gen_xor_i32(d, d, t3);
1898 
1899     tcg_temp_free_i32(t1);
1900     tcg_temp_free_i32(t2);
1901     tcg_temp_free_i32(t3);
1902 }
1903 
1904 void tcg_gen_vec_add16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1905 {
1906     TCGv_i64 m = tcg_constant_i64(dup_const(MO_16, 0x8000));
1907     gen_addv_mask(d, a, b, m);
1908 }
1909 
1910 void tcg_gen_vec_add16_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
1911 {
1912     TCGv_i32 t1 = tcg_temp_new_i32();
1913     TCGv_i32 t2 = tcg_temp_new_i32();
1914 
1915     tcg_gen_andi_i32(t1, a, ~0xffff);
1916     tcg_gen_add_i32(t2, a, b);
1917     tcg_gen_add_i32(t1, t1, b);
1918     tcg_gen_deposit_i32(d, t1, t2, 0, 16);
1919 
1920     tcg_temp_free_i32(t1);
1921     tcg_temp_free_i32(t2);
1922 }
1923 
1924 void tcg_gen_vec_add32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1925 {
1926     TCGv_i64 t1 = tcg_temp_new_i64();
1927     TCGv_i64 t2 = tcg_temp_new_i64();
1928 
1929     tcg_gen_andi_i64(t1, a, ~0xffffffffull);
1930     tcg_gen_add_i64(t2, a, b);
1931     tcg_gen_add_i64(t1, t1, b);
1932     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
1933 
1934     tcg_temp_free_i64(t1);
1935     tcg_temp_free_i64(t2);
1936 }
1937 
1938 static const TCGOpcode vecop_list_add[] = { INDEX_op_add_vec, 0 };
1939 
1940 void tcg_gen_gvec_add(unsigned vece, uint32_t dofs, uint32_t aofs,
1941                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1942 {
1943     static const GVecGen3 g[4] = {
1944         { .fni8 = tcg_gen_vec_add8_i64,
1945           .fniv = tcg_gen_add_vec,
1946           .fno = gen_helper_gvec_add8,
1947           .opt_opc = vecop_list_add,
1948           .vece = MO_8 },
1949         { .fni8 = tcg_gen_vec_add16_i64,
1950           .fniv = tcg_gen_add_vec,
1951           .fno = gen_helper_gvec_add16,
1952           .opt_opc = vecop_list_add,
1953           .vece = MO_16 },
1954         { .fni4 = tcg_gen_add_i32,
1955           .fniv = tcg_gen_add_vec,
1956           .fno = gen_helper_gvec_add32,
1957           .opt_opc = vecop_list_add,
1958           .vece = MO_32 },
1959         { .fni8 = tcg_gen_add_i64,
1960           .fniv = tcg_gen_add_vec,
1961           .fno = gen_helper_gvec_add64,
1962           .opt_opc = vecop_list_add,
1963           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1964           .vece = MO_64 },
1965     };
1966 
1967     tcg_debug_assert(vece <= MO_64);
1968     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1969 }
1970 
1971 void tcg_gen_gvec_adds(unsigned vece, uint32_t dofs, uint32_t aofs,
1972                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
1973 {
1974     static const GVecGen2s g[4] = {
1975         { .fni8 = tcg_gen_vec_add8_i64,
1976           .fniv = tcg_gen_add_vec,
1977           .fno = gen_helper_gvec_adds8,
1978           .opt_opc = vecop_list_add,
1979           .vece = MO_8 },
1980         { .fni8 = tcg_gen_vec_add16_i64,
1981           .fniv = tcg_gen_add_vec,
1982           .fno = gen_helper_gvec_adds16,
1983           .opt_opc = vecop_list_add,
1984           .vece = MO_16 },
1985         { .fni4 = tcg_gen_add_i32,
1986           .fniv = tcg_gen_add_vec,
1987           .fno = gen_helper_gvec_adds32,
1988           .opt_opc = vecop_list_add,
1989           .vece = MO_32 },
1990         { .fni8 = tcg_gen_add_i64,
1991           .fniv = tcg_gen_add_vec,
1992           .fno = gen_helper_gvec_adds64,
1993           .opt_opc = vecop_list_add,
1994           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1995           .vece = MO_64 },
1996     };
1997 
1998     tcg_debug_assert(vece <= MO_64);
1999     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
2000 }
2001 
2002 void tcg_gen_gvec_addi(unsigned vece, uint32_t dofs, uint32_t aofs,
2003                        int64_t c, uint32_t oprsz, uint32_t maxsz)
2004 {
2005     TCGv_i64 tmp = tcg_constant_i64(c);
2006     tcg_gen_gvec_adds(vece, dofs, aofs, tmp, oprsz, maxsz);
2007 }
2008 
2009 static const TCGOpcode vecop_list_sub[] = { INDEX_op_sub_vec, 0 };
2010 
2011 void tcg_gen_gvec_subs(unsigned vece, uint32_t dofs, uint32_t aofs,
2012                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2013 {
2014     static const GVecGen2s g[4] = {
2015         { .fni8 = tcg_gen_vec_sub8_i64,
2016           .fniv = tcg_gen_sub_vec,
2017           .fno = gen_helper_gvec_subs8,
2018           .opt_opc = vecop_list_sub,
2019           .vece = MO_8 },
2020         { .fni8 = tcg_gen_vec_sub16_i64,
2021           .fniv = tcg_gen_sub_vec,
2022           .fno = gen_helper_gvec_subs16,
2023           .opt_opc = vecop_list_sub,
2024           .vece = MO_16 },
2025         { .fni4 = tcg_gen_sub_i32,
2026           .fniv = tcg_gen_sub_vec,
2027           .fno = gen_helper_gvec_subs32,
2028           .opt_opc = vecop_list_sub,
2029           .vece = MO_32 },
2030         { .fni8 = tcg_gen_sub_i64,
2031           .fniv = tcg_gen_sub_vec,
2032           .fno = gen_helper_gvec_subs64,
2033           .opt_opc = vecop_list_sub,
2034           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2035           .vece = MO_64 },
2036     };
2037 
2038     tcg_debug_assert(vece <= MO_64);
2039     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
2040 }
2041 
2042 /* Perform a vector subtraction using normal subtraction and a mask.
2043    Compare gen_addv_mask above.  */
2044 static void gen_subv_mask(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 m)
2045 {
2046     TCGv_i64 t1 = tcg_temp_new_i64();
2047     TCGv_i64 t2 = tcg_temp_new_i64();
2048     TCGv_i64 t3 = tcg_temp_new_i64();
2049 
2050     tcg_gen_or_i64(t1, a, m);
2051     tcg_gen_andc_i64(t2, b, m);
2052     tcg_gen_eqv_i64(t3, a, b);
2053     tcg_gen_sub_i64(d, t1, t2);
2054     tcg_gen_and_i64(t3, t3, m);
2055     tcg_gen_xor_i64(d, d, t3);
2056 
2057     tcg_temp_free_i64(t1);
2058     tcg_temp_free_i64(t2);
2059     tcg_temp_free_i64(t3);
2060 }
2061 
2062 void tcg_gen_vec_sub8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2063 {
2064     TCGv_i64 m = tcg_constant_i64(dup_const(MO_8, 0x80));
2065     gen_subv_mask(d, a, b, m);
2066 }
2067 
2068 void tcg_gen_vec_sub8_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
2069 {
2070     TCGv_i32 m = tcg_constant_i32((int32_t)dup_const(MO_8, 0x80));
2071     TCGv_i32 t1 = tcg_temp_new_i32();
2072     TCGv_i32 t2 = tcg_temp_new_i32();
2073     TCGv_i32 t3 = tcg_temp_new_i32();
2074 
2075     tcg_gen_or_i32(t1, a, m);
2076     tcg_gen_andc_i32(t2, b, m);
2077     tcg_gen_eqv_i32(t3, a, b);
2078     tcg_gen_sub_i32(d, t1, t2);
2079     tcg_gen_and_i32(t3, t3, m);
2080     tcg_gen_xor_i32(d, d, t3);
2081 
2082     tcg_temp_free_i32(t1);
2083     tcg_temp_free_i32(t2);
2084     tcg_temp_free_i32(t3);
2085 }
2086 
2087 void tcg_gen_vec_sub16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2088 {
2089     TCGv_i64 m = tcg_constant_i64(dup_const(MO_16, 0x8000));
2090     gen_subv_mask(d, a, b, m);
2091 }
2092 
2093 void tcg_gen_vec_sub16_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
2094 {
2095     TCGv_i32 t1 = tcg_temp_new_i32();
2096     TCGv_i32 t2 = tcg_temp_new_i32();
2097 
2098     tcg_gen_andi_i32(t1, b, ~0xffff);
2099     tcg_gen_sub_i32(t2, a, b);
2100     tcg_gen_sub_i32(t1, a, t1);
2101     tcg_gen_deposit_i32(d, t1, t2, 0, 16);
2102 
2103     tcg_temp_free_i32(t1);
2104     tcg_temp_free_i32(t2);
2105 }
2106 
2107 void tcg_gen_vec_sub32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2108 {
2109     TCGv_i64 t1 = tcg_temp_new_i64();
2110     TCGv_i64 t2 = tcg_temp_new_i64();
2111 
2112     tcg_gen_andi_i64(t1, b, ~0xffffffffull);
2113     tcg_gen_sub_i64(t2, a, b);
2114     tcg_gen_sub_i64(t1, a, t1);
2115     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
2116 
2117     tcg_temp_free_i64(t1);
2118     tcg_temp_free_i64(t2);
2119 }
2120 
2121 void tcg_gen_gvec_sub(unsigned vece, uint32_t dofs, uint32_t aofs,
2122                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2123 {
2124     static const GVecGen3 g[4] = {
2125         { .fni8 = tcg_gen_vec_sub8_i64,
2126           .fniv = tcg_gen_sub_vec,
2127           .fno = gen_helper_gvec_sub8,
2128           .opt_opc = vecop_list_sub,
2129           .vece = MO_8 },
2130         { .fni8 = tcg_gen_vec_sub16_i64,
2131           .fniv = tcg_gen_sub_vec,
2132           .fno = gen_helper_gvec_sub16,
2133           .opt_opc = vecop_list_sub,
2134           .vece = MO_16 },
2135         { .fni4 = tcg_gen_sub_i32,
2136           .fniv = tcg_gen_sub_vec,
2137           .fno = gen_helper_gvec_sub32,
2138           .opt_opc = vecop_list_sub,
2139           .vece = MO_32 },
2140         { .fni8 = tcg_gen_sub_i64,
2141           .fniv = tcg_gen_sub_vec,
2142           .fno = gen_helper_gvec_sub64,
2143           .opt_opc = vecop_list_sub,
2144           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2145           .vece = MO_64 },
2146     };
2147 
2148     tcg_debug_assert(vece <= MO_64);
2149     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2150 }
2151 
2152 static const TCGOpcode vecop_list_mul[] = { INDEX_op_mul_vec, 0 };
2153 
2154 void tcg_gen_gvec_mul(unsigned vece, uint32_t dofs, uint32_t aofs,
2155                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2156 {
2157     static const GVecGen3 g[4] = {
2158         { .fniv = tcg_gen_mul_vec,
2159           .fno = gen_helper_gvec_mul8,
2160           .opt_opc = vecop_list_mul,
2161           .vece = MO_8 },
2162         { .fniv = tcg_gen_mul_vec,
2163           .fno = gen_helper_gvec_mul16,
2164           .opt_opc = vecop_list_mul,
2165           .vece = MO_16 },
2166         { .fni4 = tcg_gen_mul_i32,
2167           .fniv = tcg_gen_mul_vec,
2168           .fno = gen_helper_gvec_mul32,
2169           .opt_opc = vecop_list_mul,
2170           .vece = MO_32 },
2171         { .fni8 = tcg_gen_mul_i64,
2172           .fniv = tcg_gen_mul_vec,
2173           .fno = gen_helper_gvec_mul64,
2174           .opt_opc = vecop_list_mul,
2175           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2176           .vece = MO_64 },
2177     };
2178 
2179     tcg_debug_assert(vece <= MO_64);
2180     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2181 }
2182 
2183 void tcg_gen_gvec_muls(unsigned vece, uint32_t dofs, uint32_t aofs,
2184                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2185 {
2186     static const GVecGen2s g[4] = {
2187         { .fniv = tcg_gen_mul_vec,
2188           .fno = gen_helper_gvec_muls8,
2189           .opt_opc = vecop_list_mul,
2190           .vece = MO_8 },
2191         { .fniv = tcg_gen_mul_vec,
2192           .fno = gen_helper_gvec_muls16,
2193           .opt_opc = vecop_list_mul,
2194           .vece = MO_16 },
2195         { .fni4 = tcg_gen_mul_i32,
2196           .fniv = tcg_gen_mul_vec,
2197           .fno = gen_helper_gvec_muls32,
2198           .opt_opc = vecop_list_mul,
2199           .vece = MO_32 },
2200         { .fni8 = tcg_gen_mul_i64,
2201           .fniv = tcg_gen_mul_vec,
2202           .fno = gen_helper_gvec_muls64,
2203           .opt_opc = vecop_list_mul,
2204           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2205           .vece = MO_64 },
2206     };
2207 
2208     tcg_debug_assert(vece <= MO_64);
2209     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
2210 }
2211 
2212 void tcg_gen_gvec_muli(unsigned vece, uint32_t dofs, uint32_t aofs,
2213                        int64_t c, uint32_t oprsz, uint32_t maxsz)
2214 {
2215     TCGv_i64 tmp = tcg_constant_i64(c);
2216     tcg_gen_gvec_muls(vece, dofs, aofs, tmp, oprsz, maxsz);
2217 }
2218 
2219 void tcg_gen_gvec_ssadd(unsigned vece, uint32_t dofs, uint32_t aofs,
2220                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2221 {
2222     static const TCGOpcode vecop_list[] = { INDEX_op_ssadd_vec, 0 };
2223     static const GVecGen3 g[4] = {
2224         { .fniv = tcg_gen_ssadd_vec,
2225           .fno = gen_helper_gvec_ssadd8,
2226           .opt_opc = vecop_list,
2227           .vece = MO_8 },
2228         { .fniv = tcg_gen_ssadd_vec,
2229           .fno = gen_helper_gvec_ssadd16,
2230           .opt_opc = vecop_list,
2231           .vece = MO_16 },
2232         { .fniv = tcg_gen_ssadd_vec,
2233           .fno = gen_helper_gvec_ssadd32,
2234           .opt_opc = vecop_list,
2235           .vece = MO_32 },
2236         { .fniv = tcg_gen_ssadd_vec,
2237           .fno = gen_helper_gvec_ssadd64,
2238           .opt_opc = vecop_list,
2239           .vece = MO_64 },
2240     };
2241     tcg_debug_assert(vece <= MO_64);
2242     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2243 }
2244 
2245 void tcg_gen_gvec_sssub(unsigned vece, uint32_t dofs, uint32_t aofs,
2246                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2247 {
2248     static const TCGOpcode vecop_list[] = { INDEX_op_sssub_vec, 0 };
2249     static const GVecGen3 g[4] = {
2250         { .fniv = tcg_gen_sssub_vec,
2251           .fno = gen_helper_gvec_sssub8,
2252           .opt_opc = vecop_list,
2253           .vece = MO_8 },
2254         { .fniv = tcg_gen_sssub_vec,
2255           .fno = gen_helper_gvec_sssub16,
2256           .opt_opc = vecop_list,
2257           .vece = MO_16 },
2258         { .fniv = tcg_gen_sssub_vec,
2259           .fno = gen_helper_gvec_sssub32,
2260           .opt_opc = vecop_list,
2261           .vece = MO_32 },
2262         { .fniv = tcg_gen_sssub_vec,
2263           .fno = gen_helper_gvec_sssub64,
2264           .opt_opc = vecop_list,
2265           .vece = MO_64 },
2266     };
2267     tcg_debug_assert(vece <= MO_64);
2268     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2269 }
2270 
2271 static void tcg_gen_usadd_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
2272 {
2273     TCGv_i32 max = tcg_constant_i32(-1);
2274     tcg_gen_add_i32(d, a, b);
2275     tcg_gen_movcond_i32(TCG_COND_LTU, d, d, a, max, d);
2276 }
2277 
2278 static void tcg_gen_usadd_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2279 {
2280     TCGv_i64 max = tcg_constant_i64(-1);
2281     tcg_gen_add_i64(d, a, b);
2282     tcg_gen_movcond_i64(TCG_COND_LTU, d, d, a, max, d);
2283 }
2284 
2285 void tcg_gen_gvec_usadd(unsigned vece, uint32_t dofs, uint32_t aofs,
2286                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2287 {
2288     static const TCGOpcode vecop_list[] = { INDEX_op_usadd_vec, 0 };
2289     static const GVecGen3 g[4] = {
2290         { .fniv = tcg_gen_usadd_vec,
2291           .fno = gen_helper_gvec_usadd8,
2292           .opt_opc = vecop_list,
2293           .vece = MO_8 },
2294         { .fniv = tcg_gen_usadd_vec,
2295           .fno = gen_helper_gvec_usadd16,
2296           .opt_opc = vecop_list,
2297           .vece = MO_16 },
2298         { .fni4 = tcg_gen_usadd_i32,
2299           .fniv = tcg_gen_usadd_vec,
2300           .fno = gen_helper_gvec_usadd32,
2301           .opt_opc = vecop_list,
2302           .vece = MO_32 },
2303         { .fni8 = tcg_gen_usadd_i64,
2304           .fniv = tcg_gen_usadd_vec,
2305           .fno = gen_helper_gvec_usadd64,
2306           .opt_opc = vecop_list,
2307           .vece = MO_64 }
2308     };
2309     tcg_debug_assert(vece <= MO_64);
2310     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2311 }
2312 
2313 static void tcg_gen_ussub_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
2314 {
2315     TCGv_i32 min = tcg_constant_i32(0);
2316     tcg_gen_sub_i32(d, a, b);
2317     tcg_gen_movcond_i32(TCG_COND_LTU, d, a, b, min, d);
2318 }
2319 
2320 static void tcg_gen_ussub_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2321 {
2322     TCGv_i64 min = tcg_constant_i64(0);
2323     tcg_gen_sub_i64(d, a, b);
2324     tcg_gen_movcond_i64(TCG_COND_LTU, d, a, b, min, d);
2325 }
2326 
2327 void tcg_gen_gvec_ussub(unsigned vece, uint32_t dofs, uint32_t aofs,
2328                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2329 {
2330     static const TCGOpcode vecop_list[] = { INDEX_op_ussub_vec, 0 };
2331     static const GVecGen3 g[4] = {
2332         { .fniv = tcg_gen_ussub_vec,
2333           .fno = gen_helper_gvec_ussub8,
2334           .opt_opc = vecop_list,
2335           .vece = MO_8 },
2336         { .fniv = tcg_gen_ussub_vec,
2337           .fno = gen_helper_gvec_ussub16,
2338           .opt_opc = vecop_list,
2339           .vece = MO_16 },
2340         { .fni4 = tcg_gen_ussub_i32,
2341           .fniv = tcg_gen_ussub_vec,
2342           .fno = gen_helper_gvec_ussub32,
2343           .opt_opc = vecop_list,
2344           .vece = MO_32 },
2345         { .fni8 = tcg_gen_ussub_i64,
2346           .fniv = tcg_gen_ussub_vec,
2347           .fno = gen_helper_gvec_ussub64,
2348           .opt_opc = vecop_list,
2349           .vece = MO_64 }
2350     };
2351     tcg_debug_assert(vece <= MO_64);
2352     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2353 }
2354 
2355 void tcg_gen_gvec_smin(unsigned vece, uint32_t dofs, uint32_t aofs,
2356                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2357 {
2358     static const TCGOpcode vecop_list[] = { INDEX_op_smin_vec, 0 };
2359     static const GVecGen3 g[4] = {
2360         { .fniv = tcg_gen_smin_vec,
2361           .fno = gen_helper_gvec_smin8,
2362           .opt_opc = vecop_list,
2363           .vece = MO_8 },
2364         { .fniv = tcg_gen_smin_vec,
2365           .fno = gen_helper_gvec_smin16,
2366           .opt_opc = vecop_list,
2367           .vece = MO_16 },
2368         { .fni4 = tcg_gen_smin_i32,
2369           .fniv = tcg_gen_smin_vec,
2370           .fno = gen_helper_gvec_smin32,
2371           .opt_opc = vecop_list,
2372           .vece = MO_32 },
2373         { .fni8 = tcg_gen_smin_i64,
2374           .fniv = tcg_gen_smin_vec,
2375           .fno = gen_helper_gvec_smin64,
2376           .opt_opc = vecop_list,
2377           .vece = MO_64 }
2378     };
2379     tcg_debug_assert(vece <= MO_64);
2380     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2381 }
2382 
2383 void tcg_gen_gvec_umin(unsigned vece, uint32_t dofs, uint32_t aofs,
2384                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2385 {
2386     static const TCGOpcode vecop_list[] = { INDEX_op_umin_vec, 0 };
2387     static const GVecGen3 g[4] = {
2388         { .fniv = tcg_gen_umin_vec,
2389           .fno = gen_helper_gvec_umin8,
2390           .opt_opc = vecop_list,
2391           .vece = MO_8 },
2392         { .fniv = tcg_gen_umin_vec,
2393           .fno = gen_helper_gvec_umin16,
2394           .opt_opc = vecop_list,
2395           .vece = MO_16 },
2396         { .fni4 = tcg_gen_umin_i32,
2397           .fniv = tcg_gen_umin_vec,
2398           .fno = gen_helper_gvec_umin32,
2399           .opt_opc = vecop_list,
2400           .vece = MO_32 },
2401         { .fni8 = tcg_gen_umin_i64,
2402           .fniv = tcg_gen_umin_vec,
2403           .fno = gen_helper_gvec_umin64,
2404           .opt_opc = vecop_list,
2405           .vece = MO_64 }
2406     };
2407     tcg_debug_assert(vece <= MO_64);
2408     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2409 }
2410 
2411 void tcg_gen_gvec_smax(unsigned vece, uint32_t dofs, uint32_t aofs,
2412                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2413 {
2414     static const TCGOpcode vecop_list[] = { INDEX_op_smax_vec, 0 };
2415     static const GVecGen3 g[4] = {
2416         { .fniv = tcg_gen_smax_vec,
2417           .fno = gen_helper_gvec_smax8,
2418           .opt_opc = vecop_list,
2419           .vece = MO_8 },
2420         { .fniv = tcg_gen_smax_vec,
2421           .fno = gen_helper_gvec_smax16,
2422           .opt_opc = vecop_list,
2423           .vece = MO_16 },
2424         { .fni4 = tcg_gen_smax_i32,
2425           .fniv = tcg_gen_smax_vec,
2426           .fno = gen_helper_gvec_smax32,
2427           .opt_opc = vecop_list,
2428           .vece = MO_32 },
2429         { .fni8 = tcg_gen_smax_i64,
2430           .fniv = tcg_gen_smax_vec,
2431           .fno = gen_helper_gvec_smax64,
2432           .opt_opc = vecop_list,
2433           .vece = MO_64 }
2434     };
2435     tcg_debug_assert(vece <= MO_64);
2436     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2437 }
2438 
2439 void tcg_gen_gvec_umax(unsigned vece, uint32_t dofs, uint32_t aofs,
2440                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2441 {
2442     static const TCGOpcode vecop_list[] = { INDEX_op_umax_vec, 0 };
2443     static const GVecGen3 g[4] = {
2444         { .fniv = tcg_gen_umax_vec,
2445           .fno = gen_helper_gvec_umax8,
2446           .opt_opc = vecop_list,
2447           .vece = MO_8 },
2448         { .fniv = tcg_gen_umax_vec,
2449           .fno = gen_helper_gvec_umax16,
2450           .opt_opc = vecop_list,
2451           .vece = MO_16 },
2452         { .fni4 = tcg_gen_umax_i32,
2453           .fniv = tcg_gen_umax_vec,
2454           .fno = gen_helper_gvec_umax32,
2455           .opt_opc = vecop_list,
2456           .vece = MO_32 },
2457         { .fni8 = tcg_gen_umax_i64,
2458           .fniv = tcg_gen_umax_vec,
2459           .fno = gen_helper_gvec_umax64,
2460           .opt_opc = vecop_list,
2461           .vece = MO_64 }
2462     };
2463     tcg_debug_assert(vece <= MO_64);
2464     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2465 }
2466 
2467 /* Perform a vector negation using normal negation and a mask.
2468    Compare gen_subv_mask above.  */
2469 static void gen_negv_mask(TCGv_i64 d, TCGv_i64 b, TCGv_i64 m)
2470 {
2471     TCGv_i64 t2 = tcg_temp_new_i64();
2472     TCGv_i64 t3 = tcg_temp_new_i64();
2473 
2474     tcg_gen_andc_i64(t3, m, b);
2475     tcg_gen_andc_i64(t2, b, m);
2476     tcg_gen_sub_i64(d, m, t2);
2477     tcg_gen_xor_i64(d, d, t3);
2478 
2479     tcg_temp_free_i64(t2);
2480     tcg_temp_free_i64(t3);
2481 }
2482 
2483 void tcg_gen_vec_neg8_i64(TCGv_i64 d, TCGv_i64 b)
2484 {
2485     TCGv_i64 m = tcg_constant_i64(dup_const(MO_8, 0x80));
2486     gen_negv_mask(d, b, m);
2487 }
2488 
2489 void tcg_gen_vec_neg16_i64(TCGv_i64 d, TCGv_i64 b)
2490 {
2491     TCGv_i64 m = tcg_constant_i64(dup_const(MO_16, 0x8000));
2492     gen_negv_mask(d, b, m);
2493 }
2494 
2495 void tcg_gen_vec_neg32_i64(TCGv_i64 d, TCGv_i64 b)
2496 {
2497     TCGv_i64 t1 = tcg_temp_new_i64();
2498     TCGv_i64 t2 = tcg_temp_new_i64();
2499 
2500     tcg_gen_andi_i64(t1, b, ~0xffffffffull);
2501     tcg_gen_neg_i64(t2, b);
2502     tcg_gen_neg_i64(t1, t1);
2503     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
2504 
2505     tcg_temp_free_i64(t1);
2506     tcg_temp_free_i64(t2);
2507 }
2508 
2509 void tcg_gen_gvec_neg(unsigned vece, uint32_t dofs, uint32_t aofs,
2510                       uint32_t oprsz, uint32_t maxsz)
2511 {
2512     static const TCGOpcode vecop_list[] = { INDEX_op_neg_vec, 0 };
2513     static const GVecGen2 g[4] = {
2514         { .fni8 = tcg_gen_vec_neg8_i64,
2515           .fniv = tcg_gen_neg_vec,
2516           .fno = gen_helper_gvec_neg8,
2517           .opt_opc = vecop_list,
2518           .vece = MO_8 },
2519         { .fni8 = tcg_gen_vec_neg16_i64,
2520           .fniv = tcg_gen_neg_vec,
2521           .fno = gen_helper_gvec_neg16,
2522           .opt_opc = vecop_list,
2523           .vece = MO_16 },
2524         { .fni4 = tcg_gen_neg_i32,
2525           .fniv = tcg_gen_neg_vec,
2526           .fno = gen_helper_gvec_neg32,
2527           .opt_opc = vecop_list,
2528           .vece = MO_32 },
2529         { .fni8 = tcg_gen_neg_i64,
2530           .fniv = tcg_gen_neg_vec,
2531           .fno = gen_helper_gvec_neg64,
2532           .opt_opc = vecop_list,
2533           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2534           .vece = MO_64 },
2535     };
2536 
2537     tcg_debug_assert(vece <= MO_64);
2538     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g[vece]);
2539 }
2540 
2541 static void gen_absv_mask(TCGv_i64 d, TCGv_i64 b, unsigned vece)
2542 {
2543     TCGv_i64 t = tcg_temp_new_i64();
2544     int nbit = 8 << vece;
2545 
2546     /* Create -1 for each negative element.  */
2547     tcg_gen_shri_i64(t, b, nbit - 1);
2548     tcg_gen_andi_i64(t, t, dup_const(vece, 1));
2549     tcg_gen_muli_i64(t, t, (1 << nbit) - 1);
2550 
2551     /*
2552      * Invert (via xor -1) and add one.
2553      * Because of the ordering the msb is cleared,
2554      * so we never have carry into the next element.
2555      */
2556     tcg_gen_xor_i64(d, b, t);
2557     tcg_gen_andi_i64(t, t, dup_const(vece, 1));
2558     tcg_gen_add_i64(d, d, t);
2559 
2560     tcg_temp_free_i64(t);
2561 }
2562 
2563 static void tcg_gen_vec_abs8_i64(TCGv_i64 d, TCGv_i64 b)
2564 {
2565     gen_absv_mask(d, b, MO_8);
2566 }
2567 
2568 static void tcg_gen_vec_abs16_i64(TCGv_i64 d, TCGv_i64 b)
2569 {
2570     gen_absv_mask(d, b, MO_16);
2571 }
2572 
2573 void tcg_gen_gvec_abs(unsigned vece, uint32_t dofs, uint32_t aofs,
2574                       uint32_t oprsz, uint32_t maxsz)
2575 {
2576     static const TCGOpcode vecop_list[] = { INDEX_op_abs_vec, 0 };
2577     static const GVecGen2 g[4] = {
2578         { .fni8 = tcg_gen_vec_abs8_i64,
2579           .fniv = tcg_gen_abs_vec,
2580           .fno = gen_helper_gvec_abs8,
2581           .opt_opc = vecop_list,
2582           .vece = MO_8 },
2583         { .fni8 = tcg_gen_vec_abs16_i64,
2584           .fniv = tcg_gen_abs_vec,
2585           .fno = gen_helper_gvec_abs16,
2586           .opt_opc = vecop_list,
2587           .vece = MO_16 },
2588         { .fni4 = tcg_gen_abs_i32,
2589           .fniv = tcg_gen_abs_vec,
2590           .fno = gen_helper_gvec_abs32,
2591           .opt_opc = vecop_list,
2592           .vece = MO_32 },
2593         { .fni8 = tcg_gen_abs_i64,
2594           .fniv = tcg_gen_abs_vec,
2595           .fno = gen_helper_gvec_abs64,
2596           .opt_opc = vecop_list,
2597           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2598           .vece = MO_64 },
2599     };
2600 
2601     tcg_debug_assert(vece <= MO_64);
2602     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g[vece]);
2603 }
2604 
2605 void tcg_gen_gvec_and(unsigned vece, uint32_t dofs, uint32_t aofs,
2606                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2607 {
2608     static const GVecGen3 g = {
2609         .fni8 = tcg_gen_and_i64,
2610         .fniv = tcg_gen_and_vec,
2611         .fno = gen_helper_gvec_and,
2612         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2613     };
2614 
2615     if (aofs == bofs) {
2616         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2617     } else {
2618         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2619     }
2620 }
2621 
2622 void tcg_gen_gvec_or(unsigned vece, uint32_t dofs, uint32_t aofs,
2623                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2624 {
2625     static const GVecGen3 g = {
2626         .fni8 = tcg_gen_or_i64,
2627         .fniv = tcg_gen_or_vec,
2628         .fno = gen_helper_gvec_or,
2629         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2630     };
2631 
2632     if (aofs == bofs) {
2633         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2634     } else {
2635         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2636     }
2637 }
2638 
2639 void tcg_gen_gvec_xor(unsigned vece, uint32_t dofs, uint32_t aofs,
2640                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2641 {
2642     static const GVecGen3 g = {
2643         .fni8 = tcg_gen_xor_i64,
2644         .fniv = tcg_gen_xor_vec,
2645         .fno = gen_helper_gvec_xor,
2646         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2647     };
2648 
2649     if (aofs == bofs) {
2650         tcg_gen_gvec_dup_imm(MO_64, dofs, oprsz, maxsz, 0);
2651     } else {
2652         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2653     }
2654 }
2655 
2656 void tcg_gen_gvec_andc(unsigned vece, uint32_t dofs, uint32_t aofs,
2657                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2658 {
2659     static const GVecGen3 g = {
2660         .fni8 = tcg_gen_andc_i64,
2661         .fniv = tcg_gen_andc_vec,
2662         .fno = gen_helper_gvec_andc,
2663         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2664     };
2665 
2666     if (aofs == bofs) {
2667         tcg_gen_gvec_dup_imm(MO_64, dofs, oprsz, maxsz, 0);
2668     } else {
2669         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2670     }
2671 }
2672 
2673 void tcg_gen_gvec_orc(unsigned vece, uint32_t dofs, uint32_t aofs,
2674                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2675 {
2676     static const GVecGen3 g = {
2677         .fni8 = tcg_gen_orc_i64,
2678         .fniv = tcg_gen_orc_vec,
2679         .fno = gen_helper_gvec_orc,
2680         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2681     };
2682 
2683     if (aofs == bofs) {
2684         tcg_gen_gvec_dup_imm(MO_64, dofs, oprsz, maxsz, -1);
2685     } else {
2686         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2687     }
2688 }
2689 
2690 void tcg_gen_gvec_nand(unsigned vece, uint32_t dofs, uint32_t aofs,
2691                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2692 {
2693     static const GVecGen3 g = {
2694         .fni8 = tcg_gen_nand_i64,
2695         .fniv = tcg_gen_nand_vec,
2696         .fno = gen_helper_gvec_nand,
2697         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2698     };
2699 
2700     if (aofs == bofs) {
2701         tcg_gen_gvec_not(vece, dofs, aofs, oprsz, maxsz);
2702     } else {
2703         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2704     }
2705 }
2706 
2707 void tcg_gen_gvec_nor(unsigned vece, uint32_t dofs, uint32_t aofs,
2708                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2709 {
2710     static const GVecGen3 g = {
2711         .fni8 = tcg_gen_nor_i64,
2712         .fniv = tcg_gen_nor_vec,
2713         .fno = gen_helper_gvec_nor,
2714         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2715     };
2716 
2717     if (aofs == bofs) {
2718         tcg_gen_gvec_not(vece, dofs, aofs, oprsz, maxsz);
2719     } else {
2720         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2721     }
2722 }
2723 
2724 void tcg_gen_gvec_eqv(unsigned vece, uint32_t dofs, uint32_t aofs,
2725                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2726 {
2727     static const GVecGen3 g = {
2728         .fni8 = tcg_gen_eqv_i64,
2729         .fniv = tcg_gen_eqv_vec,
2730         .fno = gen_helper_gvec_eqv,
2731         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2732     };
2733 
2734     if (aofs == bofs) {
2735         tcg_gen_gvec_dup_imm(MO_64, dofs, oprsz, maxsz, -1);
2736     } else {
2737         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2738     }
2739 }
2740 
2741 static const GVecGen2s gop_ands = {
2742     .fni8 = tcg_gen_and_i64,
2743     .fniv = tcg_gen_and_vec,
2744     .fno = gen_helper_gvec_ands,
2745     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2746     .vece = MO_64
2747 };
2748 
2749 void tcg_gen_gvec_ands(unsigned vece, uint32_t dofs, uint32_t aofs,
2750                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2751 {
2752     TCGv_i64 tmp = tcg_temp_new_i64();
2753     tcg_gen_dup_i64(vece, tmp, c);
2754     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ands);
2755     tcg_temp_free_i64(tmp);
2756 }
2757 
2758 void tcg_gen_gvec_andi(unsigned vece, uint32_t dofs, uint32_t aofs,
2759                        int64_t c, uint32_t oprsz, uint32_t maxsz)
2760 {
2761     TCGv_i64 tmp = tcg_constant_i64(dup_const(vece, c));
2762     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ands);
2763 }
2764 
2765 static const GVecGen2s gop_xors = {
2766     .fni8 = tcg_gen_xor_i64,
2767     .fniv = tcg_gen_xor_vec,
2768     .fno = gen_helper_gvec_xors,
2769     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2770     .vece = MO_64
2771 };
2772 
2773 void tcg_gen_gvec_xors(unsigned vece, uint32_t dofs, uint32_t aofs,
2774                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2775 {
2776     TCGv_i64 tmp = tcg_temp_new_i64();
2777     tcg_gen_dup_i64(vece, tmp, c);
2778     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_xors);
2779     tcg_temp_free_i64(tmp);
2780 }
2781 
2782 void tcg_gen_gvec_xori(unsigned vece, uint32_t dofs, uint32_t aofs,
2783                        int64_t c, uint32_t oprsz, uint32_t maxsz)
2784 {
2785     TCGv_i64 tmp = tcg_constant_i64(dup_const(vece, c));
2786     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_xors);
2787 }
2788 
2789 static const GVecGen2s gop_ors = {
2790     .fni8 = tcg_gen_or_i64,
2791     .fniv = tcg_gen_or_vec,
2792     .fno = gen_helper_gvec_ors,
2793     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2794     .vece = MO_64
2795 };
2796 
2797 void tcg_gen_gvec_ors(unsigned vece, uint32_t dofs, uint32_t aofs,
2798                       TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2799 {
2800     TCGv_i64 tmp = tcg_temp_new_i64();
2801     tcg_gen_dup_i64(vece, tmp, c);
2802     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ors);
2803     tcg_temp_free_i64(tmp);
2804 }
2805 
2806 void tcg_gen_gvec_ori(unsigned vece, uint32_t dofs, uint32_t aofs,
2807                       int64_t c, uint32_t oprsz, uint32_t maxsz)
2808 {
2809     TCGv_i64 tmp = tcg_constant_i64(dup_const(vece, c));
2810     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ors);
2811 }
2812 
2813 void tcg_gen_vec_shl8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2814 {
2815     uint64_t mask = dup_const(MO_8, 0xff << c);
2816     tcg_gen_shli_i64(d, a, c);
2817     tcg_gen_andi_i64(d, d, mask);
2818 }
2819 
2820 void tcg_gen_vec_shl16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2821 {
2822     uint64_t mask = dup_const(MO_16, 0xffff << c);
2823     tcg_gen_shli_i64(d, a, c);
2824     tcg_gen_andi_i64(d, d, mask);
2825 }
2826 
2827 void tcg_gen_vec_shl8i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2828 {
2829     uint32_t mask = dup_const(MO_8, 0xff << c);
2830     tcg_gen_shli_i32(d, a, c);
2831     tcg_gen_andi_i32(d, d, mask);
2832 }
2833 
2834 void tcg_gen_vec_shl16i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2835 {
2836     uint32_t mask = dup_const(MO_16, 0xffff << c);
2837     tcg_gen_shli_i32(d, a, c);
2838     tcg_gen_andi_i32(d, d, mask);
2839 }
2840 
2841 void tcg_gen_gvec_shli(unsigned vece, uint32_t dofs, uint32_t aofs,
2842                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
2843 {
2844     static const TCGOpcode vecop_list[] = { INDEX_op_shli_vec, 0 };
2845     static const GVecGen2i g[4] = {
2846         { .fni8 = tcg_gen_vec_shl8i_i64,
2847           .fniv = tcg_gen_shli_vec,
2848           .fno = gen_helper_gvec_shl8i,
2849           .opt_opc = vecop_list,
2850           .vece = MO_8 },
2851         { .fni8 = tcg_gen_vec_shl16i_i64,
2852           .fniv = tcg_gen_shli_vec,
2853           .fno = gen_helper_gvec_shl16i,
2854           .opt_opc = vecop_list,
2855           .vece = MO_16 },
2856         { .fni4 = tcg_gen_shli_i32,
2857           .fniv = tcg_gen_shli_vec,
2858           .fno = gen_helper_gvec_shl32i,
2859           .opt_opc = vecop_list,
2860           .vece = MO_32 },
2861         { .fni8 = tcg_gen_shli_i64,
2862           .fniv = tcg_gen_shli_vec,
2863           .fno = gen_helper_gvec_shl64i,
2864           .opt_opc = vecop_list,
2865           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2866           .vece = MO_64 },
2867     };
2868 
2869     tcg_debug_assert(vece <= MO_64);
2870     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2871     if (shift == 0) {
2872         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2873     } else {
2874         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2875     }
2876 }
2877 
2878 void tcg_gen_vec_shr8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2879 {
2880     uint64_t mask = dup_const(MO_8, 0xff >> c);
2881     tcg_gen_shri_i64(d, a, c);
2882     tcg_gen_andi_i64(d, d, mask);
2883 }
2884 
2885 void tcg_gen_vec_shr16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2886 {
2887     uint64_t mask = dup_const(MO_16, 0xffff >> c);
2888     tcg_gen_shri_i64(d, a, c);
2889     tcg_gen_andi_i64(d, d, mask);
2890 }
2891 
2892 void tcg_gen_vec_shr8i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2893 {
2894     uint32_t mask = dup_const(MO_8, 0xff >> c);
2895     tcg_gen_shri_i32(d, a, c);
2896     tcg_gen_andi_i32(d, d, mask);
2897 }
2898 
2899 void tcg_gen_vec_shr16i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2900 {
2901     uint32_t mask = dup_const(MO_16, 0xffff >> c);
2902     tcg_gen_shri_i32(d, a, c);
2903     tcg_gen_andi_i32(d, d, mask);
2904 }
2905 
2906 void tcg_gen_gvec_shri(unsigned vece, uint32_t dofs, uint32_t aofs,
2907                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
2908 {
2909     static const TCGOpcode vecop_list[] = { INDEX_op_shri_vec, 0 };
2910     static const GVecGen2i g[4] = {
2911         { .fni8 = tcg_gen_vec_shr8i_i64,
2912           .fniv = tcg_gen_shri_vec,
2913           .fno = gen_helper_gvec_shr8i,
2914           .opt_opc = vecop_list,
2915           .vece = MO_8 },
2916         { .fni8 = tcg_gen_vec_shr16i_i64,
2917           .fniv = tcg_gen_shri_vec,
2918           .fno = gen_helper_gvec_shr16i,
2919           .opt_opc = vecop_list,
2920           .vece = MO_16 },
2921         { .fni4 = tcg_gen_shri_i32,
2922           .fniv = tcg_gen_shri_vec,
2923           .fno = gen_helper_gvec_shr32i,
2924           .opt_opc = vecop_list,
2925           .vece = MO_32 },
2926         { .fni8 = tcg_gen_shri_i64,
2927           .fniv = tcg_gen_shri_vec,
2928           .fno = gen_helper_gvec_shr64i,
2929           .opt_opc = vecop_list,
2930           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2931           .vece = MO_64 },
2932     };
2933 
2934     tcg_debug_assert(vece <= MO_64);
2935     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2936     if (shift == 0) {
2937         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2938     } else {
2939         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2940     }
2941 }
2942 
2943 void tcg_gen_vec_sar8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2944 {
2945     uint64_t s_mask = dup_const(MO_8, 0x80 >> c);
2946     uint64_t c_mask = dup_const(MO_8, 0xff >> c);
2947     TCGv_i64 s = tcg_temp_new_i64();
2948 
2949     tcg_gen_shri_i64(d, a, c);
2950     tcg_gen_andi_i64(s, d, s_mask);  /* isolate (shifted) sign bit */
2951     tcg_gen_muli_i64(s, s, (2 << c) - 2); /* replicate isolated signs */
2952     tcg_gen_andi_i64(d, d, c_mask);  /* clear out bits above sign  */
2953     tcg_gen_or_i64(d, d, s);         /* include sign extension */
2954     tcg_temp_free_i64(s);
2955 }
2956 
2957 void tcg_gen_vec_sar16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2958 {
2959     uint64_t s_mask = dup_const(MO_16, 0x8000 >> c);
2960     uint64_t c_mask = dup_const(MO_16, 0xffff >> c);
2961     TCGv_i64 s = tcg_temp_new_i64();
2962 
2963     tcg_gen_shri_i64(d, a, c);
2964     tcg_gen_andi_i64(s, d, s_mask);  /* isolate (shifted) sign bit */
2965     tcg_gen_andi_i64(d, d, c_mask);  /* clear out bits above sign  */
2966     tcg_gen_muli_i64(s, s, (2 << c) - 2); /* replicate isolated signs */
2967     tcg_gen_or_i64(d, d, s);         /* include sign extension */
2968     tcg_temp_free_i64(s);
2969 }
2970 
2971 void tcg_gen_vec_sar8i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2972 {
2973     uint32_t s_mask = dup_const(MO_8, 0x80 >> c);
2974     uint32_t c_mask = dup_const(MO_8, 0xff >> c);
2975     TCGv_i32 s = tcg_temp_new_i32();
2976 
2977     tcg_gen_shri_i32(d, a, c);
2978     tcg_gen_andi_i32(s, d, s_mask);  /* isolate (shifted) sign bit */
2979     tcg_gen_muli_i32(s, s, (2 << c) - 2); /* replicate isolated signs */
2980     tcg_gen_andi_i32(d, d, c_mask);  /* clear out bits above sign  */
2981     tcg_gen_or_i32(d, d, s);         /* include sign extension */
2982     tcg_temp_free_i32(s);
2983 }
2984 
2985 void tcg_gen_vec_sar16i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2986 {
2987     uint32_t s_mask = dup_const(MO_16, 0x8000 >> c);
2988     uint32_t c_mask = dup_const(MO_16, 0xffff >> c);
2989     TCGv_i32 s = tcg_temp_new_i32();
2990 
2991     tcg_gen_shri_i32(d, a, c);
2992     tcg_gen_andi_i32(s, d, s_mask);  /* isolate (shifted) sign bit */
2993     tcg_gen_andi_i32(d, d, c_mask);  /* clear out bits above sign  */
2994     tcg_gen_muli_i32(s, s, (2 << c) - 2); /* replicate isolated signs */
2995     tcg_gen_or_i32(d, d, s);         /* include sign extension */
2996     tcg_temp_free_i32(s);
2997 }
2998 
2999 void tcg_gen_gvec_sari(unsigned vece, uint32_t dofs, uint32_t aofs,
3000                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
3001 {
3002     static const TCGOpcode vecop_list[] = { INDEX_op_sari_vec, 0 };
3003     static const GVecGen2i g[4] = {
3004         { .fni8 = tcg_gen_vec_sar8i_i64,
3005           .fniv = tcg_gen_sari_vec,
3006           .fno = gen_helper_gvec_sar8i,
3007           .opt_opc = vecop_list,
3008           .vece = MO_8 },
3009         { .fni8 = tcg_gen_vec_sar16i_i64,
3010           .fniv = tcg_gen_sari_vec,
3011           .fno = gen_helper_gvec_sar16i,
3012           .opt_opc = vecop_list,
3013           .vece = MO_16 },
3014         { .fni4 = tcg_gen_sari_i32,
3015           .fniv = tcg_gen_sari_vec,
3016           .fno = gen_helper_gvec_sar32i,
3017           .opt_opc = vecop_list,
3018           .vece = MO_32 },
3019         { .fni8 = tcg_gen_sari_i64,
3020           .fniv = tcg_gen_sari_vec,
3021           .fno = gen_helper_gvec_sar64i,
3022           .opt_opc = vecop_list,
3023           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3024           .vece = MO_64 },
3025     };
3026 
3027     tcg_debug_assert(vece <= MO_64);
3028     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
3029     if (shift == 0) {
3030         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
3031     } else {
3032         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
3033     }
3034 }
3035 
3036 void tcg_gen_vec_rotl8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
3037 {
3038     uint64_t mask = dup_const(MO_8, 0xff << c);
3039 
3040     tcg_gen_shli_i64(d, a, c);
3041     tcg_gen_shri_i64(a, a, 8 - c);
3042     tcg_gen_andi_i64(d, d, mask);
3043     tcg_gen_andi_i64(a, a, ~mask);
3044     tcg_gen_or_i64(d, d, a);
3045 }
3046 
3047 void tcg_gen_vec_rotl16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
3048 {
3049     uint64_t mask = dup_const(MO_16, 0xffff << c);
3050 
3051     tcg_gen_shli_i64(d, a, c);
3052     tcg_gen_shri_i64(a, a, 16 - c);
3053     tcg_gen_andi_i64(d, d, mask);
3054     tcg_gen_andi_i64(a, a, ~mask);
3055     tcg_gen_or_i64(d, d, a);
3056 }
3057 
3058 void tcg_gen_gvec_rotli(unsigned vece, uint32_t dofs, uint32_t aofs,
3059                         int64_t shift, uint32_t oprsz, uint32_t maxsz)
3060 {
3061     static const TCGOpcode vecop_list[] = { INDEX_op_rotli_vec, 0 };
3062     static const GVecGen2i g[4] = {
3063         { .fni8 = tcg_gen_vec_rotl8i_i64,
3064           .fniv = tcg_gen_rotli_vec,
3065           .fno = gen_helper_gvec_rotl8i,
3066           .opt_opc = vecop_list,
3067           .vece = MO_8 },
3068         { .fni8 = tcg_gen_vec_rotl16i_i64,
3069           .fniv = tcg_gen_rotli_vec,
3070           .fno = gen_helper_gvec_rotl16i,
3071           .opt_opc = vecop_list,
3072           .vece = MO_16 },
3073         { .fni4 = tcg_gen_rotli_i32,
3074           .fniv = tcg_gen_rotli_vec,
3075           .fno = gen_helper_gvec_rotl32i,
3076           .opt_opc = vecop_list,
3077           .vece = MO_32 },
3078         { .fni8 = tcg_gen_rotli_i64,
3079           .fniv = tcg_gen_rotli_vec,
3080           .fno = gen_helper_gvec_rotl64i,
3081           .opt_opc = vecop_list,
3082           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3083           .vece = MO_64 },
3084     };
3085 
3086     tcg_debug_assert(vece <= MO_64);
3087     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
3088     if (shift == 0) {
3089         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
3090     } else {
3091         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
3092     }
3093 }
3094 
3095 void tcg_gen_gvec_rotri(unsigned vece, uint32_t dofs, uint32_t aofs,
3096                         int64_t shift, uint32_t oprsz, uint32_t maxsz)
3097 {
3098     tcg_debug_assert(vece <= MO_64);
3099     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
3100     tcg_gen_gvec_rotli(vece, dofs, aofs, -shift & ((8 << vece) - 1),
3101                        oprsz, maxsz);
3102 }
3103 
3104 /*
3105  * Specialized generation vector shifts by a non-constant scalar.
3106  */
3107 
3108 typedef struct {
3109     void (*fni4)(TCGv_i32, TCGv_i32, TCGv_i32);
3110     void (*fni8)(TCGv_i64, TCGv_i64, TCGv_i64);
3111     void (*fniv_s)(unsigned, TCGv_vec, TCGv_vec, TCGv_i32);
3112     void (*fniv_v)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec);
3113     gen_helper_gvec_2 *fno[4];
3114     TCGOpcode s_list[2];
3115     TCGOpcode v_list[2];
3116 } GVecGen2sh;
3117 
3118 static void expand_2sh_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
3119                            uint32_t oprsz, uint32_t tysz, TCGType type,
3120                            TCGv_i32 shift,
3121                            void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_i32))
3122 {
3123     TCGv_vec t0 = tcg_temp_new_vec(type);
3124     uint32_t i;
3125 
3126     for (i = 0; i < oprsz; i += tysz) {
3127         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
3128         fni(vece, t0, t0, shift);
3129         tcg_gen_st_vec(t0, cpu_env, dofs + i);
3130     }
3131     tcg_temp_free_vec(t0);
3132 }
3133 
3134 static void
3135 do_gvec_shifts(unsigned vece, uint32_t dofs, uint32_t aofs, TCGv_i32 shift,
3136                uint32_t oprsz, uint32_t maxsz, const GVecGen2sh *g)
3137 {
3138     TCGType type;
3139     uint32_t some;
3140 
3141     check_size_align(oprsz, maxsz, dofs | aofs);
3142     check_overlap_2(dofs, aofs, maxsz);
3143 
3144     /* If the backend has a scalar expansion, great.  */
3145     type = choose_vector_type(g->s_list, vece, oprsz, vece == MO_64);
3146     if (type) {
3147         const TCGOpcode *hold_list = tcg_swap_vecop_list(NULL);
3148         switch (type) {
3149         case TCG_TYPE_V256:
3150             some = QEMU_ALIGN_DOWN(oprsz, 32);
3151             expand_2sh_vec(vece, dofs, aofs, some, 32,
3152                            TCG_TYPE_V256, shift, g->fniv_s);
3153             if (some == oprsz) {
3154                 break;
3155             }
3156             dofs += some;
3157             aofs += some;
3158             oprsz -= some;
3159             maxsz -= some;
3160             /* fallthru */
3161         case TCG_TYPE_V128:
3162             expand_2sh_vec(vece, dofs, aofs, oprsz, 16,
3163                            TCG_TYPE_V128, shift, g->fniv_s);
3164             break;
3165         case TCG_TYPE_V64:
3166             expand_2sh_vec(vece, dofs, aofs, oprsz, 8,
3167                            TCG_TYPE_V64, shift, g->fniv_s);
3168             break;
3169         default:
3170             g_assert_not_reached();
3171         }
3172         tcg_swap_vecop_list(hold_list);
3173         goto clear_tail;
3174     }
3175 
3176     /* If the backend supports variable vector shifts, also cool.  */
3177     type = choose_vector_type(g->v_list, vece, oprsz, vece == MO_64);
3178     if (type) {
3179         const TCGOpcode *hold_list = tcg_swap_vecop_list(NULL);
3180         TCGv_vec v_shift = tcg_temp_new_vec(type);
3181 
3182         if (vece == MO_64) {
3183             TCGv_i64 sh64 = tcg_temp_new_i64();
3184             tcg_gen_extu_i32_i64(sh64, shift);
3185             tcg_gen_dup_i64_vec(MO_64, v_shift, sh64);
3186             tcg_temp_free_i64(sh64);
3187         } else {
3188             tcg_gen_dup_i32_vec(vece, v_shift, shift);
3189         }
3190 
3191         switch (type) {
3192         case TCG_TYPE_V256:
3193             some = QEMU_ALIGN_DOWN(oprsz, 32);
3194             expand_2s_vec(vece, dofs, aofs, some, 32, TCG_TYPE_V256,
3195                           v_shift, false, g->fniv_v);
3196             if (some == oprsz) {
3197                 break;
3198             }
3199             dofs += some;
3200             aofs += some;
3201             oprsz -= some;
3202             maxsz -= some;
3203             /* fallthru */
3204         case TCG_TYPE_V128:
3205             expand_2s_vec(vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
3206                           v_shift, false, g->fniv_v);
3207             break;
3208         case TCG_TYPE_V64:
3209             expand_2s_vec(vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
3210                           v_shift, false, g->fniv_v);
3211             break;
3212         default:
3213             g_assert_not_reached();
3214         }
3215         tcg_temp_free_vec(v_shift);
3216         tcg_swap_vecop_list(hold_list);
3217         goto clear_tail;
3218     }
3219 
3220     /* Otherwise fall back to integral... */
3221     if (vece == MO_32 && check_size_impl(oprsz, 4)) {
3222         expand_2s_i32(dofs, aofs, oprsz, shift, false, g->fni4);
3223     } else if (vece == MO_64 && check_size_impl(oprsz, 8)) {
3224         TCGv_i64 sh64 = tcg_temp_new_i64();
3225         tcg_gen_extu_i32_i64(sh64, shift);
3226         expand_2s_i64(dofs, aofs, oprsz, sh64, false, g->fni8);
3227         tcg_temp_free_i64(sh64);
3228     } else {
3229         TCGv_ptr a0 = tcg_temp_new_ptr();
3230         TCGv_ptr a1 = tcg_temp_new_ptr();
3231         TCGv_i32 desc = tcg_temp_new_i32();
3232 
3233         tcg_gen_shli_i32(desc, shift, SIMD_DATA_SHIFT);
3234         tcg_gen_ori_i32(desc, desc, simd_desc(oprsz, maxsz, 0));
3235         tcg_gen_addi_ptr(a0, cpu_env, dofs);
3236         tcg_gen_addi_ptr(a1, cpu_env, aofs);
3237 
3238         g->fno[vece](a0, a1, desc);
3239 
3240         tcg_temp_free_ptr(a0);
3241         tcg_temp_free_ptr(a1);
3242         tcg_temp_free_i32(desc);
3243         return;
3244     }
3245 
3246  clear_tail:
3247     if (oprsz < maxsz) {
3248         expand_clr(dofs + oprsz, maxsz - oprsz);
3249     }
3250 }
3251 
3252 void tcg_gen_gvec_shls(unsigned vece, uint32_t dofs, uint32_t aofs,
3253                        TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3254 {
3255     static const GVecGen2sh g = {
3256         .fni4 = tcg_gen_shl_i32,
3257         .fni8 = tcg_gen_shl_i64,
3258         .fniv_s = tcg_gen_shls_vec,
3259         .fniv_v = tcg_gen_shlv_vec,
3260         .fno = {
3261             gen_helper_gvec_shl8i,
3262             gen_helper_gvec_shl16i,
3263             gen_helper_gvec_shl32i,
3264             gen_helper_gvec_shl64i,
3265         },
3266         .s_list = { INDEX_op_shls_vec, 0 },
3267         .v_list = { INDEX_op_shlv_vec, 0 },
3268     };
3269 
3270     tcg_debug_assert(vece <= MO_64);
3271     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
3272 }
3273 
3274 void tcg_gen_gvec_shrs(unsigned vece, uint32_t dofs, uint32_t aofs,
3275                        TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3276 {
3277     static const GVecGen2sh g = {
3278         .fni4 = tcg_gen_shr_i32,
3279         .fni8 = tcg_gen_shr_i64,
3280         .fniv_s = tcg_gen_shrs_vec,
3281         .fniv_v = tcg_gen_shrv_vec,
3282         .fno = {
3283             gen_helper_gvec_shr8i,
3284             gen_helper_gvec_shr16i,
3285             gen_helper_gvec_shr32i,
3286             gen_helper_gvec_shr64i,
3287         },
3288         .s_list = { INDEX_op_shrs_vec, 0 },
3289         .v_list = { INDEX_op_shrv_vec, 0 },
3290     };
3291 
3292     tcg_debug_assert(vece <= MO_64);
3293     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
3294 }
3295 
3296 void tcg_gen_gvec_sars(unsigned vece, uint32_t dofs, uint32_t aofs,
3297                        TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3298 {
3299     static const GVecGen2sh g = {
3300         .fni4 = tcg_gen_sar_i32,
3301         .fni8 = tcg_gen_sar_i64,
3302         .fniv_s = tcg_gen_sars_vec,
3303         .fniv_v = tcg_gen_sarv_vec,
3304         .fno = {
3305             gen_helper_gvec_sar8i,
3306             gen_helper_gvec_sar16i,
3307             gen_helper_gvec_sar32i,
3308             gen_helper_gvec_sar64i,
3309         },
3310         .s_list = { INDEX_op_sars_vec, 0 },
3311         .v_list = { INDEX_op_sarv_vec, 0 },
3312     };
3313 
3314     tcg_debug_assert(vece <= MO_64);
3315     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
3316 }
3317 
3318 void tcg_gen_gvec_rotls(unsigned vece, uint32_t dofs, uint32_t aofs,
3319                         TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3320 {
3321     static const GVecGen2sh g = {
3322         .fni4 = tcg_gen_rotl_i32,
3323         .fni8 = tcg_gen_rotl_i64,
3324         .fniv_s = tcg_gen_rotls_vec,
3325         .fniv_v = tcg_gen_rotlv_vec,
3326         .fno = {
3327             gen_helper_gvec_rotl8i,
3328             gen_helper_gvec_rotl16i,
3329             gen_helper_gvec_rotl32i,
3330             gen_helper_gvec_rotl64i,
3331         },
3332         .s_list = { INDEX_op_rotls_vec, 0 },
3333         .v_list = { INDEX_op_rotlv_vec, 0 },
3334     };
3335 
3336     tcg_debug_assert(vece <= MO_64);
3337     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
3338 }
3339 
3340 /*
3341  * Expand D = A << (B % element bits)
3342  *
3343  * Unlike scalar shifts, where it is easy for the target front end
3344  * to include the modulo as part of the expansion.  If the target
3345  * naturally includes the modulo as part of the operation, great!
3346  * If the target has some other behaviour from out-of-range shifts,
3347  * then it could not use this function anyway, and would need to
3348  * do it's own expansion with custom functions.
3349  */
3350 static void tcg_gen_shlv_mod_vec(unsigned vece, TCGv_vec d,
3351                                  TCGv_vec a, TCGv_vec b)
3352 {
3353     TCGv_vec t = tcg_temp_new_vec_matching(d);
3354     TCGv_vec m = tcg_constant_vec_matching(d, vece, (8 << vece) - 1);
3355 
3356     tcg_gen_and_vec(vece, t, b, m);
3357     tcg_gen_shlv_vec(vece, d, a, t);
3358     tcg_temp_free_vec(t);
3359 }
3360 
3361 static void tcg_gen_shl_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3362 {
3363     TCGv_i32 t = tcg_temp_new_i32();
3364 
3365     tcg_gen_andi_i32(t, b, 31);
3366     tcg_gen_shl_i32(d, a, t);
3367     tcg_temp_free_i32(t);
3368 }
3369 
3370 static void tcg_gen_shl_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3371 {
3372     TCGv_i64 t = tcg_temp_new_i64();
3373 
3374     tcg_gen_andi_i64(t, b, 63);
3375     tcg_gen_shl_i64(d, a, t);
3376     tcg_temp_free_i64(t);
3377 }
3378 
3379 void tcg_gen_gvec_shlv(unsigned vece, uint32_t dofs, uint32_t aofs,
3380                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3381 {
3382     static const TCGOpcode vecop_list[] = { INDEX_op_shlv_vec, 0 };
3383     static const GVecGen3 g[4] = {
3384         { .fniv = tcg_gen_shlv_mod_vec,
3385           .fno = gen_helper_gvec_shl8v,
3386           .opt_opc = vecop_list,
3387           .vece = MO_8 },
3388         { .fniv = tcg_gen_shlv_mod_vec,
3389           .fno = gen_helper_gvec_shl16v,
3390           .opt_opc = vecop_list,
3391           .vece = MO_16 },
3392         { .fni4 = tcg_gen_shl_mod_i32,
3393           .fniv = tcg_gen_shlv_mod_vec,
3394           .fno = gen_helper_gvec_shl32v,
3395           .opt_opc = vecop_list,
3396           .vece = MO_32 },
3397         { .fni8 = tcg_gen_shl_mod_i64,
3398           .fniv = tcg_gen_shlv_mod_vec,
3399           .fno = gen_helper_gvec_shl64v,
3400           .opt_opc = vecop_list,
3401           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3402           .vece = MO_64 },
3403     };
3404 
3405     tcg_debug_assert(vece <= MO_64);
3406     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3407 }
3408 
3409 /*
3410  * Similarly for logical right shifts.
3411  */
3412 
3413 static void tcg_gen_shrv_mod_vec(unsigned vece, TCGv_vec d,
3414                                  TCGv_vec a, TCGv_vec b)
3415 {
3416     TCGv_vec t = tcg_temp_new_vec_matching(d);
3417     TCGv_vec m = tcg_constant_vec_matching(d, vece, (8 << vece) - 1);
3418 
3419     tcg_gen_and_vec(vece, t, b, m);
3420     tcg_gen_shrv_vec(vece, d, a, t);
3421     tcg_temp_free_vec(t);
3422 }
3423 
3424 static void tcg_gen_shr_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3425 {
3426     TCGv_i32 t = tcg_temp_new_i32();
3427 
3428     tcg_gen_andi_i32(t, b, 31);
3429     tcg_gen_shr_i32(d, a, t);
3430     tcg_temp_free_i32(t);
3431 }
3432 
3433 static void tcg_gen_shr_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3434 {
3435     TCGv_i64 t = tcg_temp_new_i64();
3436 
3437     tcg_gen_andi_i64(t, b, 63);
3438     tcg_gen_shr_i64(d, a, t);
3439     tcg_temp_free_i64(t);
3440 }
3441 
3442 void tcg_gen_gvec_shrv(unsigned vece, uint32_t dofs, uint32_t aofs,
3443                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3444 {
3445     static const TCGOpcode vecop_list[] = { INDEX_op_shrv_vec, 0 };
3446     static const GVecGen3 g[4] = {
3447         { .fniv = tcg_gen_shrv_mod_vec,
3448           .fno = gen_helper_gvec_shr8v,
3449           .opt_opc = vecop_list,
3450           .vece = MO_8 },
3451         { .fniv = tcg_gen_shrv_mod_vec,
3452           .fno = gen_helper_gvec_shr16v,
3453           .opt_opc = vecop_list,
3454           .vece = MO_16 },
3455         { .fni4 = tcg_gen_shr_mod_i32,
3456           .fniv = tcg_gen_shrv_mod_vec,
3457           .fno = gen_helper_gvec_shr32v,
3458           .opt_opc = vecop_list,
3459           .vece = MO_32 },
3460         { .fni8 = tcg_gen_shr_mod_i64,
3461           .fniv = tcg_gen_shrv_mod_vec,
3462           .fno = gen_helper_gvec_shr64v,
3463           .opt_opc = vecop_list,
3464           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3465           .vece = MO_64 },
3466     };
3467 
3468     tcg_debug_assert(vece <= MO_64);
3469     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3470 }
3471 
3472 /*
3473  * Similarly for arithmetic right shifts.
3474  */
3475 
3476 static void tcg_gen_sarv_mod_vec(unsigned vece, TCGv_vec d,
3477                                  TCGv_vec a, TCGv_vec b)
3478 {
3479     TCGv_vec t = tcg_temp_new_vec_matching(d);
3480     TCGv_vec m = tcg_constant_vec_matching(d, vece, (8 << vece) - 1);
3481 
3482     tcg_gen_and_vec(vece, t, b, m);
3483     tcg_gen_sarv_vec(vece, d, a, t);
3484     tcg_temp_free_vec(t);
3485 }
3486 
3487 static void tcg_gen_sar_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3488 {
3489     TCGv_i32 t = tcg_temp_new_i32();
3490 
3491     tcg_gen_andi_i32(t, b, 31);
3492     tcg_gen_sar_i32(d, a, t);
3493     tcg_temp_free_i32(t);
3494 }
3495 
3496 static void tcg_gen_sar_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3497 {
3498     TCGv_i64 t = tcg_temp_new_i64();
3499 
3500     tcg_gen_andi_i64(t, b, 63);
3501     tcg_gen_sar_i64(d, a, t);
3502     tcg_temp_free_i64(t);
3503 }
3504 
3505 void tcg_gen_gvec_sarv(unsigned vece, uint32_t dofs, uint32_t aofs,
3506                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3507 {
3508     static const TCGOpcode vecop_list[] = { INDEX_op_sarv_vec, 0 };
3509     static const GVecGen3 g[4] = {
3510         { .fniv = tcg_gen_sarv_mod_vec,
3511           .fno = gen_helper_gvec_sar8v,
3512           .opt_opc = vecop_list,
3513           .vece = MO_8 },
3514         { .fniv = tcg_gen_sarv_mod_vec,
3515           .fno = gen_helper_gvec_sar16v,
3516           .opt_opc = vecop_list,
3517           .vece = MO_16 },
3518         { .fni4 = tcg_gen_sar_mod_i32,
3519           .fniv = tcg_gen_sarv_mod_vec,
3520           .fno = gen_helper_gvec_sar32v,
3521           .opt_opc = vecop_list,
3522           .vece = MO_32 },
3523         { .fni8 = tcg_gen_sar_mod_i64,
3524           .fniv = tcg_gen_sarv_mod_vec,
3525           .fno = gen_helper_gvec_sar64v,
3526           .opt_opc = vecop_list,
3527           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3528           .vece = MO_64 },
3529     };
3530 
3531     tcg_debug_assert(vece <= MO_64);
3532     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3533 }
3534 
3535 /*
3536  * Similarly for rotates.
3537  */
3538 
3539 static void tcg_gen_rotlv_mod_vec(unsigned vece, TCGv_vec d,
3540                                   TCGv_vec a, TCGv_vec b)
3541 {
3542     TCGv_vec t = tcg_temp_new_vec_matching(d);
3543     TCGv_vec m = tcg_constant_vec_matching(d, vece, (8 << vece) - 1);
3544 
3545     tcg_gen_and_vec(vece, t, b, m);
3546     tcg_gen_rotlv_vec(vece, d, a, t);
3547     tcg_temp_free_vec(t);
3548 }
3549 
3550 static void tcg_gen_rotl_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3551 {
3552     TCGv_i32 t = tcg_temp_new_i32();
3553 
3554     tcg_gen_andi_i32(t, b, 31);
3555     tcg_gen_rotl_i32(d, a, t);
3556     tcg_temp_free_i32(t);
3557 }
3558 
3559 static void tcg_gen_rotl_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3560 {
3561     TCGv_i64 t = tcg_temp_new_i64();
3562 
3563     tcg_gen_andi_i64(t, b, 63);
3564     tcg_gen_rotl_i64(d, a, t);
3565     tcg_temp_free_i64(t);
3566 }
3567 
3568 void tcg_gen_gvec_rotlv(unsigned vece, uint32_t dofs, uint32_t aofs,
3569                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3570 {
3571     static const TCGOpcode vecop_list[] = { INDEX_op_rotlv_vec, 0 };
3572     static const GVecGen3 g[4] = {
3573         { .fniv = tcg_gen_rotlv_mod_vec,
3574           .fno = gen_helper_gvec_rotl8v,
3575           .opt_opc = vecop_list,
3576           .vece = MO_8 },
3577         { .fniv = tcg_gen_rotlv_mod_vec,
3578           .fno = gen_helper_gvec_rotl16v,
3579           .opt_opc = vecop_list,
3580           .vece = MO_16 },
3581         { .fni4 = tcg_gen_rotl_mod_i32,
3582           .fniv = tcg_gen_rotlv_mod_vec,
3583           .fno = gen_helper_gvec_rotl32v,
3584           .opt_opc = vecop_list,
3585           .vece = MO_32 },
3586         { .fni8 = tcg_gen_rotl_mod_i64,
3587           .fniv = tcg_gen_rotlv_mod_vec,
3588           .fno = gen_helper_gvec_rotl64v,
3589           .opt_opc = vecop_list,
3590           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3591           .vece = MO_64 },
3592     };
3593 
3594     tcg_debug_assert(vece <= MO_64);
3595     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3596 }
3597 
3598 static void tcg_gen_rotrv_mod_vec(unsigned vece, TCGv_vec d,
3599                                   TCGv_vec a, TCGv_vec b)
3600 {
3601     TCGv_vec t = tcg_temp_new_vec_matching(d);
3602     TCGv_vec m = tcg_constant_vec_matching(d, vece, (8 << vece) - 1);
3603 
3604     tcg_gen_and_vec(vece, t, b, m);
3605     tcg_gen_rotrv_vec(vece, d, a, t);
3606     tcg_temp_free_vec(t);
3607 }
3608 
3609 static void tcg_gen_rotr_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3610 {
3611     TCGv_i32 t = tcg_temp_new_i32();
3612 
3613     tcg_gen_andi_i32(t, b, 31);
3614     tcg_gen_rotr_i32(d, a, t);
3615     tcg_temp_free_i32(t);
3616 }
3617 
3618 static void tcg_gen_rotr_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3619 {
3620     TCGv_i64 t = tcg_temp_new_i64();
3621 
3622     tcg_gen_andi_i64(t, b, 63);
3623     tcg_gen_rotr_i64(d, a, t);
3624     tcg_temp_free_i64(t);
3625 }
3626 
3627 void tcg_gen_gvec_rotrv(unsigned vece, uint32_t dofs, uint32_t aofs,
3628                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3629 {
3630     static const TCGOpcode vecop_list[] = { INDEX_op_rotrv_vec, 0 };
3631     static const GVecGen3 g[4] = {
3632         { .fniv = tcg_gen_rotrv_mod_vec,
3633           .fno = gen_helper_gvec_rotr8v,
3634           .opt_opc = vecop_list,
3635           .vece = MO_8 },
3636         { .fniv = tcg_gen_rotrv_mod_vec,
3637           .fno = gen_helper_gvec_rotr16v,
3638           .opt_opc = vecop_list,
3639           .vece = MO_16 },
3640         { .fni4 = tcg_gen_rotr_mod_i32,
3641           .fniv = tcg_gen_rotrv_mod_vec,
3642           .fno = gen_helper_gvec_rotr32v,
3643           .opt_opc = vecop_list,
3644           .vece = MO_32 },
3645         { .fni8 = tcg_gen_rotr_mod_i64,
3646           .fniv = tcg_gen_rotrv_mod_vec,
3647           .fno = gen_helper_gvec_rotr64v,
3648           .opt_opc = vecop_list,
3649           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3650           .vece = MO_64 },
3651     };
3652 
3653     tcg_debug_assert(vece <= MO_64);
3654     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3655 }
3656 
3657 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
3658 static void expand_cmp_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
3659                            uint32_t oprsz, TCGCond cond)
3660 {
3661     TCGv_i32 t0 = tcg_temp_new_i32();
3662     TCGv_i32 t1 = tcg_temp_new_i32();
3663     uint32_t i;
3664 
3665     for (i = 0; i < oprsz; i += 4) {
3666         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
3667         tcg_gen_ld_i32(t1, cpu_env, bofs + i);
3668         tcg_gen_setcond_i32(cond, t0, t0, t1);
3669         tcg_gen_neg_i32(t0, t0);
3670         tcg_gen_st_i32(t0, cpu_env, dofs + i);
3671     }
3672     tcg_temp_free_i32(t1);
3673     tcg_temp_free_i32(t0);
3674 }
3675 
3676 static void expand_cmp_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
3677                            uint32_t oprsz, TCGCond cond)
3678 {
3679     TCGv_i64 t0 = tcg_temp_new_i64();
3680     TCGv_i64 t1 = tcg_temp_new_i64();
3681     uint32_t i;
3682 
3683     for (i = 0; i < oprsz; i += 8) {
3684         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
3685         tcg_gen_ld_i64(t1, cpu_env, bofs + i);
3686         tcg_gen_setcond_i64(cond, t0, t0, t1);
3687         tcg_gen_neg_i64(t0, t0);
3688         tcg_gen_st_i64(t0, cpu_env, dofs + i);
3689     }
3690     tcg_temp_free_i64(t1);
3691     tcg_temp_free_i64(t0);
3692 }
3693 
3694 static void expand_cmp_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
3695                            uint32_t bofs, uint32_t oprsz, uint32_t tysz,
3696                            TCGType type, TCGCond cond)
3697 {
3698     TCGv_vec t0 = tcg_temp_new_vec(type);
3699     TCGv_vec t1 = tcg_temp_new_vec(type);
3700     uint32_t i;
3701 
3702     for (i = 0; i < oprsz; i += tysz) {
3703         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
3704         tcg_gen_ld_vec(t1, cpu_env, bofs + i);
3705         tcg_gen_cmp_vec(cond, vece, t0, t0, t1);
3706         tcg_gen_st_vec(t0, cpu_env, dofs + i);
3707     }
3708     tcg_temp_free_vec(t1);
3709     tcg_temp_free_vec(t0);
3710 }
3711 
3712 void tcg_gen_gvec_cmp(TCGCond cond, unsigned vece, uint32_t dofs,
3713                       uint32_t aofs, uint32_t bofs,
3714                       uint32_t oprsz, uint32_t maxsz)
3715 {
3716     static const TCGOpcode cmp_list[] = { INDEX_op_cmp_vec, 0 };
3717     static gen_helper_gvec_3 * const eq_fn[4] = {
3718         gen_helper_gvec_eq8, gen_helper_gvec_eq16,
3719         gen_helper_gvec_eq32, gen_helper_gvec_eq64
3720     };
3721     static gen_helper_gvec_3 * const ne_fn[4] = {
3722         gen_helper_gvec_ne8, gen_helper_gvec_ne16,
3723         gen_helper_gvec_ne32, gen_helper_gvec_ne64
3724     };
3725     static gen_helper_gvec_3 * const lt_fn[4] = {
3726         gen_helper_gvec_lt8, gen_helper_gvec_lt16,
3727         gen_helper_gvec_lt32, gen_helper_gvec_lt64
3728     };
3729     static gen_helper_gvec_3 * const le_fn[4] = {
3730         gen_helper_gvec_le8, gen_helper_gvec_le16,
3731         gen_helper_gvec_le32, gen_helper_gvec_le64
3732     };
3733     static gen_helper_gvec_3 * const ltu_fn[4] = {
3734         gen_helper_gvec_ltu8, gen_helper_gvec_ltu16,
3735         gen_helper_gvec_ltu32, gen_helper_gvec_ltu64
3736     };
3737     static gen_helper_gvec_3 * const leu_fn[4] = {
3738         gen_helper_gvec_leu8, gen_helper_gvec_leu16,
3739         gen_helper_gvec_leu32, gen_helper_gvec_leu64
3740     };
3741     static gen_helper_gvec_3 * const * const fns[16] = {
3742         [TCG_COND_EQ] = eq_fn,
3743         [TCG_COND_NE] = ne_fn,
3744         [TCG_COND_LT] = lt_fn,
3745         [TCG_COND_LE] = le_fn,
3746         [TCG_COND_LTU] = ltu_fn,
3747         [TCG_COND_LEU] = leu_fn,
3748     };
3749 
3750     const TCGOpcode *hold_list;
3751     TCGType type;
3752     uint32_t some;
3753 
3754     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
3755     check_overlap_3(dofs, aofs, bofs, maxsz);
3756 
3757     if (cond == TCG_COND_NEVER || cond == TCG_COND_ALWAYS) {
3758         do_dup(MO_8, dofs, oprsz, maxsz,
3759                NULL, NULL, -(cond == TCG_COND_ALWAYS));
3760         return;
3761     }
3762 
3763     /*
3764      * Implement inline with a vector type, if possible.
3765      * Prefer integer when 64-bit host and 64-bit comparison.
3766      */
3767     hold_list = tcg_swap_vecop_list(cmp_list);
3768     type = choose_vector_type(cmp_list, vece, oprsz,
3769                               TCG_TARGET_REG_BITS == 64 && vece == MO_64);
3770     switch (type) {
3771     case TCG_TYPE_V256:
3772         /* Recall that ARM SVE allows vector sizes that are not a
3773          * power of 2, but always a multiple of 16.  The intent is
3774          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
3775          */
3776         some = QEMU_ALIGN_DOWN(oprsz, 32);
3777         expand_cmp_vec(vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256, cond);
3778         if (some == oprsz) {
3779             break;
3780         }
3781         dofs += some;
3782         aofs += some;
3783         bofs += some;
3784         oprsz -= some;
3785         maxsz -= some;
3786         /* fallthru */
3787     case TCG_TYPE_V128:
3788         expand_cmp_vec(vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128, cond);
3789         break;
3790     case TCG_TYPE_V64:
3791         expand_cmp_vec(vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64, cond);
3792         break;
3793 
3794     case 0:
3795         if (vece == MO_64 && check_size_impl(oprsz, 8)) {
3796             expand_cmp_i64(dofs, aofs, bofs, oprsz, cond);
3797         } else if (vece == MO_32 && check_size_impl(oprsz, 4)) {
3798             expand_cmp_i32(dofs, aofs, bofs, oprsz, cond);
3799         } else {
3800             gen_helper_gvec_3 * const *fn = fns[cond];
3801 
3802             if (fn == NULL) {
3803                 uint32_t tmp;
3804                 tmp = aofs, aofs = bofs, bofs = tmp;
3805                 cond = tcg_swap_cond(cond);
3806                 fn = fns[cond];
3807                 assert(fn != NULL);
3808             }
3809             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz, maxsz, 0, fn[vece]);
3810             oprsz = maxsz;
3811         }
3812         break;
3813 
3814     default:
3815         g_assert_not_reached();
3816     }
3817     tcg_swap_vecop_list(hold_list);
3818 
3819     if (oprsz < maxsz) {
3820         expand_clr(dofs + oprsz, maxsz - oprsz);
3821     }
3822 }
3823 
3824 static void tcg_gen_bitsel_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 c)
3825 {
3826     TCGv_i64 t = tcg_temp_new_i64();
3827 
3828     tcg_gen_and_i64(t, b, a);
3829     tcg_gen_andc_i64(d, c, a);
3830     tcg_gen_or_i64(d, d, t);
3831     tcg_temp_free_i64(t);
3832 }
3833 
3834 void tcg_gen_gvec_bitsel(unsigned vece, uint32_t dofs, uint32_t aofs,
3835                          uint32_t bofs, uint32_t cofs,
3836                          uint32_t oprsz, uint32_t maxsz)
3837 {
3838     static const GVecGen4 g = {
3839         .fni8 = tcg_gen_bitsel_i64,
3840         .fniv = tcg_gen_bitsel_vec,
3841         .fno = gen_helper_gvec_bitsel,
3842     };
3843 
3844     tcg_gen_gvec_4(dofs, aofs, bofs, cofs, oprsz, maxsz, &g);
3845 }
3846