xref: /openbmc/qemu/tcg/tcg-op-gvec.c (revision a75b180f)
1 /*
2  * Generic vector operation expansion
3  *
4  * Copyright (c) 2018 Linaro
5  *
6  * This library is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2.1 of the License, or (at your option) any later version.
10  *
11  * This library is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with this library; if not, see <http://www.gnu.org/licenses/>.
18  */
19 
20 #include "qemu/osdep.h"
21 #include "tcg/tcg.h"
22 #include "tcg/tcg-temp-internal.h"
23 #include "tcg/tcg-op-common.h"
24 #include "tcg/tcg-op-gvec-common.h"
25 #include "tcg/tcg-gvec-desc.h"
26 
27 #define MAX_UNROLL  4
28 
29 #ifdef CONFIG_DEBUG_TCG
30 static const TCGOpcode vecop_list_empty[1] = { 0 };
31 #else
32 #define vecop_list_empty NULL
33 #endif
34 
35 
36 /* Verify vector size and alignment rules.  OFS should be the OR of all
37    of the operand offsets so that we can check them all at once.  */
38 static void check_size_align(uint32_t oprsz, uint32_t maxsz, uint32_t ofs)
39 {
40     uint32_t max_align;
41 
42     switch (oprsz) {
43     case 8:
44     case 16:
45     case 32:
46         tcg_debug_assert(oprsz <= maxsz);
47         break;
48     default:
49         tcg_debug_assert(oprsz == maxsz);
50         break;
51     }
52     tcg_debug_assert(maxsz <= (8 << SIMD_MAXSZ_BITS));
53 
54     max_align = maxsz >= 16 ? 15 : 7;
55     tcg_debug_assert((maxsz & max_align) == 0);
56     tcg_debug_assert((ofs & max_align) == 0);
57 }
58 
59 /* Verify vector overlap rules for two operands.  */
60 static void check_overlap_2(uint32_t d, uint32_t a, uint32_t s)
61 {
62     tcg_debug_assert(d == a || d + s <= a || a + s <= d);
63 }
64 
65 /* Verify vector overlap rules for three operands.  */
66 static void check_overlap_3(uint32_t d, uint32_t a, uint32_t b, uint32_t s)
67 {
68     check_overlap_2(d, a, s);
69     check_overlap_2(d, b, s);
70     check_overlap_2(a, b, s);
71 }
72 
73 /* Verify vector overlap rules for four operands.  */
74 static void check_overlap_4(uint32_t d, uint32_t a, uint32_t b,
75                             uint32_t c, uint32_t s)
76 {
77     check_overlap_2(d, a, s);
78     check_overlap_2(d, b, s);
79     check_overlap_2(d, c, s);
80     check_overlap_2(a, b, s);
81     check_overlap_2(a, c, s);
82     check_overlap_2(b, c, s);
83 }
84 
85 /* Create a descriptor from components.  */
86 uint32_t simd_desc(uint32_t oprsz, uint32_t maxsz, int32_t data)
87 {
88     uint32_t desc = 0;
89 
90     check_size_align(oprsz, maxsz, 0);
91     tcg_debug_assert(data == sextract32(data, 0, SIMD_DATA_BITS));
92 
93     oprsz = (oprsz / 8) - 1;
94     maxsz = (maxsz / 8) - 1;
95 
96     /*
97      * We have just asserted in check_size_align that either
98      * oprsz is {8,16,32} or matches maxsz.  Encode the final
99      * case with '2', as that would otherwise map to 24.
100      */
101     if (oprsz == maxsz) {
102         oprsz = 2;
103     }
104 
105     desc = deposit32(desc, SIMD_OPRSZ_SHIFT, SIMD_OPRSZ_BITS, oprsz);
106     desc = deposit32(desc, SIMD_MAXSZ_SHIFT, SIMD_MAXSZ_BITS, maxsz);
107     desc = deposit32(desc, SIMD_DATA_SHIFT, SIMD_DATA_BITS, data);
108 
109     return desc;
110 }
111 
112 /* Generate a call to a gvec-style helper with two vector operands.  */
113 void tcg_gen_gvec_2_ool(uint32_t dofs, uint32_t aofs,
114                         uint32_t oprsz, uint32_t maxsz, int32_t data,
115                         gen_helper_gvec_2 *fn)
116 {
117     TCGv_ptr a0, a1;
118     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
119 
120     a0 = tcg_temp_ebb_new_ptr();
121     a1 = tcg_temp_ebb_new_ptr();
122 
123     tcg_gen_addi_ptr(a0, tcg_env, dofs);
124     tcg_gen_addi_ptr(a1, tcg_env, aofs);
125 
126     fn(a0, a1, desc);
127 
128     tcg_temp_free_ptr(a0);
129     tcg_temp_free_ptr(a1);
130 }
131 
132 /* Generate a call to a gvec-style helper with two vector operands
133    and one scalar operand.  */
134 void tcg_gen_gvec_2i_ool(uint32_t dofs, uint32_t aofs, TCGv_i64 c,
135                          uint32_t oprsz, uint32_t maxsz, int32_t data,
136                          gen_helper_gvec_2i *fn)
137 {
138     TCGv_ptr a0, a1;
139     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
140 
141     a0 = tcg_temp_ebb_new_ptr();
142     a1 = tcg_temp_ebb_new_ptr();
143 
144     tcg_gen_addi_ptr(a0, tcg_env, dofs);
145     tcg_gen_addi_ptr(a1, tcg_env, aofs);
146 
147     fn(a0, a1, c, desc);
148 
149     tcg_temp_free_ptr(a0);
150     tcg_temp_free_ptr(a1);
151 }
152 
153 /* Generate a call to a gvec-style helper with three vector operands.  */
154 void tcg_gen_gvec_3_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
155                         uint32_t oprsz, uint32_t maxsz, int32_t data,
156                         gen_helper_gvec_3 *fn)
157 {
158     TCGv_ptr a0, a1, a2;
159     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
160 
161     a0 = tcg_temp_ebb_new_ptr();
162     a1 = tcg_temp_ebb_new_ptr();
163     a2 = tcg_temp_ebb_new_ptr();
164 
165     tcg_gen_addi_ptr(a0, tcg_env, dofs);
166     tcg_gen_addi_ptr(a1, tcg_env, aofs);
167     tcg_gen_addi_ptr(a2, tcg_env, bofs);
168 
169     fn(a0, a1, a2, desc);
170 
171     tcg_temp_free_ptr(a0);
172     tcg_temp_free_ptr(a1);
173     tcg_temp_free_ptr(a2);
174 }
175 
176 /* Generate a call to a gvec-style helper with four vector operands.  */
177 void tcg_gen_gvec_4_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
178                         uint32_t cofs, uint32_t oprsz, uint32_t maxsz,
179                         int32_t data, gen_helper_gvec_4 *fn)
180 {
181     TCGv_ptr a0, a1, a2, a3;
182     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
183 
184     a0 = tcg_temp_ebb_new_ptr();
185     a1 = tcg_temp_ebb_new_ptr();
186     a2 = tcg_temp_ebb_new_ptr();
187     a3 = tcg_temp_ebb_new_ptr();
188 
189     tcg_gen_addi_ptr(a0, tcg_env, dofs);
190     tcg_gen_addi_ptr(a1, tcg_env, aofs);
191     tcg_gen_addi_ptr(a2, tcg_env, bofs);
192     tcg_gen_addi_ptr(a3, tcg_env, cofs);
193 
194     fn(a0, a1, a2, a3, desc);
195 
196     tcg_temp_free_ptr(a0);
197     tcg_temp_free_ptr(a1);
198     tcg_temp_free_ptr(a2);
199     tcg_temp_free_ptr(a3);
200 }
201 
202 /* Generate a call to a gvec-style helper with five vector operands.  */
203 void tcg_gen_gvec_5_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
204                         uint32_t cofs, uint32_t xofs, uint32_t oprsz,
205                         uint32_t maxsz, int32_t data, gen_helper_gvec_5 *fn)
206 {
207     TCGv_ptr a0, a1, a2, a3, a4;
208     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
209 
210     a0 = tcg_temp_ebb_new_ptr();
211     a1 = tcg_temp_ebb_new_ptr();
212     a2 = tcg_temp_ebb_new_ptr();
213     a3 = tcg_temp_ebb_new_ptr();
214     a4 = tcg_temp_ebb_new_ptr();
215 
216     tcg_gen_addi_ptr(a0, tcg_env, dofs);
217     tcg_gen_addi_ptr(a1, tcg_env, aofs);
218     tcg_gen_addi_ptr(a2, tcg_env, bofs);
219     tcg_gen_addi_ptr(a3, tcg_env, cofs);
220     tcg_gen_addi_ptr(a4, tcg_env, xofs);
221 
222     fn(a0, a1, a2, a3, a4, desc);
223 
224     tcg_temp_free_ptr(a0);
225     tcg_temp_free_ptr(a1);
226     tcg_temp_free_ptr(a2);
227     tcg_temp_free_ptr(a3);
228     tcg_temp_free_ptr(a4);
229 }
230 
231 /* Generate a call to a gvec-style helper with three vector operands
232    and an extra pointer operand.  */
233 void tcg_gen_gvec_2_ptr(uint32_t dofs, uint32_t aofs,
234                         TCGv_ptr ptr, uint32_t oprsz, uint32_t maxsz,
235                         int32_t data, gen_helper_gvec_2_ptr *fn)
236 {
237     TCGv_ptr a0, a1;
238     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
239 
240     a0 = tcg_temp_ebb_new_ptr();
241     a1 = tcg_temp_ebb_new_ptr();
242 
243     tcg_gen_addi_ptr(a0, tcg_env, dofs);
244     tcg_gen_addi_ptr(a1, tcg_env, aofs);
245 
246     fn(a0, a1, ptr, desc);
247 
248     tcg_temp_free_ptr(a0);
249     tcg_temp_free_ptr(a1);
250 }
251 
252 /* Generate a call to a gvec-style helper with three vector operands
253    and an extra pointer operand.  */
254 void tcg_gen_gvec_3_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
255                         TCGv_ptr ptr, uint32_t oprsz, uint32_t maxsz,
256                         int32_t data, gen_helper_gvec_3_ptr *fn)
257 {
258     TCGv_ptr a0, a1, a2;
259     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
260 
261     a0 = tcg_temp_ebb_new_ptr();
262     a1 = tcg_temp_ebb_new_ptr();
263     a2 = tcg_temp_ebb_new_ptr();
264 
265     tcg_gen_addi_ptr(a0, tcg_env, dofs);
266     tcg_gen_addi_ptr(a1, tcg_env, aofs);
267     tcg_gen_addi_ptr(a2, tcg_env, bofs);
268 
269     fn(a0, a1, a2, ptr, desc);
270 
271     tcg_temp_free_ptr(a0);
272     tcg_temp_free_ptr(a1);
273     tcg_temp_free_ptr(a2);
274 }
275 
276 /* Generate a call to a gvec-style helper with four vector operands
277    and an extra pointer operand.  */
278 void tcg_gen_gvec_4_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
279                         uint32_t cofs, TCGv_ptr ptr, uint32_t oprsz,
280                         uint32_t maxsz, int32_t data,
281                         gen_helper_gvec_4_ptr *fn)
282 {
283     TCGv_ptr a0, a1, a2, a3;
284     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
285 
286     a0 = tcg_temp_ebb_new_ptr();
287     a1 = tcg_temp_ebb_new_ptr();
288     a2 = tcg_temp_ebb_new_ptr();
289     a3 = tcg_temp_ebb_new_ptr();
290 
291     tcg_gen_addi_ptr(a0, tcg_env, dofs);
292     tcg_gen_addi_ptr(a1, tcg_env, aofs);
293     tcg_gen_addi_ptr(a2, tcg_env, bofs);
294     tcg_gen_addi_ptr(a3, tcg_env, cofs);
295 
296     fn(a0, a1, a2, a3, ptr, desc);
297 
298     tcg_temp_free_ptr(a0);
299     tcg_temp_free_ptr(a1);
300     tcg_temp_free_ptr(a2);
301     tcg_temp_free_ptr(a3);
302 }
303 
304 /* Generate a call to a gvec-style helper with five vector operands
305    and an extra pointer operand.  */
306 void tcg_gen_gvec_5_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
307                         uint32_t cofs, uint32_t eofs, TCGv_ptr ptr,
308                         uint32_t oprsz, uint32_t maxsz, int32_t data,
309                         gen_helper_gvec_5_ptr *fn)
310 {
311     TCGv_ptr a0, a1, a2, a3, a4;
312     TCGv_i32 desc = tcg_constant_i32(simd_desc(oprsz, maxsz, data));
313 
314     a0 = tcg_temp_ebb_new_ptr();
315     a1 = tcg_temp_ebb_new_ptr();
316     a2 = tcg_temp_ebb_new_ptr();
317     a3 = tcg_temp_ebb_new_ptr();
318     a4 = tcg_temp_ebb_new_ptr();
319 
320     tcg_gen_addi_ptr(a0, tcg_env, dofs);
321     tcg_gen_addi_ptr(a1, tcg_env, aofs);
322     tcg_gen_addi_ptr(a2, tcg_env, bofs);
323     tcg_gen_addi_ptr(a3, tcg_env, cofs);
324     tcg_gen_addi_ptr(a4, tcg_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, tcg_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, tcg_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, tcg_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, tcg_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         return;
565     }
566 
567     /* Otherwise, inline with an integer type, unless "large".  */
568     if (check_size_impl(oprsz, TCG_TARGET_REG_BITS / 8)) {
569         t_64 = NULL;
570         t_32 = NULL;
571 
572         if (in_32) {
573             /* We are given a 32-bit variable input.  For a 64-bit host,
574                use a 64-bit operation unless the 32-bit operation would
575                be simple enough.  */
576             if (TCG_TARGET_REG_BITS == 64
577                 && (vece != MO_32 || !check_size_impl(oprsz, 4))) {
578                 t_64 = tcg_temp_ebb_new_i64();
579                 tcg_gen_extu_i32_i64(t_64, in_32);
580                 tcg_gen_dup_i64(vece, t_64, t_64);
581             } else {
582                 t_32 = tcg_temp_ebb_new_i32();
583                 tcg_gen_dup_i32(vece, t_32, in_32);
584             }
585         } else if (in_64) {
586             /* We are given a 64-bit variable input.  */
587             t_64 = tcg_temp_ebb_new_i64();
588             tcg_gen_dup_i64(vece, t_64, in_64);
589         } else {
590             /* We are given a constant input.  */
591             /* For 64-bit hosts, use 64-bit constants for "simple" constants
592                or when we'd need too many 32-bit stores, or when a 64-bit
593                constant is really required.  */
594             if (vece == MO_64
595                 || (TCG_TARGET_REG_BITS == 64
596                     && (in_c == 0 || in_c == -1
597                         || !check_size_impl(oprsz, 4)))) {
598                 t_64 = tcg_constant_i64(in_c);
599             } else {
600                 t_32 = tcg_constant_i32(in_c);
601             }
602         }
603 
604         /* Implement inline if we picked an implementation size above.  */
605         if (t_32) {
606             for (i = 0; i < oprsz; i += 4) {
607                 tcg_gen_st_i32(t_32, tcg_env, dofs + i);
608             }
609             tcg_temp_free_i32(t_32);
610             goto done;
611         }
612         if (t_64) {
613             for (i = 0; i < oprsz; i += 8) {
614                 tcg_gen_st_i64(t_64, tcg_env, dofs + i);
615             }
616             tcg_temp_free_i64(t_64);
617             goto done;
618         }
619     }
620 
621     /* Otherwise implement out of line.  */
622     t_ptr = tcg_temp_ebb_new_ptr();
623     tcg_gen_addi_ptr(t_ptr, tcg_env, dofs);
624 
625     /*
626      * This may be expand_clr for the tail of an operation, e.g.
627      * oprsz == 8 && maxsz == 64.  The size of the clear is misaligned
628      * wrt simd_desc and will assert.  Simply pass all replicated byte
629      * stores through to memset.
630      */
631     if (oprsz == maxsz && vece == MO_8) {
632         TCGv_ptr t_size = tcg_constant_ptr(oprsz);
633         TCGv_i32 t_val;
634 
635         if (in_32) {
636             t_val = in_32;
637         } else if (in_64) {
638             t_val = tcg_temp_ebb_new_i32();
639             tcg_gen_extrl_i64_i32(t_val, in_64);
640         } else {
641             t_val = tcg_constant_i32(in_c);
642         }
643         gen_helper_memset(t_ptr, t_ptr, t_val, t_size);
644 
645         if (in_64) {
646             tcg_temp_free_i32(t_val);
647         }
648         tcg_temp_free_ptr(t_ptr);
649         return;
650     }
651 
652     t_desc = tcg_constant_i32(simd_desc(oprsz, maxsz, 0));
653 
654     if (vece == MO_64) {
655         if (in_64) {
656             gen_helper_gvec_dup64(t_ptr, t_desc, in_64);
657         } else {
658             t_64 = tcg_constant_i64(in_c);
659             gen_helper_gvec_dup64(t_ptr, t_desc, t_64);
660         }
661     } else {
662         typedef void dup_fn(TCGv_ptr, TCGv_i32, TCGv_i32);
663         static dup_fn * const fns[3] = {
664             gen_helper_gvec_dup8,
665             gen_helper_gvec_dup16,
666             gen_helper_gvec_dup32
667         };
668 
669         if (in_32) {
670             fns[vece](t_ptr, t_desc, in_32);
671         } else if (in_64) {
672             t_32 = tcg_temp_ebb_new_i32();
673             tcg_gen_extrl_i64_i32(t_32, in_64);
674             fns[vece](t_ptr, t_desc, t_32);
675             tcg_temp_free_i32(t_32);
676         } else {
677             if (vece == MO_8) {
678                 in_c &= 0xff;
679             } else if (vece == MO_16) {
680                 in_c &= 0xffff;
681             }
682             t_32 = tcg_constant_i32(in_c);
683             fns[vece](t_ptr, t_desc, t_32);
684         }
685     }
686 
687     tcg_temp_free_ptr(t_ptr);
688     return;
689 
690  done:
691     if (oprsz < maxsz) {
692         expand_clr(dofs + oprsz, maxsz - oprsz);
693     }
694 }
695 
696 /* Likewise, but with zero.  */
697 static void expand_clr(uint32_t dofs, uint32_t maxsz)
698 {
699     do_dup(MO_8, dofs, maxsz, maxsz, NULL, NULL, 0);
700 }
701 
702 /* Expand OPSZ bytes worth of two-operand operations using i32 elements.  */
703 static void expand_2_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
704                          bool load_dest, void (*fni)(TCGv_i32, TCGv_i32))
705 {
706     TCGv_i32 t0 = tcg_temp_new_i32();
707     TCGv_i32 t1 = tcg_temp_new_i32();
708     uint32_t i;
709 
710     for (i = 0; i < oprsz; i += 4) {
711         tcg_gen_ld_i32(t0, tcg_env, aofs + i);
712         if (load_dest) {
713             tcg_gen_ld_i32(t1, tcg_env, dofs + i);
714         }
715         fni(t1, t0);
716         tcg_gen_st_i32(t1, tcg_env, dofs + i);
717     }
718     tcg_temp_free_i32(t0);
719     tcg_temp_free_i32(t1);
720 }
721 
722 static void expand_2i_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
723                           int32_t c, bool load_dest,
724                           void (*fni)(TCGv_i32, TCGv_i32, int32_t))
725 {
726     TCGv_i32 t0 = tcg_temp_new_i32();
727     TCGv_i32 t1 = tcg_temp_new_i32();
728     uint32_t i;
729 
730     for (i = 0; i < oprsz; i += 4) {
731         tcg_gen_ld_i32(t0, tcg_env, aofs + i);
732         if (load_dest) {
733             tcg_gen_ld_i32(t1, tcg_env, dofs + i);
734         }
735         fni(t1, t0, c);
736         tcg_gen_st_i32(t1, tcg_env, dofs + i);
737     }
738     tcg_temp_free_i32(t0);
739     tcg_temp_free_i32(t1);
740 }
741 
742 static void expand_2s_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
743                           TCGv_i32 c, bool scalar_first,
744                           void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32))
745 {
746     TCGv_i32 t0 = tcg_temp_new_i32();
747     TCGv_i32 t1 = tcg_temp_new_i32();
748     uint32_t i;
749 
750     for (i = 0; i < oprsz; i += 4) {
751         tcg_gen_ld_i32(t0, tcg_env, aofs + i);
752         if (scalar_first) {
753             fni(t1, c, t0);
754         } else {
755             fni(t1, t0, c);
756         }
757         tcg_gen_st_i32(t1, tcg_env, dofs + i);
758     }
759     tcg_temp_free_i32(t0);
760     tcg_temp_free_i32(t1);
761 }
762 
763 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
764 static void expand_3_i32(uint32_t dofs, uint32_t aofs,
765                          uint32_t bofs, uint32_t oprsz, bool load_dest,
766                          void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32))
767 {
768     TCGv_i32 t0 = tcg_temp_new_i32();
769     TCGv_i32 t1 = tcg_temp_new_i32();
770     TCGv_i32 t2 = tcg_temp_new_i32();
771     uint32_t i;
772 
773     for (i = 0; i < oprsz; i += 4) {
774         tcg_gen_ld_i32(t0, tcg_env, aofs + i);
775         tcg_gen_ld_i32(t1, tcg_env, bofs + i);
776         if (load_dest) {
777             tcg_gen_ld_i32(t2, tcg_env, dofs + i);
778         }
779         fni(t2, t0, t1);
780         tcg_gen_st_i32(t2, tcg_env, dofs + i);
781     }
782     tcg_temp_free_i32(t2);
783     tcg_temp_free_i32(t1);
784     tcg_temp_free_i32(t0);
785 }
786 
787 static void expand_3i_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
788                           uint32_t oprsz, int32_t c, bool load_dest,
789                           void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32, int32_t))
790 {
791     TCGv_i32 t0 = tcg_temp_new_i32();
792     TCGv_i32 t1 = tcg_temp_new_i32();
793     TCGv_i32 t2 = tcg_temp_new_i32();
794     uint32_t i;
795 
796     for (i = 0; i < oprsz; i += 4) {
797         tcg_gen_ld_i32(t0, tcg_env, aofs + i);
798         tcg_gen_ld_i32(t1, tcg_env, bofs + i);
799         if (load_dest) {
800             tcg_gen_ld_i32(t2, tcg_env, dofs + i);
801         }
802         fni(t2, t0, t1, c);
803         tcg_gen_st_i32(t2, tcg_env, dofs + i);
804     }
805     tcg_temp_free_i32(t0);
806     tcg_temp_free_i32(t1);
807     tcg_temp_free_i32(t2);
808 }
809 
810 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
811 static void expand_4_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
812                          uint32_t cofs, uint32_t oprsz, bool write_aofs,
813                          void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32))
814 {
815     TCGv_i32 t0 = tcg_temp_new_i32();
816     TCGv_i32 t1 = tcg_temp_new_i32();
817     TCGv_i32 t2 = tcg_temp_new_i32();
818     TCGv_i32 t3 = tcg_temp_new_i32();
819     uint32_t i;
820 
821     for (i = 0; i < oprsz; i += 4) {
822         tcg_gen_ld_i32(t1, tcg_env, aofs + i);
823         tcg_gen_ld_i32(t2, tcg_env, bofs + i);
824         tcg_gen_ld_i32(t3, tcg_env, cofs + i);
825         fni(t0, t1, t2, t3);
826         tcg_gen_st_i32(t0, tcg_env, dofs + i);
827         if (write_aofs) {
828             tcg_gen_st_i32(t1, tcg_env, aofs + i);
829         }
830     }
831     tcg_temp_free_i32(t3);
832     tcg_temp_free_i32(t2);
833     tcg_temp_free_i32(t1);
834     tcg_temp_free_i32(t0);
835 }
836 
837 static void expand_4i_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
838                           uint32_t cofs, uint32_t oprsz, int32_t c,
839                           void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32,
840                                       int32_t))
841 {
842     TCGv_i32 t0 = tcg_temp_new_i32();
843     TCGv_i32 t1 = tcg_temp_new_i32();
844     TCGv_i32 t2 = tcg_temp_new_i32();
845     TCGv_i32 t3 = tcg_temp_new_i32();
846     uint32_t i;
847 
848     for (i = 0; i < oprsz; i += 4) {
849         tcg_gen_ld_i32(t1, tcg_env, aofs + i);
850         tcg_gen_ld_i32(t2, tcg_env, bofs + i);
851         tcg_gen_ld_i32(t3, tcg_env, cofs + i);
852         fni(t0, t1, t2, t3, c);
853         tcg_gen_st_i32(t0, tcg_env, dofs + i);
854     }
855     tcg_temp_free_i32(t3);
856     tcg_temp_free_i32(t2);
857     tcg_temp_free_i32(t1);
858     tcg_temp_free_i32(t0);
859 }
860 
861 /* Expand OPSZ bytes worth of two-operand operations using i64 elements.  */
862 static void expand_2_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
863                          bool load_dest, void (*fni)(TCGv_i64, TCGv_i64))
864 {
865     TCGv_i64 t0 = tcg_temp_new_i64();
866     TCGv_i64 t1 = tcg_temp_new_i64();
867     uint32_t i;
868 
869     for (i = 0; i < oprsz; i += 8) {
870         tcg_gen_ld_i64(t0, tcg_env, aofs + i);
871         if (load_dest) {
872             tcg_gen_ld_i64(t1, tcg_env, dofs + i);
873         }
874         fni(t1, t0);
875         tcg_gen_st_i64(t1, tcg_env, dofs + i);
876     }
877     tcg_temp_free_i64(t0);
878     tcg_temp_free_i64(t1);
879 }
880 
881 static void expand_2i_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
882                           int64_t c, bool load_dest,
883                           void (*fni)(TCGv_i64, TCGv_i64, int64_t))
884 {
885     TCGv_i64 t0 = tcg_temp_new_i64();
886     TCGv_i64 t1 = tcg_temp_new_i64();
887     uint32_t i;
888 
889     for (i = 0; i < oprsz; i += 8) {
890         tcg_gen_ld_i64(t0, tcg_env, aofs + i);
891         if (load_dest) {
892             tcg_gen_ld_i64(t1, tcg_env, dofs + i);
893         }
894         fni(t1, t0, c);
895         tcg_gen_st_i64(t1, tcg_env, dofs + i);
896     }
897     tcg_temp_free_i64(t0);
898     tcg_temp_free_i64(t1);
899 }
900 
901 static void expand_2s_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
902                           TCGv_i64 c, bool scalar_first,
903                           void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64))
904 {
905     TCGv_i64 t0 = tcg_temp_new_i64();
906     TCGv_i64 t1 = tcg_temp_new_i64();
907     uint32_t i;
908 
909     for (i = 0; i < oprsz; i += 8) {
910         tcg_gen_ld_i64(t0, tcg_env, aofs + i);
911         if (scalar_first) {
912             fni(t1, c, t0);
913         } else {
914             fni(t1, t0, c);
915         }
916         tcg_gen_st_i64(t1, tcg_env, dofs + i);
917     }
918     tcg_temp_free_i64(t0);
919     tcg_temp_free_i64(t1);
920 }
921 
922 /* Expand OPSZ bytes worth of three-operand operations using i64 elements.  */
923 static void expand_3_i64(uint32_t dofs, uint32_t aofs,
924                          uint32_t bofs, uint32_t oprsz, bool load_dest,
925                          void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64))
926 {
927     TCGv_i64 t0 = tcg_temp_new_i64();
928     TCGv_i64 t1 = tcg_temp_new_i64();
929     TCGv_i64 t2 = tcg_temp_new_i64();
930     uint32_t i;
931 
932     for (i = 0; i < oprsz; i += 8) {
933         tcg_gen_ld_i64(t0, tcg_env, aofs + i);
934         tcg_gen_ld_i64(t1, tcg_env, bofs + i);
935         if (load_dest) {
936             tcg_gen_ld_i64(t2, tcg_env, dofs + i);
937         }
938         fni(t2, t0, t1);
939         tcg_gen_st_i64(t2, tcg_env, dofs + i);
940     }
941     tcg_temp_free_i64(t2);
942     tcg_temp_free_i64(t1);
943     tcg_temp_free_i64(t0);
944 }
945 
946 static void expand_3i_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
947                           uint32_t oprsz, int64_t c, bool load_dest,
948                           void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64, int64_t))
949 {
950     TCGv_i64 t0 = tcg_temp_new_i64();
951     TCGv_i64 t1 = tcg_temp_new_i64();
952     TCGv_i64 t2 = tcg_temp_new_i64();
953     uint32_t i;
954 
955     for (i = 0; i < oprsz; i += 8) {
956         tcg_gen_ld_i64(t0, tcg_env, aofs + i);
957         tcg_gen_ld_i64(t1, tcg_env, bofs + i);
958         if (load_dest) {
959             tcg_gen_ld_i64(t2, tcg_env, dofs + i);
960         }
961         fni(t2, t0, t1, c);
962         tcg_gen_st_i64(t2, tcg_env, dofs + i);
963     }
964     tcg_temp_free_i64(t0);
965     tcg_temp_free_i64(t1);
966     tcg_temp_free_i64(t2);
967 }
968 
969 /* Expand OPSZ bytes worth of three-operand operations using i64 elements.  */
970 static void expand_4_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
971                          uint32_t cofs, uint32_t oprsz, bool write_aofs,
972                          void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64, TCGv_i64))
973 {
974     TCGv_i64 t0 = tcg_temp_new_i64();
975     TCGv_i64 t1 = tcg_temp_new_i64();
976     TCGv_i64 t2 = tcg_temp_new_i64();
977     TCGv_i64 t3 = tcg_temp_new_i64();
978     uint32_t i;
979 
980     for (i = 0; i < oprsz; i += 8) {
981         tcg_gen_ld_i64(t1, tcg_env, aofs + i);
982         tcg_gen_ld_i64(t2, tcg_env, bofs + i);
983         tcg_gen_ld_i64(t3, tcg_env, cofs + i);
984         fni(t0, t1, t2, t3);
985         tcg_gen_st_i64(t0, tcg_env, dofs + i);
986         if (write_aofs) {
987             tcg_gen_st_i64(t1, tcg_env, aofs + i);
988         }
989     }
990     tcg_temp_free_i64(t3);
991     tcg_temp_free_i64(t2);
992     tcg_temp_free_i64(t1);
993     tcg_temp_free_i64(t0);
994 }
995 
996 static void expand_4i_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
997                           uint32_t cofs, uint32_t oprsz, int64_t c,
998                           void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64, TCGv_i64,
999                                       int64_t))
1000 {
1001     TCGv_i64 t0 = tcg_temp_new_i64();
1002     TCGv_i64 t1 = tcg_temp_new_i64();
1003     TCGv_i64 t2 = tcg_temp_new_i64();
1004     TCGv_i64 t3 = tcg_temp_new_i64();
1005     uint32_t i;
1006 
1007     for (i = 0; i < oprsz; i += 8) {
1008         tcg_gen_ld_i64(t1, tcg_env, aofs + i);
1009         tcg_gen_ld_i64(t2, tcg_env, bofs + i);
1010         tcg_gen_ld_i64(t3, tcg_env, cofs + i);
1011         fni(t0, t1, t2, t3, c);
1012         tcg_gen_st_i64(t0, tcg_env, dofs + i);
1013     }
1014     tcg_temp_free_i64(t3);
1015     tcg_temp_free_i64(t2);
1016     tcg_temp_free_i64(t1);
1017     tcg_temp_free_i64(t0);
1018 }
1019 
1020 /* Expand OPSZ bytes worth of two-operand operations using host vectors.  */
1021 static void expand_2_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1022                          uint32_t oprsz, uint32_t tysz, TCGType type,
1023                          bool load_dest,
1024                          void (*fni)(unsigned, TCGv_vec, TCGv_vec))
1025 {
1026     for (uint32_t i = 0; i < oprsz; i += tysz) {
1027         TCGv_vec t0 = tcg_temp_new_vec(type);
1028         TCGv_vec t1 = tcg_temp_new_vec(type);
1029 
1030         tcg_gen_ld_vec(t0, tcg_env, aofs + i);
1031         if (load_dest) {
1032             tcg_gen_ld_vec(t1, tcg_env, dofs + i);
1033         }
1034         fni(vece, t1, t0);
1035         tcg_gen_st_vec(t1, tcg_env, dofs + i);
1036     }
1037 }
1038 
1039 /* Expand OPSZ bytes worth of two-vector operands and an immediate operand
1040    using host vectors.  */
1041 static void expand_2i_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1042                           uint32_t oprsz, uint32_t tysz, TCGType type,
1043                           int64_t c, bool load_dest,
1044                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, int64_t))
1045 {
1046     for (uint32_t i = 0; i < oprsz; i += tysz) {
1047         TCGv_vec t0 = tcg_temp_new_vec(type);
1048         TCGv_vec t1 = tcg_temp_new_vec(type);
1049 
1050         tcg_gen_ld_vec(t0, tcg_env, aofs + i);
1051         if (load_dest) {
1052             tcg_gen_ld_vec(t1, tcg_env, dofs + i);
1053         }
1054         fni(vece, t1, t0, c);
1055         tcg_gen_st_vec(t1, tcg_env, dofs + i);
1056     }
1057 }
1058 
1059 static void expand_2s_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1060                           uint32_t oprsz, uint32_t tysz, TCGType type,
1061                           TCGv_vec c, bool scalar_first,
1062                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec))
1063 {
1064     for (uint32_t i = 0; i < oprsz; i += tysz) {
1065         TCGv_vec t0 = tcg_temp_new_vec(type);
1066         TCGv_vec t1 = tcg_temp_new_vec(type);
1067 
1068         tcg_gen_ld_vec(t0, tcg_env, aofs + i);
1069         if (scalar_first) {
1070             fni(vece, t1, c, t0);
1071         } else {
1072             fni(vece, t1, t0, c);
1073         }
1074         tcg_gen_st_vec(t1, tcg_env, dofs + i);
1075     }
1076 }
1077 
1078 /* Expand OPSZ bytes worth of three-operand operations using host vectors.  */
1079 static void expand_3_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1080                          uint32_t bofs, uint32_t oprsz,
1081                          uint32_t tysz, TCGType type, bool load_dest,
1082                          void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec))
1083 {
1084     for (uint32_t i = 0; i < oprsz; i += tysz) {
1085         TCGv_vec t0 = tcg_temp_new_vec(type);
1086         TCGv_vec t1 = tcg_temp_new_vec(type);
1087         TCGv_vec t2 = tcg_temp_new_vec(type);
1088 
1089         tcg_gen_ld_vec(t0, tcg_env, aofs + i);
1090         tcg_gen_ld_vec(t1, tcg_env, bofs + i);
1091         if (load_dest) {
1092             tcg_gen_ld_vec(t2, tcg_env, dofs + i);
1093         }
1094         fni(vece, t2, t0, t1);
1095         tcg_gen_st_vec(t2, tcg_env, dofs + i);
1096     }
1097 }
1098 
1099 /*
1100  * Expand OPSZ bytes worth of three-vector operands and an immediate operand
1101  * using host vectors.
1102  */
1103 static void expand_3i_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1104                           uint32_t bofs, uint32_t oprsz, uint32_t tysz,
1105                           TCGType type, int64_t c, bool load_dest,
1106                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec,
1107                                       int64_t))
1108 {
1109     for (uint32_t i = 0; i < oprsz; i += tysz) {
1110         TCGv_vec t0 = tcg_temp_new_vec(type);
1111         TCGv_vec t1 = tcg_temp_new_vec(type);
1112         TCGv_vec t2 = tcg_temp_new_vec(type);
1113 
1114         tcg_gen_ld_vec(t0, tcg_env, aofs + i);
1115         tcg_gen_ld_vec(t1, tcg_env, bofs + i);
1116         if (load_dest) {
1117             tcg_gen_ld_vec(t2, tcg_env, dofs + i);
1118         }
1119         fni(vece, t2, t0, t1, c);
1120         tcg_gen_st_vec(t2, tcg_env, dofs + i);
1121     }
1122 }
1123 
1124 /* Expand OPSZ bytes worth of four-operand operations using host vectors.  */
1125 static void expand_4_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1126                          uint32_t bofs, uint32_t cofs, uint32_t oprsz,
1127                          uint32_t tysz, TCGType type, bool write_aofs,
1128                          void (*fni)(unsigned, TCGv_vec, TCGv_vec,
1129                                      TCGv_vec, TCGv_vec))
1130 {
1131     for (uint32_t i = 0; i < oprsz; i += tysz) {
1132         TCGv_vec t0 = tcg_temp_new_vec(type);
1133         TCGv_vec t1 = tcg_temp_new_vec(type);
1134         TCGv_vec t2 = tcg_temp_new_vec(type);
1135         TCGv_vec t3 = tcg_temp_new_vec(type);
1136 
1137         tcg_gen_ld_vec(t1, tcg_env, aofs + i);
1138         tcg_gen_ld_vec(t2, tcg_env, bofs + i);
1139         tcg_gen_ld_vec(t3, tcg_env, cofs + i);
1140         fni(vece, t0, t1, t2, t3);
1141         tcg_gen_st_vec(t0, tcg_env, dofs + i);
1142         if (write_aofs) {
1143             tcg_gen_st_vec(t1, tcg_env, aofs + i);
1144         }
1145     }
1146 }
1147 
1148 /*
1149  * Expand OPSZ bytes worth of four-vector operands and an immediate operand
1150  * using host vectors.
1151  */
1152 static void expand_4i_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1153                           uint32_t bofs, uint32_t cofs, uint32_t oprsz,
1154                           uint32_t tysz, TCGType type, int64_t c,
1155                           void (*fni)(unsigned, TCGv_vec, TCGv_vec,
1156                                      TCGv_vec, TCGv_vec, int64_t))
1157 {
1158     for (uint32_t i = 0; i < oprsz; i += tysz) {
1159         TCGv_vec t0 = tcg_temp_new_vec(type);
1160         TCGv_vec t1 = tcg_temp_new_vec(type);
1161         TCGv_vec t2 = tcg_temp_new_vec(type);
1162         TCGv_vec t3 = tcg_temp_new_vec(type);
1163 
1164         tcg_gen_ld_vec(t1, tcg_env, aofs + i);
1165         tcg_gen_ld_vec(t2, tcg_env, bofs + i);
1166         tcg_gen_ld_vec(t3, tcg_env, cofs + i);
1167         fni(vece, t0, t1, t2, t3, c);
1168         tcg_gen_st_vec(t0, tcg_env, dofs + i);
1169     }
1170 }
1171 
1172 /* Expand a vector two-operand operation.  */
1173 void tcg_gen_gvec_2(uint32_t dofs, uint32_t aofs,
1174                     uint32_t oprsz, uint32_t maxsz, const GVecGen2 *g)
1175 {
1176     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1177     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1178     TCGType type;
1179     uint32_t some;
1180 
1181     check_size_align(oprsz, maxsz, dofs | aofs);
1182     check_overlap_2(dofs, aofs, maxsz);
1183 
1184     type = 0;
1185     if (g->fniv) {
1186         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1187     }
1188     switch (type) {
1189     case TCG_TYPE_V256:
1190         /* Recall that ARM SVE allows vector sizes that are not a
1191          * power of 2, but always a multiple of 16.  The intent is
1192          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1193          */
1194         some = QEMU_ALIGN_DOWN(oprsz, 32);
1195         expand_2_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
1196                      g->load_dest, g->fniv);
1197         if (some == oprsz) {
1198             break;
1199         }
1200         dofs += some;
1201         aofs += some;
1202         oprsz -= some;
1203         maxsz -= some;
1204         /* fallthru */
1205     case TCG_TYPE_V128:
1206         expand_2_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
1207                      g->load_dest, g->fniv);
1208         break;
1209     case TCG_TYPE_V64:
1210         expand_2_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
1211                      g->load_dest, g->fniv);
1212         break;
1213 
1214     case 0:
1215         if (g->fni8 && check_size_impl(oprsz, 8)) {
1216             expand_2_i64(dofs, aofs, oprsz, g->load_dest, g->fni8);
1217         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1218             expand_2_i32(dofs, aofs, oprsz, g->load_dest, g->fni4);
1219         } else {
1220             assert(g->fno != NULL);
1221             tcg_gen_gvec_2_ool(dofs, aofs, oprsz, maxsz, g->data, g->fno);
1222             oprsz = maxsz;
1223         }
1224         break;
1225 
1226     default:
1227         g_assert_not_reached();
1228     }
1229     tcg_swap_vecop_list(hold_list);
1230 
1231     if (oprsz < maxsz) {
1232         expand_clr(dofs + oprsz, maxsz - oprsz);
1233     }
1234 }
1235 
1236 /* Expand a vector operation with two vectors and an immediate.  */
1237 void tcg_gen_gvec_2i(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
1238                      uint32_t maxsz, int64_t c, const GVecGen2i *g)
1239 {
1240     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1241     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1242     TCGType type;
1243     uint32_t some;
1244 
1245     check_size_align(oprsz, maxsz, dofs | aofs);
1246     check_overlap_2(dofs, aofs, maxsz);
1247 
1248     type = 0;
1249     if (g->fniv) {
1250         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1251     }
1252     switch (type) {
1253     case TCG_TYPE_V256:
1254         /* Recall that ARM SVE allows vector sizes that are not a
1255          * power of 2, but always a multiple of 16.  The intent is
1256          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1257          */
1258         some = QEMU_ALIGN_DOWN(oprsz, 32);
1259         expand_2i_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
1260                       c, g->load_dest, g->fniv);
1261         if (some == oprsz) {
1262             break;
1263         }
1264         dofs += some;
1265         aofs += some;
1266         oprsz -= some;
1267         maxsz -= some;
1268         /* fallthru */
1269     case TCG_TYPE_V128:
1270         expand_2i_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
1271                       c, g->load_dest, g->fniv);
1272         break;
1273     case TCG_TYPE_V64:
1274         expand_2i_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
1275                       c, g->load_dest, g->fniv);
1276         break;
1277 
1278     case 0:
1279         if (g->fni8 && check_size_impl(oprsz, 8)) {
1280             expand_2i_i64(dofs, aofs, oprsz, c, g->load_dest, g->fni8);
1281         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1282             expand_2i_i32(dofs, aofs, oprsz, c, g->load_dest, g->fni4);
1283         } else {
1284             if (g->fno) {
1285                 tcg_gen_gvec_2_ool(dofs, aofs, oprsz, maxsz, c, g->fno);
1286             } else {
1287                 TCGv_i64 tcg_c = tcg_constant_i64(c);
1288                 tcg_gen_gvec_2i_ool(dofs, aofs, tcg_c, oprsz,
1289                                     maxsz, c, g->fnoi);
1290             }
1291             oprsz = maxsz;
1292         }
1293         break;
1294 
1295     default:
1296         g_assert_not_reached();
1297     }
1298     tcg_swap_vecop_list(hold_list);
1299 
1300     if (oprsz < maxsz) {
1301         expand_clr(dofs + oprsz, maxsz - oprsz);
1302     }
1303 }
1304 
1305 /* Expand a vector operation with two vectors and a scalar.  */
1306 void tcg_gen_gvec_2s(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
1307                      uint32_t maxsz, TCGv_i64 c, const GVecGen2s *g)
1308 {
1309     TCGType type;
1310 
1311     check_size_align(oprsz, maxsz, dofs | aofs);
1312     check_overlap_2(dofs, aofs, maxsz);
1313 
1314     type = 0;
1315     if (g->fniv) {
1316         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1317     }
1318     if (type != 0) {
1319         const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1320         const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1321         TCGv_vec t_vec = tcg_temp_new_vec(type);
1322         uint32_t some;
1323 
1324         tcg_gen_dup_i64_vec(g->vece, t_vec, c);
1325 
1326         switch (type) {
1327         case TCG_TYPE_V256:
1328             /* Recall that ARM SVE allows vector sizes that are not a
1329              * power of 2, but always a multiple of 16.  The intent is
1330              * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1331              */
1332             some = QEMU_ALIGN_DOWN(oprsz, 32);
1333             expand_2s_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
1334                           t_vec, g->scalar_first, g->fniv);
1335             if (some == oprsz) {
1336                 break;
1337             }
1338             dofs += some;
1339             aofs += some;
1340             oprsz -= some;
1341             maxsz -= some;
1342             /* fallthru */
1343 
1344         case TCG_TYPE_V128:
1345             expand_2s_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
1346                           t_vec, g->scalar_first, g->fniv);
1347             break;
1348 
1349         case TCG_TYPE_V64:
1350             expand_2s_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
1351                           t_vec, g->scalar_first, g->fniv);
1352             break;
1353 
1354         default:
1355             g_assert_not_reached();
1356         }
1357         tcg_temp_free_vec(t_vec);
1358         tcg_swap_vecop_list(hold_list);
1359     } else if (g->fni8 && check_size_impl(oprsz, 8)) {
1360         TCGv_i64 t64 = tcg_temp_new_i64();
1361 
1362         tcg_gen_dup_i64(g->vece, t64, c);
1363         expand_2s_i64(dofs, aofs, oprsz, t64, g->scalar_first, g->fni8);
1364         tcg_temp_free_i64(t64);
1365     } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1366         TCGv_i32 t32 = tcg_temp_new_i32();
1367 
1368         tcg_gen_extrl_i64_i32(t32, c);
1369         tcg_gen_dup_i32(g->vece, t32, t32);
1370         expand_2s_i32(dofs, aofs, oprsz, t32, g->scalar_first, g->fni4);
1371         tcg_temp_free_i32(t32);
1372     } else {
1373         tcg_gen_gvec_2i_ool(dofs, aofs, c, oprsz, maxsz, 0, g->fno);
1374         return;
1375     }
1376 
1377     if (oprsz < maxsz) {
1378         expand_clr(dofs + oprsz, maxsz - oprsz);
1379     }
1380 }
1381 
1382 /* Expand a vector three-operand operation.  */
1383 void tcg_gen_gvec_3(uint32_t dofs, uint32_t aofs, uint32_t bofs,
1384                     uint32_t oprsz, uint32_t maxsz, const GVecGen3 *g)
1385 {
1386     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1387     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1388     TCGType type;
1389     uint32_t some;
1390 
1391     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
1392     check_overlap_3(dofs, aofs, bofs, maxsz);
1393 
1394     type = 0;
1395     if (g->fniv) {
1396         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1397     }
1398     switch (type) {
1399     case TCG_TYPE_V256:
1400         /* Recall that ARM SVE allows vector sizes that are not a
1401          * power of 2, but always a multiple of 16.  The intent is
1402          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1403          */
1404         some = QEMU_ALIGN_DOWN(oprsz, 32);
1405         expand_3_vec(g->vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256,
1406                      g->load_dest, g->fniv);
1407         if (some == oprsz) {
1408             break;
1409         }
1410         dofs += some;
1411         aofs += some;
1412         bofs += some;
1413         oprsz -= some;
1414         maxsz -= some;
1415         /* fallthru */
1416     case TCG_TYPE_V128:
1417         expand_3_vec(g->vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128,
1418                      g->load_dest, g->fniv);
1419         break;
1420     case TCG_TYPE_V64:
1421         expand_3_vec(g->vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64,
1422                      g->load_dest, g->fniv);
1423         break;
1424 
1425     case 0:
1426         if (g->fni8 && check_size_impl(oprsz, 8)) {
1427             expand_3_i64(dofs, aofs, bofs, oprsz, g->load_dest, g->fni8);
1428         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1429             expand_3_i32(dofs, aofs, bofs, oprsz, g->load_dest, g->fni4);
1430         } else {
1431             assert(g->fno != NULL);
1432             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz,
1433                                maxsz, g->data, g->fno);
1434             oprsz = maxsz;
1435         }
1436         break;
1437 
1438     default:
1439         g_assert_not_reached();
1440     }
1441     tcg_swap_vecop_list(hold_list);
1442 
1443     if (oprsz < maxsz) {
1444         expand_clr(dofs + oprsz, maxsz - oprsz);
1445     }
1446 }
1447 
1448 /* Expand a vector operation with three vectors and an immediate.  */
1449 void tcg_gen_gvec_3i(uint32_t dofs, uint32_t aofs, uint32_t bofs,
1450                      uint32_t oprsz, uint32_t maxsz, int64_t c,
1451                      const GVecGen3i *g)
1452 {
1453     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1454     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1455     TCGType type;
1456     uint32_t some;
1457 
1458     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
1459     check_overlap_3(dofs, aofs, bofs, maxsz);
1460 
1461     type = 0;
1462     if (g->fniv) {
1463         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1464     }
1465     switch (type) {
1466     case TCG_TYPE_V256:
1467         /*
1468          * Recall that ARM SVE allows vector sizes that are not a
1469          * power of 2, but always a multiple of 16.  The intent is
1470          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1471          */
1472         some = QEMU_ALIGN_DOWN(oprsz, 32);
1473         expand_3i_vec(g->vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256,
1474                       c, g->load_dest, g->fniv);
1475         if (some == oprsz) {
1476             break;
1477         }
1478         dofs += some;
1479         aofs += some;
1480         bofs += some;
1481         oprsz -= some;
1482         maxsz -= some;
1483         /* fallthru */
1484     case TCG_TYPE_V128:
1485         expand_3i_vec(g->vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128,
1486                       c, g->load_dest, g->fniv);
1487         break;
1488     case TCG_TYPE_V64:
1489         expand_3i_vec(g->vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64,
1490                       c, g->load_dest, g->fniv);
1491         break;
1492 
1493     case 0:
1494         if (g->fni8 && check_size_impl(oprsz, 8)) {
1495             expand_3i_i64(dofs, aofs, bofs, oprsz, c, g->load_dest, g->fni8);
1496         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1497             expand_3i_i32(dofs, aofs, bofs, oprsz, c, g->load_dest, g->fni4);
1498         } else {
1499             assert(g->fno != NULL);
1500             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz, maxsz, c, g->fno);
1501             oprsz = maxsz;
1502         }
1503         break;
1504 
1505     default:
1506         g_assert_not_reached();
1507     }
1508     tcg_swap_vecop_list(hold_list);
1509 
1510     if (oprsz < maxsz) {
1511         expand_clr(dofs + oprsz, maxsz - oprsz);
1512     }
1513 }
1514 
1515 /* Expand a vector four-operand operation.  */
1516 void tcg_gen_gvec_4(uint32_t dofs, uint32_t aofs, uint32_t bofs, uint32_t cofs,
1517                     uint32_t oprsz, uint32_t maxsz, const GVecGen4 *g)
1518 {
1519     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1520     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1521     TCGType type;
1522     uint32_t some;
1523 
1524     check_size_align(oprsz, maxsz, dofs | aofs | bofs | cofs);
1525     check_overlap_4(dofs, aofs, bofs, cofs, maxsz);
1526 
1527     type = 0;
1528     if (g->fniv) {
1529         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1530     }
1531     switch (type) {
1532     case TCG_TYPE_V256:
1533         /* Recall that ARM SVE allows vector sizes that are not a
1534          * power of 2, but always a multiple of 16.  The intent is
1535          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1536          */
1537         some = QEMU_ALIGN_DOWN(oprsz, 32);
1538         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, some,
1539                      32, TCG_TYPE_V256, g->write_aofs, g->fniv);
1540         if (some == oprsz) {
1541             break;
1542         }
1543         dofs += some;
1544         aofs += some;
1545         bofs += some;
1546         cofs += some;
1547         oprsz -= some;
1548         maxsz -= some;
1549         /* fallthru */
1550     case TCG_TYPE_V128:
1551         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1552                      16, TCG_TYPE_V128, g->write_aofs, g->fniv);
1553         break;
1554     case TCG_TYPE_V64:
1555         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1556                      8, TCG_TYPE_V64, g->write_aofs, g->fniv);
1557         break;
1558 
1559     case 0:
1560         if (g->fni8 && check_size_impl(oprsz, 8)) {
1561             expand_4_i64(dofs, aofs, bofs, cofs, oprsz,
1562                          g->write_aofs, g->fni8);
1563         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1564             expand_4_i32(dofs, aofs, bofs, cofs, oprsz,
1565                          g->write_aofs, g->fni4);
1566         } else {
1567             assert(g->fno != NULL);
1568             tcg_gen_gvec_4_ool(dofs, aofs, bofs, cofs,
1569                                oprsz, maxsz, g->data, g->fno);
1570             oprsz = maxsz;
1571         }
1572         break;
1573 
1574     default:
1575         g_assert_not_reached();
1576     }
1577     tcg_swap_vecop_list(hold_list);
1578 
1579     if (oprsz < maxsz) {
1580         expand_clr(dofs + oprsz, maxsz - oprsz);
1581     }
1582 }
1583 
1584 /* Expand a vector four-operand operation.  */
1585 void tcg_gen_gvec_4i(uint32_t dofs, uint32_t aofs, uint32_t bofs, uint32_t cofs,
1586                      uint32_t oprsz, uint32_t maxsz, int64_t c,
1587                      const GVecGen4i *g)
1588 {
1589     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1590     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1591     TCGType type;
1592     uint32_t some;
1593 
1594     check_size_align(oprsz, maxsz, dofs | aofs | bofs | cofs);
1595     check_overlap_4(dofs, aofs, bofs, cofs, maxsz);
1596 
1597     type = 0;
1598     if (g->fniv) {
1599         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1600     }
1601     switch (type) {
1602     case TCG_TYPE_V256:
1603         /*
1604          * Recall that ARM SVE allows vector sizes that are not a
1605          * power of 2, but always a multiple of 16.  The intent is
1606          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1607          */
1608         some = QEMU_ALIGN_DOWN(oprsz, 32);
1609         expand_4i_vec(g->vece, dofs, aofs, bofs, cofs, some,
1610                       32, TCG_TYPE_V256, c, g->fniv);
1611         if (some == oprsz) {
1612             break;
1613         }
1614         dofs += some;
1615         aofs += some;
1616         bofs += some;
1617         cofs += some;
1618         oprsz -= some;
1619         maxsz -= some;
1620         /* fallthru */
1621     case TCG_TYPE_V128:
1622         expand_4i_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1623                        16, TCG_TYPE_V128, c, g->fniv);
1624         break;
1625     case TCG_TYPE_V64:
1626         expand_4i_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1627                       8, TCG_TYPE_V64, c, g->fniv);
1628         break;
1629 
1630     case 0:
1631         if (g->fni8 && check_size_impl(oprsz, 8)) {
1632             expand_4i_i64(dofs, aofs, bofs, cofs, oprsz, c, g->fni8);
1633         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1634             expand_4i_i32(dofs, aofs, bofs, cofs, oprsz, c, g->fni4);
1635         } else {
1636             assert(g->fno != NULL);
1637             tcg_gen_gvec_4_ool(dofs, aofs, bofs, cofs,
1638                                oprsz, maxsz, c, g->fno);
1639             oprsz = maxsz;
1640         }
1641         break;
1642 
1643     default:
1644         g_assert_not_reached();
1645     }
1646     tcg_swap_vecop_list(hold_list);
1647 
1648     if (oprsz < maxsz) {
1649         expand_clr(dofs + oprsz, maxsz - oprsz);
1650     }
1651 }
1652 
1653 /*
1654  * Expand specific vector operations.
1655  */
1656 
1657 static void vec_mov2(unsigned vece, TCGv_vec a, TCGv_vec b)
1658 {
1659     tcg_gen_mov_vec(a, b);
1660 }
1661 
1662 void tcg_gen_gvec_mov(unsigned vece, uint32_t dofs, uint32_t aofs,
1663                       uint32_t oprsz, uint32_t maxsz)
1664 {
1665     static const GVecGen2 g = {
1666         .fni8 = tcg_gen_mov_i64,
1667         .fniv = vec_mov2,
1668         .fno = gen_helper_gvec_mov,
1669         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1670     };
1671     if (dofs != aofs) {
1672         tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g);
1673     } else {
1674         check_size_align(oprsz, maxsz, dofs);
1675         if (oprsz < maxsz) {
1676             expand_clr(dofs + oprsz, maxsz - oprsz);
1677         }
1678     }
1679 }
1680 
1681 void tcg_gen_gvec_dup_i32(unsigned vece, uint32_t dofs, uint32_t oprsz,
1682                           uint32_t maxsz, TCGv_i32 in)
1683 {
1684     check_size_align(oprsz, maxsz, dofs);
1685     tcg_debug_assert(vece <= MO_32);
1686     do_dup(vece, dofs, oprsz, maxsz, in, NULL, 0);
1687 }
1688 
1689 void tcg_gen_gvec_dup_i64(unsigned vece, uint32_t dofs, uint32_t oprsz,
1690                           uint32_t maxsz, TCGv_i64 in)
1691 {
1692     check_size_align(oprsz, maxsz, dofs);
1693     tcg_debug_assert(vece <= MO_64);
1694     do_dup(vece, dofs, oprsz, maxsz, NULL, in, 0);
1695 }
1696 
1697 void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t dofs, uint32_t aofs,
1698                           uint32_t oprsz, uint32_t maxsz)
1699 {
1700     check_size_align(oprsz, maxsz, dofs);
1701     if (vece <= MO_64) {
1702         TCGType type = choose_vector_type(NULL, vece, oprsz, 0);
1703         if (type != 0) {
1704             TCGv_vec t_vec = tcg_temp_new_vec(type);
1705             tcg_gen_dup_mem_vec(vece, t_vec, tcg_env, aofs);
1706             do_dup_store(type, dofs, oprsz, maxsz, t_vec);
1707         } else if (vece <= MO_32) {
1708             TCGv_i32 in = tcg_temp_ebb_new_i32();
1709             switch (vece) {
1710             case MO_8:
1711                 tcg_gen_ld8u_i32(in, tcg_env, aofs);
1712                 break;
1713             case MO_16:
1714                 tcg_gen_ld16u_i32(in, tcg_env, aofs);
1715                 break;
1716             default:
1717                 tcg_gen_ld_i32(in, tcg_env, aofs);
1718                 break;
1719             }
1720             do_dup(vece, dofs, oprsz, maxsz, in, NULL, 0);
1721             tcg_temp_free_i32(in);
1722         } else {
1723             TCGv_i64 in = tcg_temp_ebb_new_i64();
1724             tcg_gen_ld_i64(in, tcg_env, aofs);
1725             do_dup(vece, dofs, oprsz, maxsz, NULL, in, 0);
1726             tcg_temp_free_i64(in);
1727         }
1728     } else if (vece == 4) {
1729         /* 128-bit duplicate.  */
1730         int i;
1731 
1732         tcg_debug_assert(oprsz >= 16);
1733         if (TCG_TARGET_HAS_v128) {
1734             TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V128);
1735 
1736             tcg_gen_ld_vec(in, tcg_env, aofs);
1737             for (i = (aofs == dofs) * 16; i < oprsz; i += 16) {
1738                 tcg_gen_st_vec(in, tcg_env, dofs + i);
1739             }
1740         } else {
1741             TCGv_i64 in0 = tcg_temp_ebb_new_i64();
1742             TCGv_i64 in1 = tcg_temp_ebb_new_i64();
1743 
1744             tcg_gen_ld_i64(in0, tcg_env, aofs);
1745             tcg_gen_ld_i64(in1, tcg_env, aofs + 8);
1746             for (i = (aofs == dofs) * 16; i < oprsz; i += 16) {
1747                 tcg_gen_st_i64(in0, tcg_env, dofs + i);
1748                 tcg_gen_st_i64(in1, tcg_env, dofs + i + 8);
1749             }
1750             tcg_temp_free_i64(in0);
1751             tcg_temp_free_i64(in1);
1752         }
1753         if (oprsz < maxsz) {
1754             expand_clr(dofs + oprsz, maxsz - oprsz);
1755         }
1756     } else if (vece == 5) {
1757         /* 256-bit duplicate.  */
1758         int i;
1759 
1760         tcg_debug_assert(oprsz >= 32);
1761         tcg_debug_assert(oprsz % 32 == 0);
1762         if (TCG_TARGET_HAS_v256) {
1763             TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V256);
1764 
1765             tcg_gen_ld_vec(in, tcg_env, aofs);
1766             for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
1767                 tcg_gen_st_vec(in, tcg_env, dofs + i);
1768             }
1769         } else if (TCG_TARGET_HAS_v128) {
1770             TCGv_vec in0 = tcg_temp_new_vec(TCG_TYPE_V128);
1771             TCGv_vec in1 = tcg_temp_new_vec(TCG_TYPE_V128);
1772 
1773             tcg_gen_ld_vec(in0, tcg_env, aofs);
1774             tcg_gen_ld_vec(in1, tcg_env, aofs + 16);
1775             for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
1776                 tcg_gen_st_vec(in0, tcg_env, dofs + i);
1777                 tcg_gen_st_vec(in1, tcg_env, dofs + i + 16);
1778             }
1779         } else {
1780             TCGv_i64 in[4];
1781             int j;
1782 
1783             for (j = 0; j < 4; ++j) {
1784                 in[j] = tcg_temp_ebb_new_i64();
1785                 tcg_gen_ld_i64(in[j], tcg_env, aofs + j * 8);
1786             }
1787             for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
1788                 for (j = 0; j < 4; ++j) {
1789                     tcg_gen_st_i64(in[j], tcg_env, dofs + i + j * 8);
1790                 }
1791             }
1792             for (j = 0; j < 4; ++j) {
1793                 tcg_temp_free_i64(in[j]);
1794             }
1795         }
1796         if (oprsz < maxsz) {
1797             expand_clr(dofs + oprsz, maxsz - oprsz);
1798         }
1799     } else {
1800         g_assert_not_reached();
1801     }
1802 }
1803 
1804 void tcg_gen_gvec_dup_imm(unsigned vece, uint32_t dofs, uint32_t oprsz,
1805                           uint32_t maxsz, uint64_t x)
1806 {
1807     check_size_align(oprsz, maxsz, dofs);
1808     do_dup(vece, dofs, oprsz, maxsz, NULL, NULL, x);
1809 }
1810 
1811 void tcg_gen_gvec_not(unsigned vece, uint32_t dofs, uint32_t aofs,
1812                       uint32_t oprsz, uint32_t maxsz)
1813 {
1814     static const GVecGen2 g = {
1815         .fni8 = tcg_gen_not_i64,
1816         .fniv = tcg_gen_not_vec,
1817         .fno = gen_helper_gvec_not,
1818         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1819     };
1820     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g);
1821 }
1822 
1823 /* Perform a vector addition using normal addition and a mask.  The mask
1824    should be the sign bit of each lane.  This 6-operation form is more
1825    efficient than separate additions when there are 4 or more lanes in
1826    the 64-bit operation.  */
1827 static void gen_addv_mask(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 m)
1828 {
1829     TCGv_i64 t1 = tcg_temp_ebb_new_i64();
1830     TCGv_i64 t2 = tcg_temp_ebb_new_i64();
1831     TCGv_i64 t3 = tcg_temp_ebb_new_i64();
1832 
1833     tcg_gen_andc_i64(t1, a, m);
1834     tcg_gen_andc_i64(t2, b, m);
1835     tcg_gen_xor_i64(t3, a, b);
1836     tcg_gen_add_i64(d, t1, t2);
1837     tcg_gen_and_i64(t3, t3, m);
1838     tcg_gen_xor_i64(d, d, t3);
1839 
1840     tcg_temp_free_i64(t1);
1841     tcg_temp_free_i64(t2);
1842     tcg_temp_free_i64(t3);
1843 }
1844 
1845 void tcg_gen_vec_add8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1846 {
1847     TCGv_i64 m = tcg_constant_i64(dup_const(MO_8, 0x80));
1848     gen_addv_mask(d, a, b, m);
1849 }
1850 
1851 void tcg_gen_vec_add8_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
1852 {
1853     TCGv_i32 m = tcg_constant_i32((int32_t)dup_const(MO_8, 0x80));
1854     TCGv_i32 t1 = tcg_temp_ebb_new_i32();
1855     TCGv_i32 t2 = tcg_temp_ebb_new_i32();
1856     TCGv_i32 t3 = tcg_temp_ebb_new_i32();
1857 
1858     tcg_gen_andc_i32(t1, a, m);
1859     tcg_gen_andc_i32(t2, b, m);
1860     tcg_gen_xor_i32(t3, a, b);
1861     tcg_gen_add_i32(d, t1, t2);
1862     tcg_gen_and_i32(t3, t3, m);
1863     tcg_gen_xor_i32(d, d, t3);
1864 
1865     tcg_temp_free_i32(t1);
1866     tcg_temp_free_i32(t2);
1867     tcg_temp_free_i32(t3);
1868 }
1869 
1870 void tcg_gen_vec_add16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1871 {
1872     TCGv_i64 m = tcg_constant_i64(dup_const(MO_16, 0x8000));
1873     gen_addv_mask(d, a, b, m);
1874 }
1875 
1876 void tcg_gen_vec_add16_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
1877 {
1878     TCGv_i32 t1 = tcg_temp_ebb_new_i32();
1879     TCGv_i32 t2 = tcg_temp_ebb_new_i32();
1880 
1881     tcg_gen_andi_i32(t1, a, ~0xffff);
1882     tcg_gen_add_i32(t2, a, b);
1883     tcg_gen_add_i32(t1, t1, b);
1884     tcg_gen_deposit_i32(d, t1, t2, 0, 16);
1885 
1886     tcg_temp_free_i32(t1);
1887     tcg_temp_free_i32(t2);
1888 }
1889 
1890 void tcg_gen_vec_add32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1891 {
1892     TCGv_i64 t1 = tcg_temp_ebb_new_i64();
1893     TCGv_i64 t2 = tcg_temp_ebb_new_i64();
1894 
1895     tcg_gen_andi_i64(t1, a, ~0xffffffffull);
1896     tcg_gen_add_i64(t2, a, b);
1897     tcg_gen_add_i64(t1, t1, b);
1898     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
1899 
1900     tcg_temp_free_i64(t1);
1901     tcg_temp_free_i64(t2);
1902 }
1903 
1904 static const TCGOpcode vecop_list_add[] = { INDEX_op_add_vec, 0 };
1905 
1906 void tcg_gen_gvec_add(unsigned vece, uint32_t dofs, uint32_t aofs,
1907                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1908 {
1909     static const GVecGen3 g[4] = {
1910         { .fni8 = tcg_gen_vec_add8_i64,
1911           .fniv = tcg_gen_add_vec,
1912           .fno = gen_helper_gvec_add8,
1913           .opt_opc = vecop_list_add,
1914           .vece = MO_8 },
1915         { .fni8 = tcg_gen_vec_add16_i64,
1916           .fniv = tcg_gen_add_vec,
1917           .fno = gen_helper_gvec_add16,
1918           .opt_opc = vecop_list_add,
1919           .vece = MO_16 },
1920         { .fni4 = tcg_gen_add_i32,
1921           .fniv = tcg_gen_add_vec,
1922           .fno = gen_helper_gvec_add32,
1923           .opt_opc = vecop_list_add,
1924           .vece = MO_32 },
1925         { .fni8 = tcg_gen_add_i64,
1926           .fniv = tcg_gen_add_vec,
1927           .fno = gen_helper_gvec_add64,
1928           .opt_opc = vecop_list_add,
1929           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1930           .vece = MO_64 },
1931     };
1932 
1933     tcg_debug_assert(vece <= MO_64);
1934     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1935 }
1936 
1937 void tcg_gen_gvec_adds(unsigned vece, uint32_t dofs, uint32_t aofs,
1938                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
1939 {
1940     static const GVecGen2s g[4] = {
1941         { .fni8 = tcg_gen_vec_add8_i64,
1942           .fniv = tcg_gen_add_vec,
1943           .fno = gen_helper_gvec_adds8,
1944           .opt_opc = vecop_list_add,
1945           .vece = MO_8 },
1946         { .fni8 = tcg_gen_vec_add16_i64,
1947           .fniv = tcg_gen_add_vec,
1948           .fno = gen_helper_gvec_adds16,
1949           .opt_opc = vecop_list_add,
1950           .vece = MO_16 },
1951         { .fni4 = tcg_gen_add_i32,
1952           .fniv = tcg_gen_add_vec,
1953           .fno = gen_helper_gvec_adds32,
1954           .opt_opc = vecop_list_add,
1955           .vece = MO_32 },
1956         { .fni8 = tcg_gen_add_i64,
1957           .fniv = tcg_gen_add_vec,
1958           .fno = gen_helper_gvec_adds64,
1959           .opt_opc = vecop_list_add,
1960           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1961           .vece = MO_64 },
1962     };
1963 
1964     tcg_debug_assert(vece <= MO_64);
1965     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
1966 }
1967 
1968 void tcg_gen_gvec_addi(unsigned vece, uint32_t dofs, uint32_t aofs,
1969                        int64_t c, uint32_t oprsz, uint32_t maxsz)
1970 {
1971     TCGv_i64 tmp = tcg_constant_i64(c);
1972     tcg_gen_gvec_adds(vece, dofs, aofs, tmp, oprsz, maxsz);
1973 }
1974 
1975 static const TCGOpcode vecop_list_sub[] = { INDEX_op_sub_vec, 0 };
1976 
1977 void tcg_gen_gvec_subs(unsigned vece, uint32_t dofs, uint32_t aofs,
1978                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
1979 {
1980     static const GVecGen2s g[4] = {
1981         { .fni8 = tcg_gen_vec_sub8_i64,
1982           .fniv = tcg_gen_sub_vec,
1983           .fno = gen_helper_gvec_subs8,
1984           .opt_opc = vecop_list_sub,
1985           .vece = MO_8 },
1986         { .fni8 = tcg_gen_vec_sub16_i64,
1987           .fniv = tcg_gen_sub_vec,
1988           .fno = gen_helper_gvec_subs16,
1989           .opt_opc = vecop_list_sub,
1990           .vece = MO_16 },
1991         { .fni4 = tcg_gen_sub_i32,
1992           .fniv = tcg_gen_sub_vec,
1993           .fno = gen_helper_gvec_subs32,
1994           .opt_opc = vecop_list_sub,
1995           .vece = MO_32 },
1996         { .fni8 = tcg_gen_sub_i64,
1997           .fniv = tcg_gen_sub_vec,
1998           .fno = gen_helper_gvec_subs64,
1999           .opt_opc = vecop_list_sub,
2000           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2001           .vece = MO_64 },
2002     };
2003 
2004     tcg_debug_assert(vece <= MO_64);
2005     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
2006 }
2007 
2008 /* Perform a vector subtraction using normal subtraction and a mask.
2009    Compare gen_addv_mask above.  */
2010 static void gen_subv_mask(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 m)
2011 {
2012     TCGv_i64 t1 = tcg_temp_ebb_new_i64();
2013     TCGv_i64 t2 = tcg_temp_ebb_new_i64();
2014     TCGv_i64 t3 = tcg_temp_ebb_new_i64();
2015 
2016     tcg_gen_or_i64(t1, a, m);
2017     tcg_gen_andc_i64(t2, b, m);
2018     tcg_gen_eqv_i64(t3, a, b);
2019     tcg_gen_sub_i64(d, t1, t2);
2020     tcg_gen_and_i64(t3, t3, m);
2021     tcg_gen_xor_i64(d, d, t3);
2022 
2023     tcg_temp_free_i64(t1);
2024     tcg_temp_free_i64(t2);
2025     tcg_temp_free_i64(t3);
2026 }
2027 
2028 void tcg_gen_vec_sub8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2029 {
2030     TCGv_i64 m = tcg_constant_i64(dup_const(MO_8, 0x80));
2031     gen_subv_mask(d, a, b, m);
2032 }
2033 
2034 void tcg_gen_vec_sub8_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
2035 {
2036     TCGv_i32 m = tcg_constant_i32((int32_t)dup_const(MO_8, 0x80));
2037     TCGv_i32 t1 = tcg_temp_ebb_new_i32();
2038     TCGv_i32 t2 = tcg_temp_ebb_new_i32();
2039     TCGv_i32 t3 = tcg_temp_ebb_new_i32();
2040 
2041     tcg_gen_or_i32(t1, a, m);
2042     tcg_gen_andc_i32(t2, b, m);
2043     tcg_gen_eqv_i32(t3, a, b);
2044     tcg_gen_sub_i32(d, t1, t2);
2045     tcg_gen_and_i32(t3, t3, m);
2046     tcg_gen_xor_i32(d, d, t3);
2047 
2048     tcg_temp_free_i32(t1);
2049     tcg_temp_free_i32(t2);
2050     tcg_temp_free_i32(t3);
2051 }
2052 
2053 void tcg_gen_vec_sub16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2054 {
2055     TCGv_i64 m = tcg_constant_i64(dup_const(MO_16, 0x8000));
2056     gen_subv_mask(d, a, b, m);
2057 }
2058 
2059 void tcg_gen_vec_sub16_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
2060 {
2061     TCGv_i32 t1 = tcg_temp_ebb_new_i32();
2062     TCGv_i32 t2 = tcg_temp_ebb_new_i32();
2063 
2064     tcg_gen_andi_i32(t1, b, ~0xffff);
2065     tcg_gen_sub_i32(t2, a, b);
2066     tcg_gen_sub_i32(t1, a, t1);
2067     tcg_gen_deposit_i32(d, t1, t2, 0, 16);
2068 
2069     tcg_temp_free_i32(t1);
2070     tcg_temp_free_i32(t2);
2071 }
2072 
2073 void tcg_gen_vec_sub32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2074 {
2075     TCGv_i64 t1 = tcg_temp_ebb_new_i64();
2076     TCGv_i64 t2 = tcg_temp_ebb_new_i64();
2077 
2078     tcg_gen_andi_i64(t1, b, ~0xffffffffull);
2079     tcg_gen_sub_i64(t2, a, b);
2080     tcg_gen_sub_i64(t1, a, t1);
2081     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
2082 
2083     tcg_temp_free_i64(t1);
2084     tcg_temp_free_i64(t2);
2085 }
2086 
2087 void tcg_gen_gvec_sub(unsigned vece, uint32_t dofs, uint32_t aofs,
2088                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2089 {
2090     static const GVecGen3 g[4] = {
2091         { .fni8 = tcg_gen_vec_sub8_i64,
2092           .fniv = tcg_gen_sub_vec,
2093           .fno = gen_helper_gvec_sub8,
2094           .opt_opc = vecop_list_sub,
2095           .vece = MO_8 },
2096         { .fni8 = tcg_gen_vec_sub16_i64,
2097           .fniv = tcg_gen_sub_vec,
2098           .fno = gen_helper_gvec_sub16,
2099           .opt_opc = vecop_list_sub,
2100           .vece = MO_16 },
2101         { .fni4 = tcg_gen_sub_i32,
2102           .fniv = tcg_gen_sub_vec,
2103           .fno = gen_helper_gvec_sub32,
2104           .opt_opc = vecop_list_sub,
2105           .vece = MO_32 },
2106         { .fni8 = tcg_gen_sub_i64,
2107           .fniv = tcg_gen_sub_vec,
2108           .fno = gen_helper_gvec_sub64,
2109           .opt_opc = vecop_list_sub,
2110           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2111           .vece = MO_64 },
2112     };
2113 
2114     tcg_debug_assert(vece <= MO_64);
2115     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2116 }
2117 
2118 static const TCGOpcode vecop_list_mul[] = { INDEX_op_mul_vec, 0 };
2119 
2120 void tcg_gen_gvec_mul(unsigned vece, uint32_t dofs, uint32_t aofs,
2121                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2122 {
2123     static const GVecGen3 g[4] = {
2124         { .fniv = tcg_gen_mul_vec,
2125           .fno = gen_helper_gvec_mul8,
2126           .opt_opc = vecop_list_mul,
2127           .vece = MO_8 },
2128         { .fniv = tcg_gen_mul_vec,
2129           .fno = gen_helper_gvec_mul16,
2130           .opt_opc = vecop_list_mul,
2131           .vece = MO_16 },
2132         { .fni4 = tcg_gen_mul_i32,
2133           .fniv = tcg_gen_mul_vec,
2134           .fno = gen_helper_gvec_mul32,
2135           .opt_opc = vecop_list_mul,
2136           .vece = MO_32 },
2137         { .fni8 = tcg_gen_mul_i64,
2138           .fniv = tcg_gen_mul_vec,
2139           .fno = gen_helper_gvec_mul64,
2140           .opt_opc = vecop_list_mul,
2141           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2142           .vece = MO_64 },
2143     };
2144 
2145     tcg_debug_assert(vece <= MO_64);
2146     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2147 }
2148 
2149 void tcg_gen_gvec_muls(unsigned vece, uint32_t dofs, uint32_t aofs,
2150                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2151 {
2152     static const GVecGen2s g[4] = {
2153         { .fniv = tcg_gen_mul_vec,
2154           .fno = gen_helper_gvec_muls8,
2155           .opt_opc = vecop_list_mul,
2156           .vece = MO_8 },
2157         { .fniv = tcg_gen_mul_vec,
2158           .fno = gen_helper_gvec_muls16,
2159           .opt_opc = vecop_list_mul,
2160           .vece = MO_16 },
2161         { .fni4 = tcg_gen_mul_i32,
2162           .fniv = tcg_gen_mul_vec,
2163           .fno = gen_helper_gvec_muls32,
2164           .opt_opc = vecop_list_mul,
2165           .vece = MO_32 },
2166         { .fni8 = tcg_gen_mul_i64,
2167           .fniv = tcg_gen_mul_vec,
2168           .fno = gen_helper_gvec_muls64,
2169           .opt_opc = vecop_list_mul,
2170           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2171           .vece = MO_64 },
2172     };
2173 
2174     tcg_debug_assert(vece <= MO_64);
2175     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
2176 }
2177 
2178 void tcg_gen_gvec_muli(unsigned vece, uint32_t dofs, uint32_t aofs,
2179                        int64_t c, uint32_t oprsz, uint32_t maxsz)
2180 {
2181     TCGv_i64 tmp = tcg_constant_i64(c);
2182     tcg_gen_gvec_muls(vece, dofs, aofs, tmp, oprsz, maxsz);
2183 }
2184 
2185 void tcg_gen_gvec_ssadd(unsigned vece, uint32_t dofs, uint32_t aofs,
2186                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2187 {
2188     static const TCGOpcode vecop_list[] = { INDEX_op_ssadd_vec, 0 };
2189     static const GVecGen3 g[4] = {
2190         { .fniv = tcg_gen_ssadd_vec,
2191           .fno = gen_helper_gvec_ssadd8,
2192           .opt_opc = vecop_list,
2193           .vece = MO_8 },
2194         { .fniv = tcg_gen_ssadd_vec,
2195           .fno = gen_helper_gvec_ssadd16,
2196           .opt_opc = vecop_list,
2197           .vece = MO_16 },
2198         { .fniv = tcg_gen_ssadd_vec,
2199           .fno = gen_helper_gvec_ssadd32,
2200           .opt_opc = vecop_list,
2201           .vece = MO_32 },
2202         { .fniv = tcg_gen_ssadd_vec,
2203           .fno = gen_helper_gvec_ssadd64,
2204           .opt_opc = vecop_list,
2205           .vece = MO_64 },
2206     };
2207     tcg_debug_assert(vece <= MO_64);
2208     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2209 }
2210 
2211 void tcg_gen_gvec_sssub(unsigned vece, uint32_t dofs, uint32_t aofs,
2212                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2213 {
2214     static const TCGOpcode vecop_list[] = { INDEX_op_sssub_vec, 0 };
2215     static const GVecGen3 g[4] = {
2216         { .fniv = tcg_gen_sssub_vec,
2217           .fno = gen_helper_gvec_sssub8,
2218           .opt_opc = vecop_list,
2219           .vece = MO_8 },
2220         { .fniv = tcg_gen_sssub_vec,
2221           .fno = gen_helper_gvec_sssub16,
2222           .opt_opc = vecop_list,
2223           .vece = MO_16 },
2224         { .fniv = tcg_gen_sssub_vec,
2225           .fno = gen_helper_gvec_sssub32,
2226           .opt_opc = vecop_list,
2227           .vece = MO_32 },
2228         { .fniv = tcg_gen_sssub_vec,
2229           .fno = gen_helper_gvec_sssub64,
2230           .opt_opc = vecop_list,
2231           .vece = MO_64 },
2232     };
2233     tcg_debug_assert(vece <= MO_64);
2234     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2235 }
2236 
2237 static void tcg_gen_usadd_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
2238 {
2239     TCGv_i32 max = tcg_constant_i32(-1);
2240     tcg_gen_add_i32(d, a, b);
2241     tcg_gen_movcond_i32(TCG_COND_LTU, d, d, a, max, d);
2242 }
2243 
2244 static void tcg_gen_usadd_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2245 {
2246     TCGv_i64 max = tcg_constant_i64(-1);
2247     tcg_gen_add_i64(d, a, b);
2248     tcg_gen_movcond_i64(TCG_COND_LTU, d, d, a, max, d);
2249 }
2250 
2251 void tcg_gen_gvec_usadd(unsigned vece, uint32_t dofs, uint32_t aofs,
2252                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2253 {
2254     static const TCGOpcode vecop_list[] = { INDEX_op_usadd_vec, 0 };
2255     static const GVecGen3 g[4] = {
2256         { .fniv = tcg_gen_usadd_vec,
2257           .fno = gen_helper_gvec_usadd8,
2258           .opt_opc = vecop_list,
2259           .vece = MO_8 },
2260         { .fniv = tcg_gen_usadd_vec,
2261           .fno = gen_helper_gvec_usadd16,
2262           .opt_opc = vecop_list,
2263           .vece = MO_16 },
2264         { .fni4 = tcg_gen_usadd_i32,
2265           .fniv = tcg_gen_usadd_vec,
2266           .fno = gen_helper_gvec_usadd32,
2267           .opt_opc = vecop_list,
2268           .vece = MO_32 },
2269         { .fni8 = tcg_gen_usadd_i64,
2270           .fniv = tcg_gen_usadd_vec,
2271           .fno = gen_helper_gvec_usadd64,
2272           .opt_opc = vecop_list,
2273           .vece = MO_64 }
2274     };
2275     tcg_debug_assert(vece <= MO_64);
2276     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2277 }
2278 
2279 static void tcg_gen_ussub_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
2280 {
2281     TCGv_i32 min = tcg_constant_i32(0);
2282     tcg_gen_sub_i32(d, a, b);
2283     tcg_gen_movcond_i32(TCG_COND_LTU, d, a, b, min, d);
2284 }
2285 
2286 static void tcg_gen_ussub_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2287 {
2288     TCGv_i64 min = tcg_constant_i64(0);
2289     tcg_gen_sub_i64(d, a, b);
2290     tcg_gen_movcond_i64(TCG_COND_LTU, d, a, b, min, d);
2291 }
2292 
2293 void tcg_gen_gvec_ussub(unsigned vece, uint32_t dofs, uint32_t aofs,
2294                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2295 {
2296     static const TCGOpcode vecop_list[] = { INDEX_op_ussub_vec, 0 };
2297     static const GVecGen3 g[4] = {
2298         { .fniv = tcg_gen_ussub_vec,
2299           .fno = gen_helper_gvec_ussub8,
2300           .opt_opc = vecop_list,
2301           .vece = MO_8 },
2302         { .fniv = tcg_gen_ussub_vec,
2303           .fno = gen_helper_gvec_ussub16,
2304           .opt_opc = vecop_list,
2305           .vece = MO_16 },
2306         { .fni4 = tcg_gen_ussub_i32,
2307           .fniv = tcg_gen_ussub_vec,
2308           .fno = gen_helper_gvec_ussub32,
2309           .opt_opc = vecop_list,
2310           .vece = MO_32 },
2311         { .fni8 = tcg_gen_ussub_i64,
2312           .fniv = tcg_gen_ussub_vec,
2313           .fno = gen_helper_gvec_ussub64,
2314           .opt_opc = vecop_list,
2315           .vece = MO_64 }
2316     };
2317     tcg_debug_assert(vece <= MO_64);
2318     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2319 }
2320 
2321 void tcg_gen_gvec_smin(unsigned vece, uint32_t dofs, uint32_t aofs,
2322                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2323 {
2324     static const TCGOpcode vecop_list[] = { INDEX_op_smin_vec, 0 };
2325     static const GVecGen3 g[4] = {
2326         { .fniv = tcg_gen_smin_vec,
2327           .fno = gen_helper_gvec_smin8,
2328           .opt_opc = vecop_list,
2329           .vece = MO_8 },
2330         { .fniv = tcg_gen_smin_vec,
2331           .fno = gen_helper_gvec_smin16,
2332           .opt_opc = vecop_list,
2333           .vece = MO_16 },
2334         { .fni4 = tcg_gen_smin_i32,
2335           .fniv = tcg_gen_smin_vec,
2336           .fno = gen_helper_gvec_smin32,
2337           .opt_opc = vecop_list,
2338           .vece = MO_32 },
2339         { .fni8 = tcg_gen_smin_i64,
2340           .fniv = tcg_gen_smin_vec,
2341           .fno = gen_helper_gvec_smin64,
2342           .opt_opc = vecop_list,
2343           .vece = MO_64 }
2344     };
2345     tcg_debug_assert(vece <= MO_64);
2346     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2347 }
2348 
2349 void tcg_gen_gvec_umin(unsigned vece, uint32_t dofs, uint32_t aofs,
2350                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2351 {
2352     static const TCGOpcode vecop_list[] = { INDEX_op_umin_vec, 0 };
2353     static const GVecGen3 g[4] = {
2354         { .fniv = tcg_gen_umin_vec,
2355           .fno = gen_helper_gvec_umin8,
2356           .opt_opc = vecop_list,
2357           .vece = MO_8 },
2358         { .fniv = tcg_gen_umin_vec,
2359           .fno = gen_helper_gvec_umin16,
2360           .opt_opc = vecop_list,
2361           .vece = MO_16 },
2362         { .fni4 = tcg_gen_umin_i32,
2363           .fniv = tcg_gen_umin_vec,
2364           .fno = gen_helper_gvec_umin32,
2365           .opt_opc = vecop_list,
2366           .vece = MO_32 },
2367         { .fni8 = tcg_gen_umin_i64,
2368           .fniv = tcg_gen_umin_vec,
2369           .fno = gen_helper_gvec_umin64,
2370           .opt_opc = vecop_list,
2371           .vece = MO_64 }
2372     };
2373     tcg_debug_assert(vece <= MO_64);
2374     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2375 }
2376 
2377 void tcg_gen_gvec_smax(unsigned vece, uint32_t dofs, uint32_t aofs,
2378                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2379 {
2380     static const TCGOpcode vecop_list[] = { INDEX_op_smax_vec, 0 };
2381     static const GVecGen3 g[4] = {
2382         { .fniv = tcg_gen_smax_vec,
2383           .fno = gen_helper_gvec_smax8,
2384           .opt_opc = vecop_list,
2385           .vece = MO_8 },
2386         { .fniv = tcg_gen_smax_vec,
2387           .fno = gen_helper_gvec_smax16,
2388           .opt_opc = vecop_list,
2389           .vece = MO_16 },
2390         { .fni4 = tcg_gen_smax_i32,
2391           .fniv = tcg_gen_smax_vec,
2392           .fno = gen_helper_gvec_smax32,
2393           .opt_opc = vecop_list,
2394           .vece = MO_32 },
2395         { .fni8 = tcg_gen_smax_i64,
2396           .fniv = tcg_gen_smax_vec,
2397           .fno = gen_helper_gvec_smax64,
2398           .opt_opc = vecop_list,
2399           .vece = MO_64 }
2400     };
2401     tcg_debug_assert(vece <= MO_64);
2402     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2403 }
2404 
2405 void tcg_gen_gvec_umax(unsigned vece, uint32_t dofs, uint32_t aofs,
2406                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2407 {
2408     static const TCGOpcode vecop_list[] = { INDEX_op_umax_vec, 0 };
2409     static const GVecGen3 g[4] = {
2410         { .fniv = tcg_gen_umax_vec,
2411           .fno = gen_helper_gvec_umax8,
2412           .opt_opc = vecop_list,
2413           .vece = MO_8 },
2414         { .fniv = tcg_gen_umax_vec,
2415           .fno = gen_helper_gvec_umax16,
2416           .opt_opc = vecop_list,
2417           .vece = MO_16 },
2418         { .fni4 = tcg_gen_umax_i32,
2419           .fniv = tcg_gen_umax_vec,
2420           .fno = gen_helper_gvec_umax32,
2421           .opt_opc = vecop_list,
2422           .vece = MO_32 },
2423         { .fni8 = tcg_gen_umax_i64,
2424           .fniv = tcg_gen_umax_vec,
2425           .fno = gen_helper_gvec_umax64,
2426           .opt_opc = vecop_list,
2427           .vece = MO_64 }
2428     };
2429     tcg_debug_assert(vece <= MO_64);
2430     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2431 }
2432 
2433 /* Perform a vector negation using normal negation and a mask.
2434    Compare gen_subv_mask above.  */
2435 static void gen_negv_mask(TCGv_i64 d, TCGv_i64 b, TCGv_i64 m)
2436 {
2437     TCGv_i64 t2 = tcg_temp_ebb_new_i64();
2438     TCGv_i64 t3 = tcg_temp_ebb_new_i64();
2439 
2440     tcg_gen_andc_i64(t3, m, b);
2441     tcg_gen_andc_i64(t2, b, m);
2442     tcg_gen_sub_i64(d, m, t2);
2443     tcg_gen_xor_i64(d, d, t3);
2444 
2445     tcg_temp_free_i64(t2);
2446     tcg_temp_free_i64(t3);
2447 }
2448 
2449 void tcg_gen_vec_neg8_i64(TCGv_i64 d, TCGv_i64 b)
2450 {
2451     TCGv_i64 m = tcg_constant_i64(dup_const(MO_8, 0x80));
2452     gen_negv_mask(d, b, m);
2453 }
2454 
2455 void tcg_gen_vec_neg16_i64(TCGv_i64 d, TCGv_i64 b)
2456 {
2457     TCGv_i64 m = tcg_constant_i64(dup_const(MO_16, 0x8000));
2458     gen_negv_mask(d, b, m);
2459 }
2460 
2461 void tcg_gen_vec_neg32_i64(TCGv_i64 d, TCGv_i64 b)
2462 {
2463     TCGv_i64 t1 = tcg_temp_ebb_new_i64();
2464     TCGv_i64 t2 = tcg_temp_ebb_new_i64();
2465 
2466     tcg_gen_andi_i64(t1, b, ~0xffffffffull);
2467     tcg_gen_neg_i64(t2, b);
2468     tcg_gen_neg_i64(t1, t1);
2469     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
2470 
2471     tcg_temp_free_i64(t1);
2472     tcg_temp_free_i64(t2);
2473 }
2474 
2475 void tcg_gen_gvec_neg(unsigned vece, uint32_t dofs, uint32_t aofs,
2476                       uint32_t oprsz, uint32_t maxsz)
2477 {
2478     static const TCGOpcode vecop_list[] = { INDEX_op_neg_vec, 0 };
2479     static const GVecGen2 g[4] = {
2480         { .fni8 = tcg_gen_vec_neg8_i64,
2481           .fniv = tcg_gen_neg_vec,
2482           .fno = gen_helper_gvec_neg8,
2483           .opt_opc = vecop_list,
2484           .vece = MO_8 },
2485         { .fni8 = tcg_gen_vec_neg16_i64,
2486           .fniv = tcg_gen_neg_vec,
2487           .fno = gen_helper_gvec_neg16,
2488           .opt_opc = vecop_list,
2489           .vece = MO_16 },
2490         { .fni4 = tcg_gen_neg_i32,
2491           .fniv = tcg_gen_neg_vec,
2492           .fno = gen_helper_gvec_neg32,
2493           .opt_opc = vecop_list,
2494           .vece = MO_32 },
2495         { .fni8 = tcg_gen_neg_i64,
2496           .fniv = tcg_gen_neg_vec,
2497           .fno = gen_helper_gvec_neg64,
2498           .opt_opc = vecop_list,
2499           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2500           .vece = MO_64 },
2501     };
2502 
2503     tcg_debug_assert(vece <= MO_64);
2504     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g[vece]);
2505 }
2506 
2507 static void gen_absv_mask(TCGv_i64 d, TCGv_i64 b, unsigned vece)
2508 {
2509     TCGv_i64 t = tcg_temp_ebb_new_i64();
2510     int nbit = 8 << vece;
2511 
2512     /* Create -1 for each negative element.  */
2513     tcg_gen_shri_i64(t, b, nbit - 1);
2514     tcg_gen_andi_i64(t, t, dup_const(vece, 1));
2515     tcg_gen_muli_i64(t, t, (1 << nbit) - 1);
2516 
2517     /*
2518      * Invert (via xor -1) and add one.
2519      * Because of the ordering the msb is cleared,
2520      * so we never have carry into the next element.
2521      */
2522     tcg_gen_xor_i64(d, b, t);
2523     tcg_gen_andi_i64(t, t, dup_const(vece, 1));
2524     tcg_gen_add_i64(d, d, t);
2525 
2526     tcg_temp_free_i64(t);
2527 }
2528 
2529 static void tcg_gen_vec_abs8_i64(TCGv_i64 d, TCGv_i64 b)
2530 {
2531     gen_absv_mask(d, b, MO_8);
2532 }
2533 
2534 static void tcg_gen_vec_abs16_i64(TCGv_i64 d, TCGv_i64 b)
2535 {
2536     gen_absv_mask(d, b, MO_16);
2537 }
2538 
2539 void tcg_gen_gvec_abs(unsigned vece, uint32_t dofs, uint32_t aofs,
2540                       uint32_t oprsz, uint32_t maxsz)
2541 {
2542     static const TCGOpcode vecop_list[] = { INDEX_op_abs_vec, 0 };
2543     static const GVecGen2 g[4] = {
2544         { .fni8 = tcg_gen_vec_abs8_i64,
2545           .fniv = tcg_gen_abs_vec,
2546           .fno = gen_helper_gvec_abs8,
2547           .opt_opc = vecop_list,
2548           .vece = MO_8 },
2549         { .fni8 = tcg_gen_vec_abs16_i64,
2550           .fniv = tcg_gen_abs_vec,
2551           .fno = gen_helper_gvec_abs16,
2552           .opt_opc = vecop_list,
2553           .vece = MO_16 },
2554         { .fni4 = tcg_gen_abs_i32,
2555           .fniv = tcg_gen_abs_vec,
2556           .fno = gen_helper_gvec_abs32,
2557           .opt_opc = vecop_list,
2558           .vece = MO_32 },
2559         { .fni8 = tcg_gen_abs_i64,
2560           .fniv = tcg_gen_abs_vec,
2561           .fno = gen_helper_gvec_abs64,
2562           .opt_opc = vecop_list,
2563           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2564           .vece = MO_64 },
2565     };
2566 
2567     tcg_debug_assert(vece <= MO_64);
2568     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g[vece]);
2569 }
2570 
2571 void tcg_gen_gvec_and(unsigned vece, uint32_t dofs, uint32_t aofs,
2572                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2573 {
2574     static const GVecGen3 g = {
2575         .fni8 = tcg_gen_and_i64,
2576         .fniv = tcg_gen_and_vec,
2577         .fno = gen_helper_gvec_and,
2578         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2579     };
2580 
2581     if (aofs == bofs) {
2582         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2583     } else {
2584         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2585     }
2586 }
2587 
2588 void tcg_gen_gvec_or(unsigned vece, uint32_t dofs, uint32_t aofs,
2589                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2590 {
2591     static const GVecGen3 g = {
2592         .fni8 = tcg_gen_or_i64,
2593         .fniv = tcg_gen_or_vec,
2594         .fno = gen_helper_gvec_or,
2595         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2596     };
2597 
2598     if (aofs == bofs) {
2599         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2600     } else {
2601         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2602     }
2603 }
2604 
2605 void tcg_gen_gvec_xor(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_xor_i64,
2610         .fniv = tcg_gen_xor_vec,
2611         .fno = gen_helper_gvec_xor,
2612         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2613     };
2614 
2615     if (aofs == bofs) {
2616         tcg_gen_gvec_dup_imm(MO_64, dofs, oprsz, maxsz, 0);
2617     } else {
2618         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2619     }
2620 }
2621 
2622 void tcg_gen_gvec_andc(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_andc_i64,
2627         .fniv = tcg_gen_andc_vec,
2628         .fno = gen_helper_gvec_andc,
2629         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2630     };
2631 
2632     if (aofs == bofs) {
2633         tcg_gen_gvec_dup_imm(MO_64, dofs, oprsz, maxsz, 0);
2634     } else {
2635         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2636     }
2637 }
2638 
2639 void tcg_gen_gvec_orc(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_orc_i64,
2644         .fniv = tcg_gen_orc_vec,
2645         .fno = gen_helper_gvec_orc,
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, -1);
2651     } else {
2652         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2653     }
2654 }
2655 
2656 void tcg_gen_gvec_nand(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_nand_i64,
2661         .fniv = tcg_gen_nand_vec,
2662         .fno = gen_helper_gvec_nand,
2663         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2664     };
2665 
2666     if (aofs == bofs) {
2667         tcg_gen_gvec_not(vece, dofs, aofs, oprsz, maxsz);
2668     } else {
2669         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2670     }
2671 }
2672 
2673 void tcg_gen_gvec_nor(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_nor_i64,
2678         .fniv = tcg_gen_nor_vec,
2679         .fno = gen_helper_gvec_nor,
2680         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2681     };
2682 
2683     if (aofs == bofs) {
2684         tcg_gen_gvec_not(vece, dofs, aofs, oprsz, maxsz);
2685     } else {
2686         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2687     }
2688 }
2689 
2690 void tcg_gen_gvec_eqv(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_eqv_i64,
2695         .fniv = tcg_gen_eqv_vec,
2696         .fno = gen_helper_gvec_eqv,
2697         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2698     };
2699 
2700     if (aofs == bofs) {
2701         tcg_gen_gvec_dup_imm(MO_64, dofs, oprsz, maxsz, -1);
2702     } else {
2703         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2704     }
2705 }
2706 
2707 static const GVecGen2s gop_ands = {
2708     .fni8 = tcg_gen_and_i64,
2709     .fniv = tcg_gen_and_vec,
2710     .fno = gen_helper_gvec_ands,
2711     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2712     .vece = MO_64
2713 };
2714 
2715 void tcg_gen_gvec_ands(unsigned vece, uint32_t dofs, uint32_t aofs,
2716                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2717 {
2718     TCGv_i64 tmp = tcg_temp_ebb_new_i64();
2719     tcg_gen_dup_i64(vece, tmp, c);
2720     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ands);
2721     tcg_temp_free_i64(tmp);
2722 }
2723 
2724 void tcg_gen_gvec_andi(unsigned vece, uint32_t dofs, uint32_t aofs,
2725                        int64_t c, uint32_t oprsz, uint32_t maxsz)
2726 {
2727     TCGv_i64 tmp = tcg_constant_i64(dup_const(vece, c));
2728     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ands);
2729 }
2730 
2731 void tcg_gen_gvec_andcs(unsigned vece, uint32_t dofs, uint32_t aofs,
2732                         TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2733 {
2734     static GVecGen2s g = {
2735         .fni8 = tcg_gen_andc_i64,
2736         .fniv = tcg_gen_andc_vec,
2737         .fno = gen_helper_gvec_andcs,
2738         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2739         .vece = MO_64
2740     };
2741 
2742     TCGv_i64 tmp = tcg_temp_ebb_new_i64();
2743     tcg_gen_dup_i64(vece, tmp, c);
2744     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &g);
2745     tcg_temp_free_i64(tmp);
2746 }
2747 
2748 static const GVecGen2s gop_xors = {
2749     .fni8 = tcg_gen_xor_i64,
2750     .fniv = tcg_gen_xor_vec,
2751     .fno = gen_helper_gvec_xors,
2752     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2753     .vece = MO_64
2754 };
2755 
2756 void tcg_gen_gvec_xors(unsigned vece, uint32_t dofs, uint32_t aofs,
2757                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2758 {
2759     TCGv_i64 tmp = tcg_temp_ebb_new_i64();
2760     tcg_gen_dup_i64(vece, tmp, c);
2761     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_xors);
2762     tcg_temp_free_i64(tmp);
2763 }
2764 
2765 void tcg_gen_gvec_xori(unsigned vece, uint32_t dofs, uint32_t aofs,
2766                        int64_t c, uint32_t oprsz, uint32_t maxsz)
2767 {
2768     TCGv_i64 tmp = tcg_constant_i64(dup_const(vece, c));
2769     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_xors);
2770 }
2771 
2772 static const GVecGen2s gop_ors = {
2773     .fni8 = tcg_gen_or_i64,
2774     .fniv = tcg_gen_or_vec,
2775     .fno = gen_helper_gvec_ors,
2776     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2777     .vece = MO_64
2778 };
2779 
2780 void tcg_gen_gvec_ors(unsigned vece, uint32_t dofs, uint32_t aofs,
2781                       TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2782 {
2783     TCGv_i64 tmp = tcg_temp_ebb_new_i64();
2784     tcg_gen_dup_i64(vece, tmp, c);
2785     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ors);
2786     tcg_temp_free_i64(tmp);
2787 }
2788 
2789 void tcg_gen_gvec_ori(unsigned vece, uint32_t dofs, uint32_t aofs,
2790                       int64_t c, uint32_t oprsz, uint32_t maxsz)
2791 {
2792     TCGv_i64 tmp = tcg_constant_i64(dup_const(vece, c));
2793     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ors);
2794 }
2795 
2796 void tcg_gen_vec_shl8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2797 {
2798     uint64_t mask = dup_const(MO_8, 0xff << c);
2799     tcg_gen_shli_i64(d, a, c);
2800     tcg_gen_andi_i64(d, d, mask);
2801 }
2802 
2803 void tcg_gen_vec_shl16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2804 {
2805     uint64_t mask = dup_const(MO_16, 0xffff << c);
2806     tcg_gen_shli_i64(d, a, c);
2807     tcg_gen_andi_i64(d, d, mask);
2808 }
2809 
2810 void tcg_gen_vec_shl8i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2811 {
2812     uint32_t mask = dup_const(MO_8, 0xff << c);
2813     tcg_gen_shli_i32(d, a, c);
2814     tcg_gen_andi_i32(d, d, mask);
2815 }
2816 
2817 void tcg_gen_vec_shl16i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2818 {
2819     uint32_t mask = dup_const(MO_16, 0xffff << c);
2820     tcg_gen_shli_i32(d, a, c);
2821     tcg_gen_andi_i32(d, d, mask);
2822 }
2823 
2824 void tcg_gen_gvec_shli(unsigned vece, uint32_t dofs, uint32_t aofs,
2825                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
2826 {
2827     static const TCGOpcode vecop_list[] = { INDEX_op_shli_vec, 0 };
2828     static const GVecGen2i g[4] = {
2829         { .fni8 = tcg_gen_vec_shl8i_i64,
2830           .fniv = tcg_gen_shli_vec,
2831           .fno = gen_helper_gvec_shl8i,
2832           .opt_opc = vecop_list,
2833           .vece = MO_8 },
2834         { .fni8 = tcg_gen_vec_shl16i_i64,
2835           .fniv = tcg_gen_shli_vec,
2836           .fno = gen_helper_gvec_shl16i,
2837           .opt_opc = vecop_list,
2838           .vece = MO_16 },
2839         { .fni4 = tcg_gen_shli_i32,
2840           .fniv = tcg_gen_shli_vec,
2841           .fno = gen_helper_gvec_shl32i,
2842           .opt_opc = vecop_list,
2843           .vece = MO_32 },
2844         { .fni8 = tcg_gen_shli_i64,
2845           .fniv = tcg_gen_shli_vec,
2846           .fno = gen_helper_gvec_shl64i,
2847           .opt_opc = vecop_list,
2848           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2849           .vece = MO_64 },
2850     };
2851 
2852     tcg_debug_assert(vece <= MO_64);
2853     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2854     if (shift == 0) {
2855         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2856     } else {
2857         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2858     }
2859 }
2860 
2861 void tcg_gen_vec_shr8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2862 {
2863     uint64_t mask = dup_const(MO_8, 0xff >> c);
2864     tcg_gen_shri_i64(d, a, c);
2865     tcg_gen_andi_i64(d, d, mask);
2866 }
2867 
2868 void tcg_gen_vec_shr16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2869 {
2870     uint64_t mask = dup_const(MO_16, 0xffff >> c);
2871     tcg_gen_shri_i64(d, a, c);
2872     tcg_gen_andi_i64(d, d, mask);
2873 }
2874 
2875 void tcg_gen_vec_shr8i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2876 {
2877     uint32_t mask = dup_const(MO_8, 0xff >> c);
2878     tcg_gen_shri_i32(d, a, c);
2879     tcg_gen_andi_i32(d, d, mask);
2880 }
2881 
2882 void tcg_gen_vec_shr16i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2883 {
2884     uint32_t mask = dup_const(MO_16, 0xffff >> c);
2885     tcg_gen_shri_i32(d, a, c);
2886     tcg_gen_andi_i32(d, d, mask);
2887 }
2888 
2889 void tcg_gen_gvec_shri(unsigned vece, uint32_t dofs, uint32_t aofs,
2890                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
2891 {
2892     static const TCGOpcode vecop_list[] = { INDEX_op_shri_vec, 0 };
2893     static const GVecGen2i g[4] = {
2894         { .fni8 = tcg_gen_vec_shr8i_i64,
2895           .fniv = tcg_gen_shri_vec,
2896           .fno = gen_helper_gvec_shr8i,
2897           .opt_opc = vecop_list,
2898           .vece = MO_8 },
2899         { .fni8 = tcg_gen_vec_shr16i_i64,
2900           .fniv = tcg_gen_shri_vec,
2901           .fno = gen_helper_gvec_shr16i,
2902           .opt_opc = vecop_list,
2903           .vece = MO_16 },
2904         { .fni4 = tcg_gen_shri_i32,
2905           .fniv = tcg_gen_shri_vec,
2906           .fno = gen_helper_gvec_shr32i,
2907           .opt_opc = vecop_list,
2908           .vece = MO_32 },
2909         { .fni8 = tcg_gen_shri_i64,
2910           .fniv = tcg_gen_shri_vec,
2911           .fno = gen_helper_gvec_shr64i,
2912           .opt_opc = vecop_list,
2913           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2914           .vece = MO_64 },
2915     };
2916 
2917     tcg_debug_assert(vece <= MO_64);
2918     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2919     if (shift == 0) {
2920         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2921     } else {
2922         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2923     }
2924 }
2925 
2926 void tcg_gen_vec_sar8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2927 {
2928     uint64_t s_mask = dup_const(MO_8, 0x80 >> c);
2929     uint64_t c_mask = dup_const(MO_8, 0xff >> c);
2930     TCGv_i64 s = tcg_temp_ebb_new_i64();
2931 
2932     tcg_gen_shri_i64(d, a, c);
2933     tcg_gen_andi_i64(s, d, s_mask);  /* isolate (shifted) sign bit */
2934     tcg_gen_muli_i64(s, s, (2 << c) - 2); /* replicate isolated signs */
2935     tcg_gen_andi_i64(d, d, c_mask);  /* clear out bits above sign  */
2936     tcg_gen_or_i64(d, d, s);         /* include sign extension */
2937     tcg_temp_free_i64(s);
2938 }
2939 
2940 void tcg_gen_vec_sar16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2941 {
2942     uint64_t s_mask = dup_const(MO_16, 0x8000 >> c);
2943     uint64_t c_mask = dup_const(MO_16, 0xffff >> c);
2944     TCGv_i64 s = tcg_temp_ebb_new_i64();
2945 
2946     tcg_gen_shri_i64(d, a, c);
2947     tcg_gen_andi_i64(s, d, s_mask);  /* isolate (shifted) sign bit */
2948     tcg_gen_andi_i64(d, d, c_mask);  /* clear out bits above sign  */
2949     tcg_gen_muli_i64(s, s, (2 << c) - 2); /* replicate isolated signs */
2950     tcg_gen_or_i64(d, d, s);         /* include sign extension */
2951     tcg_temp_free_i64(s);
2952 }
2953 
2954 void tcg_gen_vec_sar8i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2955 {
2956     uint32_t s_mask = dup_const(MO_8, 0x80 >> c);
2957     uint32_t c_mask = dup_const(MO_8, 0xff >> c);
2958     TCGv_i32 s = tcg_temp_ebb_new_i32();
2959 
2960     tcg_gen_shri_i32(d, a, c);
2961     tcg_gen_andi_i32(s, d, s_mask);  /* isolate (shifted) sign bit */
2962     tcg_gen_muli_i32(s, s, (2 << c) - 2); /* replicate isolated signs */
2963     tcg_gen_andi_i32(d, d, c_mask);  /* clear out bits above sign  */
2964     tcg_gen_or_i32(d, d, s);         /* include sign extension */
2965     tcg_temp_free_i32(s);
2966 }
2967 
2968 void tcg_gen_vec_sar16i_i32(TCGv_i32 d, TCGv_i32 a, int32_t c)
2969 {
2970     uint32_t s_mask = dup_const(MO_16, 0x8000 >> c);
2971     uint32_t c_mask = dup_const(MO_16, 0xffff >> c);
2972     TCGv_i32 s = tcg_temp_ebb_new_i32();
2973 
2974     tcg_gen_shri_i32(d, a, c);
2975     tcg_gen_andi_i32(s, d, s_mask);  /* isolate (shifted) sign bit */
2976     tcg_gen_andi_i32(d, d, c_mask);  /* clear out bits above sign  */
2977     tcg_gen_muli_i32(s, s, (2 << c) - 2); /* replicate isolated signs */
2978     tcg_gen_or_i32(d, d, s);         /* include sign extension */
2979     tcg_temp_free_i32(s);
2980 }
2981 
2982 void tcg_gen_gvec_sari(unsigned vece, uint32_t dofs, uint32_t aofs,
2983                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
2984 {
2985     static const TCGOpcode vecop_list[] = { INDEX_op_sari_vec, 0 };
2986     static const GVecGen2i g[4] = {
2987         { .fni8 = tcg_gen_vec_sar8i_i64,
2988           .fniv = tcg_gen_sari_vec,
2989           .fno = gen_helper_gvec_sar8i,
2990           .opt_opc = vecop_list,
2991           .vece = MO_8 },
2992         { .fni8 = tcg_gen_vec_sar16i_i64,
2993           .fniv = tcg_gen_sari_vec,
2994           .fno = gen_helper_gvec_sar16i,
2995           .opt_opc = vecop_list,
2996           .vece = MO_16 },
2997         { .fni4 = tcg_gen_sari_i32,
2998           .fniv = tcg_gen_sari_vec,
2999           .fno = gen_helper_gvec_sar32i,
3000           .opt_opc = vecop_list,
3001           .vece = MO_32 },
3002         { .fni8 = tcg_gen_sari_i64,
3003           .fniv = tcg_gen_sari_vec,
3004           .fno = gen_helper_gvec_sar64i,
3005           .opt_opc = vecop_list,
3006           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3007           .vece = MO_64 },
3008     };
3009 
3010     tcg_debug_assert(vece <= MO_64);
3011     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
3012     if (shift == 0) {
3013         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
3014     } else {
3015         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
3016     }
3017 }
3018 
3019 void tcg_gen_vec_rotl8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
3020 {
3021     uint64_t mask = dup_const(MO_8, 0xff << c);
3022 
3023     tcg_gen_shli_i64(d, a, c);
3024     tcg_gen_shri_i64(a, a, 8 - c);
3025     tcg_gen_andi_i64(d, d, mask);
3026     tcg_gen_andi_i64(a, a, ~mask);
3027     tcg_gen_or_i64(d, d, a);
3028 }
3029 
3030 void tcg_gen_vec_rotl16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
3031 {
3032     uint64_t mask = dup_const(MO_16, 0xffff << c);
3033 
3034     tcg_gen_shli_i64(d, a, c);
3035     tcg_gen_shri_i64(a, a, 16 - c);
3036     tcg_gen_andi_i64(d, d, mask);
3037     tcg_gen_andi_i64(a, a, ~mask);
3038     tcg_gen_or_i64(d, d, a);
3039 }
3040 
3041 void tcg_gen_gvec_rotli(unsigned vece, uint32_t dofs, uint32_t aofs,
3042                         int64_t shift, uint32_t oprsz, uint32_t maxsz)
3043 {
3044     static const TCGOpcode vecop_list[] = { INDEX_op_rotli_vec, 0 };
3045     static const GVecGen2i g[4] = {
3046         { .fni8 = tcg_gen_vec_rotl8i_i64,
3047           .fniv = tcg_gen_rotli_vec,
3048           .fno = gen_helper_gvec_rotl8i,
3049           .opt_opc = vecop_list,
3050           .vece = MO_8 },
3051         { .fni8 = tcg_gen_vec_rotl16i_i64,
3052           .fniv = tcg_gen_rotli_vec,
3053           .fno = gen_helper_gvec_rotl16i,
3054           .opt_opc = vecop_list,
3055           .vece = MO_16 },
3056         { .fni4 = tcg_gen_rotli_i32,
3057           .fniv = tcg_gen_rotli_vec,
3058           .fno = gen_helper_gvec_rotl32i,
3059           .opt_opc = vecop_list,
3060           .vece = MO_32 },
3061         { .fni8 = tcg_gen_rotli_i64,
3062           .fniv = tcg_gen_rotli_vec,
3063           .fno = gen_helper_gvec_rotl64i,
3064           .opt_opc = vecop_list,
3065           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3066           .vece = MO_64 },
3067     };
3068 
3069     tcg_debug_assert(vece <= MO_64);
3070     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
3071     if (shift == 0) {
3072         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
3073     } else {
3074         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
3075     }
3076 }
3077 
3078 void tcg_gen_gvec_rotri(unsigned vece, uint32_t dofs, uint32_t aofs,
3079                         int64_t shift, uint32_t oprsz, uint32_t maxsz)
3080 {
3081     tcg_debug_assert(vece <= MO_64);
3082     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
3083     tcg_gen_gvec_rotli(vece, dofs, aofs, -shift & ((8 << vece) - 1),
3084                        oprsz, maxsz);
3085 }
3086 
3087 /*
3088  * Specialized generation vector shifts by a non-constant scalar.
3089  */
3090 
3091 typedef struct {
3092     void (*fni4)(TCGv_i32, TCGv_i32, TCGv_i32);
3093     void (*fni8)(TCGv_i64, TCGv_i64, TCGv_i64);
3094     void (*fniv_s)(unsigned, TCGv_vec, TCGv_vec, TCGv_i32);
3095     void (*fniv_v)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec);
3096     gen_helper_gvec_2 *fno[4];
3097     TCGOpcode s_list[2];
3098     TCGOpcode v_list[2];
3099 } GVecGen2sh;
3100 
3101 static void expand_2sh_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
3102                            uint32_t oprsz, uint32_t tysz, TCGType type,
3103                            TCGv_i32 shift,
3104                            void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_i32))
3105 {
3106     for (uint32_t i = 0; i < oprsz; i += tysz) {
3107         TCGv_vec t0 = tcg_temp_new_vec(type);
3108         TCGv_vec t1 = tcg_temp_new_vec(type);
3109 
3110         tcg_gen_ld_vec(t0, tcg_env, aofs + i);
3111         fni(vece, t1, t0, shift);
3112         tcg_gen_st_vec(t1, tcg_env, dofs + i);
3113     }
3114 }
3115 
3116 static void
3117 do_gvec_shifts(unsigned vece, uint32_t dofs, uint32_t aofs, TCGv_i32 shift,
3118                uint32_t oprsz, uint32_t maxsz, const GVecGen2sh *g)
3119 {
3120     TCGType type;
3121     uint32_t some;
3122 
3123     check_size_align(oprsz, maxsz, dofs | aofs);
3124     check_overlap_2(dofs, aofs, maxsz);
3125 
3126     /* If the backend has a scalar expansion, great.  */
3127     type = choose_vector_type(g->s_list, vece, oprsz, vece == MO_64);
3128     if (type) {
3129         const TCGOpcode *hold_list = tcg_swap_vecop_list(NULL);
3130         switch (type) {
3131         case TCG_TYPE_V256:
3132             some = QEMU_ALIGN_DOWN(oprsz, 32);
3133             expand_2sh_vec(vece, dofs, aofs, some, 32,
3134                            TCG_TYPE_V256, shift, g->fniv_s);
3135             if (some == oprsz) {
3136                 break;
3137             }
3138             dofs += some;
3139             aofs += some;
3140             oprsz -= some;
3141             maxsz -= some;
3142             /* fallthru */
3143         case TCG_TYPE_V128:
3144             expand_2sh_vec(vece, dofs, aofs, oprsz, 16,
3145                            TCG_TYPE_V128, shift, g->fniv_s);
3146             break;
3147         case TCG_TYPE_V64:
3148             expand_2sh_vec(vece, dofs, aofs, oprsz, 8,
3149                            TCG_TYPE_V64, shift, g->fniv_s);
3150             break;
3151         default:
3152             g_assert_not_reached();
3153         }
3154         tcg_swap_vecop_list(hold_list);
3155         goto clear_tail;
3156     }
3157 
3158     /* If the backend supports variable vector shifts, also cool.  */
3159     type = choose_vector_type(g->v_list, vece, oprsz, vece == MO_64);
3160     if (type) {
3161         const TCGOpcode *hold_list = tcg_swap_vecop_list(NULL);
3162         TCGv_vec v_shift = tcg_temp_new_vec(type);
3163 
3164         if (vece == MO_64) {
3165             TCGv_i64 sh64 = tcg_temp_ebb_new_i64();
3166             tcg_gen_extu_i32_i64(sh64, shift);
3167             tcg_gen_dup_i64_vec(MO_64, v_shift, sh64);
3168             tcg_temp_free_i64(sh64);
3169         } else {
3170             tcg_gen_dup_i32_vec(vece, v_shift, shift);
3171         }
3172 
3173         switch (type) {
3174         case TCG_TYPE_V256:
3175             some = QEMU_ALIGN_DOWN(oprsz, 32);
3176             expand_2s_vec(vece, dofs, aofs, some, 32, TCG_TYPE_V256,
3177                           v_shift, false, g->fniv_v);
3178             if (some == oprsz) {
3179                 break;
3180             }
3181             dofs += some;
3182             aofs += some;
3183             oprsz -= some;
3184             maxsz -= some;
3185             /* fallthru */
3186         case TCG_TYPE_V128:
3187             expand_2s_vec(vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
3188                           v_shift, false, g->fniv_v);
3189             break;
3190         case TCG_TYPE_V64:
3191             expand_2s_vec(vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
3192                           v_shift, false, g->fniv_v);
3193             break;
3194         default:
3195             g_assert_not_reached();
3196         }
3197         tcg_temp_free_vec(v_shift);
3198         tcg_swap_vecop_list(hold_list);
3199         goto clear_tail;
3200     }
3201 
3202     /* Otherwise fall back to integral... */
3203     if (vece == MO_32 && check_size_impl(oprsz, 4)) {
3204         expand_2s_i32(dofs, aofs, oprsz, shift, false, g->fni4);
3205     } else if (vece == MO_64 && check_size_impl(oprsz, 8)) {
3206         TCGv_i64 sh64 = tcg_temp_ebb_new_i64();
3207         tcg_gen_extu_i32_i64(sh64, shift);
3208         expand_2s_i64(dofs, aofs, oprsz, sh64, false, g->fni8);
3209         tcg_temp_free_i64(sh64);
3210     } else {
3211         TCGv_ptr a0 = tcg_temp_ebb_new_ptr();
3212         TCGv_ptr a1 = tcg_temp_ebb_new_ptr();
3213         TCGv_i32 desc = tcg_temp_ebb_new_i32();
3214 
3215         tcg_gen_shli_i32(desc, shift, SIMD_DATA_SHIFT);
3216         tcg_gen_ori_i32(desc, desc, simd_desc(oprsz, maxsz, 0));
3217         tcg_gen_addi_ptr(a0, tcg_env, dofs);
3218         tcg_gen_addi_ptr(a1, tcg_env, aofs);
3219 
3220         g->fno[vece](a0, a1, desc);
3221 
3222         tcg_temp_free_ptr(a0);
3223         tcg_temp_free_ptr(a1);
3224         tcg_temp_free_i32(desc);
3225         return;
3226     }
3227 
3228  clear_tail:
3229     if (oprsz < maxsz) {
3230         expand_clr(dofs + oprsz, maxsz - oprsz);
3231     }
3232 }
3233 
3234 void tcg_gen_gvec_shls(unsigned vece, uint32_t dofs, uint32_t aofs,
3235                        TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3236 {
3237     static const GVecGen2sh g = {
3238         .fni4 = tcg_gen_shl_i32,
3239         .fni8 = tcg_gen_shl_i64,
3240         .fniv_s = tcg_gen_shls_vec,
3241         .fniv_v = tcg_gen_shlv_vec,
3242         .fno = {
3243             gen_helper_gvec_shl8i,
3244             gen_helper_gvec_shl16i,
3245             gen_helper_gvec_shl32i,
3246             gen_helper_gvec_shl64i,
3247         },
3248         .s_list = { INDEX_op_shls_vec, 0 },
3249         .v_list = { INDEX_op_shlv_vec, 0 },
3250     };
3251 
3252     tcg_debug_assert(vece <= MO_64);
3253     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
3254 }
3255 
3256 void tcg_gen_gvec_shrs(unsigned vece, uint32_t dofs, uint32_t aofs,
3257                        TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3258 {
3259     static const GVecGen2sh g = {
3260         .fni4 = tcg_gen_shr_i32,
3261         .fni8 = tcg_gen_shr_i64,
3262         .fniv_s = tcg_gen_shrs_vec,
3263         .fniv_v = tcg_gen_shrv_vec,
3264         .fno = {
3265             gen_helper_gvec_shr8i,
3266             gen_helper_gvec_shr16i,
3267             gen_helper_gvec_shr32i,
3268             gen_helper_gvec_shr64i,
3269         },
3270         .s_list = { INDEX_op_shrs_vec, 0 },
3271         .v_list = { INDEX_op_shrv_vec, 0 },
3272     };
3273 
3274     tcg_debug_assert(vece <= MO_64);
3275     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
3276 }
3277 
3278 void tcg_gen_gvec_sars(unsigned vece, uint32_t dofs, uint32_t aofs,
3279                        TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3280 {
3281     static const GVecGen2sh g = {
3282         .fni4 = tcg_gen_sar_i32,
3283         .fni8 = tcg_gen_sar_i64,
3284         .fniv_s = tcg_gen_sars_vec,
3285         .fniv_v = tcg_gen_sarv_vec,
3286         .fno = {
3287             gen_helper_gvec_sar8i,
3288             gen_helper_gvec_sar16i,
3289             gen_helper_gvec_sar32i,
3290             gen_helper_gvec_sar64i,
3291         },
3292         .s_list = { INDEX_op_sars_vec, 0 },
3293         .v_list = { INDEX_op_sarv_vec, 0 },
3294     };
3295 
3296     tcg_debug_assert(vece <= MO_64);
3297     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
3298 }
3299 
3300 void tcg_gen_gvec_rotls(unsigned vece, uint32_t dofs, uint32_t aofs,
3301                         TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3302 {
3303     static const GVecGen2sh g = {
3304         .fni4 = tcg_gen_rotl_i32,
3305         .fni8 = tcg_gen_rotl_i64,
3306         .fniv_s = tcg_gen_rotls_vec,
3307         .fniv_v = tcg_gen_rotlv_vec,
3308         .fno = {
3309             gen_helper_gvec_rotl8i,
3310             gen_helper_gvec_rotl16i,
3311             gen_helper_gvec_rotl32i,
3312             gen_helper_gvec_rotl64i,
3313         },
3314         .s_list = { INDEX_op_rotls_vec, 0 },
3315         .v_list = { INDEX_op_rotlv_vec, 0 },
3316     };
3317 
3318     tcg_debug_assert(vece <= MO_64);
3319     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
3320 }
3321 
3322 void tcg_gen_gvec_rotrs(unsigned vece, uint32_t dofs, uint32_t aofs,
3323                         TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3324 {
3325     TCGv_i32 tmp = tcg_temp_ebb_new_i32();
3326 
3327     tcg_gen_neg_i32(tmp, shift);
3328     tcg_gen_andi_i32(tmp, tmp, (8 << vece) - 1);
3329     tcg_gen_gvec_rotls(vece, dofs, aofs, tmp, oprsz, maxsz);
3330     tcg_temp_free_i32(tmp);
3331 }
3332 
3333 /*
3334  * Expand D = A << (B % element bits)
3335  *
3336  * Unlike scalar shifts, where it is easy for the target front end
3337  * to include the modulo as part of the expansion.  If the target
3338  * naturally includes the modulo as part of the operation, great!
3339  * If the target has some other behaviour from out-of-range shifts,
3340  * then it could not use this function anyway, and would need to
3341  * do it's own expansion with custom functions.
3342  */
3343 static void tcg_gen_shlv_mod_vec(unsigned vece, TCGv_vec d,
3344                                  TCGv_vec a, TCGv_vec b)
3345 {
3346     TCGv_vec t = tcg_temp_new_vec_matching(d);
3347     TCGv_vec m = tcg_constant_vec_matching(d, vece, (8 << vece) - 1);
3348 
3349     tcg_gen_and_vec(vece, t, b, m);
3350     tcg_gen_shlv_vec(vece, d, a, t);
3351     tcg_temp_free_vec(t);
3352 }
3353 
3354 static void tcg_gen_shl_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3355 {
3356     TCGv_i32 t = tcg_temp_ebb_new_i32();
3357 
3358     tcg_gen_andi_i32(t, b, 31);
3359     tcg_gen_shl_i32(d, a, t);
3360     tcg_temp_free_i32(t);
3361 }
3362 
3363 static void tcg_gen_shl_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3364 {
3365     TCGv_i64 t = tcg_temp_ebb_new_i64();
3366 
3367     tcg_gen_andi_i64(t, b, 63);
3368     tcg_gen_shl_i64(d, a, t);
3369     tcg_temp_free_i64(t);
3370 }
3371 
3372 void tcg_gen_gvec_shlv(unsigned vece, uint32_t dofs, uint32_t aofs,
3373                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3374 {
3375     static const TCGOpcode vecop_list[] = { INDEX_op_shlv_vec, 0 };
3376     static const GVecGen3 g[4] = {
3377         { .fniv = tcg_gen_shlv_mod_vec,
3378           .fno = gen_helper_gvec_shl8v,
3379           .opt_opc = vecop_list,
3380           .vece = MO_8 },
3381         { .fniv = tcg_gen_shlv_mod_vec,
3382           .fno = gen_helper_gvec_shl16v,
3383           .opt_opc = vecop_list,
3384           .vece = MO_16 },
3385         { .fni4 = tcg_gen_shl_mod_i32,
3386           .fniv = tcg_gen_shlv_mod_vec,
3387           .fno = gen_helper_gvec_shl32v,
3388           .opt_opc = vecop_list,
3389           .vece = MO_32 },
3390         { .fni8 = tcg_gen_shl_mod_i64,
3391           .fniv = tcg_gen_shlv_mod_vec,
3392           .fno = gen_helper_gvec_shl64v,
3393           .opt_opc = vecop_list,
3394           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3395           .vece = MO_64 },
3396     };
3397 
3398     tcg_debug_assert(vece <= MO_64);
3399     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3400 }
3401 
3402 /*
3403  * Similarly for logical right shifts.
3404  */
3405 
3406 static void tcg_gen_shrv_mod_vec(unsigned vece, TCGv_vec d,
3407                                  TCGv_vec a, TCGv_vec b)
3408 {
3409     TCGv_vec t = tcg_temp_new_vec_matching(d);
3410     TCGv_vec m = tcg_constant_vec_matching(d, vece, (8 << vece) - 1);
3411 
3412     tcg_gen_and_vec(vece, t, b, m);
3413     tcg_gen_shrv_vec(vece, d, a, t);
3414     tcg_temp_free_vec(t);
3415 }
3416 
3417 static void tcg_gen_shr_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3418 {
3419     TCGv_i32 t = tcg_temp_ebb_new_i32();
3420 
3421     tcg_gen_andi_i32(t, b, 31);
3422     tcg_gen_shr_i32(d, a, t);
3423     tcg_temp_free_i32(t);
3424 }
3425 
3426 static void tcg_gen_shr_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3427 {
3428     TCGv_i64 t = tcg_temp_ebb_new_i64();
3429 
3430     tcg_gen_andi_i64(t, b, 63);
3431     tcg_gen_shr_i64(d, a, t);
3432     tcg_temp_free_i64(t);
3433 }
3434 
3435 void tcg_gen_gvec_shrv(unsigned vece, uint32_t dofs, uint32_t aofs,
3436                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3437 {
3438     static const TCGOpcode vecop_list[] = { INDEX_op_shrv_vec, 0 };
3439     static const GVecGen3 g[4] = {
3440         { .fniv = tcg_gen_shrv_mod_vec,
3441           .fno = gen_helper_gvec_shr8v,
3442           .opt_opc = vecop_list,
3443           .vece = MO_8 },
3444         { .fniv = tcg_gen_shrv_mod_vec,
3445           .fno = gen_helper_gvec_shr16v,
3446           .opt_opc = vecop_list,
3447           .vece = MO_16 },
3448         { .fni4 = tcg_gen_shr_mod_i32,
3449           .fniv = tcg_gen_shrv_mod_vec,
3450           .fno = gen_helper_gvec_shr32v,
3451           .opt_opc = vecop_list,
3452           .vece = MO_32 },
3453         { .fni8 = tcg_gen_shr_mod_i64,
3454           .fniv = tcg_gen_shrv_mod_vec,
3455           .fno = gen_helper_gvec_shr64v,
3456           .opt_opc = vecop_list,
3457           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3458           .vece = MO_64 },
3459     };
3460 
3461     tcg_debug_assert(vece <= MO_64);
3462     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3463 }
3464 
3465 /*
3466  * Similarly for arithmetic right shifts.
3467  */
3468 
3469 static void tcg_gen_sarv_mod_vec(unsigned vece, TCGv_vec d,
3470                                  TCGv_vec a, TCGv_vec b)
3471 {
3472     TCGv_vec t = tcg_temp_new_vec_matching(d);
3473     TCGv_vec m = tcg_constant_vec_matching(d, vece, (8 << vece) - 1);
3474 
3475     tcg_gen_and_vec(vece, t, b, m);
3476     tcg_gen_sarv_vec(vece, d, a, t);
3477     tcg_temp_free_vec(t);
3478 }
3479 
3480 static void tcg_gen_sar_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3481 {
3482     TCGv_i32 t = tcg_temp_ebb_new_i32();
3483 
3484     tcg_gen_andi_i32(t, b, 31);
3485     tcg_gen_sar_i32(d, a, t);
3486     tcg_temp_free_i32(t);
3487 }
3488 
3489 static void tcg_gen_sar_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3490 {
3491     TCGv_i64 t = tcg_temp_ebb_new_i64();
3492 
3493     tcg_gen_andi_i64(t, b, 63);
3494     tcg_gen_sar_i64(d, a, t);
3495     tcg_temp_free_i64(t);
3496 }
3497 
3498 void tcg_gen_gvec_sarv(unsigned vece, uint32_t dofs, uint32_t aofs,
3499                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3500 {
3501     static const TCGOpcode vecop_list[] = { INDEX_op_sarv_vec, 0 };
3502     static const GVecGen3 g[4] = {
3503         { .fniv = tcg_gen_sarv_mod_vec,
3504           .fno = gen_helper_gvec_sar8v,
3505           .opt_opc = vecop_list,
3506           .vece = MO_8 },
3507         { .fniv = tcg_gen_sarv_mod_vec,
3508           .fno = gen_helper_gvec_sar16v,
3509           .opt_opc = vecop_list,
3510           .vece = MO_16 },
3511         { .fni4 = tcg_gen_sar_mod_i32,
3512           .fniv = tcg_gen_sarv_mod_vec,
3513           .fno = gen_helper_gvec_sar32v,
3514           .opt_opc = vecop_list,
3515           .vece = MO_32 },
3516         { .fni8 = tcg_gen_sar_mod_i64,
3517           .fniv = tcg_gen_sarv_mod_vec,
3518           .fno = gen_helper_gvec_sar64v,
3519           .opt_opc = vecop_list,
3520           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3521           .vece = MO_64 },
3522     };
3523 
3524     tcg_debug_assert(vece <= MO_64);
3525     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3526 }
3527 
3528 /*
3529  * Similarly for rotates.
3530  */
3531 
3532 static void tcg_gen_rotlv_mod_vec(unsigned vece, TCGv_vec d,
3533                                   TCGv_vec a, TCGv_vec b)
3534 {
3535     TCGv_vec t = tcg_temp_new_vec_matching(d);
3536     TCGv_vec m = tcg_constant_vec_matching(d, vece, (8 << vece) - 1);
3537 
3538     tcg_gen_and_vec(vece, t, b, m);
3539     tcg_gen_rotlv_vec(vece, d, a, t);
3540     tcg_temp_free_vec(t);
3541 }
3542 
3543 static void tcg_gen_rotl_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3544 {
3545     TCGv_i32 t = tcg_temp_ebb_new_i32();
3546 
3547     tcg_gen_andi_i32(t, b, 31);
3548     tcg_gen_rotl_i32(d, a, t);
3549     tcg_temp_free_i32(t);
3550 }
3551 
3552 static void tcg_gen_rotl_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3553 {
3554     TCGv_i64 t = tcg_temp_ebb_new_i64();
3555 
3556     tcg_gen_andi_i64(t, b, 63);
3557     tcg_gen_rotl_i64(d, a, t);
3558     tcg_temp_free_i64(t);
3559 }
3560 
3561 void tcg_gen_gvec_rotlv(unsigned vece, uint32_t dofs, uint32_t aofs,
3562                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3563 {
3564     static const TCGOpcode vecop_list[] = { INDEX_op_rotlv_vec, 0 };
3565     static const GVecGen3 g[4] = {
3566         { .fniv = tcg_gen_rotlv_mod_vec,
3567           .fno = gen_helper_gvec_rotl8v,
3568           .opt_opc = vecop_list,
3569           .vece = MO_8 },
3570         { .fniv = tcg_gen_rotlv_mod_vec,
3571           .fno = gen_helper_gvec_rotl16v,
3572           .opt_opc = vecop_list,
3573           .vece = MO_16 },
3574         { .fni4 = tcg_gen_rotl_mod_i32,
3575           .fniv = tcg_gen_rotlv_mod_vec,
3576           .fno = gen_helper_gvec_rotl32v,
3577           .opt_opc = vecop_list,
3578           .vece = MO_32 },
3579         { .fni8 = tcg_gen_rotl_mod_i64,
3580           .fniv = tcg_gen_rotlv_mod_vec,
3581           .fno = gen_helper_gvec_rotl64v,
3582           .opt_opc = vecop_list,
3583           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3584           .vece = MO_64 },
3585     };
3586 
3587     tcg_debug_assert(vece <= MO_64);
3588     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3589 }
3590 
3591 static void tcg_gen_rotrv_mod_vec(unsigned vece, TCGv_vec d,
3592                                   TCGv_vec a, TCGv_vec b)
3593 {
3594     TCGv_vec t = tcg_temp_new_vec_matching(d);
3595     TCGv_vec m = tcg_constant_vec_matching(d, vece, (8 << vece) - 1);
3596 
3597     tcg_gen_and_vec(vece, t, b, m);
3598     tcg_gen_rotrv_vec(vece, d, a, t);
3599     tcg_temp_free_vec(t);
3600 }
3601 
3602 static void tcg_gen_rotr_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3603 {
3604     TCGv_i32 t = tcg_temp_ebb_new_i32();
3605 
3606     tcg_gen_andi_i32(t, b, 31);
3607     tcg_gen_rotr_i32(d, a, t);
3608     tcg_temp_free_i32(t);
3609 }
3610 
3611 static void tcg_gen_rotr_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3612 {
3613     TCGv_i64 t = tcg_temp_ebb_new_i64();
3614 
3615     tcg_gen_andi_i64(t, b, 63);
3616     tcg_gen_rotr_i64(d, a, t);
3617     tcg_temp_free_i64(t);
3618 }
3619 
3620 void tcg_gen_gvec_rotrv(unsigned vece, uint32_t dofs, uint32_t aofs,
3621                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3622 {
3623     static const TCGOpcode vecop_list[] = { INDEX_op_rotrv_vec, 0 };
3624     static const GVecGen3 g[4] = {
3625         { .fniv = tcg_gen_rotrv_mod_vec,
3626           .fno = gen_helper_gvec_rotr8v,
3627           .opt_opc = vecop_list,
3628           .vece = MO_8 },
3629         { .fniv = tcg_gen_rotrv_mod_vec,
3630           .fno = gen_helper_gvec_rotr16v,
3631           .opt_opc = vecop_list,
3632           .vece = MO_16 },
3633         { .fni4 = tcg_gen_rotr_mod_i32,
3634           .fniv = tcg_gen_rotrv_mod_vec,
3635           .fno = gen_helper_gvec_rotr32v,
3636           .opt_opc = vecop_list,
3637           .vece = MO_32 },
3638         { .fni8 = tcg_gen_rotr_mod_i64,
3639           .fniv = tcg_gen_rotrv_mod_vec,
3640           .fno = gen_helper_gvec_rotr64v,
3641           .opt_opc = vecop_list,
3642           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3643           .vece = MO_64 },
3644     };
3645 
3646     tcg_debug_assert(vece <= MO_64);
3647     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3648 }
3649 
3650 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
3651 static void expand_cmp_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
3652                            uint32_t oprsz, TCGCond cond)
3653 {
3654     TCGv_i32 t0 = tcg_temp_ebb_new_i32();
3655     TCGv_i32 t1 = tcg_temp_ebb_new_i32();
3656     uint32_t i;
3657 
3658     for (i = 0; i < oprsz; i += 4) {
3659         tcg_gen_ld_i32(t0, tcg_env, aofs + i);
3660         tcg_gen_ld_i32(t1, tcg_env, bofs + i);
3661         tcg_gen_negsetcond_i32(cond, t0, t0, t1);
3662         tcg_gen_st_i32(t0, tcg_env, dofs + i);
3663     }
3664     tcg_temp_free_i32(t1);
3665     tcg_temp_free_i32(t0);
3666 }
3667 
3668 static void expand_cmp_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
3669                            uint32_t oprsz, TCGCond cond)
3670 {
3671     TCGv_i64 t0 = tcg_temp_ebb_new_i64();
3672     TCGv_i64 t1 = tcg_temp_ebb_new_i64();
3673     uint32_t i;
3674 
3675     for (i = 0; i < oprsz; i += 8) {
3676         tcg_gen_ld_i64(t0, tcg_env, aofs + i);
3677         tcg_gen_ld_i64(t1, tcg_env, bofs + i);
3678         tcg_gen_negsetcond_i64(cond, t0, t0, t1);
3679         tcg_gen_st_i64(t0, tcg_env, dofs + i);
3680     }
3681     tcg_temp_free_i64(t1);
3682     tcg_temp_free_i64(t0);
3683 }
3684 
3685 static void expand_cmp_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
3686                            uint32_t bofs, uint32_t oprsz, uint32_t tysz,
3687                            TCGType type, TCGCond cond)
3688 {
3689     for (uint32_t i = 0; i < oprsz; i += tysz) {
3690         TCGv_vec t0 = tcg_temp_new_vec(type);
3691         TCGv_vec t1 = tcg_temp_new_vec(type);
3692         TCGv_vec t2 = tcg_temp_new_vec(type);
3693 
3694         tcg_gen_ld_vec(t0, tcg_env, aofs + i);
3695         tcg_gen_ld_vec(t1, tcg_env, bofs + i);
3696         tcg_gen_cmp_vec(cond, vece, t2, t0, t1);
3697         tcg_gen_st_vec(t2, tcg_env, dofs + i);
3698     }
3699 }
3700 
3701 void tcg_gen_gvec_cmp(TCGCond cond, unsigned vece, uint32_t dofs,
3702                       uint32_t aofs, uint32_t bofs,
3703                       uint32_t oprsz, uint32_t maxsz)
3704 {
3705     static const TCGOpcode cmp_list[] = { INDEX_op_cmp_vec, 0 };
3706     static gen_helper_gvec_3 * const eq_fn[4] = {
3707         gen_helper_gvec_eq8, gen_helper_gvec_eq16,
3708         gen_helper_gvec_eq32, gen_helper_gvec_eq64
3709     };
3710     static gen_helper_gvec_3 * const ne_fn[4] = {
3711         gen_helper_gvec_ne8, gen_helper_gvec_ne16,
3712         gen_helper_gvec_ne32, gen_helper_gvec_ne64
3713     };
3714     static gen_helper_gvec_3 * const lt_fn[4] = {
3715         gen_helper_gvec_lt8, gen_helper_gvec_lt16,
3716         gen_helper_gvec_lt32, gen_helper_gvec_lt64
3717     };
3718     static gen_helper_gvec_3 * const le_fn[4] = {
3719         gen_helper_gvec_le8, gen_helper_gvec_le16,
3720         gen_helper_gvec_le32, gen_helper_gvec_le64
3721     };
3722     static gen_helper_gvec_3 * const ltu_fn[4] = {
3723         gen_helper_gvec_ltu8, gen_helper_gvec_ltu16,
3724         gen_helper_gvec_ltu32, gen_helper_gvec_ltu64
3725     };
3726     static gen_helper_gvec_3 * const leu_fn[4] = {
3727         gen_helper_gvec_leu8, gen_helper_gvec_leu16,
3728         gen_helper_gvec_leu32, gen_helper_gvec_leu64
3729     };
3730     static gen_helper_gvec_3 * const * const fns[16] = {
3731         [TCG_COND_EQ] = eq_fn,
3732         [TCG_COND_NE] = ne_fn,
3733         [TCG_COND_LT] = lt_fn,
3734         [TCG_COND_LE] = le_fn,
3735         [TCG_COND_LTU] = ltu_fn,
3736         [TCG_COND_LEU] = leu_fn,
3737     };
3738 
3739     const TCGOpcode *hold_list;
3740     TCGType type;
3741     uint32_t some;
3742 
3743     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
3744     check_overlap_3(dofs, aofs, bofs, maxsz);
3745 
3746     if (cond == TCG_COND_NEVER || cond == TCG_COND_ALWAYS) {
3747         do_dup(MO_8, dofs, oprsz, maxsz,
3748                NULL, NULL, -(cond == TCG_COND_ALWAYS));
3749         return;
3750     }
3751 
3752     /*
3753      * Implement inline with a vector type, if possible.
3754      * Prefer integer when 64-bit host and 64-bit comparison.
3755      */
3756     hold_list = tcg_swap_vecop_list(cmp_list);
3757     type = choose_vector_type(cmp_list, vece, oprsz,
3758                               TCG_TARGET_REG_BITS == 64 && vece == MO_64);
3759     switch (type) {
3760     case TCG_TYPE_V256:
3761         /* Recall that ARM SVE allows vector sizes that are not a
3762          * power of 2, but always a multiple of 16.  The intent is
3763          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
3764          */
3765         some = QEMU_ALIGN_DOWN(oprsz, 32);
3766         expand_cmp_vec(vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256, cond);
3767         if (some == oprsz) {
3768             break;
3769         }
3770         dofs += some;
3771         aofs += some;
3772         bofs += some;
3773         oprsz -= some;
3774         maxsz -= some;
3775         /* fallthru */
3776     case TCG_TYPE_V128:
3777         expand_cmp_vec(vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128, cond);
3778         break;
3779     case TCG_TYPE_V64:
3780         expand_cmp_vec(vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64, cond);
3781         break;
3782 
3783     case 0:
3784         if (vece == MO_64 && check_size_impl(oprsz, 8)) {
3785             expand_cmp_i64(dofs, aofs, bofs, oprsz, cond);
3786         } else if (vece == MO_32 && check_size_impl(oprsz, 4)) {
3787             expand_cmp_i32(dofs, aofs, bofs, oprsz, cond);
3788         } else {
3789             gen_helper_gvec_3 * const *fn = fns[cond];
3790 
3791             if (fn == NULL) {
3792                 uint32_t tmp;
3793                 tmp = aofs, aofs = bofs, bofs = tmp;
3794                 cond = tcg_swap_cond(cond);
3795                 fn = fns[cond];
3796                 assert(fn != NULL);
3797             }
3798             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz, maxsz, 0, fn[vece]);
3799             oprsz = maxsz;
3800         }
3801         break;
3802 
3803     default:
3804         g_assert_not_reached();
3805     }
3806     tcg_swap_vecop_list(hold_list);
3807 
3808     if (oprsz < maxsz) {
3809         expand_clr(dofs + oprsz, maxsz - oprsz);
3810     }
3811 }
3812 
3813 static void expand_cmps_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
3814                             uint32_t oprsz, uint32_t tysz, TCGType type,
3815                             TCGCond cond, TCGv_vec c)
3816 {
3817     TCGv_vec t0 = tcg_temp_new_vec(type);
3818     TCGv_vec t1 = tcg_temp_new_vec(type);
3819     uint32_t i;
3820 
3821     for (i = 0; i < oprsz; i += tysz) {
3822         tcg_gen_ld_vec(t1, tcg_env, aofs + i);
3823         tcg_gen_cmp_vec(cond, vece, t0, t1, c);
3824         tcg_gen_st_vec(t0, tcg_env, dofs + i);
3825     }
3826 }
3827 
3828 void tcg_gen_gvec_cmps(TCGCond cond, unsigned vece, uint32_t dofs,
3829                        uint32_t aofs, TCGv_i64 c,
3830                        uint32_t oprsz, uint32_t maxsz)
3831 {
3832     static const TCGOpcode cmp_list[] = { INDEX_op_cmp_vec, 0 };
3833     static gen_helper_gvec_2i * const eq_fn[4] = {
3834         gen_helper_gvec_eqs8, gen_helper_gvec_eqs16,
3835         gen_helper_gvec_eqs32, gen_helper_gvec_eqs64
3836     };
3837     static gen_helper_gvec_2i * const lt_fn[4] = {
3838         gen_helper_gvec_lts8, gen_helper_gvec_lts16,
3839         gen_helper_gvec_lts32, gen_helper_gvec_lts64
3840     };
3841     static gen_helper_gvec_2i * const le_fn[4] = {
3842         gen_helper_gvec_les8, gen_helper_gvec_les16,
3843         gen_helper_gvec_les32, gen_helper_gvec_les64
3844     };
3845     static gen_helper_gvec_2i * const ltu_fn[4] = {
3846         gen_helper_gvec_ltus8, gen_helper_gvec_ltus16,
3847         gen_helper_gvec_ltus32, gen_helper_gvec_ltus64
3848     };
3849     static gen_helper_gvec_2i * const leu_fn[4] = {
3850         gen_helper_gvec_leus8, gen_helper_gvec_leus16,
3851         gen_helper_gvec_leus32, gen_helper_gvec_leus64
3852     };
3853     static gen_helper_gvec_2i * const * const fns[16] = {
3854         [TCG_COND_EQ] = eq_fn,
3855         [TCG_COND_LT] = lt_fn,
3856         [TCG_COND_LE] = le_fn,
3857         [TCG_COND_LTU] = ltu_fn,
3858         [TCG_COND_LEU] = leu_fn,
3859     };
3860 
3861     TCGType type;
3862 
3863     check_size_align(oprsz, maxsz, dofs | aofs);
3864     check_overlap_2(dofs, aofs, maxsz);
3865 
3866     if (cond == TCG_COND_NEVER || cond == TCG_COND_ALWAYS) {
3867         do_dup(MO_8, dofs, oprsz, maxsz,
3868                NULL, NULL, -(cond == TCG_COND_ALWAYS));
3869         return;
3870     }
3871 
3872     /*
3873      * Implement inline with a vector type, if possible.
3874      * Prefer integer when 64-bit host and 64-bit comparison.
3875      */
3876     type = choose_vector_type(cmp_list, vece, oprsz,
3877                               TCG_TARGET_REG_BITS == 64 && vece == MO_64);
3878     if (type != 0) {
3879         const TCGOpcode *hold_list = tcg_swap_vecop_list(cmp_list);
3880         TCGv_vec t_vec = tcg_temp_new_vec(type);
3881         uint32_t some;
3882 
3883         tcg_gen_dup_i64_vec(vece, t_vec, c);
3884         switch (type) {
3885         case TCG_TYPE_V256:
3886             some = QEMU_ALIGN_DOWN(oprsz, 32);
3887             expand_cmps_vec(vece, dofs, aofs, some, 32,
3888                             TCG_TYPE_V256, cond, t_vec);
3889             aofs += some;
3890             dofs += some;
3891             oprsz -= some;
3892             maxsz -= some;
3893             /* fallthru */
3894 
3895         case TCG_TYPE_V128:
3896             some = QEMU_ALIGN_DOWN(oprsz, 16);
3897             expand_cmps_vec(vece, dofs, aofs, some, 16,
3898                             TCG_TYPE_V128, cond, t_vec);
3899             break;
3900 
3901         case TCG_TYPE_V64:
3902             some = QEMU_ALIGN_DOWN(oprsz, 8);
3903             expand_cmps_vec(vece, dofs, aofs, some, 8,
3904                             TCG_TYPE_V64, cond, t_vec);
3905             break;
3906 
3907         default:
3908             g_assert_not_reached();
3909         }
3910         tcg_temp_free_vec(t_vec);
3911         tcg_swap_vecop_list(hold_list);
3912     } else if (vece == MO_64 && check_size_impl(oprsz, 8)) {
3913         TCGv_i64 t0 = tcg_temp_ebb_new_i64();
3914         uint32_t i;
3915 
3916         for (i = 0; i < oprsz; i += 8) {
3917             tcg_gen_ld_i64(t0, tcg_env, aofs + i);
3918             tcg_gen_negsetcond_i64(cond, t0, t0, c);
3919             tcg_gen_st_i64(t0, tcg_env, dofs + i);
3920         }
3921         tcg_temp_free_i64(t0);
3922     } else if (vece == MO_32 && check_size_impl(oprsz, 4)) {
3923         TCGv_i32 t0 = tcg_temp_ebb_new_i32();
3924         TCGv_i32 t1 = tcg_temp_ebb_new_i32();
3925         uint32_t i;
3926 
3927         tcg_gen_extrl_i64_i32(t1, c);
3928         for (i = 0; i < oprsz; i += 8) {
3929             tcg_gen_ld_i32(t0, tcg_env, aofs + i);
3930             tcg_gen_negsetcond_i32(cond, t0, t0, t1);
3931             tcg_gen_st_i32(t0, tcg_env, dofs + i);
3932         }
3933         tcg_temp_free_i32(t0);
3934         tcg_temp_free_i32(t1);
3935     } else {
3936         gen_helper_gvec_2i * const *fn = fns[cond];
3937         bool inv = false;
3938 
3939         if (fn == NULL) {
3940             cond = tcg_invert_cond(cond);
3941             fn = fns[cond];
3942             assert(fn != NULL);
3943             inv = true;
3944         }
3945         tcg_gen_gvec_2i_ool(dofs, aofs, c, oprsz, maxsz, inv, fn[vece]);
3946         return;
3947     }
3948 
3949     if (oprsz < maxsz) {
3950         expand_clr(dofs + oprsz, maxsz - oprsz);
3951     }
3952 }
3953 
3954 void tcg_gen_gvec_cmpi(TCGCond cond, unsigned vece, uint32_t dofs,
3955                        uint32_t aofs, int64_t c,
3956                        uint32_t oprsz, uint32_t maxsz)
3957 {
3958     TCGv_i64 tmp = tcg_constant_i64(c);
3959     tcg_gen_gvec_cmps(cond, vece, dofs, aofs, tmp, oprsz, maxsz);
3960 }
3961 
3962 static void tcg_gen_bitsel_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 c)
3963 {
3964     TCGv_i64 t = tcg_temp_ebb_new_i64();
3965 
3966     tcg_gen_and_i64(t, b, a);
3967     tcg_gen_andc_i64(d, c, a);
3968     tcg_gen_or_i64(d, d, t);
3969     tcg_temp_free_i64(t);
3970 }
3971 
3972 void tcg_gen_gvec_bitsel(unsigned vece, uint32_t dofs, uint32_t aofs,
3973                          uint32_t bofs, uint32_t cofs,
3974                          uint32_t oprsz, uint32_t maxsz)
3975 {
3976     static const GVecGen4 g = {
3977         .fni8 = tcg_gen_bitsel_i64,
3978         .fniv = tcg_gen_bitsel_vec,
3979         .fno = gen_helper_gvec_bitsel,
3980     };
3981 
3982     tcg_gen_gvec_4(dofs, aofs, bofs, cofs, oprsz, maxsz, &g);
3983 }
3984