xref: /openbmc/qemu/tcg/tcg-op-gvec.c (revision e6546342)
1 /*
2  * Generic vector operation expansion
3  *
4  * Copyright (c) 2018 Linaro
5  *
6  * This library is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2.1 of the License, or (at your option) any later version.
10  *
11  * This library is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with this library; if not, see <http://www.gnu.org/licenses/>.
18  */
19 
20 #include "qemu/osdep.h"
21 #include "tcg/tcg.h"
22 #include "tcg/tcg-op.h"
23 #include "tcg/tcg-op-gvec.h"
24 #include "qemu/main-loop.h"
25 #include "tcg/tcg-gvec-desc.h"
26 
27 #define MAX_UNROLL  4
28 
29 #ifdef CONFIG_DEBUG_TCG
30 static const TCGOpcode vecop_list_empty[1] = { 0 };
31 #else
32 #define vecop_list_empty NULL
33 #endif
34 
35 
36 /* Verify vector size and alignment rules.  OFS should be the OR of all
37    of the operand offsets so that we can check them all at once.  */
38 static void check_size_align(uint32_t oprsz, uint32_t maxsz, uint32_t ofs)
39 {
40     uint32_t opr_align = oprsz >= 16 ? 15 : 7;
41     uint32_t max_align = maxsz >= 16 || oprsz >= 16 ? 15 : 7;
42     tcg_debug_assert(oprsz > 0);
43     tcg_debug_assert(oprsz <= maxsz);
44     tcg_debug_assert((oprsz & opr_align) == 0);
45     tcg_debug_assert((maxsz & max_align) == 0);
46     tcg_debug_assert((ofs & max_align) == 0);
47 }
48 
49 /* Verify vector overlap rules for two operands.  */
50 static void check_overlap_2(uint32_t d, uint32_t a, uint32_t s)
51 {
52     tcg_debug_assert(d == a || d + s <= a || a + s <= d);
53 }
54 
55 /* Verify vector overlap rules for three operands.  */
56 static void check_overlap_3(uint32_t d, uint32_t a, uint32_t b, uint32_t s)
57 {
58     check_overlap_2(d, a, s);
59     check_overlap_2(d, b, s);
60     check_overlap_2(a, b, s);
61 }
62 
63 /* Verify vector overlap rules for four operands.  */
64 static void check_overlap_4(uint32_t d, uint32_t a, uint32_t b,
65                             uint32_t c, uint32_t s)
66 {
67     check_overlap_2(d, a, s);
68     check_overlap_2(d, b, s);
69     check_overlap_2(d, c, s);
70     check_overlap_2(a, b, s);
71     check_overlap_2(a, c, s);
72     check_overlap_2(b, c, s);
73 }
74 
75 /* Create a descriptor from components.  */
76 uint32_t simd_desc(uint32_t oprsz, uint32_t maxsz, int32_t data)
77 {
78     uint32_t desc = 0;
79 
80     assert(oprsz % 8 == 0 && oprsz <= (8 << SIMD_OPRSZ_BITS));
81     assert(maxsz % 8 == 0 && maxsz <= (8 << SIMD_MAXSZ_BITS));
82     assert(data == sextract32(data, 0, SIMD_DATA_BITS));
83 
84     oprsz = (oprsz / 8) - 1;
85     maxsz = (maxsz / 8) - 1;
86     desc = deposit32(desc, SIMD_OPRSZ_SHIFT, SIMD_OPRSZ_BITS, oprsz);
87     desc = deposit32(desc, SIMD_MAXSZ_SHIFT, SIMD_MAXSZ_BITS, maxsz);
88     desc = deposit32(desc, SIMD_DATA_SHIFT, SIMD_DATA_BITS, data);
89 
90     return desc;
91 }
92 
93 /* Generate a call to a gvec-style helper with two vector operands.  */
94 void tcg_gen_gvec_2_ool(uint32_t dofs, uint32_t aofs,
95                         uint32_t oprsz, uint32_t maxsz, int32_t data,
96                         gen_helper_gvec_2 *fn)
97 {
98     TCGv_ptr a0, a1;
99     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
100 
101     a0 = tcg_temp_new_ptr();
102     a1 = tcg_temp_new_ptr();
103 
104     tcg_gen_addi_ptr(a0, cpu_env, dofs);
105     tcg_gen_addi_ptr(a1, cpu_env, aofs);
106 
107     fn(a0, a1, desc);
108 
109     tcg_temp_free_ptr(a0);
110     tcg_temp_free_ptr(a1);
111     tcg_temp_free_i32(desc);
112 }
113 
114 /* Generate a call to a gvec-style helper with two vector operands
115    and one scalar operand.  */
116 void tcg_gen_gvec_2i_ool(uint32_t dofs, uint32_t aofs, TCGv_i64 c,
117                          uint32_t oprsz, uint32_t maxsz, int32_t data,
118                          gen_helper_gvec_2i *fn)
119 {
120     TCGv_ptr a0, a1;
121     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
122 
123     a0 = tcg_temp_new_ptr();
124     a1 = tcg_temp_new_ptr();
125 
126     tcg_gen_addi_ptr(a0, cpu_env, dofs);
127     tcg_gen_addi_ptr(a1, cpu_env, aofs);
128 
129     fn(a0, a1, c, desc);
130 
131     tcg_temp_free_ptr(a0);
132     tcg_temp_free_ptr(a1);
133     tcg_temp_free_i32(desc);
134 }
135 
136 /* Generate a call to a gvec-style helper with three vector operands.  */
137 void tcg_gen_gvec_3_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
138                         uint32_t oprsz, uint32_t maxsz, int32_t data,
139                         gen_helper_gvec_3 *fn)
140 {
141     TCGv_ptr a0, a1, a2;
142     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
143 
144     a0 = tcg_temp_new_ptr();
145     a1 = tcg_temp_new_ptr();
146     a2 = tcg_temp_new_ptr();
147 
148     tcg_gen_addi_ptr(a0, cpu_env, dofs);
149     tcg_gen_addi_ptr(a1, cpu_env, aofs);
150     tcg_gen_addi_ptr(a2, cpu_env, bofs);
151 
152     fn(a0, a1, a2, desc);
153 
154     tcg_temp_free_ptr(a0);
155     tcg_temp_free_ptr(a1);
156     tcg_temp_free_ptr(a2);
157     tcg_temp_free_i32(desc);
158 }
159 
160 /* Generate a call to a gvec-style helper with four vector operands.  */
161 void tcg_gen_gvec_4_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
162                         uint32_t cofs, uint32_t oprsz, uint32_t maxsz,
163                         int32_t data, gen_helper_gvec_4 *fn)
164 {
165     TCGv_ptr a0, a1, a2, a3;
166     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
167 
168     a0 = tcg_temp_new_ptr();
169     a1 = tcg_temp_new_ptr();
170     a2 = tcg_temp_new_ptr();
171     a3 = tcg_temp_new_ptr();
172 
173     tcg_gen_addi_ptr(a0, cpu_env, dofs);
174     tcg_gen_addi_ptr(a1, cpu_env, aofs);
175     tcg_gen_addi_ptr(a2, cpu_env, bofs);
176     tcg_gen_addi_ptr(a3, cpu_env, cofs);
177 
178     fn(a0, a1, a2, a3, desc);
179 
180     tcg_temp_free_ptr(a0);
181     tcg_temp_free_ptr(a1);
182     tcg_temp_free_ptr(a2);
183     tcg_temp_free_ptr(a3);
184     tcg_temp_free_i32(desc);
185 }
186 
187 /* Generate a call to a gvec-style helper with five vector operands.  */
188 void tcg_gen_gvec_5_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
189                         uint32_t cofs, uint32_t xofs, uint32_t oprsz,
190                         uint32_t maxsz, int32_t data, gen_helper_gvec_5 *fn)
191 {
192     TCGv_ptr a0, a1, a2, a3, a4;
193     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
194 
195     a0 = tcg_temp_new_ptr();
196     a1 = tcg_temp_new_ptr();
197     a2 = tcg_temp_new_ptr();
198     a3 = tcg_temp_new_ptr();
199     a4 = tcg_temp_new_ptr();
200 
201     tcg_gen_addi_ptr(a0, cpu_env, dofs);
202     tcg_gen_addi_ptr(a1, cpu_env, aofs);
203     tcg_gen_addi_ptr(a2, cpu_env, bofs);
204     tcg_gen_addi_ptr(a3, cpu_env, cofs);
205     tcg_gen_addi_ptr(a4, cpu_env, xofs);
206 
207     fn(a0, a1, a2, a3, a4, desc);
208 
209     tcg_temp_free_ptr(a0);
210     tcg_temp_free_ptr(a1);
211     tcg_temp_free_ptr(a2);
212     tcg_temp_free_ptr(a3);
213     tcg_temp_free_ptr(a4);
214     tcg_temp_free_i32(desc);
215 }
216 
217 /* Generate a call to a gvec-style helper with three vector operands
218    and an extra pointer operand.  */
219 void tcg_gen_gvec_2_ptr(uint32_t dofs, uint32_t aofs,
220                         TCGv_ptr ptr, uint32_t oprsz, uint32_t maxsz,
221                         int32_t data, gen_helper_gvec_2_ptr *fn)
222 {
223     TCGv_ptr a0, a1;
224     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
225 
226     a0 = tcg_temp_new_ptr();
227     a1 = tcg_temp_new_ptr();
228 
229     tcg_gen_addi_ptr(a0, cpu_env, dofs);
230     tcg_gen_addi_ptr(a1, cpu_env, aofs);
231 
232     fn(a0, a1, ptr, desc);
233 
234     tcg_temp_free_ptr(a0);
235     tcg_temp_free_ptr(a1);
236     tcg_temp_free_i32(desc);
237 }
238 
239 /* Generate a call to a gvec-style helper with three vector operands
240    and an extra pointer operand.  */
241 void tcg_gen_gvec_3_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
242                         TCGv_ptr ptr, uint32_t oprsz, uint32_t maxsz,
243                         int32_t data, gen_helper_gvec_3_ptr *fn)
244 {
245     TCGv_ptr a0, a1, a2;
246     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
247 
248     a0 = tcg_temp_new_ptr();
249     a1 = tcg_temp_new_ptr();
250     a2 = tcg_temp_new_ptr();
251 
252     tcg_gen_addi_ptr(a0, cpu_env, dofs);
253     tcg_gen_addi_ptr(a1, cpu_env, aofs);
254     tcg_gen_addi_ptr(a2, cpu_env, bofs);
255 
256     fn(a0, a1, a2, ptr, desc);
257 
258     tcg_temp_free_ptr(a0);
259     tcg_temp_free_ptr(a1);
260     tcg_temp_free_ptr(a2);
261     tcg_temp_free_i32(desc);
262 }
263 
264 /* Generate a call to a gvec-style helper with four vector operands
265    and an extra pointer operand.  */
266 void tcg_gen_gvec_4_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
267                         uint32_t cofs, TCGv_ptr ptr, uint32_t oprsz,
268                         uint32_t maxsz, int32_t data,
269                         gen_helper_gvec_4_ptr *fn)
270 {
271     TCGv_ptr a0, a1, a2, a3;
272     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
273 
274     a0 = tcg_temp_new_ptr();
275     a1 = tcg_temp_new_ptr();
276     a2 = tcg_temp_new_ptr();
277     a3 = tcg_temp_new_ptr();
278 
279     tcg_gen_addi_ptr(a0, cpu_env, dofs);
280     tcg_gen_addi_ptr(a1, cpu_env, aofs);
281     tcg_gen_addi_ptr(a2, cpu_env, bofs);
282     tcg_gen_addi_ptr(a3, cpu_env, cofs);
283 
284     fn(a0, a1, a2, a3, ptr, desc);
285 
286     tcg_temp_free_ptr(a0);
287     tcg_temp_free_ptr(a1);
288     tcg_temp_free_ptr(a2);
289     tcg_temp_free_ptr(a3);
290     tcg_temp_free_i32(desc);
291 }
292 
293 /* Generate a call to a gvec-style helper with five vector operands
294    and an extra pointer operand.  */
295 void tcg_gen_gvec_5_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
296                         uint32_t cofs, uint32_t eofs, TCGv_ptr ptr,
297                         uint32_t oprsz, uint32_t maxsz, int32_t data,
298                         gen_helper_gvec_5_ptr *fn)
299 {
300     TCGv_ptr a0, a1, a2, a3, a4;
301     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
302 
303     a0 = tcg_temp_new_ptr();
304     a1 = tcg_temp_new_ptr();
305     a2 = tcg_temp_new_ptr();
306     a3 = tcg_temp_new_ptr();
307     a4 = tcg_temp_new_ptr();
308 
309     tcg_gen_addi_ptr(a0, cpu_env, dofs);
310     tcg_gen_addi_ptr(a1, cpu_env, aofs);
311     tcg_gen_addi_ptr(a2, cpu_env, bofs);
312     tcg_gen_addi_ptr(a3, cpu_env, cofs);
313     tcg_gen_addi_ptr(a4, cpu_env, eofs);
314 
315     fn(a0, a1, a2, a3, a4, ptr, desc);
316 
317     tcg_temp_free_ptr(a0);
318     tcg_temp_free_ptr(a1);
319     tcg_temp_free_ptr(a2);
320     tcg_temp_free_ptr(a3);
321     tcg_temp_free_ptr(a4);
322     tcg_temp_free_i32(desc);
323 }
324 
325 /* Return true if we want to implement something of OPRSZ bytes
326    in units of LNSZ.  This limits the expansion of inline code.  */
327 static inline bool check_size_impl(uint32_t oprsz, uint32_t lnsz)
328 {
329     uint32_t q, r;
330 
331     if (oprsz < lnsz) {
332         return false;
333     }
334 
335     q = oprsz / lnsz;
336     r = oprsz % lnsz;
337     tcg_debug_assert((r & 7) == 0);
338 
339     if (lnsz < 16) {
340         /* For sizes below 16, accept no remainder. */
341         if (r != 0) {
342             return false;
343         }
344     } else {
345         /*
346          * Recall that ARM SVE allows vector sizes that are not a
347          * power of 2, but always a multiple of 16.  The intent is
348          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
349          * In addition, expand_clr needs to handle a multiple of 8.
350          * Thus we can handle the tail with one more operation per
351          * diminishing power of 2.
352          */
353         q += ctpop32(r);
354     }
355 
356     return q <= MAX_UNROLL;
357 }
358 
359 static void expand_clr(uint32_t dofs, uint32_t maxsz);
360 
361 /* Duplicate C as per VECE.  */
362 uint64_t (dup_const)(unsigned vece, uint64_t c)
363 {
364     switch (vece) {
365     case MO_8:
366         return 0x0101010101010101ull * (uint8_t)c;
367     case MO_16:
368         return 0x0001000100010001ull * (uint16_t)c;
369     case MO_32:
370         return 0x0000000100000001ull * (uint32_t)c;
371     case MO_64:
372         return c;
373     default:
374         g_assert_not_reached();
375     }
376 }
377 
378 /* Duplicate IN into OUT as per VECE.  */
379 static void gen_dup_i32(unsigned vece, TCGv_i32 out, TCGv_i32 in)
380 {
381     switch (vece) {
382     case MO_8:
383         tcg_gen_ext8u_i32(out, in);
384         tcg_gen_muli_i32(out, out, 0x01010101);
385         break;
386     case MO_16:
387         tcg_gen_deposit_i32(out, in, in, 16, 16);
388         break;
389     case MO_32:
390         tcg_gen_mov_i32(out, in);
391         break;
392     default:
393         g_assert_not_reached();
394     }
395 }
396 
397 static void gen_dup_i64(unsigned vece, TCGv_i64 out, TCGv_i64 in)
398 {
399     switch (vece) {
400     case MO_8:
401         tcg_gen_ext8u_i64(out, in);
402         tcg_gen_muli_i64(out, out, 0x0101010101010101ull);
403         break;
404     case MO_16:
405         tcg_gen_ext16u_i64(out, in);
406         tcg_gen_muli_i64(out, out, 0x0001000100010001ull);
407         break;
408     case MO_32:
409         tcg_gen_deposit_i64(out, in, in, 32, 32);
410         break;
411     case MO_64:
412         tcg_gen_mov_i64(out, in);
413         break;
414     default:
415         g_assert_not_reached();
416     }
417 }
418 
419 /* Select a supported vector type for implementing an operation on SIZE
420  * bytes.  If OP is 0, assume that the real operation to be performed is
421  * required by all backends.  Otherwise, make sure than OP can be performed
422  * on elements of size VECE in the selected type.  Do not select V64 if
423  * PREFER_I64 is true.  Return 0 if no vector type is selected.
424  */
425 static TCGType choose_vector_type(const TCGOpcode *list, unsigned vece,
426                                   uint32_t size, bool prefer_i64)
427 {
428     /*
429      * Recall that ARM SVE allows vector sizes that are not a
430      * power of 2, but always a multiple of 16.  The intent is
431      * that e.g. size == 80 would be expanded with 2x32 + 1x16.
432      * It is hard to imagine a case in which v256 is supported
433      * but v128 is not, but check anyway.
434      * In addition, expand_clr needs to handle a multiple of 8.
435      */
436     if (TCG_TARGET_HAS_v256 &&
437         check_size_impl(size, 32) &&
438         tcg_can_emit_vecop_list(list, TCG_TYPE_V256, vece) &&
439         (!(size & 16) ||
440          (TCG_TARGET_HAS_v128 &&
441           tcg_can_emit_vecop_list(list, TCG_TYPE_V128, vece))) &&
442         (!(size & 8) ||
443          (TCG_TARGET_HAS_v64 &&
444           tcg_can_emit_vecop_list(list, TCG_TYPE_V64, vece)))) {
445         return TCG_TYPE_V256;
446     }
447     if (TCG_TARGET_HAS_v128 &&
448         check_size_impl(size, 16) &&
449         tcg_can_emit_vecop_list(list, TCG_TYPE_V128, vece) &&
450         (!(size & 8) ||
451          (TCG_TARGET_HAS_v64 &&
452           tcg_can_emit_vecop_list(list, TCG_TYPE_V64, vece)))) {
453         return TCG_TYPE_V128;
454     }
455     if (TCG_TARGET_HAS_v64 && !prefer_i64 && check_size_impl(size, 8)
456         && tcg_can_emit_vecop_list(list, TCG_TYPE_V64, vece)) {
457         return TCG_TYPE_V64;
458     }
459     return 0;
460 }
461 
462 static void do_dup_store(TCGType type, uint32_t dofs, uint32_t oprsz,
463                          uint32_t maxsz, TCGv_vec t_vec)
464 {
465     uint32_t i = 0;
466 
467     tcg_debug_assert(oprsz >= 8);
468 
469     /*
470      * This may be expand_clr for the tail of an operation, e.g.
471      * oprsz == 8 && maxsz == 64.  The first 8 bytes of this store
472      * are misaligned wrt the maximum vector size, so do that first.
473      */
474     if (dofs & 8) {
475         tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V64);
476         i += 8;
477     }
478 
479     switch (type) {
480     case TCG_TYPE_V256:
481         /*
482          * Recall that ARM SVE allows vector sizes that are not a
483          * power of 2, but always a multiple of 16.  The intent is
484          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
485          */
486         for (; i + 32 <= oprsz; i += 32) {
487             tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V256);
488         }
489         /* fallthru */
490     case TCG_TYPE_V128:
491         for (; i + 16 <= oprsz; i += 16) {
492             tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V128);
493         }
494         break;
495     case TCG_TYPE_V64:
496         for (; i < oprsz; i += 8) {
497             tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V64);
498         }
499         break;
500     default:
501         g_assert_not_reached();
502     }
503 
504     if (oprsz < maxsz) {
505         expand_clr(dofs + oprsz, maxsz - oprsz);
506     }
507 }
508 
509 /* Set OPRSZ bytes at DOFS to replications of IN_32, IN_64 or IN_C.
510  * Only one of IN_32 or IN_64 may be set;
511  * IN_C is used if IN_32 and IN_64 are unset.
512  */
513 static void do_dup(unsigned vece, uint32_t dofs, uint32_t oprsz,
514                    uint32_t maxsz, TCGv_i32 in_32, TCGv_i64 in_64,
515                    uint64_t in_c)
516 {
517     TCGType type;
518     TCGv_i64 t_64;
519     TCGv_i32 t_32, t_desc;
520     TCGv_ptr t_ptr;
521     uint32_t i;
522 
523     assert(vece <= (in_32 ? MO_32 : MO_64));
524     assert(in_32 == NULL || in_64 == NULL);
525 
526     /* If we're storing 0, expand oprsz to maxsz.  */
527     if (in_32 == NULL && in_64 == NULL) {
528         in_c = dup_const(vece, in_c);
529         if (in_c == 0) {
530             oprsz = maxsz;
531         }
532     }
533 
534     /* Implement inline with a vector type, if possible.
535      * Prefer integer when 64-bit host and no variable dup.
536      */
537     type = choose_vector_type(NULL, vece, oprsz,
538                               (TCG_TARGET_REG_BITS == 64 && in_32 == NULL
539                                && (in_64 == NULL || vece == MO_64)));
540     if (type != 0) {
541         TCGv_vec t_vec = tcg_temp_new_vec(type);
542 
543         if (in_32) {
544             tcg_gen_dup_i32_vec(vece, t_vec, in_32);
545         } else if (in_64) {
546             tcg_gen_dup_i64_vec(vece, t_vec, in_64);
547         } else {
548             tcg_gen_dupi_vec(vece, t_vec, in_c);
549         }
550         do_dup_store(type, dofs, oprsz, maxsz, t_vec);
551         tcg_temp_free_vec(t_vec);
552         return;
553     }
554 
555     /* Otherwise, inline with an integer type, unless "large".  */
556     if (check_size_impl(oprsz, TCG_TARGET_REG_BITS / 8)) {
557         t_64 = NULL;
558         t_32 = NULL;
559 
560         if (in_32) {
561             /* We are given a 32-bit variable input.  For a 64-bit host,
562                use a 64-bit operation unless the 32-bit operation would
563                be simple enough.  */
564             if (TCG_TARGET_REG_BITS == 64
565                 && (vece != MO_32 || !check_size_impl(oprsz, 4))) {
566                 t_64 = tcg_temp_new_i64();
567                 tcg_gen_extu_i32_i64(t_64, in_32);
568                 gen_dup_i64(vece, t_64, t_64);
569             } else {
570                 t_32 = tcg_temp_new_i32();
571                 gen_dup_i32(vece, t_32, in_32);
572             }
573         } else if (in_64) {
574             /* We are given a 64-bit variable input.  */
575             t_64 = tcg_temp_new_i64();
576             gen_dup_i64(vece, t_64, in_64);
577         } else {
578             /* We are given a constant input.  */
579             /* For 64-bit hosts, use 64-bit constants for "simple" constants
580                or when we'd need too many 32-bit stores, or when a 64-bit
581                constant is really required.  */
582             if (vece == MO_64
583                 || (TCG_TARGET_REG_BITS == 64
584                     && (in_c == 0 || in_c == -1
585                         || !check_size_impl(oprsz, 4)))) {
586                 t_64 = tcg_const_i64(in_c);
587             } else {
588                 t_32 = tcg_const_i32(in_c);
589             }
590         }
591 
592         /* Implement inline if we picked an implementation size above.  */
593         if (t_32) {
594             for (i = 0; i < oprsz; i += 4) {
595                 tcg_gen_st_i32(t_32, cpu_env, dofs + i);
596             }
597             tcg_temp_free_i32(t_32);
598             goto done;
599         }
600         if (t_64) {
601             for (i = 0; i < oprsz; i += 8) {
602                 tcg_gen_st_i64(t_64, cpu_env, dofs + i);
603             }
604             tcg_temp_free_i64(t_64);
605             goto done;
606         }
607     }
608 
609     /* Otherwise implement out of line.  */
610     t_ptr = tcg_temp_new_ptr();
611     tcg_gen_addi_ptr(t_ptr, cpu_env, dofs);
612     t_desc = tcg_const_i32(simd_desc(oprsz, maxsz, 0));
613 
614     if (vece == MO_64) {
615         if (in_64) {
616             gen_helper_gvec_dup64(t_ptr, t_desc, in_64);
617         } else {
618             t_64 = tcg_const_i64(in_c);
619             gen_helper_gvec_dup64(t_ptr, t_desc, t_64);
620             tcg_temp_free_i64(t_64);
621         }
622     } else {
623         typedef void dup_fn(TCGv_ptr, TCGv_i32, TCGv_i32);
624         static dup_fn * const fns[3] = {
625             gen_helper_gvec_dup8,
626             gen_helper_gvec_dup16,
627             gen_helper_gvec_dup32
628         };
629 
630         if (in_32) {
631             fns[vece](t_ptr, t_desc, in_32);
632         } else {
633             t_32 = tcg_temp_new_i32();
634             if (in_64) {
635                 tcg_gen_extrl_i64_i32(t_32, in_64);
636             } else if (vece == MO_8) {
637                 tcg_gen_movi_i32(t_32, in_c & 0xff);
638             } else if (vece == MO_16) {
639                 tcg_gen_movi_i32(t_32, in_c & 0xffff);
640             } else {
641                 tcg_gen_movi_i32(t_32, in_c);
642             }
643             fns[vece](t_ptr, t_desc, t_32);
644             tcg_temp_free_i32(t_32);
645         }
646     }
647 
648     tcg_temp_free_ptr(t_ptr);
649     tcg_temp_free_i32(t_desc);
650     return;
651 
652  done:
653     if (oprsz < maxsz) {
654         expand_clr(dofs + oprsz, maxsz - oprsz);
655     }
656 }
657 
658 /* Likewise, but with zero.  */
659 static void expand_clr(uint32_t dofs, uint32_t maxsz)
660 {
661     do_dup(MO_8, dofs, maxsz, maxsz, NULL, NULL, 0);
662 }
663 
664 /* Expand OPSZ bytes worth of two-operand operations using i32 elements.  */
665 static void expand_2_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
666                          bool load_dest, void (*fni)(TCGv_i32, TCGv_i32))
667 {
668     TCGv_i32 t0 = tcg_temp_new_i32();
669     TCGv_i32 t1 = tcg_temp_new_i32();
670     uint32_t i;
671 
672     for (i = 0; i < oprsz; i += 4) {
673         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
674         if (load_dest) {
675             tcg_gen_ld_i32(t1, cpu_env, dofs + i);
676         }
677         fni(t1, t0);
678         tcg_gen_st_i32(t1, cpu_env, dofs + i);
679     }
680     tcg_temp_free_i32(t0);
681     tcg_temp_free_i32(t1);
682 }
683 
684 static void expand_2i_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
685                           int32_t c, bool load_dest,
686                           void (*fni)(TCGv_i32, TCGv_i32, int32_t))
687 {
688     TCGv_i32 t0 = tcg_temp_new_i32();
689     TCGv_i32 t1 = tcg_temp_new_i32();
690     uint32_t i;
691 
692     for (i = 0; i < oprsz; i += 4) {
693         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
694         if (load_dest) {
695             tcg_gen_ld_i32(t1, cpu_env, dofs + i);
696         }
697         fni(t1, t0, c);
698         tcg_gen_st_i32(t1, cpu_env, dofs + i);
699     }
700     tcg_temp_free_i32(t0);
701     tcg_temp_free_i32(t1);
702 }
703 
704 static void expand_2s_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
705                           TCGv_i32 c, bool scalar_first,
706                           void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32))
707 {
708     TCGv_i32 t0 = tcg_temp_new_i32();
709     TCGv_i32 t1 = tcg_temp_new_i32();
710     uint32_t i;
711 
712     for (i = 0; i < oprsz; i += 4) {
713         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
714         if (scalar_first) {
715             fni(t1, c, t0);
716         } else {
717             fni(t1, t0, c);
718         }
719         tcg_gen_st_i32(t1, cpu_env, dofs + i);
720     }
721     tcg_temp_free_i32(t0);
722     tcg_temp_free_i32(t1);
723 }
724 
725 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
726 static void expand_3_i32(uint32_t dofs, uint32_t aofs,
727                          uint32_t bofs, uint32_t oprsz, bool load_dest,
728                          void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32))
729 {
730     TCGv_i32 t0 = tcg_temp_new_i32();
731     TCGv_i32 t1 = tcg_temp_new_i32();
732     TCGv_i32 t2 = tcg_temp_new_i32();
733     uint32_t i;
734 
735     for (i = 0; i < oprsz; i += 4) {
736         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
737         tcg_gen_ld_i32(t1, cpu_env, bofs + i);
738         if (load_dest) {
739             tcg_gen_ld_i32(t2, cpu_env, dofs + i);
740         }
741         fni(t2, t0, t1);
742         tcg_gen_st_i32(t2, cpu_env, dofs + i);
743     }
744     tcg_temp_free_i32(t2);
745     tcg_temp_free_i32(t1);
746     tcg_temp_free_i32(t0);
747 }
748 
749 static void expand_3i_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
750                           uint32_t oprsz, int32_t c, bool load_dest,
751                           void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32, int32_t))
752 {
753     TCGv_i32 t0 = tcg_temp_new_i32();
754     TCGv_i32 t1 = tcg_temp_new_i32();
755     TCGv_i32 t2 = tcg_temp_new_i32();
756     uint32_t i;
757 
758     for (i = 0; i < oprsz; i += 4) {
759         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
760         tcg_gen_ld_i32(t1, cpu_env, bofs + i);
761         if (load_dest) {
762             tcg_gen_ld_i32(t2, cpu_env, dofs + i);
763         }
764         fni(t2, t0, t1, c);
765         tcg_gen_st_i32(t2, cpu_env, dofs + i);
766     }
767     tcg_temp_free_i32(t0);
768     tcg_temp_free_i32(t1);
769     tcg_temp_free_i32(t2);
770 }
771 
772 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
773 static void expand_4_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
774                          uint32_t cofs, uint32_t oprsz, bool write_aofs,
775                          void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32))
776 {
777     TCGv_i32 t0 = tcg_temp_new_i32();
778     TCGv_i32 t1 = tcg_temp_new_i32();
779     TCGv_i32 t2 = tcg_temp_new_i32();
780     TCGv_i32 t3 = tcg_temp_new_i32();
781     uint32_t i;
782 
783     for (i = 0; i < oprsz; i += 4) {
784         tcg_gen_ld_i32(t1, cpu_env, aofs + i);
785         tcg_gen_ld_i32(t2, cpu_env, bofs + i);
786         tcg_gen_ld_i32(t3, cpu_env, cofs + i);
787         fni(t0, t1, t2, t3);
788         tcg_gen_st_i32(t0, cpu_env, dofs + i);
789         if (write_aofs) {
790             tcg_gen_st_i32(t1, cpu_env, aofs + i);
791         }
792     }
793     tcg_temp_free_i32(t3);
794     tcg_temp_free_i32(t2);
795     tcg_temp_free_i32(t1);
796     tcg_temp_free_i32(t0);
797 }
798 
799 /* Expand OPSZ bytes worth of two-operand operations using i64 elements.  */
800 static void expand_2_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
801                          bool load_dest, void (*fni)(TCGv_i64, TCGv_i64))
802 {
803     TCGv_i64 t0 = tcg_temp_new_i64();
804     TCGv_i64 t1 = tcg_temp_new_i64();
805     uint32_t i;
806 
807     for (i = 0; i < oprsz; i += 8) {
808         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
809         if (load_dest) {
810             tcg_gen_ld_i64(t1, cpu_env, dofs + i);
811         }
812         fni(t1, t0);
813         tcg_gen_st_i64(t1, cpu_env, dofs + i);
814     }
815     tcg_temp_free_i64(t0);
816     tcg_temp_free_i64(t1);
817 }
818 
819 static void expand_2i_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
820                           int64_t c, bool load_dest,
821                           void (*fni)(TCGv_i64, TCGv_i64, int64_t))
822 {
823     TCGv_i64 t0 = tcg_temp_new_i64();
824     TCGv_i64 t1 = tcg_temp_new_i64();
825     uint32_t i;
826 
827     for (i = 0; i < oprsz; i += 8) {
828         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
829         if (load_dest) {
830             tcg_gen_ld_i64(t1, cpu_env, dofs + i);
831         }
832         fni(t1, t0, c);
833         tcg_gen_st_i64(t1, cpu_env, dofs + i);
834     }
835     tcg_temp_free_i64(t0);
836     tcg_temp_free_i64(t1);
837 }
838 
839 static void expand_2s_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
840                           TCGv_i64 c, bool scalar_first,
841                           void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64))
842 {
843     TCGv_i64 t0 = tcg_temp_new_i64();
844     TCGv_i64 t1 = tcg_temp_new_i64();
845     uint32_t i;
846 
847     for (i = 0; i < oprsz; i += 8) {
848         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
849         if (scalar_first) {
850             fni(t1, c, t0);
851         } else {
852             fni(t1, t0, c);
853         }
854         tcg_gen_st_i64(t1, cpu_env, dofs + i);
855     }
856     tcg_temp_free_i64(t0);
857     tcg_temp_free_i64(t1);
858 }
859 
860 /* Expand OPSZ bytes worth of three-operand operations using i64 elements.  */
861 static void expand_3_i64(uint32_t dofs, uint32_t aofs,
862                          uint32_t bofs, uint32_t oprsz, bool load_dest,
863                          void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64))
864 {
865     TCGv_i64 t0 = tcg_temp_new_i64();
866     TCGv_i64 t1 = tcg_temp_new_i64();
867     TCGv_i64 t2 = tcg_temp_new_i64();
868     uint32_t i;
869 
870     for (i = 0; i < oprsz; i += 8) {
871         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
872         tcg_gen_ld_i64(t1, cpu_env, bofs + i);
873         if (load_dest) {
874             tcg_gen_ld_i64(t2, cpu_env, dofs + i);
875         }
876         fni(t2, t0, t1);
877         tcg_gen_st_i64(t2, cpu_env, dofs + i);
878     }
879     tcg_temp_free_i64(t2);
880     tcg_temp_free_i64(t1);
881     tcg_temp_free_i64(t0);
882 }
883 
884 static void expand_3i_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
885                           uint32_t oprsz, int64_t c, bool load_dest,
886                           void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64, int64_t))
887 {
888     TCGv_i64 t0 = tcg_temp_new_i64();
889     TCGv_i64 t1 = tcg_temp_new_i64();
890     TCGv_i64 t2 = tcg_temp_new_i64();
891     uint32_t i;
892 
893     for (i = 0; i < oprsz; i += 8) {
894         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
895         tcg_gen_ld_i64(t1, cpu_env, bofs + i);
896         if (load_dest) {
897             tcg_gen_ld_i64(t2, cpu_env, dofs + i);
898         }
899         fni(t2, t0, t1, c);
900         tcg_gen_st_i64(t2, cpu_env, dofs + i);
901     }
902     tcg_temp_free_i64(t0);
903     tcg_temp_free_i64(t1);
904     tcg_temp_free_i64(t2);
905 }
906 
907 /* Expand OPSZ bytes worth of three-operand operations using i64 elements.  */
908 static void expand_4_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
909                          uint32_t cofs, uint32_t oprsz, bool write_aofs,
910                          void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64, TCGv_i64))
911 {
912     TCGv_i64 t0 = tcg_temp_new_i64();
913     TCGv_i64 t1 = tcg_temp_new_i64();
914     TCGv_i64 t2 = tcg_temp_new_i64();
915     TCGv_i64 t3 = tcg_temp_new_i64();
916     uint32_t i;
917 
918     for (i = 0; i < oprsz; i += 8) {
919         tcg_gen_ld_i64(t1, cpu_env, aofs + i);
920         tcg_gen_ld_i64(t2, cpu_env, bofs + i);
921         tcg_gen_ld_i64(t3, cpu_env, cofs + i);
922         fni(t0, t1, t2, t3);
923         tcg_gen_st_i64(t0, cpu_env, dofs + i);
924         if (write_aofs) {
925             tcg_gen_st_i64(t1, cpu_env, aofs + i);
926         }
927     }
928     tcg_temp_free_i64(t3);
929     tcg_temp_free_i64(t2);
930     tcg_temp_free_i64(t1);
931     tcg_temp_free_i64(t0);
932 }
933 
934 /* Expand OPSZ bytes worth of two-operand operations using host vectors.  */
935 static void expand_2_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
936                          uint32_t oprsz, uint32_t tysz, TCGType type,
937                          bool load_dest,
938                          void (*fni)(unsigned, TCGv_vec, TCGv_vec))
939 {
940     TCGv_vec t0 = tcg_temp_new_vec(type);
941     TCGv_vec t1 = tcg_temp_new_vec(type);
942     uint32_t i;
943 
944     for (i = 0; i < oprsz; i += tysz) {
945         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
946         if (load_dest) {
947             tcg_gen_ld_vec(t1, cpu_env, dofs + i);
948         }
949         fni(vece, t1, t0);
950         tcg_gen_st_vec(t1, cpu_env, dofs + i);
951     }
952     tcg_temp_free_vec(t0);
953     tcg_temp_free_vec(t1);
954 }
955 
956 /* Expand OPSZ bytes worth of two-vector operands and an immediate operand
957    using host vectors.  */
958 static void expand_2i_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
959                           uint32_t oprsz, uint32_t tysz, TCGType type,
960                           int64_t c, bool load_dest,
961                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, int64_t))
962 {
963     TCGv_vec t0 = tcg_temp_new_vec(type);
964     TCGv_vec t1 = tcg_temp_new_vec(type);
965     uint32_t i;
966 
967     for (i = 0; i < oprsz; i += tysz) {
968         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
969         if (load_dest) {
970             tcg_gen_ld_vec(t1, cpu_env, dofs + i);
971         }
972         fni(vece, t1, t0, c);
973         tcg_gen_st_vec(t1, cpu_env, dofs + i);
974     }
975     tcg_temp_free_vec(t0);
976     tcg_temp_free_vec(t1);
977 }
978 
979 static void expand_2s_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
980                           uint32_t oprsz, uint32_t tysz, TCGType type,
981                           TCGv_vec c, bool scalar_first,
982                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec))
983 {
984     TCGv_vec t0 = tcg_temp_new_vec(type);
985     TCGv_vec t1 = tcg_temp_new_vec(type);
986     uint32_t i;
987 
988     for (i = 0; i < oprsz; i += tysz) {
989         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
990         if (scalar_first) {
991             fni(vece, t1, c, t0);
992         } else {
993             fni(vece, t1, t0, c);
994         }
995         tcg_gen_st_vec(t1, cpu_env, dofs + i);
996     }
997     tcg_temp_free_vec(t0);
998     tcg_temp_free_vec(t1);
999 }
1000 
1001 /* Expand OPSZ bytes worth of three-operand operations using host vectors.  */
1002 static void expand_3_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1003                          uint32_t bofs, uint32_t oprsz,
1004                          uint32_t tysz, TCGType type, bool load_dest,
1005                          void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec))
1006 {
1007     TCGv_vec t0 = tcg_temp_new_vec(type);
1008     TCGv_vec t1 = tcg_temp_new_vec(type);
1009     TCGv_vec t2 = tcg_temp_new_vec(type);
1010     uint32_t i;
1011 
1012     for (i = 0; i < oprsz; i += tysz) {
1013         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
1014         tcg_gen_ld_vec(t1, cpu_env, bofs + i);
1015         if (load_dest) {
1016             tcg_gen_ld_vec(t2, cpu_env, dofs + i);
1017         }
1018         fni(vece, t2, t0, t1);
1019         tcg_gen_st_vec(t2, cpu_env, dofs + i);
1020     }
1021     tcg_temp_free_vec(t2);
1022     tcg_temp_free_vec(t1);
1023     tcg_temp_free_vec(t0);
1024 }
1025 
1026 /*
1027  * Expand OPSZ bytes worth of three-vector operands and an immediate operand
1028  * using host vectors.
1029  */
1030 static void expand_3i_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1031                           uint32_t bofs, uint32_t oprsz, uint32_t tysz,
1032                           TCGType type, int64_t c, bool load_dest,
1033                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec,
1034                                       int64_t))
1035 {
1036     TCGv_vec t0 = tcg_temp_new_vec(type);
1037     TCGv_vec t1 = tcg_temp_new_vec(type);
1038     TCGv_vec t2 = tcg_temp_new_vec(type);
1039     uint32_t i;
1040 
1041     for (i = 0; i < oprsz; i += tysz) {
1042         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
1043         tcg_gen_ld_vec(t1, cpu_env, bofs + i);
1044         if (load_dest) {
1045             tcg_gen_ld_vec(t2, cpu_env, dofs + i);
1046         }
1047         fni(vece, t2, t0, t1, c);
1048         tcg_gen_st_vec(t2, cpu_env, dofs + i);
1049     }
1050     tcg_temp_free_vec(t0);
1051     tcg_temp_free_vec(t1);
1052     tcg_temp_free_vec(t2);
1053 }
1054 
1055 /* Expand OPSZ bytes worth of four-operand operations using host vectors.  */
1056 static void expand_4_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
1057                          uint32_t bofs, uint32_t cofs, uint32_t oprsz,
1058                          uint32_t tysz, TCGType type, bool write_aofs,
1059                          void (*fni)(unsigned, TCGv_vec, TCGv_vec,
1060                                      TCGv_vec, TCGv_vec))
1061 {
1062     TCGv_vec t0 = tcg_temp_new_vec(type);
1063     TCGv_vec t1 = tcg_temp_new_vec(type);
1064     TCGv_vec t2 = tcg_temp_new_vec(type);
1065     TCGv_vec t3 = tcg_temp_new_vec(type);
1066     uint32_t i;
1067 
1068     for (i = 0; i < oprsz; i += tysz) {
1069         tcg_gen_ld_vec(t1, cpu_env, aofs + i);
1070         tcg_gen_ld_vec(t2, cpu_env, bofs + i);
1071         tcg_gen_ld_vec(t3, cpu_env, cofs + i);
1072         fni(vece, t0, t1, t2, t3);
1073         tcg_gen_st_vec(t0, cpu_env, dofs + i);
1074         if (write_aofs) {
1075             tcg_gen_st_vec(t1, cpu_env, aofs + i);
1076         }
1077     }
1078     tcg_temp_free_vec(t3);
1079     tcg_temp_free_vec(t2);
1080     tcg_temp_free_vec(t1);
1081     tcg_temp_free_vec(t0);
1082 }
1083 
1084 /* Expand a vector two-operand operation.  */
1085 void tcg_gen_gvec_2(uint32_t dofs, uint32_t aofs,
1086                     uint32_t oprsz, uint32_t maxsz, const GVecGen2 *g)
1087 {
1088     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1089     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1090     TCGType type;
1091     uint32_t some;
1092 
1093     check_size_align(oprsz, maxsz, dofs | aofs);
1094     check_overlap_2(dofs, aofs, maxsz);
1095 
1096     type = 0;
1097     if (g->fniv) {
1098         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1099     }
1100     switch (type) {
1101     case TCG_TYPE_V256:
1102         /* Recall that ARM SVE allows vector sizes that are not a
1103          * power of 2, but always a multiple of 16.  The intent is
1104          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1105          */
1106         some = QEMU_ALIGN_DOWN(oprsz, 32);
1107         expand_2_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
1108                      g->load_dest, g->fniv);
1109         if (some == oprsz) {
1110             break;
1111         }
1112         dofs += some;
1113         aofs += some;
1114         oprsz -= some;
1115         maxsz -= some;
1116         /* fallthru */
1117     case TCG_TYPE_V128:
1118         expand_2_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
1119                      g->load_dest, g->fniv);
1120         break;
1121     case TCG_TYPE_V64:
1122         expand_2_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
1123                      g->load_dest, g->fniv);
1124         break;
1125 
1126     case 0:
1127         if (g->fni8 && check_size_impl(oprsz, 8)) {
1128             expand_2_i64(dofs, aofs, oprsz, g->load_dest, g->fni8);
1129         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1130             expand_2_i32(dofs, aofs, oprsz, g->load_dest, g->fni4);
1131         } else {
1132             assert(g->fno != NULL);
1133             tcg_gen_gvec_2_ool(dofs, aofs, oprsz, maxsz, g->data, g->fno);
1134             oprsz = maxsz;
1135         }
1136         break;
1137 
1138     default:
1139         g_assert_not_reached();
1140     }
1141     tcg_swap_vecop_list(hold_list);
1142 
1143     if (oprsz < maxsz) {
1144         expand_clr(dofs + oprsz, maxsz - oprsz);
1145     }
1146 }
1147 
1148 /* Expand a vector operation with two vectors and an immediate.  */
1149 void tcg_gen_gvec_2i(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
1150                      uint32_t maxsz, int64_t c, const GVecGen2i *g)
1151 {
1152     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1153     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1154     TCGType type;
1155     uint32_t some;
1156 
1157     check_size_align(oprsz, maxsz, dofs | aofs);
1158     check_overlap_2(dofs, aofs, maxsz);
1159 
1160     type = 0;
1161     if (g->fniv) {
1162         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1163     }
1164     switch (type) {
1165     case TCG_TYPE_V256:
1166         /* Recall that ARM SVE allows vector sizes that are not a
1167          * power of 2, but always a multiple of 16.  The intent is
1168          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1169          */
1170         some = QEMU_ALIGN_DOWN(oprsz, 32);
1171         expand_2i_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
1172                       c, g->load_dest, g->fniv);
1173         if (some == oprsz) {
1174             break;
1175         }
1176         dofs += some;
1177         aofs += some;
1178         oprsz -= some;
1179         maxsz -= some;
1180         /* fallthru */
1181     case TCG_TYPE_V128:
1182         expand_2i_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
1183                       c, g->load_dest, g->fniv);
1184         break;
1185     case TCG_TYPE_V64:
1186         expand_2i_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
1187                       c, g->load_dest, g->fniv);
1188         break;
1189 
1190     case 0:
1191         if (g->fni8 && check_size_impl(oprsz, 8)) {
1192             expand_2i_i64(dofs, aofs, oprsz, c, g->load_dest, g->fni8);
1193         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1194             expand_2i_i32(dofs, aofs, oprsz, c, g->load_dest, g->fni4);
1195         } else {
1196             if (g->fno) {
1197                 tcg_gen_gvec_2_ool(dofs, aofs, oprsz, maxsz, c, g->fno);
1198             } else {
1199                 TCGv_i64 tcg_c = tcg_const_i64(c);
1200                 tcg_gen_gvec_2i_ool(dofs, aofs, tcg_c, oprsz,
1201                                     maxsz, c, g->fnoi);
1202                 tcg_temp_free_i64(tcg_c);
1203             }
1204             oprsz = maxsz;
1205         }
1206         break;
1207 
1208     default:
1209         g_assert_not_reached();
1210     }
1211     tcg_swap_vecop_list(hold_list);
1212 
1213     if (oprsz < maxsz) {
1214         expand_clr(dofs + oprsz, maxsz - oprsz);
1215     }
1216 }
1217 
1218 /* Expand a vector operation with two vectors and a scalar.  */
1219 void tcg_gen_gvec_2s(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
1220                      uint32_t maxsz, TCGv_i64 c, const GVecGen2s *g)
1221 {
1222     TCGType type;
1223 
1224     check_size_align(oprsz, maxsz, dofs | aofs);
1225     check_overlap_2(dofs, aofs, maxsz);
1226 
1227     type = 0;
1228     if (g->fniv) {
1229         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1230     }
1231     if (type != 0) {
1232         const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1233         const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1234         TCGv_vec t_vec = tcg_temp_new_vec(type);
1235         uint32_t some;
1236 
1237         tcg_gen_dup_i64_vec(g->vece, t_vec, c);
1238 
1239         switch (type) {
1240         case TCG_TYPE_V256:
1241             /* Recall that ARM SVE allows vector sizes that are not a
1242              * power of 2, but always a multiple of 16.  The intent is
1243              * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1244              */
1245             some = QEMU_ALIGN_DOWN(oprsz, 32);
1246             expand_2s_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
1247                           t_vec, g->scalar_first, g->fniv);
1248             if (some == oprsz) {
1249                 break;
1250             }
1251             dofs += some;
1252             aofs += some;
1253             oprsz -= some;
1254             maxsz -= some;
1255             /* fallthru */
1256 
1257         case TCG_TYPE_V128:
1258             expand_2s_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
1259                           t_vec, g->scalar_first, g->fniv);
1260             break;
1261 
1262         case TCG_TYPE_V64:
1263             expand_2s_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
1264                           t_vec, g->scalar_first, g->fniv);
1265             break;
1266 
1267         default:
1268             g_assert_not_reached();
1269         }
1270         tcg_temp_free_vec(t_vec);
1271         tcg_swap_vecop_list(hold_list);
1272     } else if (g->fni8 && check_size_impl(oprsz, 8)) {
1273         TCGv_i64 t64 = tcg_temp_new_i64();
1274 
1275         gen_dup_i64(g->vece, t64, c);
1276         expand_2s_i64(dofs, aofs, oprsz, t64, g->scalar_first, g->fni8);
1277         tcg_temp_free_i64(t64);
1278     } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1279         TCGv_i32 t32 = tcg_temp_new_i32();
1280 
1281         tcg_gen_extrl_i64_i32(t32, c);
1282         gen_dup_i32(g->vece, t32, t32);
1283         expand_2s_i32(dofs, aofs, oprsz, t32, g->scalar_first, g->fni4);
1284         tcg_temp_free_i32(t32);
1285     } else {
1286         tcg_gen_gvec_2i_ool(dofs, aofs, c, oprsz, maxsz, 0, g->fno);
1287         return;
1288     }
1289 
1290     if (oprsz < maxsz) {
1291         expand_clr(dofs + oprsz, maxsz - oprsz);
1292     }
1293 }
1294 
1295 /* Expand a vector three-operand operation.  */
1296 void tcg_gen_gvec_3(uint32_t dofs, uint32_t aofs, uint32_t bofs,
1297                     uint32_t oprsz, uint32_t maxsz, const GVecGen3 *g)
1298 {
1299     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1300     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1301     TCGType type;
1302     uint32_t some;
1303 
1304     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
1305     check_overlap_3(dofs, aofs, bofs, maxsz);
1306 
1307     type = 0;
1308     if (g->fniv) {
1309         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1310     }
1311     switch (type) {
1312     case TCG_TYPE_V256:
1313         /* Recall that ARM SVE allows vector sizes that are not a
1314          * power of 2, but always a multiple of 16.  The intent is
1315          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1316          */
1317         some = QEMU_ALIGN_DOWN(oprsz, 32);
1318         expand_3_vec(g->vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256,
1319                      g->load_dest, g->fniv);
1320         if (some == oprsz) {
1321             break;
1322         }
1323         dofs += some;
1324         aofs += some;
1325         bofs += some;
1326         oprsz -= some;
1327         maxsz -= some;
1328         /* fallthru */
1329     case TCG_TYPE_V128:
1330         expand_3_vec(g->vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128,
1331                      g->load_dest, g->fniv);
1332         break;
1333     case TCG_TYPE_V64:
1334         expand_3_vec(g->vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64,
1335                      g->load_dest, g->fniv);
1336         break;
1337 
1338     case 0:
1339         if (g->fni8 && check_size_impl(oprsz, 8)) {
1340             expand_3_i64(dofs, aofs, bofs, oprsz, g->load_dest, g->fni8);
1341         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1342             expand_3_i32(dofs, aofs, bofs, oprsz, g->load_dest, g->fni4);
1343         } else {
1344             assert(g->fno != NULL);
1345             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz,
1346                                maxsz, g->data, g->fno);
1347             oprsz = maxsz;
1348         }
1349         break;
1350 
1351     default:
1352         g_assert_not_reached();
1353     }
1354     tcg_swap_vecop_list(hold_list);
1355 
1356     if (oprsz < maxsz) {
1357         expand_clr(dofs + oprsz, maxsz - oprsz);
1358     }
1359 }
1360 
1361 /* Expand a vector operation with three vectors and an immediate.  */
1362 void tcg_gen_gvec_3i(uint32_t dofs, uint32_t aofs, uint32_t bofs,
1363                      uint32_t oprsz, uint32_t maxsz, int64_t c,
1364                      const GVecGen3i *g)
1365 {
1366     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1367     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1368     TCGType type;
1369     uint32_t some;
1370 
1371     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
1372     check_overlap_3(dofs, aofs, bofs, maxsz);
1373 
1374     type = 0;
1375     if (g->fniv) {
1376         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1377     }
1378     switch (type) {
1379     case TCG_TYPE_V256:
1380         /*
1381          * Recall that ARM SVE allows vector sizes that are not a
1382          * power of 2, but always a multiple of 16.  The intent is
1383          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1384          */
1385         some = QEMU_ALIGN_DOWN(oprsz, 32);
1386         expand_3i_vec(g->vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256,
1387                       c, g->load_dest, g->fniv);
1388         if (some == oprsz) {
1389             break;
1390         }
1391         dofs += some;
1392         aofs += some;
1393         bofs += some;
1394         oprsz -= some;
1395         maxsz -= some;
1396         /* fallthru */
1397     case TCG_TYPE_V128:
1398         expand_3i_vec(g->vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128,
1399                       c, g->load_dest, g->fniv);
1400         break;
1401     case TCG_TYPE_V64:
1402         expand_3i_vec(g->vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64,
1403                       c, g->load_dest, g->fniv);
1404         break;
1405 
1406     case 0:
1407         if (g->fni8 && check_size_impl(oprsz, 8)) {
1408             expand_3i_i64(dofs, aofs, bofs, oprsz, c, g->load_dest, g->fni8);
1409         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1410             expand_3i_i32(dofs, aofs, bofs, oprsz, c, g->load_dest, g->fni4);
1411         } else {
1412             assert(g->fno != NULL);
1413             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz, maxsz, c, g->fno);
1414             oprsz = maxsz;
1415         }
1416         break;
1417 
1418     default:
1419         g_assert_not_reached();
1420     }
1421     tcg_swap_vecop_list(hold_list);
1422 
1423     if (oprsz < maxsz) {
1424         expand_clr(dofs + oprsz, maxsz - oprsz);
1425     }
1426 }
1427 
1428 /* Expand a vector four-operand operation.  */
1429 void tcg_gen_gvec_4(uint32_t dofs, uint32_t aofs, uint32_t bofs, uint32_t cofs,
1430                     uint32_t oprsz, uint32_t maxsz, const GVecGen4 *g)
1431 {
1432     const TCGOpcode *this_list = g->opt_opc ? : vecop_list_empty;
1433     const TCGOpcode *hold_list = tcg_swap_vecop_list(this_list);
1434     TCGType type;
1435     uint32_t some;
1436 
1437     check_size_align(oprsz, maxsz, dofs | aofs | bofs | cofs);
1438     check_overlap_4(dofs, aofs, bofs, cofs, maxsz);
1439 
1440     type = 0;
1441     if (g->fniv) {
1442         type = choose_vector_type(g->opt_opc, g->vece, oprsz, g->prefer_i64);
1443     }
1444     switch (type) {
1445     case TCG_TYPE_V256:
1446         /* Recall that ARM SVE allows vector sizes that are not a
1447          * power of 2, but always a multiple of 16.  The intent is
1448          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1449          */
1450         some = QEMU_ALIGN_DOWN(oprsz, 32);
1451         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, some,
1452                      32, TCG_TYPE_V256, g->write_aofs, g->fniv);
1453         if (some == oprsz) {
1454             break;
1455         }
1456         dofs += some;
1457         aofs += some;
1458         bofs += some;
1459         cofs += some;
1460         oprsz -= some;
1461         maxsz -= some;
1462         /* fallthru */
1463     case TCG_TYPE_V128:
1464         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1465                      16, TCG_TYPE_V128, g->write_aofs, g->fniv);
1466         break;
1467     case TCG_TYPE_V64:
1468         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1469                      8, TCG_TYPE_V64, g->write_aofs, g->fniv);
1470         break;
1471 
1472     case 0:
1473         if (g->fni8 && check_size_impl(oprsz, 8)) {
1474             expand_4_i64(dofs, aofs, bofs, cofs, oprsz,
1475                          g->write_aofs, g->fni8);
1476         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1477             expand_4_i32(dofs, aofs, bofs, cofs, oprsz,
1478                          g->write_aofs, g->fni4);
1479         } else {
1480             assert(g->fno != NULL);
1481             tcg_gen_gvec_4_ool(dofs, aofs, bofs, cofs,
1482                                oprsz, maxsz, g->data, g->fno);
1483             oprsz = maxsz;
1484         }
1485         break;
1486 
1487     default:
1488         g_assert_not_reached();
1489     }
1490     tcg_swap_vecop_list(hold_list);
1491 
1492     if (oprsz < maxsz) {
1493         expand_clr(dofs + oprsz, maxsz - oprsz);
1494     }
1495 }
1496 
1497 /*
1498  * Expand specific vector operations.
1499  */
1500 
1501 static void vec_mov2(unsigned vece, TCGv_vec a, TCGv_vec b)
1502 {
1503     tcg_gen_mov_vec(a, b);
1504 }
1505 
1506 void tcg_gen_gvec_mov(unsigned vece, uint32_t dofs, uint32_t aofs,
1507                       uint32_t oprsz, uint32_t maxsz)
1508 {
1509     static const GVecGen2 g = {
1510         .fni8 = tcg_gen_mov_i64,
1511         .fniv = vec_mov2,
1512         .fno = gen_helper_gvec_mov,
1513         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1514     };
1515     if (dofs != aofs) {
1516         tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g);
1517     } else {
1518         check_size_align(oprsz, maxsz, dofs);
1519         if (oprsz < maxsz) {
1520             expand_clr(dofs + oprsz, maxsz - oprsz);
1521         }
1522     }
1523 }
1524 
1525 void tcg_gen_gvec_dup_i32(unsigned vece, uint32_t dofs, uint32_t oprsz,
1526                           uint32_t maxsz, TCGv_i32 in)
1527 {
1528     check_size_align(oprsz, maxsz, dofs);
1529     tcg_debug_assert(vece <= MO_32);
1530     do_dup(vece, dofs, oprsz, maxsz, in, NULL, 0);
1531 }
1532 
1533 void tcg_gen_gvec_dup_i64(unsigned vece, uint32_t dofs, uint32_t oprsz,
1534                           uint32_t maxsz, TCGv_i64 in)
1535 {
1536     check_size_align(oprsz, maxsz, dofs);
1537     tcg_debug_assert(vece <= MO_64);
1538     do_dup(vece, dofs, oprsz, maxsz, NULL, in, 0);
1539 }
1540 
1541 void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t dofs, uint32_t aofs,
1542                           uint32_t oprsz, uint32_t maxsz)
1543 {
1544     check_size_align(oprsz, maxsz, dofs);
1545     if (vece <= MO_64) {
1546         TCGType type = choose_vector_type(NULL, vece, oprsz, 0);
1547         if (type != 0) {
1548             TCGv_vec t_vec = tcg_temp_new_vec(type);
1549             tcg_gen_dup_mem_vec(vece, t_vec, cpu_env, aofs);
1550             do_dup_store(type, dofs, oprsz, maxsz, t_vec);
1551             tcg_temp_free_vec(t_vec);
1552         } else if (vece <= MO_32) {
1553             TCGv_i32 in = tcg_temp_new_i32();
1554             switch (vece) {
1555             case MO_8:
1556                 tcg_gen_ld8u_i32(in, cpu_env, aofs);
1557                 break;
1558             case MO_16:
1559                 tcg_gen_ld16u_i32(in, cpu_env, aofs);
1560                 break;
1561             default:
1562                 tcg_gen_ld_i32(in, cpu_env, aofs);
1563                 break;
1564             }
1565             do_dup(vece, dofs, oprsz, maxsz, in, NULL, 0);
1566             tcg_temp_free_i32(in);
1567         } else {
1568             TCGv_i64 in = tcg_temp_new_i64();
1569             tcg_gen_ld_i64(in, cpu_env, aofs);
1570             do_dup(vece, dofs, oprsz, maxsz, NULL, in, 0);
1571             tcg_temp_free_i64(in);
1572         }
1573     } else if (vece == 4) {
1574         /* 128-bit duplicate.  */
1575         int i;
1576 
1577         tcg_debug_assert(oprsz >= 16);
1578         if (TCG_TARGET_HAS_v128) {
1579             TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V128);
1580 
1581             tcg_gen_ld_vec(in, cpu_env, aofs);
1582             for (i = (aofs == dofs) * 16; i < oprsz; i += 16) {
1583                 tcg_gen_st_vec(in, cpu_env, dofs + i);
1584             }
1585             tcg_temp_free_vec(in);
1586         } else {
1587             TCGv_i64 in0 = tcg_temp_new_i64();
1588             TCGv_i64 in1 = tcg_temp_new_i64();
1589 
1590             tcg_gen_ld_i64(in0, cpu_env, aofs);
1591             tcg_gen_ld_i64(in1, cpu_env, aofs + 8);
1592             for (i = (aofs == dofs) * 16; i < oprsz; i += 16) {
1593                 tcg_gen_st_i64(in0, cpu_env, dofs + i);
1594                 tcg_gen_st_i64(in1, cpu_env, dofs + i + 8);
1595             }
1596             tcg_temp_free_i64(in0);
1597             tcg_temp_free_i64(in1);
1598         }
1599         if (oprsz < maxsz) {
1600             expand_clr(dofs + oprsz, maxsz - oprsz);
1601         }
1602     } else if (vece == 5) {
1603         /* 256-bit duplicate.  */
1604         int i;
1605 
1606         tcg_debug_assert(oprsz >= 32);
1607         tcg_debug_assert(oprsz % 32 == 0);
1608         if (TCG_TARGET_HAS_v256) {
1609             TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V256);
1610 
1611             tcg_gen_ld_vec(in, cpu_env, aofs);
1612             for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
1613                 tcg_gen_st_vec(in, cpu_env, dofs + i);
1614             }
1615             tcg_temp_free_vec(in);
1616         } else if (TCG_TARGET_HAS_v128) {
1617             TCGv_vec in0 = tcg_temp_new_vec(TCG_TYPE_V128);
1618             TCGv_vec in1 = tcg_temp_new_vec(TCG_TYPE_V128);
1619 
1620             tcg_gen_ld_vec(in0, cpu_env, aofs);
1621             tcg_gen_ld_vec(in1, cpu_env, aofs + 16);
1622             for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
1623                 tcg_gen_st_vec(in0, cpu_env, dofs + i);
1624                 tcg_gen_st_vec(in1, cpu_env, dofs + i + 16);
1625             }
1626             tcg_temp_free_vec(in0);
1627             tcg_temp_free_vec(in1);
1628         } else {
1629             TCGv_i64 in[4];
1630             int j;
1631 
1632             for (j = 0; j < 4; ++j) {
1633                 in[j] = tcg_temp_new_i64();
1634                 tcg_gen_ld_i64(in[j], cpu_env, aofs + j * 8);
1635             }
1636             for (i = (aofs == dofs) * 32; i < oprsz; i += 32) {
1637                 for (j = 0; j < 4; ++j) {
1638                     tcg_gen_st_i64(in[j], cpu_env, dofs + i + j * 8);
1639                 }
1640             }
1641             for (j = 0; j < 4; ++j) {
1642                 tcg_temp_free_i64(in[j]);
1643             }
1644         }
1645         if (oprsz < maxsz) {
1646             expand_clr(dofs + oprsz, maxsz - oprsz);
1647         }
1648     } else {
1649         g_assert_not_reached();
1650     }
1651 }
1652 
1653 void tcg_gen_gvec_dup_imm(unsigned vece, uint32_t dofs, uint32_t oprsz,
1654                           uint32_t maxsz, uint64_t x)
1655 {
1656     check_size_align(oprsz, maxsz, dofs);
1657     do_dup(vece, dofs, oprsz, maxsz, NULL, NULL, x);
1658 }
1659 
1660 void tcg_gen_gvec_not(unsigned vece, uint32_t dofs, uint32_t aofs,
1661                       uint32_t oprsz, uint32_t maxsz)
1662 {
1663     static const GVecGen2 g = {
1664         .fni8 = tcg_gen_not_i64,
1665         .fniv = tcg_gen_not_vec,
1666         .fno = gen_helper_gvec_not,
1667         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1668     };
1669     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g);
1670 }
1671 
1672 /* Perform a vector addition using normal addition and a mask.  The mask
1673    should be the sign bit of each lane.  This 6-operation form is more
1674    efficient than separate additions when there are 4 or more lanes in
1675    the 64-bit operation.  */
1676 static void gen_addv_mask(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 m)
1677 {
1678     TCGv_i64 t1 = tcg_temp_new_i64();
1679     TCGv_i64 t2 = tcg_temp_new_i64();
1680     TCGv_i64 t3 = tcg_temp_new_i64();
1681 
1682     tcg_gen_andc_i64(t1, a, m);
1683     tcg_gen_andc_i64(t2, b, m);
1684     tcg_gen_xor_i64(t3, a, b);
1685     tcg_gen_add_i64(d, t1, t2);
1686     tcg_gen_and_i64(t3, t3, m);
1687     tcg_gen_xor_i64(d, d, t3);
1688 
1689     tcg_temp_free_i64(t1);
1690     tcg_temp_free_i64(t2);
1691     tcg_temp_free_i64(t3);
1692 }
1693 
1694 void tcg_gen_vec_add8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1695 {
1696     TCGv_i64 m = tcg_const_i64(dup_const(MO_8, 0x80));
1697     gen_addv_mask(d, a, b, m);
1698     tcg_temp_free_i64(m);
1699 }
1700 
1701 void tcg_gen_vec_add16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1702 {
1703     TCGv_i64 m = tcg_const_i64(dup_const(MO_16, 0x8000));
1704     gen_addv_mask(d, a, b, m);
1705     tcg_temp_free_i64(m);
1706 }
1707 
1708 void tcg_gen_vec_add32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1709 {
1710     TCGv_i64 t1 = tcg_temp_new_i64();
1711     TCGv_i64 t2 = tcg_temp_new_i64();
1712 
1713     tcg_gen_andi_i64(t1, a, ~0xffffffffull);
1714     tcg_gen_add_i64(t2, a, b);
1715     tcg_gen_add_i64(t1, t1, b);
1716     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
1717 
1718     tcg_temp_free_i64(t1);
1719     tcg_temp_free_i64(t2);
1720 }
1721 
1722 static const TCGOpcode vecop_list_add[] = { INDEX_op_add_vec, 0 };
1723 
1724 void tcg_gen_gvec_add(unsigned vece, uint32_t dofs, uint32_t aofs,
1725                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1726 {
1727     static const GVecGen3 g[4] = {
1728         { .fni8 = tcg_gen_vec_add8_i64,
1729           .fniv = tcg_gen_add_vec,
1730           .fno = gen_helper_gvec_add8,
1731           .opt_opc = vecop_list_add,
1732           .vece = MO_8 },
1733         { .fni8 = tcg_gen_vec_add16_i64,
1734           .fniv = tcg_gen_add_vec,
1735           .fno = gen_helper_gvec_add16,
1736           .opt_opc = vecop_list_add,
1737           .vece = MO_16 },
1738         { .fni4 = tcg_gen_add_i32,
1739           .fniv = tcg_gen_add_vec,
1740           .fno = gen_helper_gvec_add32,
1741           .opt_opc = vecop_list_add,
1742           .vece = MO_32 },
1743         { .fni8 = tcg_gen_add_i64,
1744           .fniv = tcg_gen_add_vec,
1745           .fno = gen_helper_gvec_add64,
1746           .opt_opc = vecop_list_add,
1747           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1748           .vece = MO_64 },
1749     };
1750 
1751     tcg_debug_assert(vece <= MO_64);
1752     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1753 }
1754 
1755 void tcg_gen_gvec_adds(unsigned vece, uint32_t dofs, uint32_t aofs,
1756                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
1757 {
1758     static const GVecGen2s g[4] = {
1759         { .fni8 = tcg_gen_vec_add8_i64,
1760           .fniv = tcg_gen_add_vec,
1761           .fno = gen_helper_gvec_adds8,
1762           .opt_opc = vecop_list_add,
1763           .vece = MO_8 },
1764         { .fni8 = tcg_gen_vec_add16_i64,
1765           .fniv = tcg_gen_add_vec,
1766           .fno = gen_helper_gvec_adds16,
1767           .opt_opc = vecop_list_add,
1768           .vece = MO_16 },
1769         { .fni4 = tcg_gen_add_i32,
1770           .fniv = tcg_gen_add_vec,
1771           .fno = gen_helper_gvec_adds32,
1772           .opt_opc = vecop_list_add,
1773           .vece = MO_32 },
1774         { .fni8 = tcg_gen_add_i64,
1775           .fniv = tcg_gen_add_vec,
1776           .fno = gen_helper_gvec_adds64,
1777           .opt_opc = vecop_list_add,
1778           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1779           .vece = MO_64 },
1780     };
1781 
1782     tcg_debug_assert(vece <= MO_64);
1783     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
1784 }
1785 
1786 void tcg_gen_gvec_addi(unsigned vece, uint32_t dofs, uint32_t aofs,
1787                        int64_t c, uint32_t oprsz, uint32_t maxsz)
1788 {
1789     TCGv_i64 tmp = tcg_const_i64(c);
1790     tcg_gen_gvec_adds(vece, dofs, aofs, tmp, oprsz, maxsz);
1791     tcg_temp_free_i64(tmp);
1792 }
1793 
1794 static const TCGOpcode vecop_list_sub[] = { INDEX_op_sub_vec, 0 };
1795 
1796 void tcg_gen_gvec_subs(unsigned vece, uint32_t dofs, uint32_t aofs,
1797                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
1798 {
1799     static const GVecGen2s g[4] = {
1800         { .fni8 = tcg_gen_vec_sub8_i64,
1801           .fniv = tcg_gen_sub_vec,
1802           .fno = gen_helper_gvec_subs8,
1803           .opt_opc = vecop_list_sub,
1804           .vece = MO_8 },
1805         { .fni8 = tcg_gen_vec_sub16_i64,
1806           .fniv = tcg_gen_sub_vec,
1807           .fno = gen_helper_gvec_subs16,
1808           .opt_opc = vecop_list_sub,
1809           .vece = MO_16 },
1810         { .fni4 = tcg_gen_sub_i32,
1811           .fniv = tcg_gen_sub_vec,
1812           .fno = gen_helper_gvec_subs32,
1813           .opt_opc = vecop_list_sub,
1814           .vece = MO_32 },
1815         { .fni8 = tcg_gen_sub_i64,
1816           .fniv = tcg_gen_sub_vec,
1817           .fno = gen_helper_gvec_subs64,
1818           .opt_opc = vecop_list_sub,
1819           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1820           .vece = MO_64 },
1821     };
1822 
1823     tcg_debug_assert(vece <= MO_64);
1824     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
1825 }
1826 
1827 /* Perform a vector subtraction using normal subtraction and a mask.
1828    Compare gen_addv_mask above.  */
1829 static void gen_subv_mask(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 m)
1830 {
1831     TCGv_i64 t1 = tcg_temp_new_i64();
1832     TCGv_i64 t2 = tcg_temp_new_i64();
1833     TCGv_i64 t3 = tcg_temp_new_i64();
1834 
1835     tcg_gen_or_i64(t1, a, m);
1836     tcg_gen_andc_i64(t2, b, m);
1837     tcg_gen_eqv_i64(t3, a, b);
1838     tcg_gen_sub_i64(d, t1, t2);
1839     tcg_gen_and_i64(t3, t3, m);
1840     tcg_gen_xor_i64(d, d, t3);
1841 
1842     tcg_temp_free_i64(t1);
1843     tcg_temp_free_i64(t2);
1844     tcg_temp_free_i64(t3);
1845 }
1846 
1847 void tcg_gen_vec_sub8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1848 {
1849     TCGv_i64 m = tcg_const_i64(dup_const(MO_8, 0x80));
1850     gen_subv_mask(d, a, b, m);
1851     tcg_temp_free_i64(m);
1852 }
1853 
1854 void tcg_gen_vec_sub16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1855 {
1856     TCGv_i64 m = tcg_const_i64(dup_const(MO_16, 0x8000));
1857     gen_subv_mask(d, a, b, m);
1858     tcg_temp_free_i64(m);
1859 }
1860 
1861 void tcg_gen_vec_sub32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1862 {
1863     TCGv_i64 t1 = tcg_temp_new_i64();
1864     TCGv_i64 t2 = tcg_temp_new_i64();
1865 
1866     tcg_gen_andi_i64(t1, b, ~0xffffffffull);
1867     tcg_gen_sub_i64(t2, a, b);
1868     tcg_gen_sub_i64(t1, a, t1);
1869     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
1870 
1871     tcg_temp_free_i64(t1);
1872     tcg_temp_free_i64(t2);
1873 }
1874 
1875 void tcg_gen_gvec_sub(unsigned vece, uint32_t dofs, uint32_t aofs,
1876                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1877 {
1878     static const GVecGen3 g[4] = {
1879         { .fni8 = tcg_gen_vec_sub8_i64,
1880           .fniv = tcg_gen_sub_vec,
1881           .fno = gen_helper_gvec_sub8,
1882           .opt_opc = vecop_list_sub,
1883           .vece = MO_8 },
1884         { .fni8 = tcg_gen_vec_sub16_i64,
1885           .fniv = tcg_gen_sub_vec,
1886           .fno = gen_helper_gvec_sub16,
1887           .opt_opc = vecop_list_sub,
1888           .vece = MO_16 },
1889         { .fni4 = tcg_gen_sub_i32,
1890           .fniv = tcg_gen_sub_vec,
1891           .fno = gen_helper_gvec_sub32,
1892           .opt_opc = vecop_list_sub,
1893           .vece = MO_32 },
1894         { .fni8 = tcg_gen_sub_i64,
1895           .fniv = tcg_gen_sub_vec,
1896           .fno = gen_helper_gvec_sub64,
1897           .opt_opc = vecop_list_sub,
1898           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1899           .vece = MO_64 },
1900     };
1901 
1902     tcg_debug_assert(vece <= MO_64);
1903     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1904 }
1905 
1906 static const TCGOpcode vecop_list_mul[] = { INDEX_op_mul_vec, 0 };
1907 
1908 void tcg_gen_gvec_mul(unsigned vece, uint32_t dofs, uint32_t aofs,
1909                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1910 {
1911     static const GVecGen3 g[4] = {
1912         { .fniv = tcg_gen_mul_vec,
1913           .fno = gen_helper_gvec_mul8,
1914           .opt_opc = vecop_list_mul,
1915           .vece = MO_8 },
1916         { .fniv = tcg_gen_mul_vec,
1917           .fno = gen_helper_gvec_mul16,
1918           .opt_opc = vecop_list_mul,
1919           .vece = MO_16 },
1920         { .fni4 = tcg_gen_mul_i32,
1921           .fniv = tcg_gen_mul_vec,
1922           .fno = gen_helper_gvec_mul32,
1923           .opt_opc = vecop_list_mul,
1924           .vece = MO_32 },
1925         { .fni8 = tcg_gen_mul_i64,
1926           .fniv = tcg_gen_mul_vec,
1927           .fno = gen_helper_gvec_mul64,
1928           .opt_opc = vecop_list_mul,
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_muls(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         { .fniv = tcg_gen_mul_vec,
1942           .fno = gen_helper_gvec_muls8,
1943           .opt_opc = vecop_list_mul,
1944           .vece = MO_8 },
1945         { .fniv = tcg_gen_mul_vec,
1946           .fno = gen_helper_gvec_muls16,
1947           .opt_opc = vecop_list_mul,
1948           .vece = MO_16 },
1949         { .fni4 = tcg_gen_mul_i32,
1950           .fniv = tcg_gen_mul_vec,
1951           .fno = gen_helper_gvec_muls32,
1952           .opt_opc = vecop_list_mul,
1953           .vece = MO_32 },
1954         { .fni8 = tcg_gen_mul_i64,
1955           .fniv = tcg_gen_mul_vec,
1956           .fno = gen_helper_gvec_muls64,
1957           .opt_opc = vecop_list_mul,
1958           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1959           .vece = MO_64 },
1960     };
1961 
1962     tcg_debug_assert(vece <= MO_64);
1963     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
1964 }
1965 
1966 void tcg_gen_gvec_muli(unsigned vece, uint32_t dofs, uint32_t aofs,
1967                        int64_t c, uint32_t oprsz, uint32_t maxsz)
1968 {
1969     TCGv_i64 tmp = tcg_const_i64(c);
1970     tcg_gen_gvec_muls(vece, dofs, aofs, tmp, oprsz, maxsz);
1971     tcg_temp_free_i64(tmp);
1972 }
1973 
1974 void tcg_gen_gvec_ssadd(unsigned vece, uint32_t dofs, uint32_t aofs,
1975                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1976 {
1977     static const TCGOpcode vecop_list[] = { INDEX_op_ssadd_vec, 0 };
1978     static const GVecGen3 g[4] = {
1979         { .fniv = tcg_gen_ssadd_vec,
1980           .fno = gen_helper_gvec_ssadd8,
1981           .opt_opc = vecop_list,
1982           .vece = MO_8 },
1983         { .fniv = tcg_gen_ssadd_vec,
1984           .fno = gen_helper_gvec_ssadd16,
1985           .opt_opc = vecop_list,
1986           .vece = MO_16 },
1987         { .fniv = tcg_gen_ssadd_vec,
1988           .fno = gen_helper_gvec_ssadd32,
1989           .opt_opc = vecop_list,
1990           .vece = MO_32 },
1991         { .fniv = tcg_gen_ssadd_vec,
1992           .fno = gen_helper_gvec_ssadd64,
1993           .opt_opc = vecop_list,
1994           .vece = MO_64 },
1995     };
1996     tcg_debug_assert(vece <= MO_64);
1997     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1998 }
1999 
2000 void tcg_gen_gvec_sssub(unsigned vece, uint32_t dofs, uint32_t aofs,
2001                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2002 {
2003     static const TCGOpcode vecop_list[] = { INDEX_op_sssub_vec, 0 };
2004     static const GVecGen3 g[4] = {
2005         { .fniv = tcg_gen_sssub_vec,
2006           .fno = gen_helper_gvec_sssub8,
2007           .opt_opc = vecop_list,
2008           .vece = MO_8 },
2009         { .fniv = tcg_gen_sssub_vec,
2010           .fno = gen_helper_gvec_sssub16,
2011           .opt_opc = vecop_list,
2012           .vece = MO_16 },
2013         { .fniv = tcg_gen_sssub_vec,
2014           .fno = gen_helper_gvec_sssub32,
2015           .opt_opc = vecop_list,
2016           .vece = MO_32 },
2017         { .fniv = tcg_gen_sssub_vec,
2018           .fno = gen_helper_gvec_sssub64,
2019           .opt_opc = vecop_list,
2020           .vece = MO_64 },
2021     };
2022     tcg_debug_assert(vece <= MO_64);
2023     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2024 }
2025 
2026 static void tcg_gen_usadd_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
2027 {
2028     TCGv_i32 max = tcg_const_i32(-1);
2029     tcg_gen_add_i32(d, a, b);
2030     tcg_gen_movcond_i32(TCG_COND_LTU, d, d, a, max, d);
2031     tcg_temp_free_i32(max);
2032 }
2033 
2034 static void tcg_gen_usadd_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2035 {
2036     TCGv_i64 max = tcg_const_i64(-1);
2037     tcg_gen_add_i64(d, a, b);
2038     tcg_gen_movcond_i64(TCG_COND_LTU, d, d, a, max, d);
2039     tcg_temp_free_i64(max);
2040 }
2041 
2042 void tcg_gen_gvec_usadd(unsigned vece, uint32_t dofs, uint32_t aofs,
2043                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2044 {
2045     static const TCGOpcode vecop_list[] = { INDEX_op_usadd_vec, 0 };
2046     static const GVecGen3 g[4] = {
2047         { .fniv = tcg_gen_usadd_vec,
2048           .fno = gen_helper_gvec_usadd8,
2049           .opt_opc = vecop_list,
2050           .vece = MO_8 },
2051         { .fniv = tcg_gen_usadd_vec,
2052           .fno = gen_helper_gvec_usadd16,
2053           .opt_opc = vecop_list,
2054           .vece = MO_16 },
2055         { .fni4 = tcg_gen_usadd_i32,
2056           .fniv = tcg_gen_usadd_vec,
2057           .fno = gen_helper_gvec_usadd32,
2058           .opt_opc = vecop_list,
2059           .vece = MO_32 },
2060         { .fni8 = tcg_gen_usadd_i64,
2061           .fniv = tcg_gen_usadd_vec,
2062           .fno = gen_helper_gvec_usadd64,
2063           .opt_opc = vecop_list,
2064           .vece = MO_64 }
2065     };
2066     tcg_debug_assert(vece <= MO_64);
2067     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2068 }
2069 
2070 static void tcg_gen_ussub_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
2071 {
2072     TCGv_i32 min = tcg_const_i32(0);
2073     tcg_gen_sub_i32(d, a, b);
2074     tcg_gen_movcond_i32(TCG_COND_LTU, d, a, b, min, d);
2075     tcg_temp_free_i32(min);
2076 }
2077 
2078 static void tcg_gen_ussub_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
2079 {
2080     TCGv_i64 min = tcg_const_i64(0);
2081     tcg_gen_sub_i64(d, a, b);
2082     tcg_gen_movcond_i64(TCG_COND_LTU, d, a, b, min, d);
2083     tcg_temp_free_i64(min);
2084 }
2085 
2086 void tcg_gen_gvec_ussub(unsigned vece, uint32_t dofs, uint32_t aofs,
2087                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2088 {
2089     static const TCGOpcode vecop_list[] = { INDEX_op_ussub_vec, 0 };
2090     static const GVecGen3 g[4] = {
2091         { .fniv = tcg_gen_ussub_vec,
2092           .fno = gen_helper_gvec_ussub8,
2093           .opt_opc = vecop_list,
2094           .vece = MO_8 },
2095         { .fniv = tcg_gen_ussub_vec,
2096           .fno = gen_helper_gvec_ussub16,
2097           .opt_opc = vecop_list,
2098           .vece = MO_16 },
2099         { .fni4 = tcg_gen_ussub_i32,
2100           .fniv = tcg_gen_ussub_vec,
2101           .fno = gen_helper_gvec_ussub32,
2102           .opt_opc = vecop_list,
2103           .vece = MO_32 },
2104         { .fni8 = tcg_gen_ussub_i64,
2105           .fniv = tcg_gen_ussub_vec,
2106           .fno = gen_helper_gvec_ussub64,
2107           .opt_opc = vecop_list,
2108           .vece = MO_64 }
2109     };
2110     tcg_debug_assert(vece <= MO_64);
2111     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2112 }
2113 
2114 void tcg_gen_gvec_smin(unsigned vece, uint32_t dofs, uint32_t aofs,
2115                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2116 {
2117     static const TCGOpcode vecop_list[] = { INDEX_op_smin_vec, 0 };
2118     static const GVecGen3 g[4] = {
2119         { .fniv = tcg_gen_smin_vec,
2120           .fno = gen_helper_gvec_smin8,
2121           .opt_opc = vecop_list,
2122           .vece = MO_8 },
2123         { .fniv = tcg_gen_smin_vec,
2124           .fno = gen_helper_gvec_smin16,
2125           .opt_opc = vecop_list,
2126           .vece = MO_16 },
2127         { .fni4 = tcg_gen_smin_i32,
2128           .fniv = tcg_gen_smin_vec,
2129           .fno = gen_helper_gvec_smin32,
2130           .opt_opc = vecop_list,
2131           .vece = MO_32 },
2132         { .fni8 = tcg_gen_smin_i64,
2133           .fniv = tcg_gen_smin_vec,
2134           .fno = gen_helper_gvec_smin64,
2135           .opt_opc = vecop_list,
2136           .vece = MO_64 }
2137     };
2138     tcg_debug_assert(vece <= MO_64);
2139     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2140 }
2141 
2142 void tcg_gen_gvec_umin(unsigned vece, uint32_t dofs, uint32_t aofs,
2143                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2144 {
2145     static const TCGOpcode vecop_list[] = { INDEX_op_umin_vec, 0 };
2146     static const GVecGen3 g[4] = {
2147         { .fniv = tcg_gen_umin_vec,
2148           .fno = gen_helper_gvec_umin8,
2149           .opt_opc = vecop_list,
2150           .vece = MO_8 },
2151         { .fniv = tcg_gen_umin_vec,
2152           .fno = gen_helper_gvec_umin16,
2153           .opt_opc = vecop_list,
2154           .vece = MO_16 },
2155         { .fni4 = tcg_gen_umin_i32,
2156           .fniv = tcg_gen_umin_vec,
2157           .fno = gen_helper_gvec_umin32,
2158           .opt_opc = vecop_list,
2159           .vece = MO_32 },
2160         { .fni8 = tcg_gen_umin_i64,
2161           .fniv = tcg_gen_umin_vec,
2162           .fno = gen_helper_gvec_umin64,
2163           .opt_opc = vecop_list,
2164           .vece = MO_64 }
2165     };
2166     tcg_debug_assert(vece <= MO_64);
2167     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2168 }
2169 
2170 void tcg_gen_gvec_smax(unsigned vece, uint32_t dofs, uint32_t aofs,
2171                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2172 {
2173     static const TCGOpcode vecop_list[] = { INDEX_op_smax_vec, 0 };
2174     static const GVecGen3 g[4] = {
2175         { .fniv = tcg_gen_smax_vec,
2176           .fno = gen_helper_gvec_smax8,
2177           .opt_opc = vecop_list,
2178           .vece = MO_8 },
2179         { .fniv = tcg_gen_smax_vec,
2180           .fno = gen_helper_gvec_smax16,
2181           .opt_opc = vecop_list,
2182           .vece = MO_16 },
2183         { .fni4 = tcg_gen_smax_i32,
2184           .fniv = tcg_gen_smax_vec,
2185           .fno = gen_helper_gvec_smax32,
2186           .opt_opc = vecop_list,
2187           .vece = MO_32 },
2188         { .fni8 = tcg_gen_smax_i64,
2189           .fniv = tcg_gen_smax_vec,
2190           .fno = gen_helper_gvec_smax64,
2191           .opt_opc = vecop_list,
2192           .vece = MO_64 }
2193     };
2194     tcg_debug_assert(vece <= MO_64);
2195     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2196 }
2197 
2198 void tcg_gen_gvec_umax(unsigned vece, uint32_t dofs, uint32_t aofs,
2199                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2200 {
2201     static const TCGOpcode vecop_list[] = { INDEX_op_umax_vec, 0 };
2202     static const GVecGen3 g[4] = {
2203         { .fniv = tcg_gen_umax_vec,
2204           .fno = gen_helper_gvec_umax8,
2205           .opt_opc = vecop_list,
2206           .vece = MO_8 },
2207         { .fniv = tcg_gen_umax_vec,
2208           .fno = gen_helper_gvec_umax16,
2209           .opt_opc = vecop_list,
2210           .vece = MO_16 },
2211         { .fni4 = tcg_gen_umax_i32,
2212           .fniv = tcg_gen_umax_vec,
2213           .fno = gen_helper_gvec_umax32,
2214           .opt_opc = vecop_list,
2215           .vece = MO_32 },
2216         { .fni8 = tcg_gen_umax_i64,
2217           .fniv = tcg_gen_umax_vec,
2218           .fno = gen_helper_gvec_umax64,
2219           .opt_opc = vecop_list,
2220           .vece = MO_64 }
2221     };
2222     tcg_debug_assert(vece <= MO_64);
2223     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
2224 }
2225 
2226 /* Perform a vector negation using normal negation and a mask.
2227    Compare gen_subv_mask above.  */
2228 static void gen_negv_mask(TCGv_i64 d, TCGv_i64 b, TCGv_i64 m)
2229 {
2230     TCGv_i64 t2 = tcg_temp_new_i64();
2231     TCGv_i64 t3 = tcg_temp_new_i64();
2232 
2233     tcg_gen_andc_i64(t3, m, b);
2234     tcg_gen_andc_i64(t2, b, m);
2235     tcg_gen_sub_i64(d, m, t2);
2236     tcg_gen_xor_i64(d, d, t3);
2237 
2238     tcg_temp_free_i64(t2);
2239     tcg_temp_free_i64(t3);
2240 }
2241 
2242 void tcg_gen_vec_neg8_i64(TCGv_i64 d, TCGv_i64 b)
2243 {
2244     TCGv_i64 m = tcg_const_i64(dup_const(MO_8, 0x80));
2245     gen_negv_mask(d, b, m);
2246     tcg_temp_free_i64(m);
2247 }
2248 
2249 void tcg_gen_vec_neg16_i64(TCGv_i64 d, TCGv_i64 b)
2250 {
2251     TCGv_i64 m = tcg_const_i64(dup_const(MO_16, 0x8000));
2252     gen_negv_mask(d, b, m);
2253     tcg_temp_free_i64(m);
2254 }
2255 
2256 void tcg_gen_vec_neg32_i64(TCGv_i64 d, TCGv_i64 b)
2257 {
2258     TCGv_i64 t1 = tcg_temp_new_i64();
2259     TCGv_i64 t2 = tcg_temp_new_i64();
2260 
2261     tcg_gen_andi_i64(t1, b, ~0xffffffffull);
2262     tcg_gen_neg_i64(t2, b);
2263     tcg_gen_neg_i64(t1, t1);
2264     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
2265 
2266     tcg_temp_free_i64(t1);
2267     tcg_temp_free_i64(t2);
2268 }
2269 
2270 void tcg_gen_gvec_neg(unsigned vece, uint32_t dofs, uint32_t aofs,
2271                       uint32_t oprsz, uint32_t maxsz)
2272 {
2273     static const TCGOpcode vecop_list[] = { INDEX_op_neg_vec, 0 };
2274     static const GVecGen2 g[4] = {
2275         { .fni8 = tcg_gen_vec_neg8_i64,
2276           .fniv = tcg_gen_neg_vec,
2277           .fno = gen_helper_gvec_neg8,
2278           .opt_opc = vecop_list,
2279           .vece = MO_8 },
2280         { .fni8 = tcg_gen_vec_neg16_i64,
2281           .fniv = tcg_gen_neg_vec,
2282           .fno = gen_helper_gvec_neg16,
2283           .opt_opc = vecop_list,
2284           .vece = MO_16 },
2285         { .fni4 = tcg_gen_neg_i32,
2286           .fniv = tcg_gen_neg_vec,
2287           .fno = gen_helper_gvec_neg32,
2288           .opt_opc = vecop_list,
2289           .vece = MO_32 },
2290         { .fni8 = tcg_gen_neg_i64,
2291           .fniv = tcg_gen_neg_vec,
2292           .fno = gen_helper_gvec_neg64,
2293           .opt_opc = vecop_list,
2294           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2295           .vece = MO_64 },
2296     };
2297 
2298     tcg_debug_assert(vece <= MO_64);
2299     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g[vece]);
2300 }
2301 
2302 static void gen_absv_mask(TCGv_i64 d, TCGv_i64 b, unsigned vece)
2303 {
2304     TCGv_i64 t = tcg_temp_new_i64();
2305     int nbit = 8 << vece;
2306 
2307     /* Create -1 for each negative element.  */
2308     tcg_gen_shri_i64(t, b, nbit - 1);
2309     tcg_gen_andi_i64(t, t, dup_const(vece, 1));
2310     tcg_gen_muli_i64(t, t, (1 << nbit) - 1);
2311 
2312     /*
2313      * Invert (via xor -1) and add one.
2314      * Because of the ordering the msb is cleared,
2315      * so we never have carry into the next element.
2316      */
2317     tcg_gen_xor_i64(d, b, t);
2318     tcg_gen_andi_i64(t, t, dup_const(vece, 1));
2319     tcg_gen_add_i64(d, d, t);
2320 
2321     tcg_temp_free_i64(t);
2322 }
2323 
2324 static void tcg_gen_vec_abs8_i64(TCGv_i64 d, TCGv_i64 b)
2325 {
2326     gen_absv_mask(d, b, MO_8);
2327 }
2328 
2329 static void tcg_gen_vec_abs16_i64(TCGv_i64 d, TCGv_i64 b)
2330 {
2331     gen_absv_mask(d, b, MO_16);
2332 }
2333 
2334 void tcg_gen_gvec_abs(unsigned vece, uint32_t dofs, uint32_t aofs,
2335                       uint32_t oprsz, uint32_t maxsz)
2336 {
2337     static const TCGOpcode vecop_list[] = { INDEX_op_abs_vec, 0 };
2338     static const GVecGen2 g[4] = {
2339         { .fni8 = tcg_gen_vec_abs8_i64,
2340           .fniv = tcg_gen_abs_vec,
2341           .fno = gen_helper_gvec_abs8,
2342           .opt_opc = vecop_list,
2343           .vece = MO_8 },
2344         { .fni8 = tcg_gen_vec_abs16_i64,
2345           .fniv = tcg_gen_abs_vec,
2346           .fno = gen_helper_gvec_abs16,
2347           .opt_opc = vecop_list,
2348           .vece = MO_16 },
2349         { .fni4 = tcg_gen_abs_i32,
2350           .fniv = tcg_gen_abs_vec,
2351           .fno = gen_helper_gvec_abs32,
2352           .opt_opc = vecop_list,
2353           .vece = MO_32 },
2354         { .fni8 = tcg_gen_abs_i64,
2355           .fniv = tcg_gen_abs_vec,
2356           .fno = gen_helper_gvec_abs64,
2357           .opt_opc = vecop_list,
2358           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2359           .vece = MO_64 },
2360     };
2361 
2362     tcg_debug_assert(vece <= MO_64);
2363     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g[vece]);
2364 }
2365 
2366 void tcg_gen_gvec_and(unsigned vece, uint32_t dofs, uint32_t aofs,
2367                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2368 {
2369     static const GVecGen3 g = {
2370         .fni8 = tcg_gen_and_i64,
2371         .fniv = tcg_gen_and_vec,
2372         .fno = gen_helper_gvec_and,
2373         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2374     };
2375 
2376     if (aofs == bofs) {
2377         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2378     } else {
2379         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2380     }
2381 }
2382 
2383 void tcg_gen_gvec_or(unsigned vece, uint32_t dofs, uint32_t aofs,
2384                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2385 {
2386     static const GVecGen3 g = {
2387         .fni8 = tcg_gen_or_i64,
2388         .fniv = tcg_gen_or_vec,
2389         .fno = gen_helper_gvec_or,
2390         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2391     };
2392 
2393     if (aofs == bofs) {
2394         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2395     } else {
2396         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2397     }
2398 }
2399 
2400 void tcg_gen_gvec_xor(unsigned vece, uint32_t dofs, uint32_t aofs,
2401                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2402 {
2403     static const GVecGen3 g = {
2404         .fni8 = tcg_gen_xor_i64,
2405         .fniv = tcg_gen_xor_vec,
2406         .fno = gen_helper_gvec_xor,
2407         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2408     };
2409 
2410     if (aofs == bofs) {
2411         tcg_gen_gvec_dup_imm(MO_64, dofs, oprsz, maxsz, 0);
2412     } else {
2413         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2414     }
2415 }
2416 
2417 void tcg_gen_gvec_andc(unsigned vece, uint32_t dofs, uint32_t aofs,
2418                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2419 {
2420     static const GVecGen3 g = {
2421         .fni8 = tcg_gen_andc_i64,
2422         .fniv = tcg_gen_andc_vec,
2423         .fno = gen_helper_gvec_andc,
2424         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2425     };
2426 
2427     if (aofs == bofs) {
2428         tcg_gen_gvec_dup_imm(MO_64, dofs, oprsz, maxsz, 0);
2429     } else {
2430         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2431     }
2432 }
2433 
2434 void tcg_gen_gvec_orc(unsigned vece, uint32_t dofs, uint32_t aofs,
2435                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2436 {
2437     static const GVecGen3 g = {
2438         .fni8 = tcg_gen_orc_i64,
2439         .fniv = tcg_gen_orc_vec,
2440         .fno = gen_helper_gvec_orc,
2441         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2442     };
2443 
2444     if (aofs == bofs) {
2445         tcg_gen_gvec_dup_imm(MO_64, dofs, oprsz, maxsz, -1);
2446     } else {
2447         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2448     }
2449 }
2450 
2451 void tcg_gen_gvec_nand(unsigned vece, uint32_t dofs, uint32_t aofs,
2452                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2453 {
2454     static const GVecGen3 g = {
2455         .fni8 = tcg_gen_nand_i64,
2456         .fniv = tcg_gen_nand_vec,
2457         .fno = gen_helper_gvec_nand,
2458         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2459     };
2460 
2461     if (aofs == bofs) {
2462         tcg_gen_gvec_not(vece, dofs, aofs, oprsz, maxsz);
2463     } else {
2464         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2465     }
2466 }
2467 
2468 void tcg_gen_gvec_nor(unsigned vece, uint32_t dofs, uint32_t aofs,
2469                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2470 {
2471     static const GVecGen3 g = {
2472         .fni8 = tcg_gen_nor_i64,
2473         .fniv = tcg_gen_nor_vec,
2474         .fno = gen_helper_gvec_nor,
2475         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2476     };
2477 
2478     if (aofs == bofs) {
2479         tcg_gen_gvec_not(vece, dofs, aofs, oprsz, maxsz);
2480     } else {
2481         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2482     }
2483 }
2484 
2485 void tcg_gen_gvec_eqv(unsigned vece, uint32_t dofs, uint32_t aofs,
2486                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2487 {
2488     static const GVecGen3 g = {
2489         .fni8 = tcg_gen_eqv_i64,
2490         .fniv = tcg_gen_eqv_vec,
2491         .fno = gen_helper_gvec_eqv,
2492         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2493     };
2494 
2495     if (aofs == bofs) {
2496         tcg_gen_gvec_dup_imm(MO_64, dofs, oprsz, maxsz, -1);
2497     } else {
2498         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2499     }
2500 }
2501 
2502 static const GVecGen2s gop_ands = {
2503     .fni8 = tcg_gen_and_i64,
2504     .fniv = tcg_gen_and_vec,
2505     .fno = gen_helper_gvec_ands,
2506     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2507     .vece = MO_64
2508 };
2509 
2510 void tcg_gen_gvec_ands(unsigned vece, uint32_t dofs, uint32_t aofs,
2511                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2512 {
2513     TCGv_i64 tmp = tcg_temp_new_i64();
2514     gen_dup_i64(vece, tmp, c);
2515     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ands);
2516     tcg_temp_free_i64(tmp);
2517 }
2518 
2519 void tcg_gen_gvec_andi(unsigned vece, uint32_t dofs, uint32_t aofs,
2520                        int64_t c, uint32_t oprsz, uint32_t maxsz)
2521 {
2522     TCGv_i64 tmp = tcg_const_i64(dup_const(vece, c));
2523     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ands);
2524     tcg_temp_free_i64(tmp);
2525 }
2526 
2527 static const GVecGen2s gop_xors = {
2528     .fni8 = tcg_gen_xor_i64,
2529     .fniv = tcg_gen_xor_vec,
2530     .fno = gen_helper_gvec_xors,
2531     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2532     .vece = MO_64
2533 };
2534 
2535 void tcg_gen_gvec_xors(unsigned vece, uint32_t dofs, uint32_t aofs,
2536                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2537 {
2538     TCGv_i64 tmp = tcg_temp_new_i64();
2539     gen_dup_i64(vece, tmp, c);
2540     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_xors);
2541     tcg_temp_free_i64(tmp);
2542 }
2543 
2544 void tcg_gen_gvec_xori(unsigned vece, uint32_t dofs, uint32_t aofs,
2545                        int64_t c, uint32_t oprsz, uint32_t maxsz)
2546 {
2547     TCGv_i64 tmp = tcg_const_i64(dup_const(vece, c));
2548     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_xors);
2549     tcg_temp_free_i64(tmp);
2550 }
2551 
2552 static const GVecGen2s gop_ors = {
2553     .fni8 = tcg_gen_or_i64,
2554     .fniv = tcg_gen_or_vec,
2555     .fno = gen_helper_gvec_ors,
2556     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2557     .vece = MO_64
2558 };
2559 
2560 void tcg_gen_gvec_ors(unsigned vece, uint32_t dofs, uint32_t aofs,
2561                       TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2562 {
2563     TCGv_i64 tmp = tcg_temp_new_i64();
2564     gen_dup_i64(vece, tmp, c);
2565     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ors);
2566     tcg_temp_free_i64(tmp);
2567 }
2568 
2569 void tcg_gen_gvec_ori(unsigned vece, uint32_t dofs, uint32_t aofs,
2570                       int64_t c, uint32_t oprsz, uint32_t maxsz)
2571 {
2572     TCGv_i64 tmp = tcg_const_i64(dup_const(vece, c));
2573     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ors);
2574     tcg_temp_free_i64(tmp);
2575 }
2576 
2577 void tcg_gen_vec_shl8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2578 {
2579     uint64_t mask = dup_const(MO_8, 0xff << c);
2580     tcg_gen_shli_i64(d, a, c);
2581     tcg_gen_andi_i64(d, d, mask);
2582 }
2583 
2584 void tcg_gen_vec_shl16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2585 {
2586     uint64_t mask = dup_const(MO_16, 0xffff << c);
2587     tcg_gen_shli_i64(d, a, c);
2588     tcg_gen_andi_i64(d, d, mask);
2589 }
2590 
2591 void tcg_gen_gvec_shli(unsigned vece, uint32_t dofs, uint32_t aofs,
2592                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
2593 {
2594     static const TCGOpcode vecop_list[] = { INDEX_op_shli_vec, 0 };
2595     static const GVecGen2i g[4] = {
2596         { .fni8 = tcg_gen_vec_shl8i_i64,
2597           .fniv = tcg_gen_shli_vec,
2598           .fno = gen_helper_gvec_shl8i,
2599           .opt_opc = vecop_list,
2600           .vece = MO_8 },
2601         { .fni8 = tcg_gen_vec_shl16i_i64,
2602           .fniv = tcg_gen_shli_vec,
2603           .fno = gen_helper_gvec_shl16i,
2604           .opt_opc = vecop_list,
2605           .vece = MO_16 },
2606         { .fni4 = tcg_gen_shli_i32,
2607           .fniv = tcg_gen_shli_vec,
2608           .fno = gen_helper_gvec_shl32i,
2609           .opt_opc = vecop_list,
2610           .vece = MO_32 },
2611         { .fni8 = tcg_gen_shli_i64,
2612           .fniv = tcg_gen_shli_vec,
2613           .fno = gen_helper_gvec_shl64i,
2614           .opt_opc = vecop_list,
2615           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2616           .vece = MO_64 },
2617     };
2618 
2619     tcg_debug_assert(vece <= MO_64);
2620     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2621     if (shift == 0) {
2622         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2623     } else {
2624         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2625     }
2626 }
2627 
2628 void tcg_gen_vec_shr8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2629 {
2630     uint64_t mask = dup_const(MO_8, 0xff >> c);
2631     tcg_gen_shri_i64(d, a, c);
2632     tcg_gen_andi_i64(d, d, mask);
2633 }
2634 
2635 void tcg_gen_vec_shr16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2636 {
2637     uint64_t mask = dup_const(MO_16, 0xffff >> c);
2638     tcg_gen_shri_i64(d, a, c);
2639     tcg_gen_andi_i64(d, d, mask);
2640 }
2641 
2642 void tcg_gen_gvec_shri(unsigned vece, uint32_t dofs, uint32_t aofs,
2643                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
2644 {
2645     static const TCGOpcode vecop_list[] = { INDEX_op_shri_vec, 0 };
2646     static const GVecGen2i g[4] = {
2647         { .fni8 = tcg_gen_vec_shr8i_i64,
2648           .fniv = tcg_gen_shri_vec,
2649           .fno = gen_helper_gvec_shr8i,
2650           .opt_opc = vecop_list,
2651           .vece = MO_8 },
2652         { .fni8 = tcg_gen_vec_shr16i_i64,
2653           .fniv = tcg_gen_shri_vec,
2654           .fno = gen_helper_gvec_shr16i,
2655           .opt_opc = vecop_list,
2656           .vece = MO_16 },
2657         { .fni4 = tcg_gen_shri_i32,
2658           .fniv = tcg_gen_shri_vec,
2659           .fno = gen_helper_gvec_shr32i,
2660           .opt_opc = vecop_list,
2661           .vece = MO_32 },
2662         { .fni8 = tcg_gen_shri_i64,
2663           .fniv = tcg_gen_shri_vec,
2664           .fno = gen_helper_gvec_shr64i,
2665           .opt_opc = vecop_list,
2666           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2667           .vece = MO_64 },
2668     };
2669 
2670     tcg_debug_assert(vece <= MO_64);
2671     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2672     if (shift == 0) {
2673         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2674     } else {
2675         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2676     }
2677 }
2678 
2679 void tcg_gen_vec_sar8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2680 {
2681     uint64_t s_mask = dup_const(MO_8, 0x80 >> c);
2682     uint64_t c_mask = dup_const(MO_8, 0xff >> c);
2683     TCGv_i64 s = tcg_temp_new_i64();
2684 
2685     tcg_gen_shri_i64(d, a, c);
2686     tcg_gen_andi_i64(s, d, s_mask);  /* isolate (shifted) sign bit */
2687     tcg_gen_muli_i64(s, s, (2 << c) - 2); /* replicate isolated signs */
2688     tcg_gen_andi_i64(d, d, c_mask);  /* clear out bits above sign  */
2689     tcg_gen_or_i64(d, d, s);         /* include sign extension */
2690     tcg_temp_free_i64(s);
2691 }
2692 
2693 void tcg_gen_vec_sar16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2694 {
2695     uint64_t s_mask = dup_const(MO_16, 0x8000 >> c);
2696     uint64_t c_mask = dup_const(MO_16, 0xffff >> c);
2697     TCGv_i64 s = tcg_temp_new_i64();
2698 
2699     tcg_gen_shri_i64(d, a, c);
2700     tcg_gen_andi_i64(s, d, s_mask);  /* isolate (shifted) sign bit */
2701     tcg_gen_andi_i64(d, d, c_mask);  /* clear out bits above sign  */
2702     tcg_gen_muli_i64(s, s, (2 << c) - 2); /* replicate isolated signs */
2703     tcg_gen_or_i64(d, d, s);         /* include sign extension */
2704     tcg_temp_free_i64(s);
2705 }
2706 
2707 void tcg_gen_gvec_sari(unsigned vece, uint32_t dofs, uint32_t aofs,
2708                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
2709 {
2710     static const TCGOpcode vecop_list[] = { INDEX_op_sari_vec, 0 };
2711     static const GVecGen2i g[4] = {
2712         { .fni8 = tcg_gen_vec_sar8i_i64,
2713           .fniv = tcg_gen_sari_vec,
2714           .fno = gen_helper_gvec_sar8i,
2715           .opt_opc = vecop_list,
2716           .vece = MO_8 },
2717         { .fni8 = tcg_gen_vec_sar16i_i64,
2718           .fniv = tcg_gen_sari_vec,
2719           .fno = gen_helper_gvec_sar16i,
2720           .opt_opc = vecop_list,
2721           .vece = MO_16 },
2722         { .fni4 = tcg_gen_sari_i32,
2723           .fniv = tcg_gen_sari_vec,
2724           .fno = gen_helper_gvec_sar32i,
2725           .opt_opc = vecop_list,
2726           .vece = MO_32 },
2727         { .fni8 = tcg_gen_sari_i64,
2728           .fniv = tcg_gen_sari_vec,
2729           .fno = gen_helper_gvec_sar64i,
2730           .opt_opc = vecop_list,
2731           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2732           .vece = MO_64 },
2733     };
2734 
2735     tcg_debug_assert(vece <= MO_64);
2736     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2737     if (shift == 0) {
2738         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2739     } else {
2740         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2741     }
2742 }
2743 
2744 void tcg_gen_vec_rotl8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2745 {
2746     uint64_t mask = dup_const(MO_8, 0xff << c);
2747 
2748     tcg_gen_shli_i64(d, a, c);
2749     tcg_gen_shri_i64(a, a, 8 - c);
2750     tcg_gen_andi_i64(d, d, mask);
2751     tcg_gen_andi_i64(a, a, ~mask);
2752     tcg_gen_or_i64(d, d, a);
2753 }
2754 
2755 void tcg_gen_vec_rotl16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2756 {
2757     uint64_t mask = dup_const(MO_16, 0xffff << c);
2758 
2759     tcg_gen_shli_i64(d, a, c);
2760     tcg_gen_shri_i64(a, a, 16 - c);
2761     tcg_gen_andi_i64(d, d, mask);
2762     tcg_gen_andi_i64(a, a, ~mask);
2763     tcg_gen_or_i64(d, d, a);
2764 }
2765 
2766 void tcg_gen_gvec_rotli(unsigned vece, uint32_t dofs, uint32_t aofs,
2767                         int64_t shift, uint32_t oprsz, uint32_t maxsz)
2768 {
2769     static const TCGOpcode vecop_list[] = { INDEX_op_rotli_vec, 0 };
2770     static const GVecGen2i g[4] = {
2771         { .fni8 = tcg_gen_vec_rotl8i_i64,
2772           .fniv = tcg_gen_rotli_vec,
2773           .fno = gen_helper_gvec_rotl8i,
2774           .opt_opc = vecop_list,
2775           .vece = MO_8 },
2776         { .fni8 = tcg_gen_vec_rotl16i_i64,
2777           .fniv = tcg_gen_rotli_vec,
2778           .fno = gen_helper_gvec_rotl16i,
2779           .opt_opc = vecop_list,
2780           .vece = MO_16 },
2781         { .fni4 = tcg_gen_rotli_i32,
2782           .fniv = tcg_gen_rotli_vec,
2783           .fno = gen_helper_gvec_rotl32i,
2784           .opt_opc = vecop_list,
2785           .vece = MO_32 },
2786         { .fni8 = tcg_gen_rotli_i64,
2787           .fniv = tcg_gen_rotli_vec,
2788           .fno = gen_helper_gvec_rotl64i,
2789           .opt_opc = vecop_list,
2790           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2791           .vece = MO_64 },
2792     };
2793 
2794     tcg_debug_assert(vece <= MO_64);
2795     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2796     if (shift == 0) {
2797         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2798     } else {
2799         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2800     }
2801 }
2802 
2803 void tcg_gen_gvec_rotri(unsigned vece, uint32_t dofs, uint32_t aofs,
2804                         int64_t shift, uint32_t oprsz, uint32_t maxsz)
2805 {
2806     tcg_debug_assert(vece <= MO_64);
2807     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2808     tcg_gen_gvec_rotli(vece, dofs, aofs, -shift & ((8 << vece) - 1),
2809                        oprsz, maxsz);
2810 }
2811 
2812 /*
2813  * Specialized generation vector shifts by a non-constant scalar.
2814  */
2815 
2816 typedef struct {
2817     void (*fni4)(TCGv_i32, TCGv_i32, TCGv_i32);
2818     void (*fni8)(TCGv_i64, TCGv_i64, TCGv_i64);
2819     void (*fniv_s)(unsigned, TCGv_vec, TCGv_vec, TCGv_i32);
2820     void (*fniv_v)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec);
2821     gen_helper_gvec_2 *fno[4];
2822     TCGOpcode s_list[2];
2823     TCGOpcode v_list[2];
2824 } GVecGen2sh;
2825 
2826 static void expand_2sh_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
2827                            uint32_t oprsz, uint32_t tysz, TCGType type,
2828                            TCGv_i32 shift,
2829                            void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_i32))
2830 {
2831     TCGv_vec t0 = tcg_temp_new_vec(type);
2832     uint32_t i;
2833 
2834     for (i = 0; i < oprsz; i += tysz) {
2835         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
2836         fni(vece, t0, t0, shift);
2837         tcg_gen_st_vec(t0, cpu_env, dofs + i);
2838     }
2839     tcg_temp_free_vec(t0);
2840 }
2841 
2842 static void
2843 do_gvec_shifts(unsigned vece, uint32_t dofs, uint32_t aofs, TCGv_i32 shift,
2844                uint32_t oprsz, uint32_t maxsz, const GVecGen2sh *g)
2845 {
2846     TCGType type;
2847     uint32_t some;
2848 
2849     check_size_align(oprsz, maxsz, dofs | aofs);
2850     check_overlap_2(dofs, aofs, maxsz);
2851 
2852     /* If the backend has a scalar expansion, great.  */
2853     type = choose_vector_type(g->s_list, vece, oprsz, vece == MO_64);
2854     if (type) {
2855         const TCGOpcode *hold_list = tcg_swap_vecop_list(NULL);
2856         switch (type) {
2857         case TCG_TYPE_V256:
2858             some = QEMU_ALIGN_DOWN(oprsz, 32);
2859             expand_2sh_vec(vece, dofs, aofs, some, 32,
2860                            TCG_TYPE_V256, shift, g->fniv_s);
2861             if (some == oprsz) {
2862                 break;
2863             }
2864             dofs += some;
2865             aofs += some;
2866             oprsz -= some;
2867             maxsz -= some;
2868             /* fallthru */
2869         case TCG_TYPE_V128:
2870             expand_2sh_vec(vece, dofs, aofs, oprsz, 16,
2871                            TCG_TYPE_V128, shift, g->fniv_s);
2872             break;
2873         case TCG_TYPE_V64:
2874             expand_2sh_vec(vece, dofs, aofs, oprsz, 8,
2875                            TCG_TYPE_V64, shift, g->fniv_s);
2876             break;
2877         default:
2878             g_assert_not_reached();
2879         }
2880         tcg_swap_vecop_list(hold_list);
2881         goto clear_tail;
2882     }
2883 
2884     /* If the backend supports variable vector shifts, also cool.  */
2885     type = choose_vector_type(g->v_list, vece, oprsz, vece == MO_64);
2886     if (type) {
2887         const TCGOpcode *hold_list = tcg_swap_vecop_list(NULL);
2888         TCGv_vec v_shift = tcg_temp_new_vec(type);
2889 
2890         if (vece == MO_64) {
2891             TCGv_i64 sh64 = tcg_temp_new_i64();
2892             tcg_gen_extu_i32_i64(sh64, shift);
2893             tcg_gen_dup_i64_vec(MO_64, v_shift, sh64);
2894             tcg_temp_free_i64(sh64);
2895         } else {
2896             tcg_gen_dup_i32_vec(vece, v_shift, shift);
2897         }
2898 
2899         switch (type) {
2900         case TCG_TYPE_V256:
2901             some = QEMU_ALIGN_DOWN(oprsz, 32);
2902             expand_2s_vec(vece, dofs, aofs, some, 32, TCG_TYPE_V256,
2903                           v_shift, false, g->fniv_v);
2904             if (some == oprsz) {
2905                 break;
2906             }
2907             dofs += some;
2908             aofs += some;
2909             oprsz -= some;
2910             maxsz -= some;
2911             /* fallthru */
2912         case TCG_TYPE_V128:
2913             expand_2s_vec(vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
2914                           v_shift, false, g->fniv_v);
2915             break;
2916         case TCG_TYPE_V64:
2917             expand_2s_vec(vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
2918                           v_shift, false, g->fniv_v);
2919             break;
2920         default:
2921             g_assert_not_reached();
2922         }
2923         tcg_temp_free_vec(v_shift);
2924         tcg_swap_vecop_list(hold_list);
2925         goto clear_tail;
2926     }
2927 
2928     /* Otherwise fall back to integral... */
2929     if (vece == MO_32 && check_size_impl(oprsz, 4)) {
2930         expand_2s_i32(dofs, aofs, oprsz, shift, false, g->fni4);
2931     } else if (vece == MO_64 && check_size_impl(oprsz, 8)) {
2932         TCGv_i64 sh64 = tcg_temp_new_i64();
2933         tcg_gen_extu_i32_i64(sh64, shift);
2934         expand_2s_i64(dofs, aofs, oprsz, sh64, false, g->fni8);
2935         tcg_temp_free_i64(sh64);
2936     } else {
2937         TCGv_ptr a0 = tcg_temp_new_ptr();
2938         TCGv_ptr a1 = tcg_temp_new_ptr();
2939         TCGv_i32 desc = tcg_temp_new_i32();
2940 
2941         tcg_gen_shli_i32(desc, shift, SIMD_DATA_SHIFT);
2942         tcg_gen_ori_i32(desc, desc, simd_desc(oprsz, maxsz, 0));
2943         tcg_gen_addi_ptr(a0, cpu_env, dofs);
2944         tcg_gen_addi_ptr(a1, cpu_env, aofs);
2945 
2946         g->fno[vece](a0, a1, desc);
2947 
2948         tcg_temp_free_ptr(a0);
2949         tcg_temp_free_ptr(a1);
2950         tcg_temp_free_i32(desc);
2951         return;
2952     }
2953 
2954  clear_tail:
2955     if (oprsz < maxsz) {
2956         expand_clr(dofs + oprsz, maxsz - oprsz);
2957     }
2958 }
2959 
2960 void tcg_gen_gvec_shls(unsigned vece, uint32_t dofs, uint32_t aofs,
2961                        TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
2962 {
2963     static const GVecGen2sh g = {
2964         .fni4 = tcg_gen_shl_i32,
2965         .fni8 = tcg_gen_shl_i64,
2966         .fniv_s = tcg_gen_shls_vec,
2967         .fniv_v = tcg_gen_shlv_vec,
2968         .fno = {
2969             gen_helper_gvec_shl8i,
2970             gen_helper_gvec_shl16i,
2971             gen_helper_gvec_shl32i,
2972             gen_helper_gvec_shl64i,
2973         },
2974         .s_list = { INDEX_op_shls_vec, 0 },
2975         .v_list = { INDEX_op_shlv_vec, 0 },
2976     };
2977 
2978     tcg_debug_assert(vece <= MO_64);
2979     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
2980 }
2981 
2982 void tcg_gen_gvec_shrs(unsigned vece, uint32_t dofs, uint32_t aofs,
2983                        TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
2984 {
2985     static const GVecGen2sh g = {
2986         .fni4 = tcg_gen_shr_i32,
2987         .fni8 = tcg_gen_shr_i64,
2988         .fniv_s = tcg_gen_shrs_vec,
2989         .fniv_v = tcg_gen_shrv_vec,
2990         .fno = {
2991             gen_helper_gvec_shr8i,
2992             gen_helper_gvec_shr16i,
2993             gen_helper_gvec_shr32i,
2994             gen_helper_gvec_shr64i,
2995         },
2996         .s_list = { INDEX_op_shrs_vec, 0 },
2997         .v_list = { INDEX_op_shrv_vec, 0 },
2998     };
2999 
3000     tcg_debug_assert(vece <= MO_64);
3001     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
3002 }
3003 
3004 void tcg_gen_gvec_sars(unsigned vece, uint32_t dofs, uint32_t aofs,
3005                        TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3006 {
3007     static const GVecGen2sh g = {
3008         .fni4 = tcg_gen_sar_i32,
3009         .fni8 = tcg_gen_sar_i64,
3010         .fniv_s = tcg_gen_sars_vec,
3011         .fniv_v = tcg_gen_sarv_vec,
3012         .fno = {
3013             gen_helper_gvec_sar8i,
3014             gen_helper_gvec_sar16i,
3015             gen_helper_gvec_sar32i,
3016             gen_helper_gvec_sar64i,
3017         },
3018         .s_list = { INDEX_op_sars_vec, 0 },
3019         .v_list = { INDEX_op_sarv_vec, 0 },
3020     };
3021 
3022     tcg_debug_assert(vece <= MO_64);
3023     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
3024 }
3025 
3026 void tcg_gen_gvec_rotls(unsigned vece, uint32_t dofs, uint32_t aofs,
3027                         TCGv_i32 shift, uint32_t oprsz, uint32_t maxsz)
3028 {
3029     static const GVecGen2sh g = {
3030         .fni4 = tcg_gen_rotl_i32,
3031         .fni8 = tcg_gen_rotl_i64,
3032         .fniv_s = tcg_gen_rotls_vec,
3033         .fniv_v = tcg_gen_rotlv_vec,
3034         .fno = {
3035             gen_helper_gvec_rotl8i,
3036             gen_helper_gvec_rotl16i,
3037             gen_helper_gvec_rotl32i,
3038             gen_helper_gvec_rotl64i,
3039         },
3040         .s_list = { INDEX_op_rotls_vec, 0 },
3041         .v_list = { INDEX_op_rotlv_vec, 0 },
3042     };
3043 
3044     tcg_debug_assert(vece <= MO_64);
3045     do_gvec_shifts(vece, dofs, aofs, shift, oprsz, maxsz, &g);
3046 }
3047 
3048 /*
3049  * Expand D = A << (B % element bits)
3050  *
3051  * Unlike scalar shifts, where it is easy for the target front end
3052  * to include the modulo as part of the expansion.  If the target
3053  * naturally includes the modulo as part of the operation, great!
3054  * If the target has some other behaviour from out-of-range shifts,
3055  * then it could not use this function anyway, and would need to
3056  * do it's own expansion with custom functions.
3057  */
3058 static void tcg_gen_shlv_mod_vec(unsigned vece, TCGv_vec d,
3059                                  TCGv_vec a, TCGv_vec b)
3060 {
3061     TCGv_vec t = tcg_temp_new_vec_matching(d);
3062 
3063     tcg_gen_dupi_vec(vece, t, (8 << vece) - 1);
3064     tcg_gen_and_vec(vece, t, t, b);
3065     tcg_gen_shlv_vec(vece, d, a, t);
3066     tcg_temp_free_vec(t);
3067 }
3068 
3069 static void tcg_gen_shl_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3070 {
3071     TCGv_i32 t = tcg_temp_new_i32();
3072 
3073     tcg_gen_andi_i32(t, b, 31);
3074     tcg_gen_shl_i32(d, a, t);
3075     tcg_temp_free_i32(t);
3076 }
3077 
3078 static void tcg_gen_shl_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3079 {
3080     TCGv_i64 t = tcg_temp_new_i64();
3081 
3082     tcg_gen_andi_i64(t, b, 63);
3083     tcg_gen_shl_i64(d, a, t);
3084     tcg_temp_free_i64(t);
3085 }
3086 
3087 void tcg_gen_gvec_shlv(unsigned vece, uint32_t dofs, uint32_t aofs,
3088                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3089 {
3090     static const TCGOpcode vecop_list[] = { INDEX_op_shlv_vec, 0 };
3091     static const GVecGen3 g[4] = {
3092         { .fniv = tcg_gen_shlv_mod_vec,
3093           .fno = gen_helper_gvec_shl8v,
3094           .opt_opc = vecop_list,
3095           .vece = MO_8 },
3096         { .fniv = tcg_gen_shlv_mod_vec,
3097           .fno = gen_helper_gvec_shl16v,
3098           .opt_opc = vecop_list,
3099           .vece = MO_16 },
3100         { .fni4 = tcg_gen_shl_mod_i32,
3101           .fniv = tcg_gen_shlv_mod_vec,
3102           .fno = gen_helper_gvec_shl32v,
3103           .opt_opc = vecop_list,
3104           .vece = MO_32 },
3105         { .fni8 = tcg_gen_shl_mod_i64,
3106           .fniv = tcg_gen_shlv_mod_vec,
3107           .fno = gen_helper_gvec_shl64v,
3108           .opt_opc = vecop_list,
3109           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3110           .vece = MO_64 },
3111     };
3112 
3113     tcg_debug_assert(vece <= MO_64);
3114     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3115 }
3116 
3117 /*
3118  * Similarly for logical right shifts.
3119  */
3120 
3121 static void tcg_gen_shrv_mod_vec(unsigned vece, TCGv_vec d,
3122                                  TCGv_vec a, TCGv_vec b)
3123 {
3124     TCGv_vec t = tcg_temp_new_vec_matching(d);
3125 
3126     tcg_gen_dupi_vec(vece, t, (8 << vece) - 1);
3127     tcg_gen_and_vec(vece, t, t, b);
3128     tcg_gen_shrv_vec(vece, d, a, t);
3129     tcg_temp_free_vec(t);
3130 }
3131 
3132 static void tcg_gen_shr_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3133 {
3134     TCGv_i32 t = tcg_temp_new_i32();
3135 
3136     tcg_gen_andi_i32(t, b, 31);
3137     tcg_gen_shr_i32(d, a, t);
3138     tcg_temp_free_i32(t);
3139 }
3140 
3141 static void tcg_gen_shr_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3142 {
3143     TCGv_i64 t = tcg_temp_new_i64();
3144 
3145     tcg_gen_andi_i64(t, b, 63);
3146     tcg_gen_shr_i64(d, a, t);
3147     tcg_temp_free_i64(t);
3148 }
3149 
3150 void tcg_gen_gvec_shrv(unsigned vece, uint32_t dofs, uint32_t aofs,
3151                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3152 {
3153     static const TCGOpcode vecop_list[] = { INDEX_op_shrv_vec, 0 };
3154     static const GVecGen3 g[4] = {
3155         { .fniv = tcg_gen_shrv_mod_vec,
3156           .fno = gen_helper_gvec_shr8v,
3157           .opt_opc = vecop_list,
3158           .vece = MO_8 },
3159         { .fniv = tcg_gen_shrv_mod_vec,
3160           .fno = gen_helper_gvec_shr16v,
3161           .opt_opc = vecop_list,
3162           .vece = MO_16 },
3163         { .fni4 = tcg_gen_shr_mod_i32,
3164           .fniv = tcg_gen_shrv_mod_vec,
3165           .fno = gen_helper_gvec_shr32v,
3166           .opt_opc = vecop_list,
3167           .vece = MO_32 },
3168         { .fni8 = tcg_gen_shr_mod_i64,
3169           .fniv = tcg_gen_shrv_mod_vec,
3170           .fno = gen_helper_gvec_shr64v,
3171           .opt_opc = vecop_list,
3172           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3173           .vece = MO_64 },
3174     };
3175 
3176     tcg_debug_assert(vece <= MO_64);
3177     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3178 }
3179 
3180 /*
3181  * Similarly for arithmetic right shifts.
3182  */
3183 
3184 static void tcg_gen_sarv_mod_vec(unsigned vece, TCGv_vec d,
3185                                  TCGv_vec a, TCGv_vec b)
3186 {
3187     TCGv_vec t = tcg_temp_new_vec_matching(d);
3188 
3189     tcg_gen_dupi_vec(vece, t, (8 << vece) - 1);
3190     tcg_gen_and_vec(vece, t, t, b);
3191     tcg_gen_sarv_vec(vece, d, a, t);
3192     tcg_temp_free_vec(t);
3193 }
3194 
3195 static void tcg_gen_sar_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3196 {
3197     TCGv_i32 t = tcg_temp_new_i32();
3198 
3199     tcg_gen_andi_i32(t, b, 31);
3200     tcg_gen_sar_i32(d, a, t);
3201     tcg_temp_free_i32(t);
3202 }
3203 
3204 static void tcg_gen_sar_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3205 {
3206     TCGv_i64 t = tcg_temp_new_i64();
3207 
3208     tcg_gen_andi_i64(t, b, 63);
3209     tcg_gen_sar_i64(d, a, t);
3210     tcg_temp_free_i64(t);
3211 }
3212 
3213 void tcg_gen_gvec_sarv(unsigned vece, uint32_t dofs, uint32_t aofs,
3214                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3215 {
3216     static const TCGOpcode vecop_list[] = { INDEX_op_sarv_vec, 0 };
3217     static const GVecGen3 g[4] = {
3218         { .fniv = tcg_gen_sarv_mod_vec,
3219           .fno = gen_helper_gvec_sar8v,
3220           .opt_opc = vecop_list,
3221           .vece = MO_8 },
3222         { .fniv = tcg_gen_sarv_mod_vec,
3223           .fno = gen_helper_gvec_sar16v,
3224           .opt_opc = vecop_list,
3225           .vece = MO_16 },
3226         { .fni4 = tcg_gen_sar_mod_i32,
3227           .fniv = tcg_gen_sarv_mod_vec,
3228           .fno = gen_helper_gvec_sar32v,
3229           .opt_opc = vecop_list,
3230           .vece = MO_32 },
3231         { .fni8 = tcg_gen_sar_mod_i64,
3232           .fniv = tcg_gen_sarv_mod_vec,
3233           .fno = gen_helper_gvec_sar64v,
3234           .opt_opc = vecop_list,
3235           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3236           .vece = MO_64 },
3237     };
3238 
3239     tcg_debug_assert(vece <= MO_64);
3240     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3241 }
3242 
3243 /*
3244  * Similarly for rotates.
3245  */
3246 
3247 static void tcg_gen_rotlv_mod_vec(unsigned vece, TCGv_vec d,
3248                                   TCGv_vec a, TCGv_vec b)
3249 {
3250     TCGv_vec t = tcg_temp_new_vec_matching(d);
3251 
3252     tcg_gen_dupi_vec(vece, t, (8 << vece) - 1);
3253     tcg_gen_and_vec(vece, t, t, b);
3254     tcg_gen_rotlv_vec(vece, d, a, t);
3255     tcg_temp_free_vec(t);
3256 }
3257 
3258 static void tcg_gen_rotl_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3259 {
3260     TCGv_i32 t = tcg_temp_new_i32();
3261 
3262     tcg_gen_andi_i32(t, b, 31);
3263     tcg_gen_rotl_i32(d, a, t);
3264     tcg_temp_free_i32(t);
3265 }
3266 
3267 static void tcg_gen_rotl_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3268 {
3269     TCGv_i64 t = tcg_temp_new_i64();
3270 
3271     tcg_gen_andi_i64(t, b, 63);
3272     tcg_gen_rotl_i64(d, a, t);
3273     tcg_temp_free_i64(t);
3274 }
3275 
3276 void tcg_gen_gvec_rotlv(unsigned vece, uint32_t dofs, uint32_t aofs,
3277                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3278 {
3279     static const TCGOpcode vecop_list[] = { INDEX_op_rotlv_vec, 0 };
3280     static const GVecGen3 g[4] = {
3281         { .fniv = tcg_gen_rotlv_mod_vec,
3282           .fno = gen_helper_gvec_rotl8v,
3283           .opt_opc = vecop_list,
3284           .vece = MO_8 },
3285         { .fniv = tcg_gen_rotlv_mod_vec,
3286           .fno = gen_helper_gvec_rotl16v,
3287           .opt_opc = vecop_list,
3288           .vece = MO_16 },
3289         { .fni4 = tcg_gen_rotl_mod_i32,
3290           .fniv = tcg_gen_rotlv_mod_vec,
3291           .fno = gen_helper_gvec_rotl32v,
3292           .opt_opc = vecop_list,
3293           .vece = MO_32 },
3294         { .fni8 = tcg_gen_rotl_mod_i64,
3295           .fniv = tcg_gen_rotlv_mod_vec,
3296           .fno = gen_helper_gvec_rotl64v,
3297           .opt_opc = vecop_list,
3298           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3299           .vece = MO_64 },
3300     };
3301 
3302     tcg_debug_assert(vece <= MO_64);
3303     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3304 }
3305 
3306 static void tcg_gen_rotrv_mod_vec(unsigned vece, TCGv_vec d,
3307                                   TCGv_vec a, TCGv_vec b)
3308 {
3309     TCGv_vec t = tcg_temp_new_vec_matching(d);
3310 
3311     tcg_gen_dupi_vec(vece, t, (8 << vece) - 1);
3312     tcg_gen_and_vec(vece, t, t, b);
3313     tcg_gen_rotrv_vec(vece, d, a, t);
3314     tcg_temp_free_vec(t);
3315 }
3316 
3317 static void tcg_gen_rotr_mod_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
3318 {
3319     TCGv_i32 t = tcg_temp_new_i32();
3320 
3321     tcg_gen_andi_i32(t, b, 31);
3322     tcg_gen_rotr_i32(d, a, t);
3323     tcg_temp_free_i32(t);
3324 }
3325 
3326 static void tcg_gen_rotr_mod_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
3327 {
3328     TCGv_i64 t = tcg_temp_new_i64();
3329 
3330     tcg_gen_andi_i64(t, b, 63);
3331     tcg_gen_rotr_i64(d, a, t);
3332     tcg_temp_free_i64(t);
3333 }
3334 
3335 void tcg_gen_gvec_rotrv(unsigned vece, uint32_t dofs, uint32_t aofs,
3336                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
3337 {
3338     static const TCGOpcode vecop_list[] = { INDEX_op_rotrv_vec, 0 };
3339     static const GVecGen3 g[4] = {
3340         { .fniv = tcg_gen_rotrv_mod_vec,
3341           .fno = gen_helper_gvec_rotr8v,
3342           .opt_opc = vecop_list,
3343           .vece = MO_8 },
3344         { .fniv = tcg_gen_rotrv_mod_vec,
3345           .fno = gen_helper_gvec_rotr16v,
3346           .opt_opc = vecop_list,
3347           .vece = MO_16 },
3348         { .fni4 = tcg_gen_rotr_mod_i32,
3349           .fniv = tcg_gen_rotrv_mod_vec,
3350           .fno = gen_helper_gvec_rotr32v,
3351           .opt_opc = vecop_list,
3352           .vece = MO_32 },
3353         { .fni8 = tcg_gen_rotr_mod_i64,
3354           .fniv = tcg_gen_rotrv_mod_vec,
3355           .fno = gen_helper_gvec_rotr64v,
3356           .opt_opc = vecop_list,
3357           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
3358           .vece = MO_64 },
3359     };
3360 
3361     tcg_debug_assert(vece <= MO_64);
3362     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
3363 }
3364 
3365 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
3366 static void expand_cmp_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
3367                            uint32_t oprsz, TCGCond cond)
3368 {
3369     TCGv_i32 t0 = tcg_temp_new_i32();
3370     TCGv_i32 t1 = tcg_temp_new_i32();
3371     uint32_t i;
3372 
3373     for (i = 0; i < oprsz; i += 4) {
3374         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
3375         tcg_gen_ld_i32(t1, cpu_env, bofs + i);
3376         tcg_gen_setcond_i32(cond, t0, t0, t1);
3377         tcg_gen_neg_i32(t0, t0);
3378         tcg_gen_st_i32(t0, cpu_env, dofs + i);
3379     }
3380     tcg_temp_free_i32(t1);
3381     tcg_temp_free_i32(t0);
3382 }
3383 
3384 static void expand_cmp_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
3385                            uint32_t oprsz, TCGCond cond)
3386 {
3387     TCGv_i64 t0 = tcg_temp_new_i64();
3388     TCGv_i64 t1 = tcg_temp_new_i64();
3389     uint32_t i;
3390 
3391     for (i = 0; i < oprsz; i += 8) {
3392         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
3393         tcg_gen_ld_i64(t1, cpu_env, bofs + i);
3394         tcg_gen_setcond_i64(cond, t0, t0, t1);
3395         tcg_gen_neg_i64(t0, t0);
3396         tcg_gen_st_i64(t0, cpu_env, dofs + i);
3397     }
3398     tcg_temp_free_i64(t1);
3399     tcg_temp_free_i64(t0);
3400 }
3401 
3402 static void expand_cmp_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
3403                            uint32_t bofs, uint32_t oprsz, uint32_t tysz,
3404                            TCGType type, TCGCond cond)
3405 {
3406     TCGv_vec t0 = tcg_temp_new_vec(type);
3407     TCGv_vec t1 = tcg_temp_new_vec(type);
3408     uint32_t i;
3409 
3410     for (i = 0; i < oprsz; i += tysz) {
3411         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
3412         tcg_gen_ld_vec(t1, cpu_env, bofs + i);
3413         tcg_gen_cmp_vec(cond, vece, t0, t0, t1);
3414         tcg_gen_st_vec(t0, cpu_env, dofs + i);
3415     }
3416     tcg_temp_free_vec(t1);
3417     tcg_temp_free_vec(t0);
3418 }
3419 
3420 void tcg_gen_gvec_cmp(TCGCond cond, unsigned vece, uint32_t dofs,
3421                       uint32_t aofs, uint32_t bofs,
3422                       uint32_t oprsz, uint32_t maxsz)
3423 {
3424     static const TCGOpcode cmp_list[] = { INDEX_op_cmp_vec, 0 };
3425     static gen_helper_gvec_3 * const eq_fn[4] = {
3426         gen_helper_gvec_eq8, gen_helper_gvec_eq16,
3427         gen_helper_gvec_eq32, gen_helper_gvec_eq64
3428     };
3429     static gen_helper_gvec_3 * const ne_fn[4] = {
3430         gen_helper_gvec_ne8, gen_helper_gvec_ne16,
3431         gen_helper_gvec_ne32, gen_helper_gvec_ne64
3432     };
3433     static gen_helper_gvec_3 * const lt_fn[4] = {
3434         gen_helper_gvec_lt8, gen_helper_gvec_lt16,
3435         gen_helper_gvec_lt32, gen_helper_gvec_lt64
3436     };
3437     static gen_helper_gvec_3 * const le_fn[4] = {
3438         gen_helper_gvec_le8, gen_helper_gvec_le16,
3439         gen_helper_gvec_le32, gen_helper_gvec_le64
3440     };
3441     static gen_helper_gvec_3 * const ltu_fn[4] = {
3442         gen_helper_gvec_ltu8, gen_helper_gvec_ltu16,
3443         gen_helper_gvec_ltu32, gen_helper_gvec_ltu64
3444     };
3445     static gen_helper_gvec_3 * const leu_fn[4] = {
3446         gen_helper_gvec_leu8, gen_helper_gvec_leu16,
3447         gen_helper_gvec_leu32, gen_helper_gvec_leu64
3448     };
3449     static gen_helper_gvec_3 * const * const fns[16] = {
3450         [TCG_COND_EQ] = eq_fn,
3451         [TCG_COND_NE] = ne_fn,
3452         [TCG_COND_LT] = lt_fn,
3453         [TCG_COND_LE] = le_fn,
3454         [TCG_COND_LTU] = ltu_fn,
3455         [TCG_COND_LEU] = leu_fn,
3456     };
3457 
3458     const TCGOpcode *hold_list;
3459     TCGType type;
3460     uint32_t some;
3461 
3462     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
3463     check_overlap_3(dofs, aofs, bofs, maxsz);
3464 
3465     if (cond == TCG_COND_NEVER || cond == TCG_COND_ALWAYS) {
3466         do_dup(MO_8, dofs, oprsz, maxsz,
3467                NULL, NULL, -(cond == TCG_COND_ALWAYS));
3468         return;
3469     }
3470 
3471     /*
3472      * Implement inline with a vector type, if possible.
3473      * Prefer integer when 64-bit host and 64-bit comparison.
3474      */
3475     hold_list = tcg_swap_vecop_list(cmp_list);
3476     type = choose_vector_type(cmp_list, vece, oprsz,
3477                               TCG_TARGET_REG_BITS == 64 && vece == MO_64);
3478     switch (type) {
3479     case TCG_TYPE_V256:
3480         /* Recall that ARM SVE allows vector sizes that are not a
3481          * power of 2, but always a multiple of 16.  The intent is
3482          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
3483          */
3484         some = QEMU_ALIGN_DOWN(oprsz, 32);
3485         expand_cmp_vec(vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256, cond);
3486         if (some == oprsz) {
3487             break;
3488         }
3489         dofs += some;
3490         aofs += some;
3491         bofs += some;
3492         oprsz -= some;
3493         maxsz -= some;
3494         /* fallthru */
3495     case TCG_TYPE_V128:
3496         expand_cmp_vec(vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128, cond);
3497         break;
3498     case TCG_TYPE_V64:
3499         expand_cmp_vec(vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64, cond);
3500         break;
3501 
3502     case 0:
3503         if (vece == MO_64 && check_size_impl(oprsz, 8)) {
3504             expand_cmp_i64(dofs, aofs, bofs, oprsz, cond);
3505         } else if (vece == MO_32 && check_size_impl(oprsz, 4)) {
3506             expand_cmp_i32(dofs, aofs, bofs, oprsz, cond);
3507         } else {
3508             gen_helper_gvec_3 * const *fn = fns[cond];
3509 
3510             if (fn == NULL) {
3511                 uint32_t tmp;
3512                 tmp = aofs, aofs = bofs, bofs = tmp;
3513                 cond = tcg_swap_cond(cond);
3514                 fn = fns[cond];
3515                 assert(fn != NULL);
3516             }
3517             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz, maxsz, 0, fn[vece]);
3518             oprsz = maxsz;
3519         }
3520         break;
3521 
3522     default:
3523         g_assert_not_reached();
3524     }
3525     tcg_swap_vecop_list(hold_list);
3526 
3527     if (oprsz < maxsz) {
3528         expand_clr(dofs + oprsz, maxsz - oprsz);
3529     }
3530 }
3531 
3532 static void tcg_gen_bitsel_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 c)
3533 {
3534     TCGv_i64 t = tcg_temp_new_i64();
3535 
3536     tcg_gen_and_i64(t, b, a);
3537     tcg_gen_andc_i64(d, c, a);
3538     tcg_gen_or_i64(d, d, t);
3539     tcg_temp_free_i64(t);
3540 }
3541 
3542 void tcg_gen_gvec_bitsel(unsigned vece, uint32_t dofs, uint32_t aofs,
3543                          uint32_t bofs, uint32_t cofs,
3544                          uint32_t oprsz, uint32_t maxsz)
3545 {
3546     static const GVecGen4 g = {
3547         .fni8 = tcg_gen_bitsel_i64,
3548         .fniv = tcg_gen_bitsel_vec,
3549         .fno = gen_helper_gvec_bitsel,
3550     };
3551 
3552     tcg_gen_gvec_4(dofs, aofs, bofs, cofs, oprsz, maxsz, &g);
3553 }
3554