xref: /openbmc/qemu/tcg/tcg-op-gvec.c (revision 40d6ee94)
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 "qemu-common.h"
22 #include "tcg.h"
23 #include "tcg-op.h"
24 #include "tcg-op-gvec.h"
25 #include "tcg-gvec-desc.h"
26 
27 #define MAX_UNROLL  4
28 
29 /* Verify vector size and alignment rules.  OFS should be the OR of all
30    of the operand offsets so that we can check them all at once.  */
31 static void check_size_align(uint32_t oprsz, uint32_t maxsz, uint32_t ofs)
32 {
33     uint32_t opr_align = oprsz >= 16 ? 15 : 7;
34     uint32_t max_align = maxsz >= 16 || oprsz >= 16 ? 15 : 7;
35     tcg_debug_assert(oprsz > 0);
36     tcg_debug_assert(oprsz <= maxsz);
37     tcg_debug_assert((oprsz & opr_align) == 0);
38     tcg_debug_assert((maxsz & max_align) == 0);
39     tcg_debug_assert((ofs & max_align) == 0);
40 }
41 
42 /* Verify vector overlap rules for two operands.  */
43 static void check_overlap_2(uint32_t d, uint32_t a, uint32_t s)
44 {
45     tcg_debug_assert(d == a || d + s <= a || a + s <= d);
46 }
47 
48 /* Verify vector overlap rules for three operands.  */
49 static void check_overlap_3(uint32_t d, uint32_t a, uint32_t b, uint32_t s)
50 {
51     check_overlap_2(d, a, s);
52     check_overlap_2(d, b, s);
53     check_overlap_2(a, b, s);
54 }
55 
56 /* Verify vector overlap rules for four operands.  */
57 static void check_overlap_4(uint32_t d, uint32_t a, uint32_t b,
58                             uint32_t c, uint32_t s)
59 {
60     check_overlap_2(d, a, s);
61     check_overlap_2(d, b, s);
62     check_overlap_2(d, c, s);
63     check_overlap_2(a, b, s);
64     check_overlap_2(a, c, s);
65     check_overlap_2(b, c, s);
66 }
67 
68 /* Create a descriptor from components.  */
69 uint32_t simd_desc(uint32_t oprsz, uint32_t maxsz, int32_t data)
70 {
71     uint32_t desc = 0;
72 
73     assert(oprsz % 8 == 0 && oprsz <= (8 << SIMD_OPRSZ_BITS));
74     assert(maxsz % 8 == 0 && maxsz <= (8 << SIMD_MAXSZ_BITS));
75     assert(data == sextract32(data, 0, SIMD_DATA_BITS));
76 
77     oprsz = (oprsz / 8) - 1;
78     maxsz = (maxsz / 8) - 1;
79     desc = deposit32(desc, SIMD_OPRSZ_SHIFT, SIMD_OPRSZ_BITS, oprsz);
80     desc = deposit32(desc, SIMD_MAXSZ_SHIFT, SIMD_MAXSZ_BITS, maxsz);
81     desc = deposit32(desc, SIMD_DATA_SHIFT, SIMD_DATA_BITS, data);
82 
83     return desc;
84 }
85 
86 /* Generate a call to a gvec-style helper with two vector operands.  */
87 void tcg_gen_gvec_2_ool(uint32_t dofs, uint32_t aofs,
88                         uint32_t oprsz, uint32_t maxsz, int32_t data,
89                         gen_helper_gvec_2 *fn)
90 {
91     TCGv_ptr a0, a1;
92     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
93 
94     a0 = tcg_temp_new_ptr();
95     a1 = tcg_temp_new_ptr();
96 
97     tcg_gen_addi_ptr(a0, cpu_env, dofs);
98     tcg_gen_addi_ptr(a1, cpu_env, aofs);
99 
100     fn(a0, a1, desc);
101 
102     tcg_temp_free_ptr(a0);
103     tcg_temp_free_ptr(a1);
104     tcg_temp_free_i32(desc);
105 }
106 
107 /* Generate a call to a gvec-style helper with two vector operands
108    and one scalar operand.  */
109 void tcg_gen_gvec_2i_ool(uint32_t dofs, uint32_t aofs, TCGv_i64 c,
110                          uint32_t oprsz, uint32_t maxsz, int32_t data,
111                          gen_helper_gvec_2i *fn)
112 {
113     TCGv_ptr a0, a1;
114     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
115 
116     a0 = tcg_temp_new_ptr();
117     a1 = tcg_temp_new_ptr();
118 
119     tcg_gen_addi_ptr(a0, cpu_env, dofs);
120     tcg_gen_addi_ptr(a1, cpu_env, aofs);
121 
122     fn(a0, a1, c, desc);
123 
124     tcg_temp_free_ptr(a0);
125     tcg_temp_free_ptr(a1);
126     tcg_temp_free_i32(desc);
127 }
128 
129 /* Generate a call to a gvec-style helper with three vector operands.  */
130 void tcg_gen_gvec_3_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
131                         uint32_t oprsz, uint32_t maxsz, int32_t data,
132                         gen_helper_gvec_3 *fn)
133 {
134     TCGv_ptr a0, a1, a2;
135     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
136 
137     a0 = tcg_temp_new_ptr();
138     a1 = tcg_temp_new_ptr();
139     a2 = tcg_temp_new_ptr();
140 
141     tcg_gen_addi_ptr(a0, cpu_env, dofs);
142     tcg_gen_addi_ptr(a1, cpu_env, aofs);
143     tcg_gen_addi_ptr(a2, cpu_env, bofs);
144 
145     fn(a0, a1, a2, desc);
146 
147     tcg_temp_free_ptr(a0);
148     tcg_temp_free_ptr(a1);
149     tcg_temp_free_ptr(a2);
150     tcg_temp_free_i32(desc);
151 }
152 
153 /* Generate a call to a gvec-style helper with four vector operands.  */
154 void tcg_gen_gvec_4_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
155                         uint32_t cofs, uint32_t oprsz, uint32_t maxsz,
156                         int32_t data, gen_helper_gvec_4 *fn)
157 {
158     TCGv_ptr a0, a1, a2, a3;
159     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
160 
161     a0 = tcg_temp_new_ptr();
162     a1 = tcg_temp_new_ptr();
163     a2 = tcg_temp_new_ptr();
164     a3 = tcg_temp_new_ptr();
165 
166     tcg_gen_addi_ptr(a0, cpu_env, dofs);
167     tcg_gen_addi_ptr(a1, cpu_env, aofs);
168     tcg_gen_addi_ptr(a2, cpu_env, bofs);
169     tcg_gen_addi_ptr(a3, cpu_env, cofs);
170 
171     fn(a0, a1, a2, a3, desc);
172 
173     tcg_temp_free_ptr(a0);
174     tcg_temp_free_ptr(a1);
175     tcg_temp_free_ptr(a2);
176     tcg_temp_free_ptr(a3);
177     tcg_temp_free_i32(desc);
178 }
179 
180 /* Generate a call to a gvec-style helper with five vector operands.  */
181 void tcg_gen_gvec_5_ool(uint32_t dofs, uint32_t aofs, uint32_t bofs,
182                         uint32_t cofs, uint32_t xofs, uint32_t oprsz,
183                         uint32_t maxsz, int32_t data, gen_helper_gvec_5 *fn)
184 {
185     TCGv_ptr a0, a1, a2, a3, a4;
186     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
187 
188     a0 = tcg_temp_new_ptr();
189     a1 = tcg_temp_new_ptr();
190     a2 = tcg_temp_new_ptr();
191     a3 = tcg_temp_new_ptr();
192     a4 = tcg_temp_new_ptr();
193 
194     tcg_gen_addi_ptr(a0, cpu_env, dofs);
195     tcg_gen_addi_ptr(a1, cpu_env, aofs);
196     tcg_gen_addi_ptr(a2, cpu_env, bofs);
197     tcg_gen_addi_ptr(a3, cpu_env, cofs);
198     tcg_gen_addi_ptr(a4, cpu_env, xofs);
199 
200     fn(a0, a1, a2, a3, a4, desc);
201 
202     tcg_temp_free_ptr(a0);
203     tcg_temp_free_ptr(a1);
204     tcg_temp_free_ptr(a2);
205     tcg_temp_free_ptr(a3);
206     tcg_temp_free_ptr(a4);
207     tcg_temp_free_i32(desc);
208 }
209 
210 /* Generate a call to a gvec-style helper with three vector operands
211    and an extra pointer operand.  */
212 void tcg_gen_gvec_2_ptr(uint32_t dofs, uint32_t aofs,
213                         TCGv_ptr ptr, uint32_t oprsz, uint32_t maxsz,
214                         int32_t data, gen_helper_gvec_2_ptr *fn)
215 {
216     TCGv_ptr a0, a1;
217     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
218 
219     a0 = tcg_temp_new_ptr();
220     a1 = tcg_temp_new_ptr();
221 
222     tcg_gen_addi_ptr(a0, cpu_env, dofs);
223     tcg_gen_addi_ptr(a1, cpu_env, aofs);
224 
225     fn(a0, a1, ptr, desc);
226 
227     tcg_temp_free_ptr(a0);
228     tcg_temp_free_ptr(a1);
229     tcg_temp_free_i32(desc);
230 }
231 
232 /* Generate a call to a gvec-style helper with three vector operands
233    and an extra pointer operand.  */
234 void tcg_gen_gvec_3_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
235                         TCGv_ptr ptr, uint32_t oprsz, uint32_t maxsz,
236                         int32_t data, gen_helper_gvec_3_ptr *fn)
237 {
238     TCGv_ptr a0, a1, a2;
239     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
240 
241     a0 = tcg_temp_new_ptr();
242     a1 = tcg_temp_new_ptr();
243     a2 = tcg_temp_new_ptr();
244 
245     tcg_gen_addi_ptr(a0, cpu_env, dofs);
246     tcg_gen_addi_ptr(a1, cpu_env, aofs);
247     tcg_gen_addi_ptr(a2, cpu_env, bofs);
248 
249     fn(a0, a1, a2, ptr, desc);
250 
251     tcg_temp_free_ptr(a0);
252     tcg_temp_free_ptr(a1);
253     tcg_temp_free_ptr(a2);
254     tcg_temp_free_i32(desc);
255 }
256 
257 /* Generate a call to a gvec-style helper with four vector operands
258    and an extra pointer operand.  */
259 void tcg_gen_gvec_4_ptr(uint32_t dofs, uint32_t aofs, uint32_t bofs,
260                         uint32_t cofs, TCGv_ptr ptr, uint32_t oprsz,
261                         uint32_t maxsz, int32_t data,
262                         gen_helper_gvec_4_ptr *fn)
263 {
264     TCGv_ptr a0, a1, a2, a3;
265     TCGv_i32 desc = tcg_const_i32(simd_desc(oprsz, maxsz, data));
266 
267     a0 = tcg_temp_new_ptr();
268     a1 = tcg_temp_new_ptr();
269     a2 = tcg_temp_new_ptr();
270     a3 = tcg_temp_new_ptr();
271 
272     tcg_gen_addi_ptr(a0, cpu_env, dofs);
273     tcg_gen_addi_ptr(a1, cpu_env, aofs);
274     tcg_gen_addi_ptr(a2, cpu_env, bofs);
275     tcg_gen_addi_ptr(a3, cpu_env, cofs);
276 
277     fn(a0, a1, a2, a3, ptr, desc);
278 
279     tcg_temp_free_ptr(a0);
280     tcg_temp_free_ptr(a1);
281     tcg_temp_free_ptr(a2);
282     tcg_temp_free_ptr(a3);
283     tcg_temp_free_i32(desc);
284 }
285 
286 /* Return true if we want to implement something of OPRSZ bytes
287    in units of LNSZ.  This limits the expansion of inline code.  */
288 static inline bool check_size_impl(uint32_t oprsz, uint32_t lnsz)
289 {
290     if (oprsz % lnsz == 0) {
291         uint32_t lnct = oprsz / lnsz;
292         return lnct >= 1 && lnct <= MAX_UNROLL;
293     }
294     return false;
295 }
296 
297 static void expand_clr(uint32_t dofs, uint32_t maxsz);
298 
299 /* Duplicate C as per VECE.  */
300 uint64_t (dup_const)(unsigned vece, uint64_t c)
301 {
302     switch (vece) {
303     case MO_8:
304         return 0x0101010101010101ull * (uint8_t)c;
305     case MO_16:
306         return 0x0001000100010001ull * (uint16_t)c;
307     case MO_32:
308         return 0x0000000100000001ull * (uint32_t)c;
309     case MO_64:
310         return c;
311     default:
312         g_assert_not_reached();
313     }
314 }
315 
316 /* Duplicate IN into OUT as per VECE.  */
317 static void gen_dup_i32(unsigned vece, TCGv_i32 out, TCGv_i32 in)
318 {
319     switch (vece) {
320     case MO_8:
321         tcg_gen_ext8u_i32(out, in);
322         tcg_gen_muli_i32(out, out, 0x01010101);
323         break;
324     case MO_16:
325         tcg_gen_deposit_i32(out, in, in, 16, 16);
326         break;
327     case MO_32:
328         tcg_gen_mov_i32(out, in);
329         break;
330     default:
331         g_assert_not_reached();
332     }
333 }
334 
335 static void gen_dup_i64(unsigned vece, TCGv_i64 out, TCGv_i64 in)
336 {
337     switch (vece) {
338     case MO_8:
339         tcg_gen_ext8u_i64(out, in);
340         tcg_gen_muli_i64(out, out, 0x0101010101010101ull);
341         break;
342     case MO_16:
343         tcg_gen_ext16u_i64(out, in);
344         tcg_gen_muli_i64(out, out, 0x0001000100010001ull);
345         break;
346     case MO_32:
347         tcg_gen_deposit_i64(out, in, in, 32, 32);
348         break;
349     case MO_64:
350         tcg_gen_mov_i64(out, in);
351         break;
352     default:
353         g_assert_not_reached();
354     }
355 }
356 
357 /* Select a supported vector type for implementing an operation on SIZE
358  * bytes.  If OP is 0, assume that the real operation to be performed is
359  * required by all backends.  Otherwise, make sure than OP can be performed
360  * on elements of size VECE in the selected type.  Do not select V64 if
361  * PREFER_I64 is true.  Return 0 if no vector type is selected.
362  */
363 static TCGType choose_vector_type(TCGOpcode op, unsigned vece, uint32_t size,
364                                   bool prefer_i64)
365 {
366     if (TCG_TARGET_HAS_v256 && check_size_impl(size, 32)) {
367         if (op == 0) {
368             return TCG_TYPE_V256;
369         }
370         /* Recall that ARM SVE allows vector sizes that are not a
371          * power of 2, but always a multiple of 16.  The intent is
372          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
373          * It is hard to imagine a case in which v256 is supported
374          * but v128 is not, but check anyway.
375          */
376         if (tcg_can_emit_vec_op(op, TCG_TYPE_V256, vece)
377             && (size % 32 == 0
378                 || tcg_can_emit_vec_op(op, TCG_TYPE_V128, vece))) {
379             return TCG_TYPE_V256;
380         }
381     }
382     if (TCG_TARGET_HAS_v128 && check_size_impl(size, 16)
383         && (op == 0 || tcg_can_emit_vec_op(op, TCG_TYPE_V128, vece))) {
384         return TCG_TYPE_V128;
385     }
386     if (TCG_TARGET_HAS_v64 && !prefer_i64 && check_size_impl(size, 8)
387         && (op == 0 || tcg_can_emit_vec_op(op, TCG_TYPE_V64, vece))) {
388         return TCG_TYPE_V64;
389     }
390     return 0;
391 }
392 
393 /* Set OPRSZ bytes at DOFS to replications of IN_32, IN_64 or IN_C.
394  * Only one of IN_32 or IN_64 may be set;
395  * IN_C is used if IN_32 and IN_64 are unset.
396  */
397 static void do_dup(unsigned vece, uint32_t dofs, uint32_t oprsz,
398                    uint32_t maxsz, TCGv_i32 in_32, TCGv_i64 in_64,
399                    uint64_t in_c)
400 {
401     TCGType type;
402     TCGv_i64 t_64;
403     TCGv_i32 t_32, t_desc;
404     TCGv_ptr t_ptr;
405     uint32_t i;
406 
407     assert(vece <= (in_32 ? MO_32 : MO_64));
408     assert(in_32 == NULL || in_64 == NULL);
409 
410     /* If we're storing 0, expand oprsz to maxsz.  */
411     if (in_32 == NULL && in_64 == NULL) {
412         in_c = dup_const(vece, in_c);
413         if (in_c == 0) {
414             oprsz = maxsz;
415         }
416     }
417 
418     /* Implement inline with a vector type, if possible.
419      * Prefer integer when 64-bit host and no variable dup.
420      */
421     type = choose_vector_type(0, vece, oprsz,
422                               (TCG_TARGET_REG_BITS == 64 && in_32 == NULL
423                                && (in_64 == NULL || vece == MO_64)));
424     if (type != 0) {
425         TCGv_vec t_vec = tcg_temp_new_vec(type);
426 
427         if (in_32) {
428             tcg_gen_dup_i32_vec(vece, t_vec, in_32);
429         } else if (in_64) {
430             tcg_gen_dup_i64_vec(vece, t_vec, in_64);
431         } else {
432             switch (vece) {
433             case MO_8:
434                 tcg_gen_dup8i_vec(t_vec, in_c);
435                 break;
436             case MO_16:
437                 tcg_gen_dup16i_vec(t_vec, in_c);
438                 break;
439             case MO_32:
440                 tcg_gen_dup32i_vec(t_vec, in_c);
441                 break;
442             default:
443                 tcg_gen_dup64i_vec(t_vec, in_c);
444                 break;
445             }
446         }
447 
448         i = 0;
449         switch (type) {
450         case TCG_TYPE_V256:
451             /* Recall that ARM SVE allows vector sizes that are not a
452              * power of 2, but always a multiple of 16.  The intent is
453              * that e.g. size == 80 would be expanded with 2x32 + 1x16.
454              */
455             for (; i + 32 <= oprsz; i += 32) {
456                 tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V256);
457             }
458             /* fallthru */
459         case TCG_TYPE_V128:
460             for (; i + 16 <= oprsz; i += 16) {
461                 tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V128);
462             }
463             break;
464         case TCG_TYPE_V64:
465             for (; i < oprsz; i += 8) {
466                 tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V64);
467             }
468             break;
469         default:
470             g_assert_not_reached();
471         }
472 
473         tcg_temp_free_vec(t_vec);
474         goto done;
475     }
476 
477     /* Otherwise, inline with an integer type, unless "large".  */
478     if (check_size_impl(oprsz, TCG_TARGET_REG_BITS / 8)) {
479         t_64 = NULL;
480         t_32 = NULL;
481 
482         if (in_32) {
483             /* We are given a 32-bit variable input.  For a 64-bit host,
484                use a 64-bit operation unless the 32-bit operation would
485                be simple enough.  */
486             if (TCG_TARGET_REG_BITS == 64
487                 && (vece != MO_32 || !check_size_impl(oprsz, 4))) {
488                 t_64 = tcg_temp_new_i64();
489                 tcg_gen_extu_i32_i64(t_64, in_32);
490                 gen_dup_i64(vece, t_64, t_64);
491             } else {
492                 t_32 = tcg_temp_new_i32();
493                 gen_dup_i32(vece, t_32, in_32);
494             }
495         } else if (in_64) {
496             /* We are given a 64-bit variable input.  */
497             t_64 = tcg_temp_new_i64();
498             gen_dup_i64(vece, t_64, in_64);
499         } else {
500             /* We are given a constant input.  */
501             /* For 64-bit hosts, use 64-bit constants for "simple" constants
502                or when we'd need too many 32-bit stores, or when a 64-bit
503                constant is really required.  */
504             if (vece == MO_64
505                 || (TCG_TARGET_REG_BITS == 64
506                     && (in_c == 0 || in_c == -1
507                         || !check_size_impl(oprsz, 4)))) {
508                 t_64 = tcg_const_i64(in_c);
509             } else {
510                 t_32 = tcg_const_i32(in_c);
511             }
512         }
513 
514         /* Implement inline if we picked an implementation size above.  */
515         if (t_32) {
516             for (i = 0; i < oprsz; i += 4) {
517                 tcg_gen_st_i32(t_32, cpu_env, dofs + i);
518             }
519             tcg_temp_free_i32(t_32);
520             goto done;
521         }
522         if (t_64) {
523             for (i = 0; i < oprsz; i += 8) {
524                 tcg_gen_st_i64(t_64, cpu_env, dofs + i);
525             }
526             tcg_temp_free_i64(t_64);
527             goto done;
528         }
529     }
530 
531     /* Otherwise implement out of line.  */
532     t_ptr = tcg_temp_new_ptr();
533     tcg_gen_addi_ptr(t_ptr, cpu_env, dofs);
534     t_desc = tcg_const_i32(simd_desc(oprsz, maxsz, 0));
535 
536     if (vece == MO_64) {
537         if (in_64) {
538             gen_helper_gvec_dup64(t_ptr, t_desc, in_64);
539         } else {
540             t_64 = tcg_const_i64(in_c);
541             gen_helper_gvec_dup64(t_ptr, t_desc, t_64);
542             tcg_temp_free_i64(t_64);
543         }
544     } else {
545         typedef void dup_fn(TCGv_ptr, TCGv_i32, TCGv_i32);
546         static dup_fn * const fns[3] = {
547             gen_helper_gvec_dup8,
548             gen_helper_gvec_dup16,
549             gen_helper_gvec_dup32
550         };
551 
552         if (in_32) {
553             fns[vece](t_ptr, t_desc, in_32);
554         } else {
555             t_32 = tcg_temp_new_i32();
556             if (in_64) {
557                 tcg_gen_extrl_i64_i32(t_32, in_64);
558             } else if (vece == MO_8) {
559                 tcg_gen_movi_i32(t_32, in_c & 0xff);
560             } else if (vece == MO_16) {
561                 tcg_gen_movi_i32(t_32, in_c & 0xffff);
562             } else {
563                 tcg_gen_movi_i32(t_32, in_c);
564             }
565             fns[vece](t_ptr, t_desc, t_32);
566             tcg_temp_free_i32(t_32);
567         }
568     }
569 
570     tcg_temp_free_ptr(t_ptr);
571     tcg_temp_free_i32(t_desc);
572     return;
573 
574  done:
575     if (oprsz < maxsz) {
576         expand_clr(dofs + oprsz, maxsz - oprsz);
577     }
578 }
579 
580 /* Likewise, but with zero.  */
581 static void expand_clr(uint32_t dofs, uint32_t maxsz)
582 {
583     do_dup(MO_8, dofs, maxsz, maxsz, NULL, NULL, 0);
584 }
585 
586 /* Expand OPSZ bytes worth of two-operand operations using i32 elements.  */
587 static void expand_2_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
588                          void (*fni)(TCGv_i32, TCGv_i32))
589 {
590     TCGv_i32 t0 = tcg_temp_new_i32();
591     uint32_t i;
592 
593     for (i = 0; i < oprsz; i += 4) {
594         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
595         fni(t0, t0);
596         tcg_gen_st_i32(t0, cpu_env, dofs + i);
597     }
598     tcg_temp_free_i32(t0);
599 }
600 
601 static void expand_2i_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
602                           int32_t c, bool load_dest,
603                           void (*fni)(TCGv_i32, TCGv_i32, int32_t))
604 {
605     TCGv_i32 t0 = tcg_temp_new_i32();
606     TCGv_i32 t1 = tcg_temp_new_i32();
607     uint32_t i;
608 
609     for (i = 0; i < oprsz; i += 4) {
610         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
611         if (load_dest) {
612             tcg_gen_ld_i32(t1, cpu_env, dofs + i);
613         }
614         fni(t1, t0, c);
615         tcg_gen_st_i32(t1, cpu_env, dofs + i);
616     }
617     tcg_temp_free_i32(t0);
618     tcg_temp_free_i32(t1);
619 }
620 
621 static void expand_2s_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
622                           TCGv_i32 c, bool scalar_first,
623                           void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32))
624 {
625     TCGv_i32 t0 = tcg_temp_new_i32();
626     TCGv_i32 t1 = tcg_temp_new_i32();
627     uint32_t i;
628 
629     for (i = 0; i < oprsz; i += 4) {
630         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
631         if (scalar_first) {
632             fni(t1, c, t0);
633         } else {
634             fni(t1, t0, c);
635         }
636         tcg_gen_st_i32(t1, cpu_env, dofs + i);
637     }
638     tcg_temp_free_i32(t0);
639     tcg_temp_free_i32(t1);
640 }
641 
642 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
643 static void expand_3_i32(uint32_t dofs, uint32_t aofs,
644                          uint32_t bofs, uint32_t oprsz, bool load_dest,
645                          void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32))
646 {
647     TCGv_i32 t0 = tcg_temp_new_i32();
648     TCGv_i32 t1 = tcg_temp_new_i32();
649     TCGv_i32 t2 = tcg_temp_new_i32();
650     uint32_t i;
651 
652     for (i = 0; i < oprsz; i += 4) {
653         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
654         tcg_gen_ld_i32(t1, cpu_env, bofs + i);
655         if (load_dest) {
656             tcg_gen_ld_i32(t2, cpu_env, dofs + i);
657         }
658         fni(t2, t0, t1);
659         tcg_gen_st_i32(t2, cpu_env, dofs + i);
660     }
661     tcg_temp_free_i32(t2);
662     tcg_temp_free_i32(t1);
663     tcg_temp_free_i32(t0);
664 }
665 
666 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
667 static void expand_4_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
668                          uint32_t cofs, uint32_t oprsz, bool write_aofs,
669                          void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32))
670 {
671     TCGv_i32 t0 = tcg_temp_new_i32();
672     TCGv_i32 t1 = tcg_temp_new_i32();
673     TCGv_i32 t2 = tcg_temp_new_i32();
674     TCGv_i32 t3 = tcg_temp_new_i32();
675     uint32_t i;
676 
677     for (i = 0; i < oprsz; i += 4) {
678         tcg_gen_ld_i32(t1, cpu_env, aofs + i);
679         tcg_gen_ld_i32(t2, cpu_env, bofs + i);
680         tcg_gen_ld_i32(t3, cpu_env, cofs + i);
681         fni(t0, t1, t2, t3);
682         tcg_gen_st_i32(t0, cpu_env, dofs + i);
683         if (write_aofs) {
684             tcg_gen_st_i32(t1, cpu_env, aofs + i);
685         }
686     }
687     tcg_temp_free_i32(t3);
688     tcg_temp_free_i32(t2);
689     tcg_temp_free_i32(t1);
690     tcg_temp_free_i32(t0);
691 }
692 
693 /* Expand OPSZ bytes worth of two-operand operations using i64 elements.  */
694 static void expand_2_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
695                          void (*fni)(TCGv_i64, TCGv_i64))
696 {
697     TCGv_i64 t0 = tcg_temp_new_i64();
698     uint32_t i;
699 
700     for (i = 0; i < oprsz; i += 8) {
701         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
702         fni(t0, t0);
703         tcg_gen_st_i64(t0, cpu_env, dofs + i);
704     }
705     tcg_temp_free_i64(t0);
706 }
707 
708 static void expand_2i_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
709                           int64_t c, bool load_dest,
710                           void (*fni)(TCGv_i64, TCGv_i64, int64_t))
711 {
712     TCGv_i64 t0 = tcg_temp_new_i64();
713     TCGv_i64 t1 = tcg_temp_new_i64();
714     uint32_t i;
715 
716     for (i = 0; i < oprsz; i += 8) {
717         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
718         if (load_dest) {
719             tcg_gen_ld_i64(t1, cpu_env, dofs + i);
720         }
721         fni(t1, t0, c);
722         tcg_gen_st_i64(t1, cpu_env, dofs + i);
723     }
724     tcg_temp_free_i64(t0);
725     tcg_temp_free_i64(t1);
726 }
727 
728 static void expand_2s_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
729                           TCGv_i64 c, bool scalar_first,
730                           void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64))
731 {
732     TCGv_i64 t0 = tcg_temp_new_i64();
733     TCGv_i64 t1 = tcg_temp_new_i64();
734     uint32_t i;
735 
736     for (i = 0; i < oprsz; i += 8) {
737         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
738         if (scalar_first) {
739             fni(t1, c, t0);
740         } else {
741             fni(t1, t0, c);
742         }
743         tcg_gen_st_i64(t1, cpu_env, dofs + i);
744     }
745     tcg_temp_free_i64(t0);
746     tcg_temp_free_i64(t1);
747 }
748 
749 /* Expand OPSZ bytes worth of three-operand operations using i64 elements.  */
750 static void expand_3_i64(uint32_t dofs, uint32_t aofs,
751                          uint32_t bofs, uint32_t oprsz, bool load_dest,
752                          void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64))
753 {
754     TCGv_i64 t0 = tcg_temp_new_i64();
755     TCGv_i64 t1 = tcg_temp_new_i64();
756     TCGv_i64 t2 = tcg_temp_new_i64();
757     uint32_t i;
758 
759     for (i = 0; i < oprsz; i += 8) {
760         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
761         tcg_gen_ld_i64(t1, cpu_env, bofs + i);
762         if (load_dest) {
763             tcg_gen_ld_i64(t2, cpu_env, dofs + i);
764         }
765         fni(t2, t0, t1);
766         tcg_gen_st_i64(t2, cpu_env, dofs + i);
767     }
768     tcg_temp_free_i64(t2);
769     tcg_temp_free_i64(t1);
770     tcg_temp_free_i64(t0);
771 }
772 
773 /* Expand OPSZ bytes worth of three-operand operations using i64 elements.  */
774 static void expand_4_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
775                          uint32_t cofs, uint32_t oprsz, bool write_aofs,
776                          void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64, TCGv_i64))
777 {
778     TCGv_i64 t0 = tcg_temp_new_i64();
779     TCGv_i64 t1 = tcg_temp_new_i64();
780     TCGv_i64 t2 = tcg_temp_new_i64();
781     TCGv_i64 t3 = tcg_temp_new_i64();
782     uint32_t i;
783 
784     for (i = 0; i < oprsz; i += 8) {
785         tcg_gen_ld_i64(t1, cpu_env, aofs + i);
786         tcg_gen_ld_i64(t2, cpu_env, bofs + i);
787         tcg_gen_ld_i64(t3, cpu_env, cofs + i);
788         fni(t0, t1, t2, t3);
789         tcg_gen_st_i64(t0, cpu_env, dofs + i);
790         if (write_aofs) {
791             tcg_gen_st_i64(t1, cpu_env, aofs + i);
792         }
793     }
794     tcg_temp_free_i64(t3);
795     tcg_temp_free_i64(t2);
796     tcg_temp_free_i64(t1);
797     tcg_temp_free_i64(t0);
798 }
799 
800 /* Expand OPSZ bytes worth of two-operand operations using host vectors.  */
801 static void expand_2_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
802                          uint32_t oprsz, uint32_t tysz, TCGType type,
803                          void (*fni)(unsigned, TCGv_vec, TCGv_vec))
804 {
805     TCGv_vec t0 = tcg_temp_new_vec(type);
806     uint32_t i;
807 
808     for (i = 0; i < oprsz; i += tysz) {
809         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
810         fni(vece, t0, t0);
811         tcg_gen_st_vec(t0, cpu_env, dofs + i);
812     }
813     tcg_temp_free_vec(t0);
814 }
815 
816 /* Expand OPSZ bytes worth of two-vector operands and an immediate operand
817    using host vectors.  */
818 static void expand_2i_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
819                           uint32_t oprsz, uint32_t tysz, TCGType type,
820                           int64_t c, bool load_dest,
821                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, int64_t))
822 {
823     TCGv_vec t0 = tcg_temp_new_vec(type);
824     TCGv_vec t1 = tcg_temp_new_vec(type);
825     uint32_t i;
826 
827     for (i = 0; i < oprsz; i += tysz) {
828         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
829         if (load_dest) {
830             tcg_gen_ld_vec(t1, cpu_env, dofs + i);
831         }
832         fni(vece, t1, t0, c);
833         tcg_gen_st_vec(t1, cpu_env, dofs + i);
834     }
835     tcg_temp_free_vec(t0);
836     tcg_temp_free_vec(t1);
837 }
838 
839 static void expand_2s_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
840                           uint32_t oprsz, uint32_t tysz, TCGType type,
841                           TCGv_vec c, bool scalar_first,
842                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec))
843 {
844     TCGv_vec t0 = tcg_temp_new_vec(type);
845     TCGv_vec t1 = tcg_temp_new_vec(type);
846     uint32_t i;
847 
848     for (i = 0; i < oprsz; i += tysz) {
849         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
850         if (scalar_first) {
851             fni(vece, t1, c, t0);
852         } else {
853             fni(vece, t1, t0, c);
854         }
855         tcg_gen_st_vec(t1, cpu_env, dofs + i);
856     }
857     tcg_temp_free_vec(t0);
858     tcg_temp_free_vec(t1);
859 }
860 
861 /* Expand OPSZ bytes worth of three-operand operations using host vectors.  */
862 static void expand_3_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
863                          uint32_t bofs, uint32_t oprsz,
864                          uint32_t tysz, TCGType type, bool load_dest,
865                          void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec))
866 {
867     TCGv_vec t0 = tcg_temp_new_vec(type);
868     TCGv_vec t1 = tcg_temp_new_vec(type);
869     TCGv_vec t2 = tcg_temp_new_vec(type);
870     uint32_t i;
871 
872     for (i = 0; i < oprsz; i += tysz) {
873         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
874         tcg_gen_ld_vec(t1, cpu_env, bofs + i);
875         if (load_dest) {
876             tcg_gen_ld_vec(t2, cpu_env, dofs + i);
877         }
878         fni(vece, t2, t0, t1);
879         tcg_gen_st_vec(t2, cpu_env, dofs + i);
880     }
881     tcg_temp_free_vec(t2);
882     tcg_temp_free_vec(t1);
883     tcg_temp_free_vec(t0);
884 }
885 
886 /* Expand OPSZ bytes worth of four-operand operations using host vectors.  */
887 static void expand_4_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
888                          uint32_t bofs, uint32_t cofs, uint32_t oprsz,
889                          uint32_t tysz, TCGType type, bool write_aofs,
890                          void (*fni)(unsigned, TCGv_vec, TCGv_vec,
891                                      TCGv_vec, TCGv_vec))
892 {
893     TCGv_vec t0 = tcg_temp_new_vec(type);
894     TCGv_vec t1 = tcg_temp_new_vec(type);
895     TCGv_vec t2 = tcg_temp_new_vec(type);
896     TCGv_vec t3 = tcg_temp_new_vec(type);
897     uint32_t i;
898 
899     for (i = 0; i < oprsz; i += tysz) {
900         tcg_gen_ld_vec(t1, cpu_env, aofs + i);
901         tcg_gen_ld_vec(t2, cpu_env, bofs + i);
902         tcg_gen_ld_vec(t3, cpu_env, cofs + i);
903         fni(vece, t0, t1, t2, t3);
904         tcg_gen_st_vec(t0, cpu_env, dofs + i);
905         if (write_aofs) {
906             tcg_gen_st_vec(t1, cpu_env, aofs + i);
907         }
908     }
909     tcg_temp_free_vec(t3);
910     tcg_temp_free_vec(t2);
911     tcg_temp_free_vec(t1);
912     tcg_temp_free_vec(t0);
913 }
914 
915 /* Expand a vector two-operand operation.  */
916 void tcg_gen_gvec_2(uint32_t dofs, uint32_t aofs,
917                     uint32_t oprsz, uint32_t maxsz, const GVecGen2 *g)
918 {
919     TCGType type;
920     uint32_t some;
921 
922     check_size_align(oprsz, maxsz, dofs | aofs);
923     check_overlap_2(dofs, aofs, maxsz);
924 
925     type = 0;
926     if (g->fniv) {
927         type = choose_vector_type(g->opc, g->vece, oprsz, g->prefer_i64);
928     }
929     switch (type) {
930     case TCG_TYPE_V256:
931         /* Recall that ARM SVE allows vector sizes that are not a
932          * power of 2, but always a multiple of 16.  The intent is
933          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
934          */
935         some = QEMU_ALIGN_DOWN(oprsz, 32);
936         expand_2_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256, g->fniv);
937         if (some == oprsz) {
938             break;
939         }
940         dofs += some;
941         aofs += some;
942         oprsz -= some;
943         maxsz -= some;
944         /* fallthru */
945     case TCG_TYPE_V128:
946         expand_2_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128, g->fniv);
947         break;
948     case TCG_TYPE_V64:
949         expand_2_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64, g->fniv);
950         break;
951 
952     case 0:
953         if (g->fni8 && check_size_impl(oprsz, 8)) {
954             expand_2_i64(dofs, aofs, oprsz, g->fni8);
955         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
956             expand_2_i32(dofs, aofs, oprsz, g->fni4);
957         } else {
958             assert(g->fno != NULL);
959             tcg_gen_gvec_2_ool(dofs, aofs, oprsz, maxsz, g->data, g->fno);
960             return;
961         }
962         break;
963 
964     default:
965         g_assert_not_reached();
966     }
967 
968     if (oprsz < maxsz) {
969         expand_clr(dofs + oprsz, maxsz - oprsz);
970     }
971 }
972 
973 /* Expand a vector operation with two vectors and an immediate.  */
974 void tcg_gen_gvec_2i(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
975                      uint32_t maxsz, int64_t c, const GVecGen2i *g)
976 {
977     TCGType type;
978     uint32_t some;
979 
980     check_size_align(oprsz, maxsz, dofs | aofs);
981     check_overlap_2(dofs, aofs, maxsz);
982 
983     type = 0;
984     if (g->fniv) {
985         type = choose_vector_type(g->opc, g->vece, oprsz, g->prefer_i64);
986     }
987     switch (type) {
988     case TCG_TYPE_V256:
989         /* Recall that ARM SVE allows vector sizes that are not a
990          * power of 2, but always a multiple of 16.  The intent is
991          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
992          */
993         some = QEMU_ALIGN_DOWN(oprsz, 32);
994         expand_2i_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
995                       c, g->load_dest, g->fniv);
996         if (some == oprsz) {
997             break;
998         }
999         dofs += some;
1000         aofs += some;
1001         oprsz -= some;
1002         maxsz -= some;
1003         /* fallthru */
1004     case TCG_TYPE_V128:
1005         expand_2i_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
1006                       c, g->load_dest, g->fniv);
1007         break;
1008     case TCG_TYPE_V64:
1009         expand_2i_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
1010                       c, g->load_dest, g->fniv);
1011         break;
1012 
1013     case 0:
1014         if (g->fni8 && check_size_impl(oprsz, 8)) {
1015             expand_2i_i64(dofs, aofs, oprsz, c, g->load_dest, g->fni8);
1016         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1017             expand_2i_i32(dofs, aofs, oprsz, c, g->load_dest, g->fni4);
1018         } else {
1019             if (g->fno) {
1020                 tcg_gen_gvec_2_ool(dofs, aofs, oprsz, maxsz, c, g->fno);
1021             } else {
1022                 TCGv_i64 tcg_c = tcg_const_i64(c);
1023                 tcg_gen_gvec_2i_ool(dofs, aofs, tcg_c, oprsz,
1024                                     maxsz, c, g->fnoi);
1025                 tcg_temp_free_i64(tcg_c);
1026             }
1027             return;
1028         }
1029         break;
1030 
1031     default:
1032         g_assert_not_reached();
1033     }
1034 
1035     if (oprsz < maxsz) {
1036         expand_clr(dofs + oprsz, maxsz - oprsz);
1037     }
1038 }
1039 
1040 /* Expand a vector operation with two vectors and a scalar.  */
1041 void tcg_gen_gvec_2s(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
1042                      uint32_t maxsz, TCGv_i64 c, const GVecGen2s *g)
1043 {
1044     TCGType type;
1045 
1046     check_size_align(oprsz, maxsz, dofs | aofs);
1047     check_overlap_2(dofs, aofs, maxsz);
1048 
1049     type = 0;
1050     if (g->fniv) {
1051         type = choose_vector_type(g->opc, g->vece, oprsz, g->prefer_i64);
1052     }
1053     if (type != 0) {
1054         TCGv_vec t_vec = tcg_temp_new_vec(type);
1055         uint32_t some;
1056 
1057         tcg_gen_dup_i64_vec(g->vece, t_vec, c);
1058 
1059         switch (type) {
1060         case TCG_TYPE_V256:
1061             /* Recall that ARM SVE allows vector sizes that are not a
1062              * power of 2, but always a multiple of 16.  The intent is
1063              * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1064              */
1065             some = QEMU_ALIGN_DOWN(oprsz, 32);
1066             expand_2s_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
1067                           t_vec, g->scalar_first, g->fniv);
1068             if (some == oprsz) {
1069                 break;
1070             }
1071             dofs += some;
1072             aofs += some;
1073             oprsz -= some;
1074             maxsz -= some;
1075             /* fallthru */
1076 
1077         case TCG_TYPE_V128:
1078             expand_2s_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
1079                           t_vec, g->scalar_first, g->fniv);
1080             break;
1081 
1082         case TCG_TYPE_V64:
1083             expand_2s_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
1084                           t_vec, g->scalar_first, g->fniv);
1085             break;
1086 
1087         default:
1088             g_assert_not_reached();
1089         }
1090         tcg_temp_free_vec(t_vec);
1091     } else if (g->fni8 && check_size_impl(oprsz, 8)) {
1092         TCGv_i64 t64 = tcg_temp_new_i64();
1093 
1094         gen_dup_i64(g->vece, t64, c);
1095         expand_2s_i64(dofs, aofs, oprsz, t64, g->scalar_first, g->fni8);
1096         tcg_temp_free_i64(t64);
1097     } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1098         TCGv_i32 t32 = tcg_temp_new_i32();
1099 
1100         tcg_gen_extrl_i64_i32(t32, c);
1101         gen_dup_i32(g->vece, t32, t32);
1102         expand_2s_i32(dofs, aofs, oprsz, t32, g->scalar_first, g->fni4);
1103         tcg_temp_free_i32(t32);
1104     } else {
1105         tcg_gen_gvec_2i_ool(dofs, aofs, c, oprsz, maxsz, 0, g->fno);
1106         return;
1107     }
1108 
1109     if (oprsz < maxsz) {
1110         expand_clr(dofs + oprsz, maxsz - oprsz);
1111     }
1112 }
1113 
1114 /* Expand a vector three-operand operation.  */
1115 void tcg_gen_gvec_3(uint32_t dofs, uint32_t aofs, uint32_t bofs,
1116                     uint32_t oprsz, uint32_t maxsz, const GVecGen3 *g)
1117 {
1118     TCGType type;
1119     uint32_t some;
1120 
1121     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
1122     check_overlap_3(dofs, aofs, bofs, maxsz);
1123 
1124     type = 0;
1125     if (g->fniv) {
1126         type = choose_vector_type(g->opc, g->vece, oprsz, g->prefer_i64);
1127     }
1128     switch (type) {
1129     case TCG_TYPE_V256:
1130         /* Recall that ARM SVE allows vector sizes that are not a
1131          * power of 2, but always a multiple of 16.  The intent is
1132          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1133          */
1134         some = QEMU_ALIGN_DOWN(oprsz, 32);
1135         expand_3_vec(g->vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256,
1136                      g->load_dest, g->fniv);
1137         if (some == oprsz) {
1138             break;
1139         }
1140         dofs += some;
1141         aofs += some;
1142         bofs += some;
1143         oprsz -= some;
1144         maxsz -= some;
1145         /* fallthru */
1146     case TCG_TYPE_V128:
1147         expand_3_vec(g->vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128,
1148                      g->load_dest, g->fniv);
1149         break;
1150     case TCG_TYPE_V64:
1151         expand_3_vec(g->vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64,
1152                      g->load_dest, g->fniv);
1153         break;
1154 
1155     case 0:
1156         if (g->fni8 && check_size_impl(oprsz, 8)) {
1157             expand_3_i64(dofs, aofs, bofs, oprsz, g->load_dest, g->fni8);
1158         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1159             expand_3_i32(dofs, aofs, bofs, oprsz, g->load_dest, g->fni4);
1160         } else {
1161             assert(g->fno != NULL);
1162             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz,
1163                                maxsz, g->data, g->fno);
1164             return;
1165         }
1166         break;
1167 
1168     default:
1169         g_assert_not_reached();
1170     }
1171 
1172     if (oprsz < maxsz) {
1173         expand_clr(dofs + oprsz, maxsz - oprsz);
1174     }
1175 }
1176 
1177 /* Expand a vector four-operand operation.  */
1178 void tcg_gen_gvec_4(uint32_t dofs, uint32_t aofs, uint32_t bofs, uint32_t cofs,
1179                     uint32_t oprsz, uint32_t maxsz, const GVecGen4 *g)
1180 {
1181     TCGType type;
1182     uint32_t some;
1183 
1184     check_size_align(oprsz, maxsz, dofs | aofs | bofs | cofs);
1185     check_overlap_4(dofs, aofs, bofs, cofs, maxsz);
1186 
1187     type = 0;
1188     if (g->fniv) {
1189         type = choose_vector_type(g->opc, g->vece, oprsz, g->prefer_i64);
1190     }
1191     switch (type) {
1192     case TCG_TYPE_V256:
1193         /* Recall that ARM SVE allows vector sizes that are not a
1194          * power of 2, but always a multiple of 16.  The intent is
1195          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1196          */
1197         some = QEMU_ALIGN_DOWN(oprsz, 32);
1198         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, some,
1199                      32, TCG_TYPE_V256, g->write_aofs, g->fniv);
1200         if (some == oprsz) {
1201             break;
1202         }
1203         dofs += some;
1204         aofs += some;
1205         bofs += some;
1206         cofs += some;
1207         oprsz -= some;
1208         maxsz -= some;
1209         /* fallthru */
1210     case TCG_TYPE_V128:
1211         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1212                      16, TCG_TYPE_V128, g->write_aofs, g->fniv);
1213         break;
1214     case TCG_TYPE_V64:
1215         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1216                      8, TCG_TYPE_V64, g->write_aofs, g->fniv);
1217         break;
1218 
1219     case 0:
1220         if (g->fni8 && check_size_impl(oprsz, 8)) {
1221             expand_4_i64(dofs, aofs, bofs, cofs, oprsz,
1222                          g->write_aofs, g->fni8);
1223         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1224             expand_4_i32(dofs, aofs, bofs, cofs, oprsz,
1225                          g->write_aofs, g->fni4);
1226         } else {
1227             assert(g->fno != NULL);
1228             tcg_gen_gvec_4_ool(dofs, aofs, bofs, cofs,
1229                                oprsz, maxsz, g->data, g->fno);
1230             return;
1231         }
1232         break;
1233 
1234     default:
1235         g_assert_not_reached();
1236     }
1237 
1238     if (oprsz < maxsz) {
1239         expand_clr(dofs + oprsz, maxsz - oprsz);
1240     }
1241 }
1242 
1243 /*
1244  * Expand specific vector operations.
1245  */
1246 
1247 static void vec_mov2(unsigned vece, TCGv_vec a, TCGv_vec b)
1248 {
1249     tcg_gen_mov_vec(a, b);
1250 }
1251 
1252 void tcg_gen_gvec_mov(unsigned vece, uint32_t dofs, uint32_t aofs,
1253                       uint32_t oprsz, uint32_t maxsz)
1254 {
1255     static const GVecGen2 g = {
1256         .fni8 = tcg_gen_mov_i64,
1257         .fniv = vec_mov2,
1258         .fno = gen_helper_gvec_mov,
1259         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1260     };
1261     if (dofs != aofs) {
1262         tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g);
1263     } else {
1264         check_size_align(oprsz, maxsz, dofs);
1265         if (oprsz < maxsz) {
1266             expand_clr(dofs + oprsz, maxsz - oprsz);
1267         }
1268     }
1269 }
1270 
1271 void tcg_gen_gvec_dup_i32(unsigned vece, uint32_t dofs, uint32_t oprsz,
1272                           uint32_t maxsz, TCGv_i32 in)
1273 {
1274     check_size_align(oprsz, maxsz, dofs);
1275     tcg_debug_assert(vece <= MO_32);
1276     do_dup(vece, dofs, oprsz, maxsz, in, NULL, 0);
1277 }
1278 
1279 void tcg_gen_gvec_dup_i64(unsigned vece, uint32_t dofs, uint32_t oprsz,
1280                           uint32_t maxsz, TCGv_i64 in)
1281 {
1282     check_size_align(oprsz, maxsz, dofs);
1283     tcg_debug_assert(vece <= MO_64);
1284     do_dup(vece, dofs, oprsz, maxsz, NULL, in, 0);
1285 }
1286 
1287 void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t dofs, uint32_t aofs,
1288                           uint32_t oprsz, uint32_t maxsz)
1289 {
1290     if (vece <= MO_32) {
1291         TCGv_i32 in = tcg_temp_new_i32();
1292         switch (vece) {
1293         case MO_8:
1294             tcg_gen_ld8u_i32(in, cpu_env, aofs);
1295             break;
1296         case MO_16:
1297             tcg_gen_ld16u_i32(in, cpu_env, aofs);
1298             break;
1299         case MO_32:
1300             tcg_gen_ld_i32(in, cpu_env, aofs);
1301             break;
1302         }
1303         tcg_gen_gvec_dup_i32(vece, dofs, oprsz, maxsz, in);
1304         tcg_temp_free_i32(in);
1305     } else if (vece == MO_64) {
1306         TCGv_i64 in = tcg_temp_new_i64();
1307         tcg_gen_ld_i64(in, cpu_env, aofs);
1308         tcg_gen_gvec_dup_i64(MO_64, dofs, oprsz, maxsz, in);
1309         tcg_temp_free_i64(in);
1310     } else {
1311         /* 128-bit duplicate.  */
1312         /* ??? Dup to 256-bit vector.  */
1313         int i;
1314 
1315         tcg_debug_assert(vece == 4);
1316         tcg_debug_assert(oprsz >= 16);
1317         if (TCG_TARGET_HAS_v128) {
1318             TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V128);
1319 
1320             tcg_gen_ld_vec(in, cpu_env, aofs);
1321             for (i = 0; i < oprsz; i += 16) {
1322                 tcg_gen_st_vec(in, cpu_env, dofs + i);
1323             }
1324             tcg_temp_free_vec(in);
1325         } else {
1326             TCGv_i64 in0 = tcg_temp_new_i64();
1327             TCGv_i64 in1 = tcg_temp_new_i64();
1328 
1329             tcg_gen_ld_i64(in0, cpu_env, aofs);
1330             tcg_gen_ld_i64(in1, cpu_env, aofs + 8);
1331             for (i = 0; i < oprsz; i += 16) {
1332                 tcg_gen_st_i64(in0, cpu_env, dofs + i);
1333                 tcg_gen_st_i64(in1, cpu_env, dofs + i + 8);
1334             }
1335             tcg_temp_free_i64(in0);
1336             tcg_temp_free_i64(in1);
1337         }
1338     }
1339 }
1340 
1341 void tcg_gen_gvec_dup64i(uint32_t dofs, uint32_t oprsz,
1342                          uint32_t maxsz, uint64_t x)
1343 {
1344     check_size_align(oprsz, maxsz, dofs);
1345     do_dup(MO_64, dofs, oprsz, maxsz, NULL, NULL, x);
1346 }
1347 
1348 void tcg_gen_gvec_dup32i(uint32_t dofs, uint32_t oprsz,
1349                          uint32_t maxsz, uint32_t x)
1350 {
1351     check_size_align(oprsz, maxsz, dofs);
1352     do_dup(MO_32, dofs, oprsz, maxsz, NULL, NULL, x);
1353 }
1354 
1355 void tcg_gen_gvec_dup16i(uint32_t dofs, uint32_t oprsz,
1356                          uint32_t maxsz, uint16_t x)
1357 {
1358     check_size_align(oprsz, maxsz, dofs);
1359     do_dup(MO_16, dofs, oprsz, maxsz, NULL, NULL, x);
1360 }
1361 
1362 void tcg_gen_gvec_dup8i(uint32_t dofs, uint32_t oprsz,
1363                          uint32_t maxsz, uint8_t x)
1364 {
1365     check_size_align(oprsz, maxsz, dofs);
1366     do_dup(MO_8, dofs, oprsz, maxsz, NULL, NULL, x);
1367 }
1368 
1369 void tcg_gen_gvec_not(unsigned vece, uint32_t dofs, uint32_t aofs,
1370                       uint32_t oprsz, uint32_t maxsz)
1371 {
1372     static const GVecGen2 g = {
1373         .fni8 = tcg_gen_not_i64,
1374         .fniv = tcg_gen_not_vec,
1375         .fno = gen_helper_gvec_not,
1376         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1377     };
1378     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g);
1379 }
1380 
1381 /* Perform a vector addition using normal addition and a mask.  The mask
1382    should be the sign bit of each lane.  This 6-operation form is more
1383    efficient than separate additions when there are 4 or more lanes in
1384    the 64-bit operation.  */
1385 static void gen_addv_mask(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 m)
1386 {
1387     TCGv_i64 t1 = tcg_temp_new_i64();
1388     TCGv_i64 t2 = tcg_temp_new_i64();
1389     TCGv_i64 t3 = tcg_temp_new_i64();
1390 
1391     tcg_gen_andc_i64(t1, a, m);
1392     tcg_gen_andc_i64(t2, b, m);
1393     tcg_gen_xor_i64(t3, a, b);
1394     tcg_gen_add_i64(d, t1, t2);
1395     tcg_gen_and_i64(t3, t3, m);
1396     tcg_gen_xor_i64(d, d, t3);
1397 
1398     tcg_temp_free_i64(t1);
1399     tcg_temp_free_i64(t2);
1400     tcg_temp_free_i64(t3);
1401 }
1402 
1403 void tcg_gen_vec_add8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1404 {
1405     TCGv_i64 m = tcg_const_i64(dup_const(MO_8, 0x80));
1406     gen_addv_mask(d, a, b, m);
1407     tcg_temp_free_i64(m);
1408 }
1409 
1410 void tcg_gen_vec_add16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1411 {
1412     TCGv_i64 m = tcg_const_i64(dup_const(MO_16, 0x8000));
1413     gen_addv_mask(d, a, b, m);
1414     tcg_temp_free_i64(m);
1415 }
1416 
1417 void tcg_gen_vec_add32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1418 {
1419     TCGv_i64 t1 = tcg_temp_new_i64();
1420     TCGv_i64 t2 = tcg_temp_new_i64();
1421 
1422     tcg_gen_andi_i64(t1, a, ~0xffffffffull);
1423     tcg_gen_add_i64(t2, a, b);
1424     tcg_gen_add_i64(t1, t1, b);
1425     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
1426 
1427     tcg_temp_free_i64(t1);
1428     tcg_temp_free_i64(t2);
1429 }
1430 
1431 void tcg_gen_gvec_add(unsigned vece, uint32_t dofs, uint32_t aofs,
1432                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1433 {
1434     static const GVecGen3 g[4] = {
1435         { .fni8 = tcg_gen_vec_add8_i64,
1436           .fniv = tcg_gen_add_vec,
1437           .fno = gen_helper_gvec_add8,
1438           .opc = INDEX_op_add_vec,
1439           .vece = MO_8 },
1440         { .fni8 = tcg_gen_vec_add16_i64,
1441           .fniv = tcg_gen_add_vec,
1442           .fno = gen_helper_gvec_add16,
1443           .opc = INDEX_op_add_vec,
1444           .vece = MO_16 },
1445         { .fni4 = tcg_gen_add_i32,
1446           .fniv = tcg_gen_add_vec,
1447           .fno = gen_helper_gvec_add32,
1448           .opc = INDEX_op_add_vec,
1449           .vece = MO_32 },
1450         { .fni8 = tcg_gen_add_i64,
1451           .fniv = tcg_gen_add_vec,
1452           .fno = gen_helper_gvec_add64,
1453           .opc = INDEX_op_add_vec,
1454           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1455           .vece = MO_64 },
1456     };
1457 
1458     tcg_debug_assert(vece <= MO_64);
1459     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1460 }
1461 
1462 void tcg_gen_gvec_adds(unsigned vece, uint32_t dofs, uint32_t aofs,
1463                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
1464 {
1465     static const GVecGen2s g[4] = {
1466         { .fni8 = tcg_gen_vec_add8_i64,
1467           .fniv = tcg_gen_add_vec,
1468           .fno = gen_helper_gvec_adds8,
1469           .opc = INDEX_op_add_vec,
1470           .vece = MO_8 },
1471         { .fni8 = tcg_gen_vec_add16_i64,
1472           .fniv = tcg_gen_add_vec,
1473           .fno = gen_helper_gvec_adds16,
1474           .opc = INDEX_op_add_vec,
1475           .vece = MO_16 },
1476         { .fni4 = tcg_gen_add_i32,
1477           .fniv = tcg_gen_add_vec,
1478           .fno = gen_helper_gvec_adds32,
1479           .opc = INDEX_op_add_vec,
1480           .vece = MO_32 },
1481         { .fni8 = tcg_gen_add_i64,
1482           .fniv = tcg_gen_add_vec,
1483           .fno = gen_helper_gvec_adds64,
1484           .opc = INDEX_op_add_vec,
1485           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1486           .vece = MO_64 },
1487     };
1488 
1489     tcg_debug_assert(vece <= MO_64);
1490     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
1491 }
1492 
1493 void tcg_gen_gvec_addi(unsigned vece, uint32_t dofs, uint32_t aofs,
1494                        int64_t c, uint32_t oprsz, uint32_t maxsz)
1495 {
1496     TCGv_i64 tmp = tcg_const_i64(c);
1497     tcg_gen_gvec_adds(vece, dofs, aofs, tmp, oprsz, maxsz);
1498     tcg_temp_free_i64(tmp);
1499 }
1500 
1501 void tcg_gen_gvec_subs(unsigned vece, uint32_t dofs, uint32_t aofs,
1502                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
1503 {
1504     static const GVecGen2s g[4] = {
1505         { .fni8 = tcg_gen_vec_sub8_i64,
1506           .fniv = tcg_gen_sub_vec,
1507           .fno = gen_helper_gvec_subs8,
1508           .opc = INDEX_op_sub_vec,
1509           .vece = MO_8 },
1510         { .fni8 = tcg_gen_vec_sub16_i64,
1511           .fniv = tcg_gen_sub_vec,
1512           .fno = gen_helper_gvec_subs16,
1513           .opc = INDEX_op_sub_vec,
1514           .vece = MO_16 },
1515         { .fni4 = tcg_gen_sub_i32,
1516           .fniv = tcg_gen_sub_vec,
1517           .fno = gen_helper_gvec_subs32,
1518           .opc = INDEX_op_sub_vec,
1519           .vece = MO_32 },
1520         { .fni8 = tcg_gen_sub_i64,
1521           .fniv = tcg_gen_sub_vec,
1522           .fno = gen_helper_gvec_subs64,
1523           .opc = INDEX_op_sub_vec,
1524           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1525           .vece = MO_64 },
1526     };
1527 
1528     tcg_debug_assert(vece <= MO_64);
1529     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
1530 }
1531 
1532 /* Perform a vector subtraction using normal subtraction and a mask.
1533    Compare gen_addv_mask above.  */
1534 static void gen_subv_mask(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 m)
1535 {
1536     TCGv_i64 t1 = tcg_temp_new_i64();
1537     TCGv_i64 t2 = tcg_temp_new_i64();
1538     TCGv_i64 t3 = tcg_temp_new_i64();
1539 
1540     tcg_gen_or_i64(t1, a, m);
1541     tcg_gen_andc_i64(t2, b, m);
1542     tcg_gen_eqv_i64(t3, a, b);
1543     tcg_gen_sub_i64(d, t1, t2);
1544     tcg_gen_and_i64(t3, t3, m);
1545     tcg_gen_xor_i64(d, d, t3);
1546 
1547     tcg_temp_free_i64(t1);
1548     tcg_temp_free_i64(t2);
1549     tcg_temp_free_i64(t3);
1550 }
1551 
1552 void tcg_gen_vec_sub8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1553 {
1554     TCGv_i64 m = tcg_const_i64(dup_const(MO_8, 0x80));
1555     gen_subv_mask(d, a, b, m);
1556     tcg_temp_free_i64(m);
1557 }
1558 
1559 void tcg_gen_vec_sub16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1560 {
1561     TCGv_i64 m = tcg_const_i64(dup_const(MO_16, 0x8000));
1562     gen_subv_mask(d, a, b, m);
1563     tcg_temp_free_i64(m);
1564 }
1565 
1566 void tcg_gen_vec_sub32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1567 {
1568     TCGv_i64 t1 = tcg_temp_new_i64();
1569     TCGv_i64 t2 = tcg_temp_new_i64();
1570 
1571     tcg_gen_andi_i64(t1, b, ~0xffffffffull);
1572     tcg_gen_sub_i64(t2, a, b);
1573     tcg_gen_sub_i64(t1, a, t1);
1574     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
1575 
1576     tcg_temp_free_i64(t1);
1577     tcg_temp_free_i64(t2);
1578 }
1579 
1580 void tcg_gen_gvec_sub(unsigned vece, uint32_t dofs, uint32_t aofs,
1581                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1582 {
1583     static const GVecGen3 g[4] = {
1584         { .fni8 = tcg_gen_vec_sub8_i64,
1585           .fniv = tcg_gen_sub_vec,
1586           .fno = gen_helper_gvec_sub8,
1587           .opc = INDEX_op_sub_vec,
1588           .vece = MO_8 },
1589         { .fni8 = tcg_gen_vec_sub16_i64,
1590           .fniv = tcg_gen_sub_vec,
1591           .fno = gen_helper_gvec_sub16,
1592           .opc = INDEX_op_sub_vec,
1593           .vece = MO_16 },
1594         { .fni4 = tcg_gen_sub_i32,
1595           .fniv = tcg_gen_sub_vec,
1596           .fno = gen_helper_gvec_sub32,
1597           .opc = INDEX_op_sub_vec,
1598           .vece = MO_32 },
1599         { .fni8 = tcg_gen_sub_i64,
1600           .fniv = tcg_gen_sub_vec,
1601           .fno = gen_helper_gvec_sub64,
1602           .opc = INDEX_op_sub_vec,
1603           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1604           .vece = MO_64 },
1605     };
1606 
1607     tcg_debug_assert(vece <= MO_64);
1608     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1609 }
1610 
1611 void tcg_gen_gvec_mul(unsigned vece, uint32_t dofs, uint32_t aofs,
1612                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1613 {
1614     static const GVecGen3 g[4] = {
1615         { .fniv = tcg_gen_mul_vec,
1616           .fno = gen_helper_gvec_mul8,
1617           .opc = INDEX_op_mul_vec,
1618           .vece = MO_8 },
1619         { .fniv = tcg_gen_mul_vec,
1620           .fno = gen_helper_gvec_mul16,
1621           .opc = INDEX_op_mul_vec,
1622           .vece = MO_16 },
1623         { .fni4 = tcg_gen_mul_i32,
1624           .fniv = tcg_gen_mul_vec,
1625           .fno = gen_helper_gvec_mul32,
1626           .opc = INDEX_op_mul_vec,
1627           .vece = MO_32 },
1628         { .fni8 = tcg_gen_mul_i64,
1629           .fniv = tcg_gen_mul_vec,
1630           .fno = gen_helper_gvec_mul64,
1631           .opc = INDEX_op_mul_vec,
1632           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1633           .vece = MO_64 },
1634     };
1635 
1636     tcg_debug_assert(vece <= MO_64);
1637     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1638 }
1639 
1640 void tcg_gen_gvec_muls(unsigned vece, uint32_t dofs, uint32_t aofs,
1641                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
1642 {
1643     static const GVecGen2s g[4] = {
1644         { .fniv = tcg_gen_mul_vec,
1645           .fno = gen_helper_gvec_muls8,
1646           .opc = INDEX_op_mul_vec,
1647           .vece = MO_8 },
1648         { .fniv = tcg_gen_mul_vec,
1649           .fno = gen_helper_gvec_muls16,
1650           .opc = INDEX_op_mul_vec,
1651           .vece = MO_16 },
1652         { .fni4 = tcg_gen_mul_i32,
1653           .fniv = tcg_gen_mul_vec,
1654           .fno = gen_helper_gvec_muls32,
1655           .opc = INDEX_op_mul_vec,
1656           .vece = MO_32 },
1657         { .fni8 = tcg_gen_mul_i64,
1658           .fniv = tcg_gen_mul_vec,
1659           .fno = gen_helper_gvec_muls64,
1660           .opc = INDEX_op_mul_vec,
1661           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1662           .vece = MO_64 },
1663     };
1664 
1665     tcg_debug_assert(vece <= MO_64);
1666     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
1667 }
1668 
1669 void tcg_gen_gvec_muli(unsigned vece, uint32_t dofs, uint32_t aofs,
1670                        int64_t c, uint32_t oprsz, uint32_t maxsz)
1671 {
1672     TCGv_i64 tmp = tcg_const_i64(c);
1673     tcg_gen_gvec_muls(vece, dofs, aofs, tmp, oprsz, maxsz);
1674     tcg_temp_free_i64(tmp);
1675 }
1676 
1677 void tcg_gen_gvec_ssadd(unsigned vece, uint32_t dofs, uint32_t aofs,
1678                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1679 {
1680     static const GVecGen3 g[4] = {
1681         { .fniv = tcg_gen_ssadd_vec,
1682           .fno = gen_helper_gvec_ssadd8,
1683           .opc = INDEX_op_ssadd_vec,
1684           .vece = MO_8 },
1685         { .fniv = tcg_gen_ssadd_vec,
1686           .fno = gen_helper_gvec_ssadd16,
1687           .opc = INDEX_op_ssadd_vec,
1688           .vece = MO_16 },
1689         { .fniv = tcg_gen_ssadd_vec,
1690           .fno = gen_helper_gvec_ssadd32,
1691           .opc = INDEX_op_ssadd_vec,
1692           .vece = MO_32 },
1693         { .fniv = tcg_gen_ssadd_vec,
1694           .fno = gen_helper_gvec_ssadd64,
1695           .opc = INDEX_op_ssadd_vec,
1696           .vece = MO_64 },
1697     };
1698     tcg_debug_assert(vece <= MO_64);
1699     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1700 }
1701 
1702 void tcg_gen_gvec_sssub(unsigned vece, uint32_t dofs, uint32_t aofs,
1703                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1704 {
1705     static const GVecGen3 g[4] = {
1706         { .fniv = tcg_gen_sssub_vec,
1707           .fno = gen_helper_gvec_sssub8,
1708           .opc = INDEX_op_sssub_vec,
1709           .vece = MO_8 },
1710         { .fniv = tcg_gen_sssub_vec,
1711           .fno = gen_helper_gvec_sssub16,
1712           .opc = INDEX_op_sssub_vec,
1713           .vece = MO_16 },
1714         { .fniv = tcg_gen_sssub_vec,
1715           .fno = gen_helper_gvec_sssub32,
1716           .opc = INDEX_op_sssub_vec,
1717           .vece = MO_32 },
1718         { .fniv = tcg_gen_sssub_vec,
1719           .fno = gen_helper_gvec_sssub64,
1720           .opc = INDEX_op_sssub_vec,
1721           .vece = MO_64 },
1722     };
1723     tcg_debug_assert(vece <= MO_64);
1724     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1725 }
1726 
1727 static void tcg_gen_usadd_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
1728 {
1729     TCGv_i32 max = tcg_const_i32(-1);
1730     tcg_gen_add_i32(d, a, b);
1731     tcg_gen_movcond_i32(TCG_COND_LTU, d, d, a, max, d);
1732     tcg_temp_free_i32(max);
1733 }
1734 
1735 static void tcg_gen_usadd_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1736 {
1737     TCGv_i64 max = tcg_const_i64(-1);
1738     tcg_gen_add_i64(d, a, b);
1739     tcg_gen_movcond_i64(TCG_COND_LTU, d, d, a, max, d);
1740     tcg_temp_free_i64(max);
1741 }
1742 
1743 void tcg_gen_gvec_usadd(unsigned vece, uint32_t dofs, uint32_t aofs,
1744                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1745 {
1746     static const GVecGen3 g[4] = {
1747         { .fniv = tcg_gen_usadd_vec,
1748           .fno = gen_helper_gvec_usadd8,
1749           .opc = INDEX_op_usadd_vec,
1750           .vece = MO_8 },
1751         { .fniv = tcg_gen_usadd_vec,
1752           .fno = gen_helper_gvec_usadd16,
1753           .opc = INDEX_op_usadd_vec,
1754           .vece = MO_16 },
1755         { .fni4 = tcg_gen_usadd_i32,
1756           .fniv = tcg_gen_usadd_vec,
1757           .fno = gen_helper_gvec_usadd32,
1758           .opc = INDEX_op_usadd_vec,
1759           .vece = MO_32 },
1760         { .fni8 = tcg_gen_usadd_i64,
1761           .fniv = tcg_gen_usadd_vec,
1762           .fno = gen_helper_gvec_usadd64,
1763           .opc = INDEX_op_usadd_vec,
1764           .vece = MO_64 }
1765     };
1766     tcg_debug_assert(vece <= MO_64);
1767     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1768 }
1769 
1770 static void tcg_gen_ussub_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
1771 {
1772     TCGv_i32 min = tcg_const_i32(0);
1773     tcg_gen_sub_i32(d, a, b);
1774     tcg_gen_movcond_i32(TCG_COND_LTU, d, a, b, min, d);
1775     tcg_temp_free_i32(min);
1776 }
1777 
1778 static void tcg_gen_ussub_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1779 {
1780     TCGv_i64 min = tcg_const_i64(0);
1781     tcg_gen_sub_i64(d, a, b);
1782     tcg_gen_movcond_i64(TCG_COND_LTU, d, a, b, min, d);
1783     tcg_temp_free_i64(min);
1784 }
1785 
1786 void tcg_gen_gvec_ussub(unsigned vece, uint32_t dofs, uint32_t aofs,
1787                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1788 {
1789     static const GVecGen3 g[4] = {
1790         { .fniv = tcg_gen_ussub_vec,
1791           .fno = gen_helper_gvec_ussub8,
1792           .opc = INDEX_op_ussub_vec,
1793           .vece = MO_8 },
1794         { .fniv = tcg_gen_ussub_vec,
1795           .fno = gen_helper_gvec_ussub16,
1796           .opc = INDEX_op_ussub_vec,
1797           .vece = MO_16 },
1798         { .fni4 = tcg_gen_ussub_i32,
1799           .fniv = tcg_gen_ussub_vec,
1800           .fno = gen_helper_gvec_ussub32,
1801           .opc = INDEX_op_ussub_vec,
1802           .vece = MO_32 },
1803         { .fni8 = tcg_gen_ussub_i64,
1804           .fniv = tcg_gen_ussub_vec,
1805           .fno = gen_helper_gvec_ussub64,
1806           .opc = INDEX_op_ussub_vec,
1807           .vece = MO_64 }
1808     };
1809     tcg_debug_assert(vece <= MO_64);
1810     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1811 }
1812 
1813 void tcg_gen_gvec_smin(unsigned vece, uint32_t dofs, uint32_t aofs,
1814                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1815 {
1816     static const GVecGen3 g[4] = {
1817         { .fniv = tcg_gen_smin_vec,
1818           .fno = gen_helper_gvec_smin8,
1819           .opc = INDEX_op_smin_vec,
1820           .vece = MO_8 },
1821         { .fniv = tcg_gen_smin_vec,
1822           .fno = gen_helper_gvec_smin16,
1823           .opc = INDEX_op_smin_vec,
1824           .vece = MO_16 },
1825         { .fni4 = tcg_gen_smin_i32,
1826           .fniv = tcg_gen_smin_vec,
1827           .fno = gen_helper_gvec_smin32,
1828           .opc = INDEX_op_smin_vec,
1829           .vece = MO_32 },
1830         { .fni8 = tcg_gen_smin_i64,
1831           .fniv = tcg_gen_smin_vec,
1832           .fno = gen_helper_gvec_smin64,
1833           .opc = INDEX_op_smin_vec,
1834           .vece = MO_64 }
1835     };
1836     tcg_debug_assert(vece <= MO_64);
1837     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1838 }
1839 
1840 void tcg_gen_gvec_umin(unsigned vece, uint32_t dofs, uint32_t aofs,
1841                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1842 {
1843     static const GVecGen3 g[4] = {
1844         { .fniv = tcg_gen_umin_vec,
1845           .fno = gen_helper_gvec_umin8,
1846           .opc = INDEX_op_umin_vec,
1847           .vece = MO_8 },
1848         { .fniv = tcg_gen_umin_vec,
1849           .fno = gen_helper_gvec_umin16,
1850           .opc = INDEX_op_umin_vec,
1851           .vece = MO_16 },
1852         { .fni4 = tcg_gen_umin_i32,
1853           .fniv = tcg_gen_umin_vec,
1854           .fno = gen_helper_gvec_umin32,
1855           .opc = INDEX_op_umin_vec,
1856           .vece = MO_32 },
1857         { .fni8 = tcg_gen_umin_i64,
1858           .fniv = tcg_gen_umin_vec,
1859           .fno = gen_helper_gvec_umin64,
1860           .opc = INDEX_op_umin_vec,
1861           .vece = MO_64 }
1862     };
1863     tcg_debug_assert(vece <= MO_64);
1864     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1865 }
1866 
1867 void tcg_gen_gvec_smax(unsigned vece, uint32_t dofs, uint32_t aofs,
1868                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1869 {
1870     static const GVecGen3 g[4] = {
1871         { .fniv = tcg_gen_smax_vec,
1872           .fno = gen_helper_gvec_smax8,
1873           .opc = INDEX_op_smax_vec,
1874           .vece = MO_8 },
1875         { .fniv = tcg_gen_smax_vec,
1876           .fno = gen_helper_gvec_smax16,
1877           .opc = INDEX_op_smax_vec,
1878           .vece = MO_16 },
1879         { .fni4 = tcg_gen_smax_i32,
1880           .fniv = tcg_gen_smax_vec,
1881           .fno = gen_helper_gvec_smax32,
1882           .opc = INDEX_op_smax_vec,
1883           .vece = MO_32 },
1884         { .fni8 = tcg_gen_smax_i64,
1885           .fniv = tcg_gen_smax_vec,
1886           .fno = gen_helper_gvec_smax64,
1887           .opc = INDEX_op_smax_vec,
1888           .vece = MO_64 }
1889     };
1890     tcg_debug_assert(vece <= MO_64);
1891     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1892 }
1893 
1894 void tcg_gen_gvec_umax(unsigned vece, uint32_t dofs, uint32_t aofs,
1895                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1896 {
1897     static const GVecGen3 g[4] = {
1898         { .fniv = tcg_gen_umax_vec,
1899           .fno = gen_helper_gvec_umax8,
1900           .opc = INDEX_op_umax_vec,
1901           .vece = MO_8 },
1902         { .fniv = tcg_gen_umax_vec,
1903           .fno = gen_helper_gvec_umax16,
1904           .opc = INDEX_op_umax_vec,
1905           .vece = MO_16 },
1906         { .fni4 = tcg_gen_umax_i32,
1907           .fniv = tcg_gen_umax_vec,
1908           .fno = gen_helper_gvec_umax32,
1909           .opc = INDEX_op_umax_vec,
1910           .vece = MO_32 },
1911         { .fni8 = tcg_gen_umax_i64,
1912           .fniv = tcg_gen_umax_vec,
1913           .fno = gen_helper_gvec_umax64,
1914           .opc = INDEX_op_umax_vec,
1915           .vece = MO_64 }
1916     };
1917     tcg_debug_assert(vece <= MO_64);
1918     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1919 }
1920 
1921 /* Perform a vector negation using normal negation and a mask.
1922    Compare gen_subv_mask above.  */
1923 static void gen_negv_mask(TCGv_i64 d, TCGv_i64 b, TCGv_i64 m)
1924 {
1925     TCGv_i64 t2 = tcg_temp_new_i64();
1926     TCGv_i64 t3 = tcg_temp_new_i64();
1927 
1928     tcg_gen_andc_i64(t3, m, b);
1929     tcg_gen_andc_i64(t2, b, m);
1930     tcg_gen_sub_i64(d, m, t2);
1931     tcg_gen_xor_i64(d, d, t3);
1932 
1933     tcg_temp_free_i64(t2);
1934     tcg_temp_free_i64(t3);
1935 }
1936 
1937 void tcg_gen_vec_neg8_i64(TCGv_i64 d, TCGv_i64 b)
1938 {
1939     TCGv_i64 m = tcg_const_i64(dup_const(MO_8, 0x80));
1940     gen_negv_mask(d, b, m);
1941     tcg_temp_free_i64(m);
1942 }
1943 
1944 void tcg_gen_vec_neg16_i64(TCGv_i64 d, TCGv_i64 b)
1945 {
1946     TCGv_i64 m = tcg_const_i64(dup_const(MO_16, 0x8000));
1947     gen_negv_mask(d, b, m);
1948     tcg_temp_free_i64(m);
1949 }
1950 
1951 void tcg_gen_vec_neg32_i64(TCGv_i64 d, TCGv_i64 b)
1952 {
1953     TCGv_i64 t1 = tcg_temp_new_i64();
1954     TCGv_i64 t2 = tcg_temp_new_i64();
1955 
1956     tcg_gen_andi_i64(t1, b, ~0xffffffffull);
1957     tcg_gen_neg_i64(t2, b);
1958     tcg_gen_neg_i64(t1, t1);
1959     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
1960 
1961     tcg_temp_free_i64(t1);
1962     tcg_temp_free_i64(t2);
1963 }
1964 
1965 void tcg_gen_gvec_neg(unsigned vece, uint32_t dofs, uint32_t aofs,
1966                       uint32_t oprsz, uint32_t maxsz)
1967 {
1968     static const GVecGen2 g[4] = {
1969         { .fni8 = tcg_gen_vec_neg8_i64,
1970           .fniv = tcg_gen_neg_vec,
1971           .fno = gen_helper_gvec_neg8,
1972           .opc = INDEX_op_neg_vec,
1973           .vece = MO_8 },
1974         { .fni8 = tcg_gen_vec_neg16_i64,
1975           .fniv = tcg_gen_neg_vec,
1976           .fno = gen_helper_gvec_neg16,
1977           .opc = INDEX_op_neg_vec,
1978           .vece = MO_16 },
1979         { .fni4 = tcg_gen_neg_i32,
1980           .fniv = tcg_gen_neg_vec,
1981           .fno = gen_helper_gvec_neg32,
1982           .opc = INDEX_op_neg_vec,
1983           .vece = MO_32 },
1984         { .fni8 = tcg_gen_neg_i64,
1985           .fniv = tcg_gen_neg_vec,
1986           .fno = gen_helper_gvec_neg64,
1987           .opc = INDEX_op_neg_vec,
1988           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1989           .vece = MO_64 },
1990     };
1991 
1992     tcg_debug_assert(vece <= MO_64);
1993     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g[vece]);
1994 }
1995 
1996 void tcg_gen_gvec_and(unsigned vece, uint32_t dofs, uint32_t aofs,
1997                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1998 {
1999     static const GVecGen3 g = {
2000         .fni8 = tcg_gen_and_i64,
2001         .fniv = tcg_gen_and_vec,
2002         .fno = gen_helper_gvec_and,
2003         .opc = INDEX_op_and_vec,
2004         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2005     };
2006 
2007     if (aofs == bofs) {
2008         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2009     } else {
2010         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2011     }
2012 }
2013 
2014 void tcg_gen_gvec_or(unsigned vece, uint32_t dofs, uint32_t aofs,
2015                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2016 {
2017     static const GVecGen3 g = {
2018         .fni8 = tcg_gen_or_i64,
2019         .fniv = tcg_gen_or_vec,
2020         .fno = gen_helper_gvec_or,
2021         .opc = INDEX_op_or_vec,
2022         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2023     };
2024 
2025     if (aofs == bofs) {
2026         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2027     } else {
2028         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2029     }
2030 }
2031 
2032 void tcg_gen_gvec_xor(unsigned vece, uint32_t dofs, uint32_t aofs,
2033                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2034 {
2035     static const GVecGen3 g = {
2036         .fni8 = tcg_gen_xor_i64,
2037         .fniv = tcg_gen_xor_vec,
2038         .fno = gen_helper_gvec_xor,
2039         .opc = INDEX_op_xor_vec,
2040         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2041     };
2042 
2043     if (aofs == bofs) {
2044         tcg_gen_gvec_dup8i(dofs, oprsz, maxsz, 0);
2045     } else {
2046         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2047     }
2048 }
2049 
2050 void tcg_gen_gvec_andc(unsigned vece, uint32_t dofs, uint32_t aofs,
2051                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2052 {
2053     static const GVecGen3 g = {
2054         .fni8 = tcg_gen_andc_i64,
2055         .fniv = tcg_gen_andc_vec,
2056         .fno = gen_helper_gvec_andc,
2057         .opc = INDEX_op_andc_vec,
2058         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2059     };
2060 
2061     if (aofs == bofs) {
2062         tcg_gen_gvec_dup8i(dofs, oprsz, maxsz, 0);
2063     } else {
2064         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2065     }
2066 }
2067 
2068 void tcg_gen_gvec_orc(unsigned vece, uint32_t dofs, uint32_t aofs,
2069                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2070 {
2071     static const GVecGen3 g = {
2072         .fni8 = tcg_gen_orc_i64,
2073         .fniv = tcg_gen_orc_vec,
2074         .fno = gen_helper_gvec_orc,
2075         .opc = INDEX_op_orc_vec,
2076         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2077     };
2078 
2079     if (aofs == bofs) {
2080         tcg_gen_gvec_dup8i(dofs, oprsz, maxsz, -1);
2081     } else {
2082         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2083     }
2084 }
2085 
2086 void tcg_gen_gvec_nand(unsigned vece, uint32_t dofs, uint32_t aofs,
2087                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2088 {
2089     static const GVecGen3 g = {
2090         .fni8 = tcg_gen_nand_i64,
2091         .fniv = tcg_gen_nand_vec,
2092         .fno = gen_helper_gvec_nand,
2093         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2094     };
2095 
2096     if (aofs == bofs) {
2097         tcg_gen_gvec_not(vece, dofs, aofs, oprsz, maxsz);
2098     } else {
2099         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2100     }
2101 }
2102 
2103 void tcg_gen_gvec_nor(unsigned vece, uint32_t dofs, uint32_t aofs,
2104                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2105 {
2106     static const GVecGen3 g = {
2107         .fni8 = tcg_gen_nor_i64,
2108         .fniv = tcg_gen_nor_vec,
2109         .fno = gen_helper_gvec_nor,
2110         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2111     };
2112 
2113     if (aofs == bofs) {
2114         tcg_gen_gvec_not(vece, dofs, aofs, oprsz, maxsz);
2115     } else {
2116         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2117     }
2118 }
2119 
2120 void tcg_gen_gvec_eqv(unsigned vece, uint32_t dofs, uint32_t aofs,
2121                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
2122 {
2123     static const GVecGen3 g = {
2124         .fni8 = tcg_gen_eqv_i64,
2125         .fniv = tcg_gen_eqv_vec,
2126         .fno = gen_helper_gvec_eqv,
2127         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2128     };
2129 
2130     if (aofs == bofs) {
2131         tcg_gen_gvec_dup8i(dofs, oprsz, maxsz, -1);
2132     } else {
2133         tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
2134     }
2135 }
2136 
2137 static const GVecGen2s gop_ands = {
2138     .fni8 = tcg_gen_and_i64,
2139     .fniv = tcg_gen_and_vec,
2140     .fno = gen_helper_gvec_ands,
2141     .opc = INDEX_op_and_vec,
2142     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2143     .vece = MO_64
2144 };
2145 
2146 void tcg_gen_gvec_ands(unsigned vece, uint32_t dofs, uint32_t aofs,
2147                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2148 {
2149     TCGv_i64 tmp = tcg_temp_new_i64();
2150     gen_dup_i64(vece, tmp, c);
2151     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ands);
2152     tcg_temp_free_i64(tmp);
2153 }
2154 
2155 void tcg_gen_gvec_andi(unsigned vece, uint32_t dofs, uint32_t aofs,
2156                        int64_t c, uint32_t oprsz, uint32_t maxsz)
2157 {
2158     TCGv_i64 tmp = tcg_const_i64(dup_const(vece, c));
2159     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ands);
2160     tcg_temp_free_i64(tmp);
2161 }
2162 
2163 static const GVecGen2s gop_xors = {
2164     .fni8 = tcg_gen_xor_i64,
2165     .fniv = tcg_gen_xor_vec,
2166     .fno = gen_helper_gvec_xors,
2167     .opc = INDEX_op_xor_vec,
2168     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2169     .vece = MO_64
2170 };
2171 
2172 void tcg_gen_gvec_xors(unsigned vece, uint32_t dofs, uint32_t aofs,
2173                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2174 {
2175     TCGv_i64 tmp = tcg_temp_new_i64();
2176     gen_dup_i64(vece, tmp, c);
2177     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_xors);
2178     tcg_temp_free_i64(tmp);
2179 }
2180 
2181 void tcg_gen_gvec_xori(unsigned vece, uint32_t dofs, uint32_t aofs,
2182                        int64_t c, uint32_t oprsz, uint32_t maxsz)
2183 {
2184     TCGv_i64 tmp = tcg_const_i64(dup_const(vece, c));
2185     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_xors);
2186     tcg_temp_free_i64(tmp);
2187 }
2188 
2189 static const GVecGen2s gop_ors = {
2190     .fni8 = tcg_gen_or_i64,
2191     .fniv = tcg_gen_or_vec,
2192     .fno = gen_helper_gvec_ors,
2193     .opc = INDEX_op_or_vec,
2194     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2195     .vece = MO_64
2196 };
2197 
2198 void tcg_gen_gvec_ors(unsigned vece, uint32_t dofs, uint32_t aofs,
2199                       TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
2200 {
2201     TCGv_i64 tmp = tcg_temp_new_i64();
2202     gen_dup_i64(vece, tmp, c);
2203     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ors);
2204     tcg_temp_free_i64(tmp);
2205 }
2206 
2207 void tcg_gen_gvec_ori(unsigned vece, uint32_t dofs, uint32_t aofs,
2208                       int64_t c, uint32_t oprsz, uint32_t maxsz)
2209 {
2210     TCGv_i64 tmp = tcg_const_i64(dup_const(vece, c));
2211     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ors);
2212     tcg_temp_free_i64(tmp);
2213 }
2214 
2215 void tcg_gen_vec_shl8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2216 {
2217     uint64_t mask = dup_const(MO_8, 0xff << c);
2218     tcg_gen_shli_i64(d, a, c);
2219     tcg_gen_andi_i64(d, d, mask);
2220 }
2221 
2222 void tcg_gen_vec_shl16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2223 {
2224     uint64_t mask = dup_const(MO_16, 0xffff << c);
2225     tcg_gen_shli_i64(d, a, c);
2226     tcg_gen_andi_i64(d, d, mask);
2227 }
2228 
2229 void tcg_gen_gvec_shli(unsigned vece, uint32_t dofs, uint32_t aofs,
2230                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
2231 {
2232     static const GVecGen2i g[4] = {
2233         { .fni8 = tcg_gen_vec_shl8i_i64,
2234           .fniv = tcg_gen_shli_vec,
2235           .fno = gen_helper_gvec_shl8i,
2236           .opc = INDEX_op_shli_vec,
2237           .vece = MO_8 },
2238         { .fni8 = tcg_gen_vec_shl16i_i64,
2239           .fniv = tcg_gen_shli_vec,
2240           .fno = gen_helper_gvec_shl16i,
2241           .opc = INDEX_op_shli_vec,
2242           .vece = MO_16 },
2243         { .fni4 = tcg_gen_shli_i32,
2244           .fniv = tcg_gen_shli_vec,
2245           .fno = gen_helper_gvec_shl32i,
2246           .opc = INDEX_op_shli_vec,
2247           .vece = MO_32 },
2248         { .fni8 = tcg_gen_shli_i64,
2249           .fniv = tcg_gen_shli_vec,
2250           .fno = gen_helper_gvec_shl64i,
2251           .opc = INDEX_op_shli_vec,
2252           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2253           .vece = MO_64 },
2254     };
2255 
2256     tcg_debug_assert(vece <= MO_64);
2257     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2258     if (shift == 0) {
2259         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2260     } else {
2261         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2262     }
2263 }
2264 
2265 void tcg_gen_vec_shr8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2266 {
2267     uint64_t mask = dup_const(MO_8, 0xff >> c);
2268     tcg_gen_shri_i64(d, a, c);
2269     tcg_gen_andi_i64(d, d, mask);
2270 }
2271 
2272 void tcg_gen_vec_shr16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2273 {
2274     uint64_t mask = dup_const(MO_16, 0xffff >> c);
2275     tcg_gen_shri_i64(d, a, c);
2276     tcg_gen_andi_i64(d, d, mask);
2277 }
2278 
2279 void tcg_gen_gvec_shri(unsigned vece, uint32_t dofs, uint32_t aofs,
2280                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
2281 {
2282     static const GVecGen2i g[4] = {
2283         { .fni8 = tcg_gen_vec_shr8i_i64,
2284           .fniv = tcg_gen_shri_vec,
2285           .fno = gen_helper_gvec_shr8i,
2286           .opc = INDEX_op_shri_vec,
2287           .vece = MO_8 },
2288         { .fni8 = tcg_gen_vec_shr16i_i64,
2289           .fniv = tcg_gen_shri_vec,
2290           .fno = gen_helper_gvec_shr16i,
2291           .opc = INDEX_op_shri_vec,
2292           .vece = MO_16 },
2293         { .fni4 = tcg_gen_shri_i32,
2294           .fniv = tcg_gen_shri_vec,
2295           .fno = gen_helper_gvec_shr32i,
2296           .opc = INDEX_op_shri_vec,
2297           .vece = MO_32 },
2298         { .fni8 = tcg_gen_shri_i64,
2299           .fniv = tcg_gen_shri_vec,
2300           .fno = gen_helper_gvec_shr64i,
2301           .opc = INDEX_op_shri_vec,
2302           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2303           .vece = MO_64 },
2304     };
2305 
2306     tcg_debug_assert(vece <= MO_64);
2307     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2308     if (shift == 0) {
2309         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2310     } else {
2311         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2312     }
2313 }
2314 
2315 void tcg_gen_vec_sar8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2316 {
2317     uint64_t s_mask = dup_const(MO_8, 0x80 >> c);
2318     uint64_t c_mask = dup_const(MO_8, 0xff >> c);
2319     TCGv_i64 s = tcg_temp_new_i64();
2320 
2321     tcg_gen_shri_i64(d, a, c);
2322     tcg_gen_andi_i64(s, d, s_mask);  /* isolate (shifted) sign bit */
2323     tcg_gen_muli_i64(s, s, (2 << c) - 2); /* replicate isolated signs */
2324     tcg_gen_andi_i64(d, d, c_mask);  /* clear out bits above sign  */
2325     tcg_gen_or_i64(d, d, s);         /* include sign extension */
2326     tcg_temp_free_i64(s);
2327 }
2328 
2329 void tcg_gen_vec_sar16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2330 {
2331     uint64_t s_mask = dup_const(MO_16, 0x8000 >> c);
2332     uint64_t c_mask = dup_const(MO_16, 0xffff >> c);
2333     TCGv_i64 s = tcg_temp_new_i64();
2334 
2335     tcg_gen_shri_i64(d, a, c);
2336     tcg_gen_andi_i64(s, d, s_mask);  /* isolate (shifted) sign bit */
2337     tcg_gen_andi_i64(d, d, c_mask);  /* clear out bits above sign  */
2338     tcg_gen_muli_i64(s, s, (2 << c) - 2); /* replicate isolated signs */
2339     tcg_gen_or_i64(d, d, s);         /* include sign extension */
2340     tcg_temp_free_i64(s);
2341 }
2342 
2343 void tcg_gen_gvec_sari(unsigned vece, uint32_t dofs, uint32_t aofs,
2344                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
2345 {
2346     static const GVecGen2i g[4] = {
2347         { .fni8 = tcg_gen_vec_sar8i_i64,
2348           .fniv = tcg_gen_sari_vec,
2349           .fno = gen_helper_gvec_sar8i,
2350           .opc = INDEX_op_sari_vec,
2351           .vece = MO_8 },
2352         { .fni8 = tcg_gen_vec_sar16i_i64,
2353           .fniv = tcg_gen_sari_vec,
2354           .fno = gen_helper_gvec_sar16i,
2355           .opc = INDEX_op_sari_vec,
2356           .vece = MO_16 },
2357         { .fni4 = tcg_gen_sari_i32,
2358           .fniv = tcg_gen_sari_vec,
2359           .fno = gen_helper_gvec_sar32i,
2360           .opc = INDEX_op_sari_vec,
2361           .vece = MO_32 },
2362         { .fni8 = tcg_gen_sari_i64,
2363           .fniv = tcg_gen_sari_vec,
2364           .fno = gen_helper_gvec_sar64i,
2365           .opc = INDEX_op_sari_vec,
2366           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2367           .vece = MO_64 },
2368     };
2369 
2370     tcg_debug_assert(vece <= MO_64);
2371     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2372     if (shift == 0) {
2373         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2374     } else {
2375         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2376     }
2377 }
2378 
2379 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
2380 static void expand_cmp_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
2381                            uint32_t oprsz, TCGCond cond)
2382 {
2383     TCGv_i32 t0 = tcg_temp_new_i32();
2384     TCGv_i32 t1 = tcg_temp_new_i32();
2385     uint32_t i;
2386 
2387     for (i = 0; i < oprsz; i += 4) {
2388         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
2389         tcg_gen_ld_i32(t1, cpu_env, bofs + i);
2390         tcg_gen_setcond_i32(cond, t0, t0, t1);
2391         tcg_gen_neg_i32(t0, t0);
2392         tcg_gen_st_i32(t0, cpu_env, dofs + i);
2393     }
2394     tcg_temp_free_i32(t1);
2395     tcg_temp_free_i32(t0);
2396 }
2397 
2398 static void expand_cmp_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
2399                            uint32_t oprsz, TCGCond cond)
2400 {
2401     TCGv_i64 t0 = tcg_temp_new_i64();
2402     TCGv_i64 t1 = tcg_temp_new_i64();
2403     uint32_t i;
2404 
2405     for (i = 0; i < oprsz; i += 8) {
2406         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
2407         tcg_gen_ld_i64(t1, cpu_env, bofs + i);
2408         tcg_gen_setcond_i64(cond, t0, t0, t1);
2409         tcg_gen_neg_i64(t0, t0);
2410         tcg_gen_st_i64(t0, cpu_env, dofs + i);
2411     }
2412     tcg_temp_free_i64(t1);
2413     tcg_temp_free_i64(t0);
2414 }
2415 
2416 static void expand_cmp_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
2417                            uint32_t bofs, uint32_t oprsz, uint32_t tysz,
2418                            TCGType type, TCGCond cond)
2419 {
2420     TCGv_vec t0 = tcg_temp_new_vec(type);
2421     TCGv_vec t1 = tcg_temp_new_vec(type);
2422     uint32_t i;
2423 
2424     for (i = 0; i < oprsz; i += tysz) {
2425         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
2426         tcg_gen_ld_vec(t1, cpu_env, bofs + i);
2427         tcg_gen_cmp_vec(cond, vece, t0, t0, t1);
2428         tcg_gen_st_vec(t0, cpu_env, dofs + i);
2429     }
2430     tcg_temp_free_vec(t1);
2431     tcg_temp_free_vec(t0);
2432 }
2433 
2434 void tcg_gen_gvec_cmp(TCGCond cond, unsigned vece, uint32_t dofs,
2435                       uint32_t aofs, uint32_t bofs,
2436                       uint32_t oprsz, uint32_t maxsz)
2437 {
2438     static gen_helper_gvec_3 * const eq_fn[4] = {
2439         gen_helper_gvec_eq8, gen_helper_gvec_eq16,
2440         gen_helper_gvec_eq32, gen_helper_gvec_eq64
2441     };
2442     static gen_helper_gvec_3 * const ne_fn[4] = {
2443         gen_helper_gvec_ne8, gen_helper_gvec_ne16,
2444         gen_helper_gvec_ne32, gen_helper_gvec_ne64
2445     };
2446     static gen_helper_gvec_3 * const lt_fn[4] = {
2447         gen_helper_gvec_lt8, gen_helper_gvec_lt16,
2448         gen_helper_gvec_lt32, gen_helper_gvec_lt64
2449     };
2450     static gen_helper_gvec_3 * const le_fn[4] = {
2451         gen_helper_gvec_le8, gen_helper_gvec_le16,
2452         gen_helper_gvec_le32, gen_helper_gvec_le64
2453     };
2454     static gen_helper_gvec_3 * const ltu_fn[4] = {
2455         gen_helper_gvec_ltu8, gen_helper_gvec_ltu16,
2456         gen_helper_gvec_ltu32, gen_helper_gvec_ltu64
2457     };
2458     static gen_helper_gvec_3 * const leu_fn[4] = {
2459         gen_helper_gvec_leu8, gen_helper_gvec_leu16,
2460         gen_helper_gvec_leu32, gen_helper_gvec_leu64
2461     };
2462     static gen_helper_gvec_3 * const * const fns[16] = {
2463         [TCG_COND_EQ] = eq_fn,
2464         [TCG_COND_NE] = ne_fn,
2465         [TCG_COND_LT] = lt_fn,
2466         [TCG_COND_LE] = le_fn,
2467         [TCG_COND_LTU] = ltu_fn,
2468         [TCG_COND_LEU] = leu_fn,
2469     };
2470     TCGType type;
2471     uint32_t some;
2472 
2473     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
2474     check_overlap_3(dofs, aofs, bofs, maxsz);
2475 
2476     if (cond == TCG_COND_NEVER || cond == TCG_COND_ALWAYS) {
2477         do_dup(MO_8, dofs, oprsz, maxsz,
2478                NULL, NULL, -(cond == TCG_COND_ALWAYS));
2479         return;
2480     }
2481 
2482     /* Implement inline with a vector type, if possible.
2483      * Prefer integer when 64-bit host and 64-bit comparison.
2484      */
2485     type = choose_vector_type(INDEX_op_cmp_vec, vece, oprsz,
2486                               TCG_TARGET_REG_BITS == 64 && vece == MO_64);
2487     switch (type) {
2488     case TCG_TYPE_V256:
2489         /* Recall that ARM SVE allows vector sizes that are not a
2490          * power of 2, but always a multiple of 16.  The intent is
2491          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
2492          */
2493         some = QEMU_ALIGN_DOWN(oprsz, 32);
2494         expand_cmp_vec(vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256, cond);
2495         if (some == oprsz) {
2496             break;
2497         }
2498         dofs += some;
2499         aofs += some;
2500         bofs += some;
2501         oprsz -= some;
2502         maxsz -= some;
2503         /* fallthru */
2504     case TCG_TYPE_V128:
2505         expand_cmp_vec(vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128, cond);
2506         break;
2507     case TCG_TYPE_V64:
2508         expand_cmp_vec(vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64, cond);
2509         break;
2510 
2511     case 0:
2512         if (vece == MO_64 && check_size_impl(oprsz, 8)) {
2513             expand_cmp_i64(dofs, aofs, bofs, oprsz, cond);
2514         } else if (vece == MO_32 && check_size_impl(oprsz, 4)) {
2515             expand_cmp_i32(dofs, aofs, bofs, oprsz, cond);
2516         } else {
2517             gen_helper_gvec_3 * const *fn = fns[cond];
2518 
2519             if (fn == NULL) {
2520                 uint32_t tmp;
2521                 tmp = aofs, aofs = bofs, bofs = tmp;
2522                 cond = tcg_swap_cond(cond);
2523                 fn = fns[cond];
2524                 assert(fn != NULL);
2525             }
2526             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz, maxsz, 0, fn[vece]);
2527             return;
2528         }
2529         break;
2530 
2531     default:
2532         g_assert_not_reached();
2533     }
2534 
2535     if (oprsz < maxsz) {
2536         expand_clr(dofs + oprsz, maxsz - oprsz);
2537     }
2538 }
2539