xref: /openbmc/qemu/tcg/tcg-op-gvec.c (revision d1fd31f8)
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 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     uint32_t lnct = oprsz / lnsz;
291     return lnct >= 1 && lnct <= MAX_UNROLL;
292 }
293 
294 static void expand_clr(uint32_t dofs, uint32_t maxsz);
295 
296 /* Duplicate C as per VECE.  */
297 uint64_t (dup_const)(unsigned vece, uint64_t c)
298 {
299     switch (vece) {
300     case MO_8:
301         return 0x0101010101010101ull * (uint8_t)c;
302     case MO_16:
303         return 0x0001000100010001ull * (uint16_t)c;
304     case MO_32:
305         return 0x0000000100000001ull * (uint32_t)c;
306     case MO_64:
307         return c;
308     default:
309         g_assert_not_reached();
310     }
311 }
312 
313 /* Duplicate IN into OUT as per VECE.  */
314 static void gen_dup_i32(unsigned vece, TCGv_i32 out, TCGv_i32 in)
315 {
316     switch (vece) {
317     case MO_8:
318         tcg_gen_ext8u_i32(out, in);
319         tcg_gen_muli_i32(out, out, 0x01010101);
320         break;
321     case MO_16:
322         tcg_gen_deposit_i32(out, in, in, 16, 16);
323         break;
324     case MO_32:
325         tcg_gen_mov_i32(out, in);
326         break;
327     default:
328         g_assert_not_reached();
329     }
330 }
331 
332 static void gen_dup_i64(unsigned vece, TCGv_i64 out, TCGv_i64 in)
333 {
334     switch (vece) {
335     case MO_8:
336         tcg_gen_ext8u_i64(out, in);
337         tcg_gen_muli_i64(out, out, 0x0101010101010101ull);
338         break;
339     case MO_16:
340         tcg_gen_ext16u_i64(out, in);
341         tcg_gen_muli_i64(out, out, 0x0001000100010001ull);
342         break;
343     case MO_32:
344         tcg_gen_deposit_i64(out, in, in, 32, 32);
345         break;
346     case MO_64:
347         tcg_gen_mov_i64(out, in);
348         break;
349     default:
350         g_assert_not_reached();
351     }
352 }
353 
354 /* Select a supported vector type for implementing an operation on SIZE
355  * bytes.  If OP is 0, assume that the real operation to be performed is
356  * required by all backends.  Otherwise, make sure than OP can be performed
357  * on elements of size VECE in the selected type.  Do not select V64 if
358  * PREFER_I64 is true.  Return 0 if no vector type is selected.
359  */
360 static TCGType choose_vector_type(TCGOpcode op, unsigned vece, uint32_t size,
361                                   bool prefer_i64)
362 {
363     if (TCG_TARGET_HAS_v256 && check_size_impl(size, 32)) {
364         if (op == 0) {
365             return TCG_TYPE_V256;
366         }
367         /* Recall that ARM SVE allows vector sizes that are not a
368          * power of 2, but always a multiple of 16.  The intent is
369          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
370          * It is hard to imagine a case in which v256 is supported
371          * but v128 is not, but check anyway.
372          */
373         if (tcg_can_emit_vec_op(op, TCG_TYPE_V256, vece)
374             && (size % 32 == 0
375                 || tcg_can_emit_vec_op(op, TCG_TYPE_V128, vece))) {
376             return TCG_TYPE_V256;
377         }
378     }
379     if (TCG_TARGET_HAS_v128 && check_size_impl(size, 16)
380         && (op == 0 || tcg_can_emit_vec_op(op, TCG_TYPE_V128, vece))) {
381         return TCG_TYPE_V128;
382     }
383     if (TCG_TARGET_HAS_v64 && !prefer_i64 && check_size_impl(size, 8)
384         && (op == 0 || tcg_can_emit_vec_op(op, TCG_TYPE_V64, vece))) {
385         return TCG_TYPE_V64;
386     }
387     return 0;
388 }
389 
390 /* Set OPRSZ bytes at DOFS to replications of IN_32, IN_64 or IN_C.
391  * Only one of IN_32 or IN_64 may be set;
392  * IN_C is used if IN_32 and IN_64 are unset.
393  */
394 static void do_dup(unsigned vece, uint32_t dofs, uint32_t oprsz,
395                    uint32_t maxsz, TCGv_i32 in_32, TCGv_i64 in_64,
396                    uint64_t in_c)
397 {
398     TCGType type;
399     TCGv_i64 t_64;
400     TCGv_i32 t_32, t_desc;
401     TCGv_ptr t_ptr;
402     uint32_t i;
403 
404     assert(vece <= (in_32 ? MO_32 : MO_64));
405     assert(in_32 == NULL || in_64 == NULL);
406 
407     /* If we're storing 0, expand oprsz to maxsz.  */
408     if (in_32 == NULL && in_64 == NULL) {
409         in_c = dup_const(vece, in_c);
410         if (in_c == 0) {
411             oprsz = maxsz;
412         }
413     }
414 
415     /* Implement inline with a vector type, if possible.
416      * Prefer integer when 64-bit host and no variable dup.
417      */
418     type = choose_vector_type(0, vece, oprsz,
419                               (TCG_TARGET_REG_BITS == 64 && in_32 == NULL
420                                && (in_64 == NULL || vece == MO_64)));
421     if (type != 0) {
422         TCGv_vec t_vec = tcg_temp_new_vec(type);
423 
424         if (in_32) {
425             tcg_gen_dup_i32_vec(vece, t_vec, in_32);
426         } else if (in_64) {
427             tcg_gen_dup_i64_vec(vece, t_vec, in_64);
428         } else {
429             switch (vece) {
430             case MO_8:
431                 tcg_gen_dup8i_vec(t_vec, in_c);
432                 break;
433             case MO_16:
434                 tcg_gen_dup16i_vec(t_vec, in_c);
435                 break;
436             case MO_32:
437                 tcg_gen_dup32i_vec(t_vec, in_c);
438                 break;
439             default:
440                 tcg_gen_dup64i_vec(t_vec, in_c);
441                 break;
442             }
443         }
444 
445         i = 0;
446         switch (type) {
447         case TCG_TYPE_V256:
448             /* Recall that ARM SVE allows vector sizes that are not a
449              * power of 2, but always a multiple of 16.  The intent is
450              * that e.g. size == 80 would be expanded with 2x32 + 1x16.
451              */
452             for (; i + 32 <= oprsz; i += 32) {
453                 tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V256);
454             }
455             /* fallthru */
456         case TCG_TYPE_V128:
457             for (; i + 16 <= oprsz; i += 16) {
458                 tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V128);
459             }
460             break;
461         case TCG_TYPE_V64:
462             for (; i < oprsz; i += 8) {
463                 tcg_gen_stl_vec(t_vec, cpu_env, dofs + i, TCG_TYPE_V64);
464             }
465             break;
466         default:
467             g_assert_not_reached();
468         }
469 
470         tcg_temp_free_vec(t_vec);
471         goto done;
472     }
473 
474     /* Otherwise, inline with an integer type, unless "large".  */
475     if (check_size_impl(oprsz, TCG_TARGET_REG_BITS / 8)) {
476         t_64 = NULL;
477         t_32 = NULL;
478 
479         if (in_32) {
480             /* We are given a 32-bit variable input.  For a 64-bit host,
481                use a 64-bit operation unless the 32-bit operation would
482                be simple enough.  */
483             if (TCG_TARGET_REG_BITS == 64
484                 && (vece != MO_32 || !check_size_impl(oprsz, 4))) {
485                 t_64 = tcg_temp_new_i64();
486                 tcg_gen_extu_i32_i64(t_64, in_32);
487                 gen_dup_i64(vece, t_64, t_64);
488             } else {
489                 t_32 = tcg_temp_new_i32();
490                 gen_dup_i32(vece, t_32, in_32);
491             }
492         } else if (in_64) {
493             /* We are given a 64-bit variable input.  */
494             t_64 = tcg_temp_new_i64();
495             gen_dup_i64(vece, t_64, in_64);
496         } else {
497             /* We are given a constant input.  */
498             /* For 64-bit hosts, use 64-bit constants for "simple" constants
499                or when we'd need too many 32-bit stores, or when a 64-bit
500                constant is really required.  */
501             if (vece == MO_64
502                 || (TCG_TARGET_REG_BITS == 64
503                     && (in_c == 0 || in_c == -1
504                         || !check_size_impl(oprsz, 4)))) {
505                 t_64 = tcg_const_i64(in_c);
506             } else {
507                 t_32 = tcg_const_i32(in_c);
508             }
509         }
510 
511         /* Implement inline if we picked an implementation size above.  */
512         if (t_32) {
513             for (i = 0; i < oprsz; i += 4) {
514                 tcg_gen_st_i32(t_32, cpu_env, dofs + i);
515             }
516             tcg_temp_free_i32(t_32);
517             goto done;
518         }
519         if (t_64) {
520             for (i = 0; i < oprsz; i += 8) {
521                 tcg_gen_st_i64(t_64, cpu_env, dofs + i);
522             }
523             tcg_temp_free_i64(t_64);
524             goto done;
525         }
526     }
527 
528     /* Otherwise implement out of line.  */
529     t_ptr = tcg_temp_new_ptr();
530     tcg_gen_addi_ptr(t_ptr, cpu_env, dofs);
531     t_desc = tcg_const_i32(simd_desc(oprsz, maxsz, 0));
532 
533     if (vece == MO_64) {
534         if (in_64) {
535             gen_helper_gvec_dup64(t_ptr, t_desc, in_64);
536         } else {
537             t_64 = tcg_const_i64(in_c);
538             gen_helper_gvec_dup64(t_ptr, t_desc, t_64);
539             tcg_temp_free_i64(t_64);
540         }
541     } else {
542         typedef void dup_fn(TCGv_ptr, TCGv_i32, TCGv_i32);
543         static dup_fn * const fns[3] = {
544             gen_helper_gvec_dup8,
545             gen_helper_gvec_dup16,
546             gen_helper_gvec_dup32
547         };
548 
549         if (in_32) {
550             fns[vece](t_ptr, t_desc, in_32);
551         } else {
552             t_32 = tcg_temp_new_i32();
553             if (in_64) {
554                 tcg_gen_extrl_i64_i32(t_32, in_64);
555             } else if (vece == MO_8) {
556                 tcg_gen_movi_i32(t_32, in_c & 0xff);
557             } else if (vece == MO_16) {
558                 tcg_gen_movi_i32(t_32, in_c & 0xffff);
559             } else {
560                 tcg_gen_movi_i32(t_32, in_c);
561             }
562             fns[vece](t_ptr, t_desc, t_32);
563             tcg_temp_free_i32(t_32);
564         }
565     }
566 
567     tcg_temp_free_ptr(t_ptr);
568     tcg_temp_free_i32(t_desc);
569     return;
570 
571  done:
572     if (oprsz < maxsz) {
573         expand_clr(dofs + oprsz, maxsz - oprsz);
574     }
575 }
576 
577 /* Likewise, but with zero.  */
578 static void expand_clr(uint32_t dofs, uint32_t maxsz)
579 {
580     do_dup(MO_8, dofs, maxsz, maxsz, NULL, NULL, 0);
581 }
582 
583 /* Expand OPSZ bytes worth of two-operand operations using i32 elements.  */
584 static void expand_2_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
585                          void (*fni)(TCGv_i32, TCGv_i32))
586 {
587     TCGv_i32 t0 = tcg_temp_new_i32();
588     uint32_t i;
589 
590     for (i = 0; i < oprsz; i += 4) {
591         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
592         fni(t0, t0);
593         tcg_gen_st_i32(t0, cpu_env, dofs + i);
594     }
595     tcg_temp_free_i32(t0);
596 }
597 
598 static void expand_2i_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
599                           int32_t c, bool load_dest,
600                           void (*fni)(TCGv_i32, TCGv_i32, int32_t))
601 {
602     TCGv_i32 t0 = tcg_temp_new_i32();
603     TCGv_i32 t1 = tcg_temp_new_i32();
604     uint32_t i;
605 
606     for (i = 0; i < oprsz; i += 4) {
607         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
608         if (load_dest) {
609             tcg_gen_ld_i32(t1, cpu_env, dofs + i);
610         }
611         fni(t1, t0, c);
612         tcg_gen_st_i32(t1, cpu_env, dofs + i);
613     }
614     tcg_temp_free_i32(t0);
615     tcg_temp_free_i32(t1);
616 }
617 
618 static void expand_2s_i32(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
619                           TCGv_i32 c, bool scalar_first,
620                           void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32))
621 {
622     TCGv_i32 t0 = tcg_temp_new_i32();
623     TCGv_i32 t1 = tcg_temp_new_i32();
624     uint32_t i;
625 
626     for (i = 0; i < oprsz; i += 4) {
627         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
628         if (scalar_first) {
629             fni(t1, c, t0);
630         } else {
631             fni(t1, t0, c);
632         }
633         tcg_gen_st_i32(t1, cpu_env, dofs + i);
634     }
635     tcg_temp_free_i32(t0);
636     tcg_temp_free_i32(t1);
637 }
638 
639 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
640 static void expand_3_i32(uint32_t dofs, uint32_t aofs,
641                          uint32_t bofs, uint32_t oprsz, bool load_dest,
642                          void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32))
643 {
644     TCGv_i32 t0 = tcg_temp_new_i32();
645     TCGv_i32 t1 = tcg_temp_new_i32();
646     TCGv_i32 t2 = tcg_temp_new_i32();
647     uint32_t i;
648 
649     for (i = 0; i < oprsz; i += 4) {
650         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
651         tcg_gen_ld_i32(t1, cpu_env, bofs + i);
652         if (load_dest) {
653             tcg_gen_ld_i32(t2, cpu_env, dofs + i);
654         }
655         fni(t2, t0, t1);
656         tcg_gen_st_i32(t2, cpu_env, dofs + i);
657     }
658     tcg_temp_free_i32(t2);
659     tcg_temp_free_i32(t1);
660     tcg_temp_free_i32(t0);
661 }
662 
663 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
664 static void expand_4_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
665                          uint32_t cofs, uint32_t oprsz,
666                          void (*fni)(TCGv_i32, TCGv_i32, TCGv_i32, TCGv_i32))
667 {
668     TCGv_i32 t0 = tcg_temp_new_i32();
669     TCGv_i32 t1 = tcg_temp_new_i32();
670     TCGv_i32 t2 = tcg_temp_new_i32();
671     TCGv_i32 t3 = tcg_temp_new_i32();
672     uint32_t i;
673 
674     for (i = 0; i < oprsz; i += 4) {
675         tcg_gen_ld_i32(t1, cpu_env, aofs + i);
676         tcg_gen_ld_i32(t2, cpu_env, bofs + i);
677         tcg_gen_ld_i32(t3, cpu_env, cofs + i);
678         fni(t0, t1, t2, t3);
679         tcg_gen_st_i32(t0, cpu_env, dofs + i);
680     }
681     tcg_temp_free_i32(t3);
682     tcg_temp_free_i32(t2);
683     tcg_temp_free_i32(t1);
684     tcg_temp_free_i32(t0);
685 }
686 
687 /* Expand OPSZ bytes worth of two-operand operations using i64 elements.  */
688 static void expand_2_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
689                          void (*fni)(TCGv_i64, TCGv_i64))
690 {
691     TCGv_i64 t0 = tcg_temp_new_i64();
692     uint32_t i;
693 
694     for (i = 0; i < oprsz; i += 8) {
695         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
696         fni(t0, t0);
697         tcg_gen_st_i64(t0, cpu_env, dofs + i);
698     }
699     tcg_temp_free_i64(t0);
700 }
701 
702 static void expand_2i_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
703                           int64_t c, bool load_dest,
704                           void (*fni)(TCGv_i64, TCGv_i64, int64_t))
705 {
706     TCGv_i64 t0 = tcg_temp_new_i64();
707     TCGv_i64 t1 = tcg_temp_new_i64();
708     uint32_t i;
709 
710     for (i = 0; i < oprsz; i += 8) {
711         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
712         if (load_dest) {
713             tcg_gen_ld_i64(t1, cpu_env, dofs + i);
714         }
715         fni(t1, t0, c);
716         tcg_gen_st_i64(t1, cpu_env, dofs + i);
717     }
718     tcg_temp_free_i64(t0);
719     tcg_temp_free_i64(t1);
720 }
721 
722 static void expand_2s_i64(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
723                           TCGv_i64 c, bool scalar_first,
724                           void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64))
725 {
726     TCGv_i64 t0 = tcg_temp_new_i64();
727     TCGv_i64 t1 = tcg_temp_new_i64();
728     uint32_t i;
729 
730     for (i = 0; i < oprsz; i += 8) {
731         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
732         if (scalar_first) {
733             fni(t1, c, t0);
734         } else {
735             fni(t1, t0, c);
736         }
737         tcg_gen_st_i64(t1, cpu_env, dofs + i);
738     }
739     tcg_temp_free_i64(t0);
740     tcg_temp_free_i64(t1);
741 }
742 
743 /* Expand OPSZ bytes worth of three-operand operations using i64 elements.  */
744 static void expand_3_i64(uint32_t dofs, uint32_t aofs,
745                          uint32_t bofs, uint32_t oprsz, bool load_dest,
746                          void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64))
747 {
748     TCGv_i64 t0 = tcg_temp_new_i64();
749     TCGv_i64 t1 = tcg_temp_new_i64();
750     TCGv_i64 t2 = tcg_temp_new_i64();
751     uint32_t i;
752 
753     for (i = 0; i < oprsz; i += 8) {
754         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
755         tcg_gen_ld_i64(t1, cpu_env, bofs + i);
756         if (load_dest) {
757             tcg_gen_ld_i64(t2, cpu_env, dofs + i);
758         }
759         fni(t2, t0, t1);
760         tcg_gen_st_i64(t2, cpu_env, dofs + i);
761     }
762     tcg_temp_free_i64(t2);
763     tcg_temp_free_i64(t1);
764     tcg_temp_free_i64(t0);
765 }
766 
767 /* Expand OPSZ bytes worth of three-operand operations using i64 elements.  */
768 static void expand_4_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
769                          uint32_t cofs, uint32_t oprsz,
770                          void (*fni)(TCGv_i64, TCGv_i64, TCGv_i64, TCGv_i64))
771 {
772     TCGv_i64 t0 = tcg_temp_new_i64();
773     TCGv_i64 t1 = tcg_temp_new_i64();
774     TCGv_i64 t2 = tcg_temp_new_i64();
775     TCGv_i64 t3 = tcg_temp_new_i64();
776     uint32_t i;
777 
778     for (i = 0; i < oprsz; i += 8) {
779         tcg_gen_ld_i64(t1, cpu_env, aofs + i);
780         tcg_gen_ld_i64(t2, cpu_env, bofs + i);
781         tcg_gen_ld_i64(t3, cpu_env, cofs + i);
782         fni(t0, t1, t2, t3);
783         tcg_gen_st_i64(t0, cpu_env, dofs + i);
784     }
785     tcg_temp_free_i64(t3);
786     tcg_temp_free_i64(t2);
787     tcg_temp_free_i64(t1);
788     tcg_temp_free_i64(t0);
789 }
790 
791 /* Expand OPSZ bytes worth of two-operand operations using host vectors.  */
792 static void expand_2_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
793                          uint32_t oprsz, uint32_t tysz, TCGType type,
794                          void (*fni)(unsigned, TCGv_vec, TCGv_vec))
795 {
796     TCGv_vec t0 = tcg_temp_new_vec(type);
797     uint32_t i;
798 
799     for (i = 0; i < oprsz; i += tysz) {
800         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
801         fni(vece, t0, t0);
802         tcg_gen_st_vec(t0, cpu_env, dofs + i);
803     }
804     tcg_temp_free_vec(t0);
805 }
806 
807 /* Expand OPSZ bytes worth of two-vector operands and an immediate operand
808    using host vectors.  */
809 static void expand_2i_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
810                           uint32_t oprsz, uint32_t tysz, TCGType type,
811                           int64_t c, bool load_dest,
812                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, int64_t))
813 {
814     TCGv_vec t0 = tcg_temp_new_vec(type);
815     TCGv_vec t1 = tcg_temp_new_vec(type);
816     uint32_t i;
817 
818     for (i = 0; i < oprsz; i += tysz) {
819         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
820         if (load_dest) {
821             tcg_gen_ld_vec(t1, cpu_env, dofs + i);
822         }
823         fni(vece, t1, t0, c);
824         tcg_gen_st_vec(t1, cpu_env, dofs + i);
825     }
826     tcg_temp_free_vec(t0);
827     tcg_temp_free_vec(t1);
828 }
829 
830 static void expand_2s_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
831                           uint32_t oprsz, uint32_t tysz, TCGType type,
832                           TCGv_vec c, bool scalar_first,
833                           void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec))
834 {
835     TCGv_vec t0 = tcg_temp_new_vec(type);
836     TCGv_vec t1 = tcg_temp_new_vec(type);
837     uint32_t i;
838 
839     for (i = 0; i < oprsz; i += tysz) {
840         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
841         if (scalar_first) {
842             fni(vece, t1, c, t0);
843         } else {
844             fni(vece, t1, t0, c);
845         }
846         tcg_gen_st_vec(t1, cpu_env, dofs + i);
847     }
848     tcg_temp_free_vec(t0);
849     tcg_temp_free_vec(t1);
850 }
851 
852 /* Expand OPSZ bytes worth of three-operand operations using host vectors.  */
853 static void expand_3_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
854                          uint32_t bofs, uint32_t oprsz,
855                          uint32_t tysz, TCGType type, bool load_dest,
856                          void (*fni)(unsigned, TCGv_vec, TCGv_vec, TCGv_vec))
857 {
858     TCGv_vec t0 = tcg_temp_new_vec(type);
859     TCGv_vec t1 = tcg_temp_new_vec(type);
860     TCGv_vec t2 = tcg_temp_new_vec(type);
861     uint32_t i;
862 
863     for (i = 0; i < oprsz; i += tysz) {
864         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
865         tcg_gen_ld_vec(t1, cpu_env, bofs + i);
866         if (load_dest) {
867             tcg_gen_ld_vec(t2, cpu_env, dofs + i);
868         }
869         fni(vece, t2, t0, t1);
870         tcg_gen_st_vec(t2, cpu_env, dofs + i);
871     }
872     tcg_temp_free_vec(t2);
873     tcg_temp_free_vec(t1);
874     tcg_temp_free_vec(t0);
875 }
876 
877 /* Expand OPSZ bytes worth of four-operand operations using host vectors.  */
878 static void expand_4_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
879                          uint32_t bofs, uint32_t cofs, uint32_t oprsz,
880                          uint32_t tysz, TCGType type,
881                          void (*fni)(unsigned, TCGv_vec, TCGv_vec,
882                                      TCGv_vec, TCGv_vec))
883 {
884     TCGv_vec t0 = tcg_temp_new_vec(type);
885     TCGv_vec t1 = tcg_temp_new_vec(type);
886     TCGv_vec t2 = tcg_temp_new_vec(type);
887     TCGv_vec t3 = tcg_temp_new_vec(type);
888     uint32_t i;
889 
890     for (i = 0; i < oprsz; i += tysz) {
891         tcg_gen_ld_vec(t1, cpu_env, aofs + i);
892         tcg_gen_ld_vec(t2, cpu_env, bofs + i);
893         tcg_gen_ld_vec(t3, cpu_env, cofs + i);
894         fni(vece, t0, t1, t2, t3);
895         tcg_gen_st_vec(t0, cpu_env, dofs + i);
896     }
897     tcg_temp_free_vec(t3);
898     tcg_temp_free_vec(t2);
899     tcg_temp_free_vec(t1);
900     tcg_temp_free_vec(t0);
901 }
902 
903 /* Expand a vector two-operand operation.  */
904 void tcg_gen_gvec_2(uint32_t dofs, uint32_t aofs,
905                     uint32_t oprsz, uint32_t maxsz, const GVecGen2 *g)
906 {
907     TCGType type;
908     uint32_t some;
909 
910     check_size_align(oprsz, maxsz, dofs | aofs);
911     check_overlap_2(dofs, aofs, maxsz);
912 
913     type = 0;
914     if (g->fniv) {
915         type = choose_vector_type(g->opc, g->vece, oprsz, g->prefer_i64);
916     }
917     switch (type) {
918     case TCG_TYPE_V256:
919         /* Recall that ARM SVE allows vector sizes that are not a
920          * power of 2, but always a multiple of 16.  The intent is
921          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
922          */
923         some = QEMU_ALIGN_DOWN(oprsz, 32);
924         expand_2_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256, g->fniv);
925         if (some == oprsz) {
926             break;
927         }
928         dofs += some;
929         aofs += some;
930         oprsz -= some;
931         maxsz -= some;
932         /* fallthru */
933     case TCG_TYPE_V128:
934         expand_2_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128, g->fniv);
935         break;
936     case TCG_TYPE_V64:
937         expand_2_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64, g->fniv);
938         break;
939 
940     case 0:
941         if (g->fni8 && check_size_impl(oprsz, 8)) {
942             expand_2_i64(dofs, aofs, oprsz, g->fni8);
943         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
944             expand_2_i32(dofs, aofs, oprsz, g->fni4);
945         } else {
946             assert(g->fno != NULL);
947             tcg_gen_gvec_2_ool(dofs, aofs, oprsz, maxsz, g->data, g->fno);
948             return;
949         }
950         break;
951 
952     default:
953         g_assert_not_reached();
954     }
955 
956     if (oprsz < maxsz) {
957         expand_clr(dofs + oprsz, maxsz - oprsz);
958     }
959 }
960 
961 /* Expand a vector operation with two vectors and an immediate.  */
962 void tcg_gen_gvec_2i(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
963                      uint32_t maxsz, int64_t c, const GVecGen2i *g)
964 {
965     TCGType type;
966     uint32_t some;
967 
968     check_size_align(oprsz, maxsz, dofs | aofs);
969     check_overlap_2(dofs, aofs, maxsz);
970 
971     type = 0;
972     if (g->fniv) {
973         type = choose_vector_type(g->opc, g->vece, oprsz, g->prefer_i64);
974     }
975     switch (type) {
976     case TCG_TYPE_V256:
977         /* Recall that ARM SVE allows vector sizes that are not a
978          * power of 2, but always a multiple of 16.  The intent is
979          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
980          */
981         some = QEMU_ALIGN_DOWN(oprsz, 32);
982         expand_2i_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
983                       c, g->load_dest, g->fniv);
984         if (some == oprsz) {
985             break;
986         }
987         dofs += some;
988         aofs += some;
989         oprsz -= some;
990         maxsz -= some;
991         /* fallthru */
992     case TCG_TYPE_V128:
993         expand_2i_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
994                       c, g->load_dest, g->fniv);
995         break;
996     case TCG_TYPE_V64:
997         expand_2i_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
998                       c, g->load_dest, g->fniv);
999         break;
1000 
1001     case 0:
1002         if (g->fni8 && check_size_impl(oprsz, 8)) {
1003             expand_2i_i64(dofs, aofs, oprsz, c, g->load_dest, g->fni8);
1004         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1005             expand_2i_i32(dofs, aofs, oprsz, c, g->load_dest, g->fni4);
1006         } else {
1007             if (g->fno) {
1008                 tcg_gen_gvec_2_ool(dofs, aofs, oprsz, maxsz, c, g->fno);
1009             } else {
1010                 TCGv_i64 tcg_c = tcg_const_i64(c);
1011                 tcg_gen_gvec_2i_ool(dofs, aofs, tcg_c, oprsz,
1012                                     maxsz, c, g->fnoi);
1013                 tcg_temp_free_i64(tcg_c);
1014             }
1015             return;
1016         }
1017         break;
1018 
1019     default:
1020         g_assert_not_reached();
1021     }
1022 
1023     if (oprsz < maxsz) {
1024         expand_clr(dofs + oprsz, maxsz - oprsz);
1025     }
1026 }
1027 
1028 /* Expand a vector operation with two vectors and a scalar.  */
1029 void tcg_gen_gvec_2s(uint32_t dofs, uint32_t aofs, uint32_t oprsz,
1030                      uint32_t maxsz, TCGv_i64 c, const GVecGen2s *g)
1031 {
1032     TCGType type;
1033 
1034     check_size_align(oprsz, maxsz, dofs | aofs);
1035     check_overlap_2(dofs, aofs, maxsz);
1036 
1037     type = 0;
1038     if (g->fniv) {
1039         type = choose_vector_type(g->opc, g->vece, oprsz, g->prefer_i64);
1040     }
1041     if (type != 0) {
1042         TCGv_vec t_vec = tcg_temp_new_vec(type);
1043         uint32_t some;
1044 
1045         tcg_gen_dup_i64_vec(g->vece, t_vec, c);
1046 
1047         switch (type) {
1048         case TCG_TYPE_V256:
1049             /* Recall that ARM SVE allows vector sizes that are not a
1050              * power of 2, but always a multiple of 16.  The intent is
1051              * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1052              */
1053             some = QEMU_ALIGN_DOWN(oprsz, 32);
1054             expand_2s_vec(g->vece, dofs, aofs, some, 32, TCG_TYPE_V256,
1055                           t_vec, g->scalar_first, g->fniv);
1056             if (some == oprsz) {
1057                 break;
1058             }
1059             dofs += some;
1060             aofs += some;
1061             oprsz -= some;
1062             maxsz -= some;
1063             /* fallthru */
1064 
1065         case TCG_TYPE_V128:
1066             expand_2s_vec(g->vece, dofs, aofs, oprsz, 16, TCG_TYPE_V128,
1067                           t_vec, g->scalar_first, g->fniv);
1068             break;
1069 
1070         case TCG_TYPE_V64:
1071             expand_2s_vec(g->vece, dofs, aofs, oprsz, 8, TCG_TYPE_V64,
1072                           t_vec, g->scalar_first, g->fniv);
1073             break;
1074 
1075         default:
1076             g_assert_not_reached();
1077         }
1078         tcg_temp_free_vec(t_vec);
1079     } else if (g->fni8 && check_size_impl(oprsz, 8)) {
1080         TCGv_i64 t64 = tcg_temp_new_i64();
1081 
1082         gen_dup_i64(g->vece, t64, c);
1083         expand_2s_i64(dofs, aofs, oprsz, t64, g->scalar_first, g->fni8);
1084         tcg_temp_free_i64(t64);
1085     } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1086         TCGv_i32 t32 = tcg_temp_new_i32();
1087 
1088         tcg_gen_extrl_i64_i32(t32, c);
1089         gen_dup_i32(g->vece, t32, t32);
1090         expand_2s_i32(dofs, aofs, oprsz, t32, g->scalar_first, g->fni4);
1091         tcg_temp_free_i32(t32);
1092     } else {
1093         tcg_gen_gvec_2i_ool(dofs, aofs, c, oprsz, maxsz, 0, g->fno);
1094         return;
1095     }
1096 
1097     if (oprsz < maxsz) {
1098         expand_clr(dofs + oprsz, maxsz - oprsz);
1099     }
1100 }
1101 
1102 /* Expand a vector three-operand operation.  */
1103 void tcg_gen_gvec_3(uint32_t dofs, uint32_t aofs, uint32_t bofs,
1104                     uint32_t oprsz, uint32_t maxsz, const GVecGen3 *g)
1105 {
1106     TCGType type;
1107     uint32_t some;
1108 
1109     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
1110     check_overlap_3(dofs, aofs, bofs, maxsz);
1111 
1112     type = 0;
1113     if (g->fniv) {
1114         type = choose_vector_type(g->opc, g->vece, oprsz, g->prefer_i64);
1115     }
1116     switch (type) {
1117     case TCG_TYPE_V256:
1118         /* Recall that ARM SVE allows vector sizes that are not a
1119          * power of 2, but always a multiple of 16.  The intent is
1120          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1121          */
1122         some = QEMU_ALIGN_DOWN(oprsz, 32);
1123         expand_3_vec(g->vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256,
1124                      g->load_dest, g->fniv);
1125         if (some == oprsz) {
1126             break;
1127         }
1128         dofs += some;
1129         aofs += some;
1130         bofs += some;
1131         oprsz -= some;
1132         maxsz -= some;
1133         /* fallthru */
1134     case TCG_TYPE_V128:
1135         expand_3_vec(g->vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128,
1136                      g->load_dest, g->fniv);
1137         break;
1138     case TCG_TYPE_V64:
1139         expand_3_vec(g->vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64,
1140                      g->load_dest, g->fniv);
1141         break;
1142 
1143     case 0:
1144         if (g->fni8 && check_size_impl(oprsz, 8)) {
1145             expand_3_i64(dofs, aofs, bofs, oprsz, g->load_dest, g->fni8);
1146         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1147             expand_3_i32(dofs, aofs, bofs, oprsz, g->load_dest, g->fni4);
1148         } else {
1149             assert(g->fno != NULL);
1150             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz,
1151                                maxsz, g->data, g->fno);
1152             return;
1153         }
1154         break;
1155 
1156     default:
1157         g_assert_not_reached();
1158     }
1159 
1160     if (oprsz < maxsz) {
1161         expand_clr(dofs + oprsz, maxsz - oprsz);
1162     }
1163 }
1164 
1165 /* Expand a vector four-operand operation.  */
1166 void tcg_gen_gvec_4(uint32_t dofs, uint32_t aofs, uint32_t bofs, uint32_t cofs,
1167                     uint32_t oprsz, uint32_t maxsz, const GVecGen4 *g)
1168 {
1169     TCGType type;
1170     uint32_t some;
1171 
1172     check_size_align(oprsz, maxsz, dofs | aofs | bofs | cofs);
1173     check_overlap_4(dofs, aofs, bofs, cofs, maxsz);
1174 
1175     type = 0;
1176     if (g->fniv) {
1177         type = choose_vector_type(g->opc, g->vece, oprsz, g->prefer_i64);
1178     }
1179     switch (type) {
1180     case TCG_TYPE_V256:
1181         /* Recall that ARM SVE allows vector sizes that are not a
1182          * power of 2, but always a multiple of 16.  The intent is
1183          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
1184          */
1185         some = QEMU_ALIGN_DOWN(oprsz, 32);
1186         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, some,
1187                      32, TCG_TYPE_V256, g->fniv);
1188         if (some == oprsz) {
1189             break;
1190         }
1191         dofs += some;
1192         aofs += some;
1193         bofs += some;
1194         cofs += some;
1195         oprsz -= some;
1196         maxsz -= some;
1197         /* fallthru */
1198     case TCG_TYPE_V128:
1199         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1200                      16, TCG_TYPE_V128, g->fniv);
1201         break;
1202     case TCG_TYPE_V64:
1203         expand_4_vec(g->vece, dofs, aofs, bofs, cofs, oprsz,
1204                      8, TCG_TYPE_V64, g->fniv);
1205         break;
1206 
1207     case 0:
1208         if (g->fni8 && check_size_impl(oprsz, 8)) {
1209             expand_4_i64(dofs, aofs, bofs, cofs, oprsz, g->fni8);
1210         } else if (g->fni4 && check_size_impl(oprsz, 4)) {
1211             expand_4_i32(dofs, aofs, bofs, cofs, oprsz, g->fni4);
1212         } else {
1213             assert(g->fno != NULL);
1214             tcg_gen_gvec_4_ool(dofs, aofs, bofs, cofs,
1215                                oprsz, maxsz, g->data, g->fno);
1216             return;
1217         }
1218         break;
1219 
1220     default:
1221         g_assert_not_reached();
1222     }
1223 
1224     if (oprsz < maxsz) {
1225         expand_clr(dofs + oprsz, maxsz - oprsz);
1226     }
1227 }
1228 
1229 /*
1230  * Expand specific vector operations.
1231  */
1232 
1233 static void vec_mov2(unsigned vece, TCGv_vec a, TCGv_vec b)
1234 {
1235     tcg_gen_mov_vec(a, b);
1236 }
1237 
1238 void tcg_gen_gvec_mov(unsigned vece, uint32_t dofs, uint32_t aofs,
1239                       uint32_t oprsz, uint32_t maxsz)
1240 {
1241     static const GVecGen2 g = {
1242         .fni8 = tcg_gen_mov_i64,
1243         .fniv = vec_mov2,
1244         .fno = gen_helper_gvec_mov,
1245         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1246     };
1247     if (dofs != aofs) {
1248         tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g);
1249     } else {
1250         check_size_align(oprsz, maxsz, dofs);
1251         if (oprsz < maxsz) {
1252             expand_clr(dofs + oprsz, maxsz - oprsz);
1253         }
1254     }
1255 }
1256 
1257 void tcg_gen_gvec_dup_i32(unsigned vece, uint32_t dofs, uint32_t oprsz,
1258                           uint32_t maxsz, TCGv_i32 in)
1259 {
1260     check_size_align(oprsz, maxsz, dofs);
1261     tcg_debug_assert(vece <= MO_32);
1262     do_dup(vece, dofs, oprsz, maxsz, in, NULL, 0);
1263 }
1264 
1265 void tcg_gen_gvec_dup_i64(unsigned vece, uint32_t dofs, uint32_t oprsz,
1266                           uint32_t maxsz, TCGv_i64 in)
1267 {
1268     check_size_align(oprsz, maxsz, dofs);
1269     tcg_debug_assert(vece <= MO_64);
1270     do_dup(vece, dofs, oprsz, maxsz, NULL, in, 0);
1271 }
1272 
1273 void tcg_gen_gvec_dup_mem(unsigned vece, uint32_t dofs, uint32_t aofs,
1274                           uint32_t oprsz, uint32_t maxsz)
1275 {
1276     if (vece <= MO_32) {
1277         TCGv_i32 in = tcg_temp_new_i32();
1278         switch (vece) {
1279         case MO_8:
1280             tcg_gen_ld8u_i32(in, cpu_env, aofs);
1281             break;
1282         case MO_16:
1283             tcg_gen_ld16u_i32(in, cpu_env, aofs);
1284             break;
1285         case MO_32:
1286             tcg_gen_ld_i32(in, cpu_env, aofs);
1287             break;
1288         }
1289         tcg_gen_gvec_dup_i32(vece, dofs, oprsz, maxsz, in);
1290         tcg_temp_free_i32(in);
1291     } else if (vece == MO_64) {
1292         TCGv_i64 in = tcg_temp_new_i64();
1293         tcg_gen_ld_i64(in, cpu_env, aofs);
1294         tcg_gen_gvec_dup_i64(MO_64, dofs, oprsz, maxsz, in);
1295         tcg_temp_free_i64(in);
1296     } else {
1297         /* 128-bit duplicate.  */
1298         /* ??? Dup to 256-bit vector.  */
1299         int i;
1300 
1301         tcg_debug_assert(vece == 4);
1302         tcg_debug_assert(oprsz >= 16);
1303         if (TCG_TARGET_HAS_v128) {
1304             TCGv_vec in = tcg_temp_new_vec(TCG_TYPE_V128);
1305 
1306             tcg_gen_ld_vec(in, cpu_env, aofs);
1307             for (i = 0; i < oprsz; i += 16) {
1308                 tcg_gen_st_vec(in, cpu_env, dofs + i);
1309             }
1310             tcg_temp_free_vec(in);
1311         } else {
1312             TCGv_i64 in0 = tcg_temp_new_i64();
1313             TCGv_i64 in1 = tcg_temp_new_i64();
1314 
1315             tcg_gen_ld_i64(in0, cpu_env, aofs);
1316             tcg_gen_ld_i64(in1, cpu_env, aofs + 8);
1317             for (i = 0; i < oprsz; i += 16) {
1318                 tcg_gen_st_i64(in0, cpu_env, dofs + i);
1319                 tcg_gen_st_i64(in1, cpu_env, dofs + i + 8);
1320             }
1321             tcg_temp_free_i64(in0);
1322             tcg_temp_free_i64(in1);
1323         }
1324     }
1325 }
1326 
1327 void tcg_gen_gvec_dup64i(uint32_t dofs, uint32_t oprsz,
1328                          uint32_t maxsz, uint64_t x)
1329 {
1330     check_size_align(oprsz, maxsz, dofs);
1331     do_dup(MO_64, dofs, oprsz, maxsz, NULL, NULL, x);
1332 }
1333 
1334 void tcg_gen_gvec_dup32i(uint32_t dofs, uint32_t oprsz,
1335                          uint32_t maxsz, uint32_t x)
1336 {
1337     check_size_align(oprsz, maxsz, dofs);
1338     do_dup(MO_32, dofs, oprsz, maxsz, NULL, NULL, x);
1339 }
1340 
1341 void tcg_gen_gvec_dup16i(uint32_t dofs, uint32_t oprsz,
1342                          uint32_t maxsz, uint16_t x)
1343 {
1344     check_size_align(oprsz, maxsz, dofs);
1345     do_dup(MO_16, dofs, oprsz, maxsz, NULL, NULL, x);
1346 }
1347 
1348 void tcg_gen_gvec_dup8i(uint32_t dofs, uint32_t oprsz,
1349                          uint32_t maxsz, uint8_t x)
1350 {
1351     check_size_align(oprsz, maxsz, dofs);
1352     do_dup(MO_8, dofs, oprsz, maxsz, NULL, NULL, x);
1353 }
1354 
1355 void tcg_gen_gvec_not(unsigned vece, uint32_t dofs, uint32_t aofs,
1356                       uint32_t oprsz, uint32_t maxsz)
1357 {
1358     static const GVecGen2 g = {
1359         .fni8 = tcg_gen_not_i64,
1360         .fniv = tcg_gen_not_vec,
1361         .fno = gen_helper_gvec_not,
1362         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1363     };
1364     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g);
1365 }
1366 
1367 /* Perform a vector addition using normal addition and a mask.  The mask
1368    should be the sign bit of each lane.  This 6-operation form is more
1369    efficient than separate additions when there are 4 or more lanes in
1370    the 64-bit operation.  */
1371 static void gen_addv_mask(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 m)
1372 {
1373     TCGv_i64 t1 = tcg_temp_new_i64();
1374     TCGv_i64 t2 = tcg_temp_new_i64();
1375     TCGv_i64 t3 = tcg_temp_new_i64();
1376 
1377     tcg_gen_andc_i64(t1, a, m);
1378     tcg_gen_andc_i64(t2, b, m);
1379     tcg_gen_xor_i64(t3, a, b);
1380     tcg_gen_add_i64(d, t1, t2);
1381     tcg_gen_and_i64(t3, t3, m);
1382     tcg_gen_xor_i64(d, d, t3);
1383 
1384     tcg_temp_free_i64(t1);
1385     tcg_temp_free_i64(t2);
1386     tcg_temp_free_i64(t3);
1387 }
1388 
1389 void tcg_gen_vec_add8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1390 {
1391     TCGv_i64 m = tcg_const_i64(dup_const(MO_8, 0x80));
1392     gen_addv_mask(d, a, b, m);
1393     tcg_temp_free_i64(m);
1394 }
1395 
1396 void tcg_gen_vec_add16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1397 {
1398     TCGv_i64 m = tcg_const_i64(dup_const(MO_16, 0x8000));
1399     gen_addv_mask(d, a, b, m);
1400     tcg_temp_free_i64(m);
1401 }
1402 
1403 void tcg_gen_vec_add32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1404 {
1405     TCGv_i64 t1 = tcg_temp_new_i64();
1406     TCGv_i64 t2 = tcg_temp_new_i64();
1407 
1408     tcg_gen_andi_i64(t1, a, ~0xffffffffull);
1409     tcg_gen_add_i64(t2, a, b);
1410     tcg_gen_add_i64(t1, t1, b);
1411     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
1412 
1413     tcg_temp_free_i64(t1);
1414     tcg_temp_free_i64(t2);
1415 }
1416 
1417 void tcg_gen_gvec_add(unsigned vece, uint32_t dofs, uint32_t aofs,
1418                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1419 {
1420     static const GVecGen3 g[4] = {
1421         { .fni8 = tcg_gen_vec_add8_i64,
1422           .fniv = tcg_gen_add_vec,
1423           .fno = gen_helper_gvec_add8,
1424           .opc = INDEX_op_add_vec,
1425           .vece = MO_8 },
1426         { .fni8 = tcg_gen_vec_add16_i64,
1427           .fniv = tcg_gen_add_vec,
1428           .fno = gen_helper_gvec_add16,
1429           .opc = INDEX_op_add_vec,
1430           .vece = MO_16 },
1431         { .fni4 = tcg_gen_add_i32,
1432           .fniv = tcg_gen_add_vec,
1433           .fno = gen_helper_gvec_add32,
1434           .opc = INDEX_op_add_vec,
1435           .vece = MO_32 },
1436         { .fni8 = tcg_gen_add_i64,
1437           .fniv = tcg_gen_add_vec,
1438           .fno = gen_helper_gvec_add64,
1439           .opc = INDEX_op_add_vec,
1440           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1441           .vece = MO_64 },
1442     };
1443 
1444     tcg_debug_assert(vece <= MO_64);
1445     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1446 }
1447 
1448 void tcg_gen_gvec_adds(unsigned vece, uint32_t dofs, uint32_t aofs,
1449                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
1450 {
1451     static const GVecGen2s g[4] = {
1452         { .fni8 = tcg_gen_vec_add8_i64,
1453           .fniv = tcg_gen_add_vec,
1454           .fno = gen_helper_gvec_adds8,
1455           .opc = INDEX_op_add_vec,
1456           .vece = MO_8 },
1457         { .fni8 = tcg_gen_vec_add16_i64,
1458           .fniv = tcg_gen_add_vec,
1459           .fno = gen_helper_gvec_adds16,
1460           .opc = INDEX_op_add_vec,
1461           .vece = MO_16 },
1462         { .fni4 = tcg_gen_add_i32,
1463           .fniv = tcg_gen_add_vec,
1464           .fno = gen_helper_gvec_adds32,
1465           .opc = INDEX_op_add_vec,
1466           .vece = MO_32 },
1467         { .fni8 = tcg_gen_add_i64,
1468           .fniv = tcg_gen_add_vec,
1469           .fno = gen_helper_gvec_adds64,
1470           .opc = INDEX_op_add_vec,
1471           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1472           .vece = MO_64 },
1473     };
1474 
1475     tcg_debug_assert(vece <= MO_64);
1476     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
1477 }
1478 
1479 void tcg_gen_gvec_addi(unsigned vece, uint32_t dofs, uint32_t aofs,
1480                        int64_t c, uint32_t oprsz, uint32_t maxsz)
1481 {
1482     TCGv_i64 tmp = tcg_const_i64(c);
1483     tcg_gen_gvec_adds(vece, dofs, aofs, tmp, oprsz, maxsz);
1484     tcg_temp_free_i64(tmp);
1485 }
1486 
1487 void tcg_gen_gvec_subs(unsigned vece, uint32_t dofs, uint32_t aofs,
1488                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
1489 {
1490     static const GVecGen2s g[4] = {
1491         { .fni8 = tcg_gen_vec_sub8_i64,
1492           .fniv = tcg_gen_sub_vec,
1493           .fno = gen_helper_gvec_subs8,
1494           .opc = INDEX_op_sub_vec,
1495           .vece = MO_8 },
1496         { .fni8 = tcg_gen_vec_sub16_i64,
1497           .fniv = tcg_gen_sub_vec,
1498           .fno = gen_helper_gvec_subs16,
1499           .opc = INDEX_op_sub_vec,
1500           .vece = MO_16 },
1501         { .fni4 = tcg_gen_sub_i32,
1502           .fniv = tcg_gen_sub_vec,
1503           .fno = gen_helper_gvec_subs32,
1504           .opc = INDEX_op_sub_vec,
1505           .vece = MO_32 },
1506         { .fni8 = tcg_gen_sub_i64,
1507           .fniv = tcg_gen_sub_vec,
1508           .fno = gen_helper_gvec_subs64,
1509           .opc = INDEX_op_sub_vec,
1510           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1511           .vece = MO_64 },
1512     };
1513 
1514     tcg_debug_assert(vece <= MO_64);
1515     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
1516 }
1517 
1518 /* Perform a vector subtraction using normal subtraction and a mask.
1519    Compare gen_addv_mask above.  */
1520 static void gen_subv_mask(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b, TCGv_i64 m)
1521 {
1522     TCGv_i64 t1 = tcg_temp_new_i64();
1523     TCGv_i64 t2 = tcg_temp_new_i64();
1524     TCGv_i64 t3 = tcg_temp_new_i64();
1525 
1526     tcg_gen_or_i64(t1, a, m);
1527     tcg_gen_andc_i64(t2, b, m);
1528     tcg_gen_eqv_i64(t3, a, b);
1529     tcg_gen_sub_i64(d, t1, t2);
1530     tcg_gen_and_i64(t3, t3, m);
1531     tcg_gen_xor_i64(d, d, t3);
1532 
1533     tcg_temp_free_i64(t1);
1534     tcg_temp_free_i64(t2);
1535     tcg_temp_free_i64(t3);
1536 }
1537 
1538 void tcg_gen_vec_sub8_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1539 {
1540     TCGv_i64 m = tcg_const_i64(dup_const(MO_8, 0x80));
1541     gen_subv_mask(d, a, b, m);
1542     tcg_temp_free_i64(m);
1543 }
1544 
1545 void tcg_gen_vec_sub16_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1546 {
1547     TCGv_i64 m = tcg_const_i64(dup_const(MO_16, 0x8000));
1548     gen_subv_mask(d, a, b, m);
1549     tcg_temp_free_i64(m);
1550 }
1551 
1552 void tcg_gen_vec_sub32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1553 {
1554     TCGv_i64 t1 = tcg_temp_new_i64();
1555     TCGv_i64 t2 = tcg_temp_new_i64();
1556 
1557     tcg_gen_andi_i64(t1, b, ~0xffffffffull);
1558     tcg_gen_sub_i64(t2, a, b);
1559     tcg_gen_sub_i64(t1, a, t1);
1560     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
1561 
1562     tcg_temp_free_i64(t1);
1563     tcg_temp_free_i64(t2);
1564 }
1565 
1566 void tcg_gen_gvec_sub(unsigned vece, uint32_t dofs, uint32_t aofs,
1567                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1568 {
1569     static const GVecGen3 g[4] = {
1570         { .fni8 = tcg_gen_vec_sub8_i64,
1571           .fniv = tcg_gen_sub_vec,
1572           .fno = gen_helper_gvec_sub8,
1573           .opc = INDEX_op_sub_vec,
1574           .vece = MO_8 },
1575         { .fni8 = tcg_gen_vec_sub16_i64,
1576           .fniv = tcg_gen_sub_vec,
1577           .fno = gen_helper_gvec_sub16,
1578           .opc = INDEX_op_sub_vec,
1579           .vece = MO_16 },
1580         { .fni4 = tcg_gen_sub_i32,
1581           .fniv = tcg_gen_sub_vec,
1582           .fno = gen_helper_gvec_sub32,
1583           .opc = INDEX_op_sub_vec,
1584           .vece = MO_32 },
1585         { .fni8 = tcg_gen_sub_i64,
1586           .fniv = tcg_gen_sub_vec,
1587           .fno = gen_helper_gvec_sub64,
1588           .opc = INDEX_op_sub_vec,
1589           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1590           .vece = MO_64 },
1591     };
1592 
1593     tcg_debug_assert(vece <= MO_64);
1594     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1595 }
1596 
1597 void tcg_gen_gvec_mul(unsigned vece, uint32_t dofs, uint32_t aofs,
1598                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1599 {
1600     static const GVecGen3 g[4] = {
1601         { .fniv = tcg_gen_mul_vec,
1602           .fno = gen_helper_gvec_mul8,
1603           .opc = INDEX_op_mul_vec,
1604           .vece = MO_8 },
1605         { .fniv = tcg_gen_mul_vec,
1606           .fno = gen_helper_gvec_mul16,
1607           .opc = INDEX_op_mul_vec,
1608           .vece = MO_16 },
1609         { .fni4 = tcg_gen_mul_i32,
1610           .fniv = tcg_gen_mul_vec,
1611           .fno = gen_helper_gvec_mul32,
1612           .opc = INDEX_op_mul_vec,
1613           .vece = MO_32 },
1614         { .fni8 = tcg_gen_mul_i64,
1615           .fniv = tcg_gen_mul_vec,
1616           .fno = gen_helper_gvec_mul64,
1617           .opc = INDEX_op_mul_vec,
1618           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1619           .vece = MO_64 },
1620     };
1621 
1622     tcg_debug_assert(vece <= MO_64);
1623     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1624 }
1625 
1626 void tcg_gen_gvec_muls(unsigned vece, uint32_t dofs, uint32_t aofs,
1627                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
1628 {
1629     static const GVecGen2s g[4] = {
1630         { .fniv = tcg_gen_mul_vec,
1631           .fno = gen_helper_gvec_muls8,
1632           .opc = INDEX_op_mul_vec,
1633           .vece = MO_8 },
1634         { .fniv = tcg_gen_mul_vec,
1635           .fno = gen_helper_gvec_muls16,
1636           .opc = INDEX_op_mul_vec,
1637           .vece = MO_16 },
1638         { .fni4 = tcg_gen_mul_i32,
1639           .fniv = tcg_gen_mul_vec,
1640           .fno = gen_helper_gvec_muls32,
1641           .opc = INDEX_op_mul_vec,
1642           .vece = MO_32 },
1643         { .fni8 = tcg_gen_mul_i64,
1644           .fniv = tcg_gen_mul_vec,
1645           .fno = gen_helper_gvec_muls64,
1646           .opc = INDEX_op_mul_vec,
1647           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1648           .vece = MO_64 },
1649     };
1650 
1651     tcg_debug_assert(vece <= MO_64);
1652     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, c, &g[vece]);
1653 }
1654 
1655 void tcg_gen_gvec_muli(unsigned vece, uint32_t dofs, uint32_t aofs,
1656                        int64_t c, uint32_t oprsz, uint32_t maxsz)
1657 {
1658     TCGv_i64 tmp = tcg_const_i64(c);
1659     tcg_gen_gvec_muls(vece, dofs, aofs, tmp, oprsz, maxsz);
1660     tcg_temp_free_i64(tmp);
1661 }
1662 
1663 void tcg_gen_gvec_ssadd(unsigned vece, uint32_t dofs, uint32_t aofs,
1664                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1665 {
1666     static const GVecGen3 g[4] = {
1667         { .fno = gen_helper_gvec_ssadd8, .vece = MO_8 },
1668         { .fno = gen_helper_gvec_ssadd16, .vece = MO_16 },
1669         { .fno = gen_helper_gvec_ssadd32, .vece = MO_32 },
1670         { .fno = gen_helper_gvec_ssadd64, .vece = MO_64 }
1671     };
1672     tcg_debug_assert(vece <= MO_64);
1673     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1674 }
1675 
1676 void tcg_gen_gvec_sssub(unsigned vece, uint32_t dofs, uint32_t aofs,
1677                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1678 {
1679     static const GVecGen3 g[4] = {
1680         { .fno = gen_helper_gvec_sssub8, .vece = MO_8 },
1681         { .fno = gen_helper_gvec_sssub16, .vece = MO_16 },
1682         { .fno = gen_helper_gvec_sssub32, .vece = MO_32 },
1683         { .fno = gen_helper_gvec_sssub64, .vece = MO_64 }
1684     };
1685     tcg_debug_assert(vece <= MO_64);
1686     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1687 }
1688 
1689 static void tcg_gen_vec_usadd32_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
1690 {
1691     TCGv_i32 max = tcg_const_i32(-1);
1692     tcg_gen_add_i32(d, a, b);
1693     tcg_gen_movcond_i32(TCG_COND_LTU, d, d, a, max, d);
1694     tcg_temp_free_i32(max);
1695 }
1696 
1697 static void tcg_gen_vec_usadd32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1698 {
1699     TCGv_i64 max = tcg_const_i64(-1);
1700     tcg_gen_add_i64(d, a, b);
1701     tcg_gen_movcond_i64(TCG_COND_LTU, d, d, a, max, d);
1702     tcg_temp_free_i64(max);
1703 }
1704 
1705 void tcg_gen_gvec_usadd(unsigned vece, uint32_t dofs, uint32_t aofs,
1706                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1707 {
1708     static const GVecGen3 g[4] = {
1709         { .fno = gen_helper_gvec_usadd8, .vece = MO_8 },
1710         { .fno = gen_helper_gvec_usadd16, .vece = MO_16 },
1711         { .fni4 = tcg_gen_vec_usadd32_i32,
1712           .fno = gen_helper_gvec_usadd32,
1713           .vece = MO_32 },
1714         { .fni8 = tcg_gen_vec_usadd32_i64,
1715           .fno = gen_helper_gvec_usadd64,
1716           .vece = MO_64 }
1717     };
1718     tcg_debug_assert(vece <= MO_64);
1719     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1720 }
1721 
1722 static void tcg_gen_vec_ussub32_i32(TCGv_i32 d, TCGv_i32 a, TCGv_i32 b)
1723 {
1724     TCGv_i32 min = tcg_const_i32(0);
1725     tcg_gen_sub_i32(d, a, b);
1726     tcg_gen_movcond_i32(TCG_COND_LTU, d, a, b, min, d);
1727     tcg_temp_free_i32(min);
1728 }
1729 
1730 static void tcg_gen_vec_ussub32_i64(TCGv_i64 d, TCGv_i64 a, TCGv_i64 b)
1731 {
1732     TCGv_i64 min = tcg_const_i64(0);
1733     tcg_gen_sub_i64(d, a, b);
1734     tcg_gen_movcond_i64(TCG_COND_LTU, d, a, b, min, d);
1735     tcg_temp_free_i64(min);
1736 }
1737 
1738 void tcg_gen_gvec_ussub(unsigned vece, uint32_t dofs, uint32_t aofs,
1739                         uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1740 {
1741     static const GVecGen3 g[4] = {
1742         { .fno = gen_helper_gvec_ussub8, .vece = MO_8 },
1743         { .fno = gen_helper_gvec_ussub16, .vece = MO_16 },
1744         { .fni4 = tcg_gen_vec_ussub32_i32,
1745           .fno = gen_helper_gvec_ussub32,
1746           .vece = MO_32 },
1747         { .fni8 = tcg_gen_vec_ussub32_i64,
1748           .fno = gen_helper_gvec_ussub64,
1749           .vece = MO_64 }
1750     };
1751     tcg_debug_assert(vece <= MO_64);
1752     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g[vece]);
1753 }
1754 
1755 /* Perform a vector negation using normal negation and a mask.
1756    Compare gen_subv_mask above.  */
1757 static void gen_negv_mask(TCGv_i64 d, TCGv_i64 b, TCGv_i64 m)
1758 {
1759     TCGv_i64 t2 = tcg_temp_new_i64();
1760     TCGv_i64 t3 = tcg_temp_new_i64();
1761 
1762     tcg_gen_andc_i64(t3, m, b);
1763     tcg_gen_andc_i64(t2, b, m);
1764     tcg_gen_sub_i64(d, m, t2);
1765     tcg_gen_xor_i64(d, d, t3);
1766 
1767     tcg_temp_free_i64(t2);
1768     tcg_temp_free_i64(t3);
1769 }
1770 
1771 void tcg_gen_vec_neg8_i64(TCGv_i64 d, TCGv_i64 b)
1772 {
1773     TCGv_i64 m = tcg_const_i64(dup_const(MO_8, 0x80));
1774     gen_negv_mask(d, b, m);
1775     tcg_temp_free_i64(m);
1776 }
1777 
1778 void tcg_gen_vec_neg16_i64(TCGv_i64 d, TCGv_i64 b)
1779 {
1780     TCGv_i64 m = tcg_const_i64(dup_const(MO_16, 0x8000));
1781     gen_negv_mask(d, b, m);
1782     tcg_temp_free_i64(m);
1783 }
1784 
1785 void tcg_gen_vec_neg32_i64(TCGv_i64 d, TCGv_i64 b)
1786 {
1787     TCGv_i64 t1 = tcg_temp_new_i64();
1788     TCGv_i64 t2 = tcg_temp_new_i64();
1789 
1790     tcg_gen_andi_i64(t1, b, ~0xffffffffull);
1791     tcg_gen_neg_i64(t2, b);
1792     tcg_gen_neg_i64(t1, t1);
1793     tcg_gen_deposit_i64(d, t1, t2, 0, 32);
1794 
1795     tcg_temp_free_i64(t1);
1796     tcg_temp_free_i64(t2);
1797 }
1798 
1799 void tcg_gen_gvec_neg(unsigned vece, uint32_t dofs, uint32_t aofs,
1800                       uint32_t oprsz, uint32_t maxsz)
1801 {
1802     static const GVecGen2 g[4] = {
1803         { .fni8 = tcg_gen_vec_neg8_i64,
1804           .fniv = tcg_gen_neg_vec,
1805           .fno = gen_helper_gvec_neg8,
1806           .opc = INDEX_op_neg_vec,
1807           .vece = MO_8 },
1808         { .fni8 = tcg_gen_vec_neg16_i64,
1809           .fniv = tcg_gen_neg_vec,
1810           .fno = gen_helper_gvec_neg16,
1811           .opc = INDEX_op_neg_vec,
1812           .vece = MO_16 },
1813         { .fni4 = tcg_gen_neg_i32,
1814           .fniv = tcg_gen_neg_vec,
1815           .fno = gen_helper_gvec_neg32,
1816           .opc = INDEX_op_neg_vec,
1817           .vece = MO_32 },
1818         { .fni8 = tcg_gen_neg_i64,
1819           .fniv = tcg_gen_neg_vec,
1820           .fno = gen_helper_gvec_neg64,
1821           .opc = INDEX_op_neg_vec,
1822           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1823           .vece = MO_64 },
1824     };
1825 
1826     tcg_debug_assert(vece <= MO_64);
1827     tcg_gen_gvec_2(dofs, aofs, oprsz, maxsz, &g[vece]);
1828 }
1829 
1830 void tcg_gen_gvec_and(unsigned vece, uint32_t dofs, uint32_t aofs,
1831                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1832 {
1833     static const GVecGen3 g = {
1834         .fni8 = tcg_gen_and_i64,
1835         .fniv = tcg_gen_and_vec,
1836         .fno = gen_helper_gvec_and,
1837         .opc = INDEX_op_and_vec,
1838         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1839     };
1840     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
1841 }
1842 
1843 void tcg_gen_gvec_or(unsigned vece, uint32_t dofs, uint32_t aofs,
1844                      uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1845 {
1846     static const GVecGen3 g = {
1847         .fni8 = tcg_gen_or_i64,
1848         .fniv = tcg_gen_or_vec,
1849         .fno = gen_helper_gvec_or,
1850         .opc = INDEX_op_or_vec,
1851         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1852     };
1853     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
1854 }
1855 
1856 void tcg_gen_gvec_xor(unsigned vece, uint32_t dofs, uint32_t aofs,
1857                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1858 {
1859     static const GVecGen3 g = {
1860         .fni8 = tcg_gen_xor_i64,
1861         .fniv = tcg_gen_xor_vec,
1862         .fno = gen_helper_gvec_xor,
1863         .opc = INDEX_op_xor_vec,
1864         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1865     };
1866     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
1867 }
1868 
1869 void tcg_gen_gvec_andc(unsigned vece, uint32_t dofs, uint32_t aofs,
1870                        uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1871 {
1872     static const GVecGen3 g = {
1873         .fni8 = tcg_gen_andc_i64,
1874         .fniv = tcg_gen_andc_vec,
1875         .fno = gen_helper_gvec_andc,
1876         .opc = INDEX_op_andc_vec,
1877         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1878     };
1879     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
1880 }
1881 
1882 void tcg_gen_gvec_orc(unsigned vece, uint32_t dofs, uint32_t aofs,
1883                       uint32_t bofs, uint32_t oprsz, uint32_t maxsz)
1884 {
1885     static const GVecGen3 g = {
1886         .fni8 = tcg_gen_orc_i64,
1887         .fniv = tcg_gen_orc_vec,
1888         .fno = gen_helper_gvec_orc,
1889         .opc = INDEX_op_orc_vec,
1890         .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1891     };
1892     tcg_gen_gvec_3(dofs, aofs, bofs, oprsz, maxsz, &g);
1893 }
1894 
1895 static const GVecGen2s gop_ands = {
1896     .fni8 = tcg_gen_and_i64,
1897     .fniv = tcg_gen_and_vec,
1898     .fno = gen_helper_gvec_ands,
1899     .opc = INDEX_op_and_vec,
1900     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1901     .vece = MO_64
1902 };
1903 
1904 void tcg_gen_gvec_ands(unsigned vece, uint32_t dofs, uint32_t aofs,
1905                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
1906 {
1907     TCGv_i64 tmp = tcg_temp_new_i64();
1908     gen_dup_i64(vece, tmp, c);
1909     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ands);
1910     tcg_temp_free_i64(tmp);
1911 }
1912 
1913 void tcg_gen_gvec_andi(unsigned vece, uint32_t dofs, uint32_t aofs,
1914                        int64_t c, uint32_t oprsz, uint32_t maxsz)
1915 {
1916     TCGv_i64 tmp = tcg_const_i64(dup_const(vece, c));
1917     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ands);
1918     tcg_temp_free_i64(tmp);
1919 }
1920 
1921 static const GVecGen2s gop_xors = {
1922     .fni8 = tcg_gen_xor_i64,
1923     .fniv = tcg_gen_xor_vec,
1924     .fno = gen_helper_gvec_xors,
1925     .opc = INDEX_op_xor_vec,
1926     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1927     .vece = MO_64
1928 };
1929 
1930 void tcg_gen_gvec_xors(unsigned vece, uint32_t dofs, uint32_t aofs,
1931                        TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
1932 {
1933     TCGv_i64 tmp = tcg_temp_new_i64();
1934     gen_dup_i64(vece, tmp, c);
1935     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_xors);
1936     tcg_temp_free_i64(tmp);
1937 }
1938 
1939 void tcg_gen_gvec_xori(unsigned vece, uint32_t dofs, uint32_t aofs,
1940                        int64_t c, uint32_t oprsz, uint32_t maxsz)
1941 {
1942     TCGv_i64 tmp = tcg_const_i64(dup_const(vece, c));
1943     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_xors);
1944     tcg_temp_free_i64(tmp);
1945 }
1946 
1947 static const GVecGen2s gop_ors = {
1948     .fni8 = tcg_gen_or_i64,
1949     .fniv = tcg_gen_or_vec,
1950     .fno = gen_helper_gvec_ors,
1951     .opc = INDEX_op_or_vec,
1952     .prefer_i64 = TCG_TARGET_REG_BITS == 64,
1953     .vece = MO_64
1954 };
1955 
1956 void tcg_gen_gvec_ors(unsigned vece, uint32_t dofs, uint32_t aofs,
1957                       TCGv_i64 c, uint32_t oprsz, uint32_t maxsz)
1958 {
1959     TCGv_i64 tmp = tcg_temp_new_i64();
1960     gen_dup_i64(vece, tmp, c);
1961     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ors);
1962     tcg_temp_free_i64(tmp);
1963 }
1964 
1965 void tcg_gen_gvec_ori(unsigned vece, uint32_t dofs, uint32_t aofs,
1966                       int64_t c, uint32_t oprsz, uint32_t maxsz)
1967 {
1968     TCGv_i64 tmp = tcg_const_i64(dup_const(vece, c));
1969     tcg_gen_gvec_2s(dofs, aofs, oprsz, maxsz, tmp, &gop_ors);
1970     tcg_temp_free_i64(tmp);
1971 }
1972 
1973 void tcg_gen_vec_shl8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
1974 {
1975     uint64_t mask = dup_const(MO_8, 0xff << c);
1976     tcg_gen_shli_i64(d, a, c);
1977     tcg_gen_andi_i64(d, d, mask);
1978 }
1979 
1980 void tcg_gen_vec_shl16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
1981 {
1982     uint64_t mask = dup_const(MO_16, 0xffff << c);
1983     tcg_gen_shli_i64(d, a, c);
1984     tcg_gen_andi_i64(d, d, mask);
1985 }
1986 
1987 void tcg_gen_gvec_shli(unsigned vece, uint32_t dofs, uint32_t aofs,
1988                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
1989 {
1990     static const GVecGen2i g[4] = {
1991         { .fni8 = tcg_gen_vec_shl8i_i64,
1992           .fniv = tcg_gen_shli_vec,
1993           .fno = gen_helper_gvec_shl8i,
1994           .opc = INDEX_op_shli_vec,
1995           .vece = MO_8 },
1996         { .fni8 = tcg_gen_vec_shl16i_i64,
1997           .fniv = tcg_gen_shli_vec,
1998           .fno = gen_helper_gvec_shl16i,
1999           .opc = INDEX_op_shli_vec,
2000           .vece = MO_16 },
2001         { .fni4 = tcg_gen_shli_i32,
2002           .fniv = tcg_gen_shli_vec,
2003           .fno = gen_helper_gvec_shl32i,
2004           .opc = INDEX_op_shli_vec,
2005           .vece = MO_32 },
2006         { .fni8 = tcg_gen_shli_i64,
2007           .fniv = tcg_gen_shli_vec,
2008           .fno = gen_helper_gvec_shl64i,
2009           .opc = INDEX_op_shli_vec,
2010           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2011           .vece = MO_64 },
2012     };
2013 
2014     tcg_debug_assert(vece <= MO_64);
2015     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2016     if (shift == 0) {
2017         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2018     } else {
2019         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2020     }
2021 }
2022 
2023 void tcg_gen_vec_shr8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2024 {
2025     uint64_t mask = dup_const(MO_8, 0xff >> c);
2026     tcg_gen_shri_i64(d, a, c);
2027     tcg_gen_andi_i64(d, d, mask);
2028 }
2029 
2030 void tcg_gen_vec_shr16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2031 {
2032     uint64_t mask = dup_const(MO_16, 0xffff >> c);
2033     tcg_gen_shri_i64(d, a, c);
2034     tcg_gen_andi_i64(d, d, mask);
2035 }
2036 
2037 void tcg_gen_gvec_shri(unsigned vece, uint32_t dofs, uint32_t aofs,
2038                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
2039 {
2040     static const GVecGen2i g[4] = {
2041         { .fni8 = tcg_gen_vec_shr8i_i64,
2042           .fniv = tcg_gen_shri_vec,
2043           .fno = gen_helper_gvec_shr8i,
2044           .opc = INDEX_op_shri_vec,
2045           .vece = MO_8 },
2046         { .fni8 = tcg_gen_vec_shr16i_i64,
2047           .fniv = tcg_gen_shri_vec,
2048           .fno = gen_helper_gvec_shr16i,
2049           .opc = INDEX_op_shri_vec,
2050           .vece = MO_16 },
2051         { .fni4 = tcg_gen_shri_i32,
2052           .fniv = tcg_gen_shri_vec,
2053           .fno = gen_helper_gvec_shr32i,
2054           .opc = INDEX_op_shri_vec,
2055           .vece = MO_32 },
2056         { .fni8 = tcg_gen_shri_i64,
2057           .fniv = tcg_gen_shri_vec,
2058           .fno = gen_helper_gvec_shr64i,
2059           .opc = INDEX_op_shri_vec,
2060           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2061           .vece = MO_64 },
2062     };
2063 
2064     tcg_debug_assert(vece <= MO_64);
2065     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2066     if (shift == 0) {
2067         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2068     } else {
2069         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2070     }
2071 }
2072 
2073 void tcg_gen_vec_sar8i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2074 {
2075     uint64_t s_mask = dup_const(MO_8, 0x80 >> c);
2076     uint64_t c_mask = dup_const(MO_8, 0xff >> c);
2077     TCGv_i64 s = tcg_temp_new_i64();
2078 
2079     tcg_gen_shri_i64(d, a, c);
2080     tcg_gen_andi_i64(s, d, s_mask);  /* isolate (shifted) sign bit */
2081     tcg_gen_muli_i64(s, s, (2 << c) - 2); /* replicate isolated signs */
2082     tcg_gen_andi_i64(d, d, c_mask);  /* clear out bits above sign  */
2083     tcg_gen_or_i64(d, d, s);         /* include sign extension */
2084     tcg_temp_free_i64(s);
2085 }
2086 
2087 void tcg_gen_vec_sar16i_i64(TCGv_i64 d, TCGv_i64 a, int64_t c)
2088 {
2089     uint64_t s_mask = dup_const(MO_16, 0x8000 >> c);
2090     uint64_t c_mask = dup_const(MO_16, 0xffff >> c);
2091     TCGv_i64 s = tcg_temp_new_i64();
2092 
2093     tcg_gen_shri_i64(d, a, c);
2094     tcg_gen_andi_i64(s, d, s_mask);  /* isolate (shifted) sign bit */
2095     tcg_gen_andi_i64(d, d, c_mask);  /* clear out bits above sign  */
2096     tcg_gen_muli_i64(s, s, (2 << c) - 2); /* replicate isolated signs */
2097     tcg_gen_or_i64(d, d, s);         /* include sign extension */
2098     tcg_temp_free_i64(s);
2099 }
2100 
2101 void tcg_gen_gvec_sari(unsigned vece, uint32_t dofs, uint32_t aofs,
2102                        int64_t shift, uint32_t oprsz, uint32_t maxsz)
2103 {
2104     static const GVecGen2i g[4] = {
2105         { .fni8 = tcg_gen_vec_sar8i_i64,
2106           .fniv = tcg_gen_sari_vec,
2107           .fno = gen_helper_gvec_sar8i,
2108           .opc = INDEX_op_sari_vec,
2109           .vece = MO_8 },
2110         { .fni8 = tcg_gen_vec_sar16i_i64,
2111           .fniv = tcg_gen_sari_vec,
2112           .fno = gen_helper_gvec_sar16i,
2113           .opc = INDEX_op_sari_vec,
2114           .vece = MO_16 },
2115         { .fni4 = tcg_gen_sari_i32,
2116           .fniv = tcg_gen_sari_vec,
2117           .fno = gen_helper_gvec_sar32i,
2118           .opc = INDEX_op_sari_vec,
2119           .vece = MO_32 },
2120         { .fni8 = tcg_gen_sari_i64,
2121           .fniv = tcg_gen_sari_vec,
2122           .fno = gen_helper_gvec_sar64i,
2123           .opc = INDEX_op_sari_vec,
2124           .prefer_i64 = TCG_TARGET_REG_BITS == 64,
2125           .vece = MO_64 },
2126     };
2127 
2128     tcg_debug_assert(vece <= MO_64);
2129     tcg_debug_assert(shift >= 0 && shift < (8 << vece));
2130     if (shift == 0) {
2131         tcg_gen_gvec_mov(vece, dofs, aofs, oprsz, maxsz);
2132     } else {
2133         tcg_gen_gvec_2i(dofs, aofs, oprsz, maxsz, shift, &g[vece]);
2134     }
2135 }
2136 
2137 /* Expand OPSZ bytes worth of three-operand operations using i32 elements.  */
2138 static void expand_cmp_i32(uint32_t dofs, uint32_t aofs, uint32_t bofs,
2139                            uint32_t oprsz, TCGCond cond)
2140 {
2141     TCGv_i32 t0 = tcg_temp_new_i32();
2142     TCGv_i32 t1 = tcg_temp_new_i32();
2143     uint32_t i;
2144 
2145     for (i = 0; i < oprsz; i += 4) {
2146         tcg_gen_ld_i32(t0, cpu_env, aofs + i);
2147         tcg_gen_ld_i32(t1, cpu_env, bofs + i);
2148         tcg_gen_setcond_i32(cond, t0, t0, t1);
2149         tcg_gen_neg_i32(t0, t0);
2150         tcg_gen_st_i32(t0, cpu_env, dofs + i);
2151     }
2152     tcg_temp_free_i32(t1);
2153     tcg_temp_free_i32(t0);
2154 }
2155 
2156 static void expand_cmp_i64(uint32_t dofs, uint32_t aofs, uint32_t bofs,
2157                            uint32_t oprsz, TCGCond cond)
2158 {
2159     TCGv_i64 t0 = tcg_temp_new_i64();
2160     TCGv_i64 t1 = tcg_temp_new_i64();
2161     uint32_t i;
2162 
2163     for (i = 0; i < oprsz; i += 8) {
2164         tcg_gen_ld_i64(t0, cpu_env, aofs + i);
2165         tcg_gen_ld_i64(t1, cpu_env, bofs + i);
2166         tcg_gen_setcond_i64(cond, t0, t0, t1);
2167         tcg_gen_neg_i64(t0, t0);
2168         tcg_gen_st_i64(t0, cpu_env, dofs + i);
2169     }
2170     tcg_temp_free_i64(t1);
2171     tcg_temp_free_i64(t0);
2172 }
2173 
2174 static void expand_cmp_vec(unsigned vece, uint32_t dofs, uint32_t aofs,
2175                            uint32_t bofs, uint32_t oprsz, uint32_t tysz,
2176                            TCGType type, TCGCond cond)
2177 {
2178     TCGv_vec t0 = tcg_temp_new_vec(type);
2179     TCGv_vec t1 = tcg_temp_new_vec(type);
2180     uint32_t i;
2181 
2182     for (i = 0; i < oprsz; i += tysz) {
2183         tcg_gen_ld_vec(t0, cpu_env, aofs + i);
2184         tcg_gen_ld_vec(t1, cpu_env, bofs + i);
2185         tcg_gen_cmp_vec(cond, vece, t0, t0, t1);
2186         tcg_gen_st_vec(t0, cpu_env, dofs + i);
2187     }
2188     tcg_temp_free_vec(t1);
2189     tcg_temp_free_vec(t0);
2190 }
2191 
2192 void tcg_gen_gvec_cmp(TCGCond cond, unsigned vece, uint32_t dofs,
2193                       uint32_t aofs, uint32_t bofs,
2194                       uint32_t oprsz, uint32_t maxsz)
2195 {
2196     static gen_helper_gvec_3 * const eq_fn[4] = {
2197         gen_helper_gvec_eq8, gen_helper_gvec_eq16,
2198         gen_helper_gvec_eq32, gen_helper_gvec_eq64
2199     };
2200     static gen_helper_gvec_3 * const ne_fn[4] = {
2201         gen_helper_gvec_ne8, gen_helper_gvec_ne16,
2202         gen_helper_gvec_ne32, gen_helper_gvec_ne64
2203     };
2204     static gen_helper_gvec_3 * const lt_fn[4] = {
2205         gen_helper_gvec_lt8, gen_helper_gvec_lt16,
2206         gen_helper_gvec_lt32, gen_helper_gvec_lt64
2207     };
2208     static gen_helper_gvec_3 * const le_fn[4] = {
2209         gen_helper_gvec_le8, gen_helper_gvec_le16,
2210         gen_helper_gvec_le32, gen_helper_gvec_le64
2211     };
2212     static gen_helper_gvec_3 * const ltu_fn[4] = {
2213         gen_helper_gvec_ltu8, gen_helper_gvec_ltu16,
2214         gen_helper_gvec_ltu32, gen_helper_gvec_ltu64
2215     };
2216     static gen_helper_gvec_3 * const leu_fn[4] = {
2217         gen_helper_gvec_leu8, gen_helper_gvec_leu16,
2218         gen_helper_gvec_leu32, gen_helper_gvec_leu64
2219     };
2220     static gen_helper_gvec_3 * const * const fns[16] = {
2221         [TCG_COND_EQ] = eq_fn,
2222         [TCG_COND_NE] = ne_fn,
2223         [TCG_COND_LT] = lt_fn,
2224         [TCG_COND_LE] = le_fn,
2225         [TCG_COND_LTU] = ltu_fn,
2226         [TCG_COND_LEU] = leu_fn,
2227     };
2228     TCGType type;
2229     uint32_t some;
2230 
2231     check_size_align(oprsz, maxsz, dofs | aofs | bofs);
2232     check_overlap_3(dofs, aofs, bofs, maxsz);
2233 
2234     if (cond == TCG_COND_NEVER || cond == TCG_COND_ALWAYS) {
2235         do_dup(MO_8, dofs, oprsz, maxsz,
2236                NULL, NULL, -(cond == TCG_COND_ALWAYS));
2237         return;
2238     }
2239 
2240     /* Implement inline with a vector type, if possible.
2241      * Prefer integer when 64-bit host and 64-bit comparison.
2242      */
2243     type = choose_vector_type(INDEX_op_cmp_vec, vece, oprsz,
2244                               TCG_TARGET_REG_BITS == 64 && vece == MO_64);
2245     switch (type) {
2246     case TCG_TYPE_V256:
2247         /* Recall that ARM SVE allows vector sizes that are not a
2248          * power of 2, but always a multiple of 16.  The intent is
2249          * that e.g. size == 80 would be expanded with 2x32 + 1x16.
2250          */
2251         some = QEMU_ALIGN_DOWN(oprsz, 32);
2252         expand_cmp_vec(vece, dofs, aofs, bofs, some, 32, TCG_TYPE_V256, cond);
2253         if (some == oprsz) {
2254             break;
2255         }
2256         dofs += some;
2257         aofs += some;
2258         bofs += some;
2259         oprsz -= some;
2260         maxsz -= some;
2261         /* fallthru */
2262     case TCG_TYPE_V128:
2263         expand_cmp_vec(vece, dofs, aofs, bofs, oprsz, 16, TCG_TYPE_V128, cond);
2264         break;
2265     case TCG_TYPE_V64:
2266         expand_cmp_vec(vece, dofs, aofs, bofs, oprsz, 8, TCG_TYPE_V64, cond);
2267         break;
2268 
2269     case 0:
2270         if (vece == MO_64 && check_size_impl(oprsz, 8)) {
2271             expand_cmp_i64(dofs, aofs, bofs, oprsz, cond);
2272         } else if (vece == MO_32 && check_size_impl(oprsz, 4)) {
2273             expand_cmp_i32(dofs, aofs, bofs, oprsz, cond);
2274         } else {
2275             gen_helper_gvec_3 * const *fn = fns[cond];
2276 
2277             if (fn == NULL) {
2278                 uint32_t tmp;
2279                 tmp = aofs, aofs = bofs, bofs = tmp;
2280                 cond = tcg_swap_cond(cond);
2281                 fn = fns[cond];
2282                 assert(fn != NULL);
2283             }
2284             tcg_gen_gvec_3_ool(dofs, aofs, bofs, oprsz, maxsz, 0, fn[vece]);
2285             return;
2286         }
2287         break;
2288 
2289     default:
2290         g_assert_not_reached();
2291     }
2292 
2293     if (oprsz < maxsz) {
2294         expand_clr(dofs + oprsz, maxsz - oprsz);
2295     }
2296 }
2297