]> git.proxmox.com Git - mirror_qemu.git/blob - tcg/tcg-op.c
projects / mirror_qemu.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge remote-tracking branch 'remotes/kraxel/tags/vga-20190705-pull-request' into...
[mirror_qemu.git] / tcg / tcg-op.c
1 /*
2 * Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2008 Fabrice Bellard
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to deal
8 * in the Software without restriction, including without limitation the rights
9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10 * copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 * THE SOFTWARE.
23 */
24
25 #include "qemu/osdep.h"
26 #include "cpu.h"
27 #include "exec/exec-all.h"
28 #include "tcg.h"
29 #include "tcg-op.h"
30 #include "tcg-mo.h"
31 #include "trace-tcg.h"
32 #include "trace/mem.h"
33
34 /* Reduce the number of ifdefs below. This assumes that all uses of
35 TCGV_HIGH and TCGV_LOW are properly protected by a conditional that
36 the compiler can eliminate. */
37 #if TCG_TARGET_REG_BITS == 64
38 extern TCGv_i32 TCGV_LOW_link_error(TCGv_i64);
39 extern TCGv_i32 TCGV_HIGH_link_error(TCGv_i64);
40 #define TCGV_LOW TCGV_LOW_link_error
41 #define TCGV_HIGH TCGV_HIGH_link_error
42 #endif
43
44 void tcg_gen_op1(TCGOpcode opc, TCGArg a1)
45 {
46 TCGOp *op = tcg_emit_op(opc);
47 op->args[0] = a1;
48 }
49
50 void tcg_gen_op2(TCGOpcode opc, TCGArg a1, TCGArg a2)
51 {
52 TCGOp *op = tcg_emit_op(opc);
53 op->args[0] = a1;
54 op->args[1] = a2;
55 }
56
57 void tcg_gen_op3(TCGOpcode opc, TCGArg a1, TCGArg a2, TCGArg a3)
58 {
59 TCGOp *op = tcg_emit_op(opc);
60 op->args[0] = a1;
61 op->args[1] = a2;
62 op->args[2] = a3;
63 }
64
65 void tcg_gen_op4(TCGOpcode opc, TCGArg a1, TCGArg a2, TCGArg a3, TCGArg a4)
66 {
67 TCGOp *op = tcg_emit_op(opc);
68 op->args[0] = a1;
69 op->args[1] = a2;
70 op->args[2] = a3;
71 op->args[3] = a4;
72 }
73
74 void tcg_gen_op5(TCGOpcode opc, TCGArg a1, TCGArg a2, TCGArg a3,
75 TCGArg a4, TCGArg a5)
76 {
77 TCGOp *op = tcg_emit_op(opc);
78 op->args[0] = a1;
79 op->args[1] = a2;
80 op->args[2] = a3;
81 op->args[3] = a4;
82 op->args[4] = a5;
83 }
84
85 void tcg_gen_op6(TCGOpcode opc, TCGArg a1, TCGArg a2, TCGArg a3,
86 TCGArg a4, TCGArg a5, TCGArg a6)
87 {
88 TCGOp *op = tcg_emit_op(opc);
89 op->args[0] = a1;
90 op->args[1] = a2;
91 op->args[2] = a3;
92 op->args[3] = a4;
93 op->args[4] = a5;
94 op->args[5] = a6;
95 }
96
97 void tcg_gen_mb(TCGBar mb_type)
98 {
99 if (tcg_ctx->tb_cflags & CF_PARALLEL) {
100 tcg_gen_op1(INDEX_op_mb, mb_type);
101 }
102 }
103
104 /* 32 bit ops */
105
106 void tcg_gen_addi_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
107 {
108 /* some cases can be optimized here */
109 if (arg2 == 0) {
110 tcg_gen_mov_i32(ret, arg1);
111 } else {
112 TCGv_i32 t0 = tcg_const_i32(arg2);
113 tcg_gen_add_i32(ret, arg1, t0);
114 tcg_temp_free_i32(t0);
115 }
116 }
117
118 void tcg_gen_subfi_i32(TCGv_i32 ret, int32_t arg1, TCGv_i32 arg2)
119 {
120 if (arg1 == 0 && TCG_TARGET_HAS_neg_i32) {
121 /* Don't recurse with tcg_gen_neg_i32. */
122 tcg_gen_op2_i32(INDEX_op_neg_i32, ret, arg2);
123 } else {
124 TCGv_i32 t0 = tcg_const_i32(arg1);
125 tcg_gen_sub_i32(ret, t0, arg2);
126 tcg_temp_free_i32(t0);
127 }
128 }
129
130 void tcg_gen_subi_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
131 {
132 /* some cases can be optimized here */
133 if (arg2 == 0) {
134 tcg_gen_mov_i32(ret, arg1);
135 } else {
136 TCGv_i32 t0 = tcg_const_i32(arg2);
137 tcg_gen_sub_i32(ret, arg1, t0);
138 tcg_temp_free_i32(t0);
139 }
140 }
141
142 void tcg_gen_andi_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
143 {
144 TCGv_i32 t0;
145 /* Some cases can be optimized here. */
146 switch (arg2) {
147 case 0:
148 tcg_gen_movi_i32(ret, 0);
149 return;
150 case -1:
151 tcg_gen_mov_i32(ret, arg1);
152 return;
153 case 0xff:
154 /* Don't recurse with tcg_gen_ext8u_i32. */
155 if (TCG_TARGET_HAS_ext8u_i32) {
156 tcg_gen_op2_i32(INDEX_op_ext8u_i32, ret, arg1);
157 return;
158 }
159 break;
160 case 0xffff:
161 if (TCG_TARGET_HAS_ext16u_i32) {
162 tcg_gen_op2_i32(INDEX_op_ext16u_i32, ret, arg1);
163 return;
164 }
165 break;
166 }
167 t0 = tcg_const_i32(arg2);
168 tcg_gen_and_i32(ret, arg1, t0);
169 tcg_temp_free_i32(t0);
170 }
171
172 void tcg_gen_ori_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
173 {
174 /* Some cases can be optimized here. */
175 if (arg2 == -1) {
176 tcg_gen_movi_i32(ret, -1);
177 } else if (arg2 == 0) {
178 tcg_gen_mov_i32(ret, arg1);
179 } else {
180 TCGv_i32 t0 = tcg_const_i32(arg2);
181 tcg_gen_or_i32(ret, arg1, t0);
182 tcg_temp_free_i32(t0);
183 }
184 }
185
186 void tcg_gen_xori_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
187 {
188 /* Some cases can be optimized here. */
189 if (arg2 == 0) {
190 tcg_gen_mov_i32(ret, arg1);
191 } else if (arg2 == -1 && TCG_TARGET_HAS_not_i32) {
192 /* Don't recurse with tcg_gen_not_i32. */
193 tcg_gen_op2_i32(INDEX_op_not_i32, ret, arg1);
194 } else {
195 TCGv_i32 t0 = tcg_const_i32(arg2);
196 tcg_gen_xor_i32(ret, arg1, t0);
197 tcg_temp_free_i32(t0);
198 }
199 }
200
201 void tcg_gen_shli_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
202 {
203 tcg_debug_assert(arg2 >= 0 && arg2 < 32);
204 if (arg2 == 0) {
205 tcg_gen_mov_i32(ret, arg1);
206 } else {
207 TCGv_i32 t0 = tcg_const_i32(arg2);
208 tcg_gen_shl_i32(ret, arg1, t0);
209 tcg_temp_free_i32(t0);
210 }
211 }
212
213 void tcg_gen_shri_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
214 {
215 tcg_debug_assert(arg2 >= 0 && arg2 < 32);
216 if (arg2 == 0) {
217 tcg_gen_mov_i32(ret, arg1);
218 } else {
219 TCGv_i32 t0 = tcg_const_i32(arg2);
220 tcg_gen_shr_i32(ret, arg1, t0);
221 tcg_temp_free_i32(t0);
222 }
223 }
224
225 void tcg_gen_sari_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
226 {
227 tcg_debug_assert(arg2 >= 0 && arg2 < 32);
228 if (arg2 == 0) {
229 tcg_gen_mov_i32(ret, arg1);
230 } else {
231 TCGv_i32 t0 = tcg_const_i32(arg2);
232 tcg_gen_sar_i32(ret, arg1, t0);
233 tcg_temp_free_i32(t0);
234 }
235 }
236
237 void tcg_gen_brcond_i32(TCGCond cond, TCGv_i32 arg1, TCGv_i32 arg2, TCGLabel *l)
238 {
239 if (cond == TCG_COND_ALWAYS) {
240 tcg_gen_br(l);
241 } else if (cond != TCG_COND_NEVER) {
242 l->refs++;
243 tcg_gen_op4ii_i32(INDEX_op_brcond_i32, arg1, arg2, cond, label_arg(l));
244 }
245 }
246
247 void tcg_gen_brcondi_i32(TCGCond cond, TCGv_i32 arg1, int32_t arg2, TCGLabel *l)
248 {
249 if (cond == TCG_COND_ALWAYS) {
250 tcg_gen_br(l);
251 } else if (cond != TCG_COND_NEVER) {
252 TCGv_i32 t0 = tcg_const_i32(arg2);
253 tcg_gen_brcond_i32(cond, arg1, t0, l);
254 tcg_temp_free_i32(t0);
255 }
256 }
257
258 void tcg_gen_setcond_i32(TCGCond cond, TCGv_i32 ret,
259 TCGv_i32 arg1, TCGv_i32 arg2)
260 {
261 if (cond == TCG_COND_ALWAYS) {
262 tcg_gen_movi_i32(ret, 1);
263 } else if (cond == TCG_COND_NEVER) {
264 tcg_gen_movi_i32(ret, 0);
265 } else {
266 tcg_gen_op4i_i32(INDEX_op_setcond_i32, ret, arg1, arg2, cond);
267 }
268 }
269
270 void tcg_gen_setcondi_i32(TCGCond cond, TCGv_i32 ret,
271 TCGv_i32 arg1, int32_t arg2)
272 {
273 TCGv_i32 t0 = tcg_const_i32(arg2);
274 tcg_gen_setcond_i32(cond, ret, arg1, t0);
275 tcg_temp_free_i32(t0);
276 }
277
278 void tcg_gen_muli_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
279 {
280 if (arg2 == 0) {
281 tcg_gen_movi_i32(ret, 0);
282 } else if (is_power_of_2(arg2)) {
283 tcg_gen_shli_i32(ret, arg1, ctz32(arg2));
284 } else {
285 TCGv_i32 t0 = tcg_const_i32(arg2);
286 tcg_gen_mul_i32(ret, arg1, t0);
287 tcg_temp_free_i32(t0);
288 }
289 }
290
291 void tcg_gen_div_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
292 {
293 if (TCG_TARGET_HAS_div_i32) {
294 tcg_gen_op3_i32(INDEX_op_div_i32, ret, arg1, arg2);
295 } else if (TCG_TARGET_HAS_div2_i32) {
296 TCGv_i32 t0 = tcg_temp_new_i32();
297 tcg_gen_sari_i32(t0, arg1, 31);
298 tcg_gen_op5_i32(INDEX_op_div2_i32, ret, t0, arg1, t0, arg2);
299 tcg_temp_free_i32(t0);
300 } else {
301 gen_helper_div_i32(ret, arg1, arg2);
302 }
303 }
304
305 void tcg_gen_rem_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
306 {
307 if (TCG_TARGET_HAS_rem_i32) {
308 tcg_gen_op3_i32(INDEX_op_rem_i32, ret, arg1, arg2);
309 } else if (TCG_TARGET_HAS_div_i32) {
310 TCGv_i32 t0 = tcg_temp_new_i32();
311 tcg_gen_op3_i32(INDEX_op_div_i32, t0, arg1, arg2);
312 tcg_gen_mul_i32(t0, t0, arg2);
313 tcg_gen_sub_i32(ret, arg1, t0);
314 tcg_temp_free_i32(t0);
315 } else if (TCG_TARGET_HAS_div2_i32) {
316 TCGv_i32 t0 = tcg_temp_new_i32();
317 tcg_gen_sari_i32(t0, arg1, 31);
318 tcg_gen_op5_i32(INDEX_op_div2_i32, t0, ret, arg1, t0, arg2);
319 tcg_temp_free_i32(t0);
320 } else {
321 gen_helper_rem_i32(ret, arg1, arg2);
322 }
323 }
324
325 void tcg_gen_divu_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
326 {
327 if (TCG_TARGET_HAS_div_i32) {
328 tcg_gen_op3_i32(INDEX_op_divu_i32, ret, arg1, arg2);
329 } else if (TCG_TARGET_HAS_div2_i32) {
330 TCGv_i32 t0 = tcg_temp_new_i32();
331 tcg_gen_movi_i32(t0, 0);
332 tcg_gen_op5_i32(INDEX_op_divu2_i32, ret, t0, arg1, t0, arg2);
333 tcg_temp_free_i32(t0);
334 } else {
335 gen_helper_divu_i32(ret, arg1, arg2);
336 }
337 }
338
339 void tcg_gen_remu_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
340 {
341 if (TCG_TARGET_HAS_rem_i32) {
342 tcg_gen_op3_i32(INDEX_op_remu_i32, ret, arg1, arg2);
343 } else if (TCG_TARGET_HAS_div_i32) {
344 TCGv_i32 t0 = tcg_temp_new_i32();
345 tcg_gen_op3_i32(INDEX_op_divu_i32, t0, arg1, arg2);
346 tcg_gen_mul_i32(t0, t0, arg2);
347 tcg_gen_sub_i32(ret, arg1, t0);
348 tcg_temp_free_i32(t0);
349 } else if (TCG_TARGET_HAS_div2_i32) {
350 TCGv_i32 t0 = tcg_temp_new_i32();
351 tcg_gen_movi_i32(t0, 0);
352 tcg_gen_op5_i32(INDEX_op_divu2_i32, t0, ret, arg1, t0, arg2);
353 tcg_temp_free_i32(t0);
354 } else {
355 gen_helper_remu_i32(ret, arg1, arg2);
356 }
357 }
358
359 void tcg_gen_andc_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
360 {
361 if (TCG_TARGET_HAS_andc_i32) {
362 tcg_gen_op3_i32(INDEX_op_andc_i32, ret, arg1, arg2);
363 } else {
364 TCGv_i32 t0 = tcg_temp_new_i32();
365 tcg_gen_not_i32(t0, arg2);
366 tcg_gen_and_i32(ret, arg1, t0);
367 tcg_temp_free_i32(t0);
368 }
369 }
370
371 void tcg_gen_eqv_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
372 {
373 if (TCG_TARGET_HAS_eqv_i32) {
374 tcg_gen_op3_i32(INDEX_op_eqv_i32, ret, arg1, arg2);
375 } else {
376 tcg_gen_xor_i32(ret, arg1, arg2);
377 tcg_gen_not_i32(ret, ret);
378 }
379 }
380
381 void tcg_gen_nand_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
382 {
383 if (TCG_TARGET_HAS_nand_i32) {
384 tcg_gen_op3_i32(INDEX_op_nand_i32, ret, arg1, arg2);
385 } else {
386 tcg_gen_and_i32(ret, arg1, arg2);
387 tcg_gen_not_i32(ret, ret);
388 }
389 }
390
391 void tcg_gen_nor_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
392 {
393 if (TCG_TARGET_HAS_nor_i32) {
394 tcg_gen_op3_i32(INDEX_op_nor_i32, ret, arg1, arg2);
395 } else {
396 tcg_gen_or_i32(ret, arg1, arg2);
397 tcg_gen_not_i32(ret, ret);
398 }
399 }
400
401 void tcg_gen_orc_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
402 {
403 if (TCG_TARGET_HAS_orc_i32) {
404 tcg_gen_op3_i32(INDEX_op_orc_i32, ret, arg1, arg2);
405 } else {
406 TCGv_i32 t0 = tcg_temp_new_i32();
407 tcg_gen_not_i32(t0, arg2);
408 tcg_gen_or_i32(ret, arg1, t0);
409 tcg_temp_free_i32(t0);
410 }
411 }
412
413 void tcg_gen_clz_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
414 {
415 if (TCG_TARGET_HAS_clz_i32) {
416 tcg_gen_op3_i32(INDEX_op_clz_i32, ret, arg1, arg2);
417 } else if (TCG_TARGET_HAS_clz_i64) {
418 TCGv_i64 t1 = tcg_temp_new_i64();
419 TCGv_i64 t2 = tcg_temp_new_i64();
420 tcg_gen_extu_i32_i64(t1, arg1);
421 tcg_gen_extu_i32_i64(t2, arg2);
422 tcg_gen_addi_i64(t2, t2, 32);
423 tcg_gen_clz_i64(t1, t1, t2);
424 tcg_gen_extrl_i64_i32(ret, t1);
425 tcg_temp_free_i64(t1);
426 tcg_temp_free_i64(t2);
427 tcg_gen_subi_i32(ret, ret, 32);
428 } else {
429 gen_helper_clz_i32(ret, arg1, arg2);
430 }
431 }
432
433 void tcg_gen_clzi_i32(TCGv_i32 ret, TCGv_i32 arg1, uint32_t arg2)
434 {
435 TCGv_i32 t = tcg_const_i32(arg2);
436 tcg_gen_clz_i32(ret, arg1, t);
437 tcg_temp_free_i32(t);
438 }
439
440 void tcg_gen_ctz_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
441 {
442 if (TCG_TARGET_HAS_ctz_i32) {
443 tcg_gen_op3_i32(INDEX_op_ctz_i32, ret, arg1, arg2);
444 } else if (TCG_TARGET_HAS_ctz_i64) {
445 TCGv_i64 t1 = tcg_temp_new_i64();
446 TCGv_i64 t2 = tcg_temp_new_i64();
447 tcg_gen_extu_i32_i64(t1, arg1);
448 tcg_gen_extu_i32_i64(t2, arg2);
449 tcg_gen_ctz_i64(t1, t1, t2);
450 tcg_gen_extrl_i64_i32(ret, t1);
451 tcg_temp_free_i64(t1);
452 tcg_temp_free_i64(t2);
453 } else if (TCG_TARGET_HAS_ctpop_i32
454 || TCG_TARGET_HAS_ctpop_i64
455 || TCG_TARGET_HAS_clz_i32
456 || TCG_TARGET_HAS_clz_i64) {
457 TCGv_i32 z, t = tcg_temp_new_i32();
458
459 if (TCG_TARGET_HAS_ctpop_i32 || TCG_TARGET_HAS_ctpop_i64) {
460 tcg_gen_subi_i32(t, arg1, 1);
461 tcg_gen_andc_i32(t, t, arg1);
462 tcg_gen_ctpop_i32(t, t);
463 } else {
464 /* Since all non-x86 hosts have clz(0) == 32, don't fight it. */
465 tcg_gen_neg_i32(t, arg1);
466 tcg_gen_and_i32(t, t, arg1);
467 tcg_gen_clzi_i32(t, t, 32);
468 tcg_gen_xori_i32(t, t, 31);
469 }
470 z = tcg_const_i32(0);
471 tcg_gen_movcond_i32(TCG_COND_EQ, ret, arg1, z, arg2, t);
472 tcg_temp_free_i32(t);
473 tcg_temp_free_i32(z);
474 } else {
475 gen_helper_ctz_i32(ret, arg1, arg2);
476 }
477 }
478
479 void tcg_gen_ctzi_i32(TCGv_i32 ret, TCGv_i32 arg1, uint32_t arg2)
480 {
481 if (!TCG_TARGET_HAS_ctz_i32 && TCG_TARGET_HAS_ctpop_i32 && arg2 == 32) {
482 /* This equivalence has the advantage of not requiring a fixup. */
483 TCGv_i32 t = tcg_temp_new_i32();
484 tcg_gen_subi_i32(t, arg1, 1);
485 tcg_gen_andc_i32(t, t, arg1);
486 tcg_gen_ctpop_i32(ret, t);
487 tcg_temp_free_i32(t);
488 } else {
489 TCGv_i32 t = tcg_const_i32(arg2);
490 tcg_gen_ctz_i32(ret, arg1, t);
491 tcg_temp_free_i32(t);
492 }
493 }
494
495 void tcg_gen_clrsb_i32(TCGv_i32 ret, TCGv_i32 arg)
496 {
497 if (TCG_TARGET_HAS_clz_i32) {
498 TCGv_i32 t = tcg_temp_new_i32();
499 tcg_gen_sari_i32(t, arg, 31);
500 tcg_gen_xor_i32(t, t, arg);
501 tcg_gen_clzi_i32(t, t, 32);
502 tcg_gen_subi_i32(ret, t, 1);
503 tcg_temp_free_i32(t);
504 } else {
505 gen_helper_clrsb_i32(ret, arg);
506 }
507 }
508
509 void tcg_gen_ctpop_i32(TCGv_i32 ret, TCGv_i32 arg1)
510 {
511 if (TCG_TARGET_HAS_ctpop_i32) {
512 tcg_gen_op2_i32(INDEX_op_ctpop_i32, ret, arg1);
513 } else if (TCG_TARGET_HAS_ctpop_i64) {
514 TCGv_i64 t = tcg_temp_new_i64();
515 tcg_gen_extu_i32_i64(t, arg1);
516 tcg_gen_ctpop_i64(t, t);
517 tcg_gen_extrl_i64_i32(ret, t);
518 tcg_temp_free_i64(t);
519 } else {
520 gen_helper_ctpop_i32(ret, arg1);
521 }
522 }
523
524 void tcg_gen_rotl_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
525 {
526 if (TCG_TARGET_HAS_rot_i32) {
527 tcg_gen_op3_i32(INDEX_op_rotl_i32, ret, arg1, arg2);
528 } else {
529 TCGv_i32 t0, t1;
530
531 t0 = tcg_temp_new_i32();
532 t1 = tcg_temp_new_i32();
533 tcg_gen_shl_i32(t0, arg1, arg2);
534 tcg_gen_subfi_i32(t1, 32, arg2);
535 tcg_gen_shr_i32(t1, arg1, t1);
536 tcg_gen_or_i32(ret, t0, t1);
537 tcg_temp_free_i32(t0);
538 tcg_temp_free_i32(t1);
539 }
540 }
541
542 void tcg_gen_rotli_i32(TCGv_i32 ret, TCGv_i32 arg1, unsigned arg2)
543 {
544 tcg_debug_assert(arg2 < 32);
545 /* some cases can be optimized here */
546 if (arg2 == 0) {
547 tcg_gen_mov_i32(ret, arg1);
548 } else if (TCG_TARGET_HAS_rot_i32) {
549 TCGv_i32 t0 = tcg_const_i32(arg2);
550 tcg_gen_rotl_i32(ret, arg1, t0);
551 tcg_temp_free_i32(t0);
552 } else {
553 TCGv_i32 t0, t1;
554 t0 = tcg_temp_new_i32();
555 t1 = tcg_temp_new_i32();
556 tcg_gen_shli_i32(t0, arg1, arg2);
557 tcg_gen_shri_i32(t1, arg1, 32 - arg2);
558 tcg_gen_or_i32(ret, t0, t1);
559 tcg_temp_free_i32(t0);
560 tcg_temp_free_i32(t1);
561 }
562 }
563
564 void tcg_gen_rotr_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
565 {
566 if (TCG_TARGET_HAS_rot_i32) {
567 tcg_gen_op3_i32(INDEX_op_rotr_i32, ret, arg1, arg2);
568 } else {
569 TCGv_i32 t0, t1;
570
571 t0 = tcg_temp_new_i32();
572 t1 = tcg_temp_new_i32();
573 tcg_gen_shr_i32(t0, arg1, arg2);
574 tcg_gen_subfi_i32(t1, 32, arg2);
575 tcg_gen_shl_i32(t1, arg1, t1);
576 tcg_gen_or_i32(ret, t0, t1);
577 tcg_temp_free_i32(t0);
578 tcg_temp_free_i32(t1);
579 }
580 }
581
582 void tcg_gen_rotri_i32(TCGv_i32 ret, TCGv_i32 arg1, unsigned arg2)
583 {
584 tcg_debug_assert(arg2 < 32);
585 /* some cases can be optimized here */
586 if (arg2 == 0) {
587 tcg_gen_mov_i32(ret, arg1);
588 } else {
589 tcg_gen_rotli_i32(ret, arg1, 32 - arg2);
590 }
591 }
592
593 void tcg_gen_deposit_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2,
594 unsigned int ofs, unsigned int len)
595 {
596 uint32_t mask;
597 TCGv_i32 t1;
598
599 tcg_debug_assert(ofs < 32);
600 tcg_debug_assert(len > 0);
601 tcg_debug_assert(len <= 32);
602 tcg_debug_assert(ofs + len <= 32);
603
604 if (len == 32) {
605 tcg_gen_mov_i32(ret, arg2);
606 return;
607 }
608 if (TCG_TARGET_HAS_deposit_i32 && TCG_TARGET_deposit_i32_valid(ofs, len)) {
609 tcg_gen_op5ii_i32(INDEX_op_deposit_i32, ret, arg1, arg2, ofs, len);
610 return;
611 }
612
613 t1 = tcg_temp_new_i32();
614
615 if (TCG_TARGET_HAS_extract2_i32) {
616 if (ofs + len == 32) {
617 tcg_gen_shli_i32(t1, arg1, len);
618 tcg_gen_extract2_i32(ret, t1, arg2, len);
619 goto done;
620 }
621 if (ofs == 0) {
622 tcg_gen_extract2_i32(ret, arg1, arg2, len);
623 tcg_gen_rotli_i32(ret, ret, len);
624 goto done;
625 }
626 }
627
628 mask = (1u << len) - 1;
629 if (ofs + len < 32) {
630 tcg_gen_andi_i32(t1, arg2, mask);
631 tcg_gen_shli_i32(t1, t1, ofs);
632 } else {
633 tcg_gen_shli_i32(t1, arg2, ofs);
634 }
635 tcg_gen_andi_i32(ret, arg1, ~(mask << ofs));
636 tcg_gen_or_i32(ret, ret, t1);
637 done:
638 tcg_temp_free_i32(t1);
639 }
640
641 void tcg_gen_deposit_z_i32(TCGv_i32 ret, TCGv_i32 arg,
642 unsigned int ofs, unsigned int len)
643 {
644 tcg_debug_assert(ofs < 32);
645 tcg_debug_assert(len > 0);
646 tcg_debug_assert(len <= 32);
647 tcg_debug_assert(ofs + len <= 32);
648
649 if (ofs + len == 32) {
650 tcg_gen_shli_i32(ret, arg, ofs);
651 } else if (ofs == 0) {
652 tcg_gen_andi_i32(ret, arg, (1u << len) - 1);
653 } else if (TCG_TARGET_HAS_deposit_i32
654 && TCG_TARGET_deposit_i32_valid(ofs, len)) {
655 TCGv_i32 zero = tcg_const_i32(0);
656 tcg_gen_op5ii_i32(INDEX_op_deposit_i32, ret, zero, arg, ofs, len);
657 tcg_temp_free_i32(zero);
658 } else {
659 /* To help two-operand hosts we prefer to zero-extend first,
660 which allows ARG to stay live. */
661 switch (len) {
662 case 16:
663 if (TCG_TARGET_HAS_ext16u_i32) {
664 tcg_gen_ext16u_i32(ret, arg);
665 tcg_gen_shli_i32(ret, ret, ofs);
666 return;
667 }
668 break;
669 case 8:
670 if (TCG_TARGET_HAS_ext8u_i32) {
671 tcg_gen_ext8u_i32(ret, arg);
672 tcg_gen_shli_i32(ret, ret, ofs);
673 return;
674 }
675 break;
676 }
677 /* Otherwise prefer zero-extension over AND for code size. */
678 switch (ofs + len) {
679 case 16:
680 if (TCG_TARGET_HAS_ext16u_i32) {
681 tcg_gen_shli_i32(ret, arg, ofs);
682 tcg_gen_ext16u_i32(ret, ret);
683 return;
684 }
685 break;
686 case 8:
687 if (TCG_TARGET_HAS_ext8u_i32) {
688 tcg_gen_shli_i32(ret, arg, ofs);
689 tcg_gen_ext8u_i32(ret, ret);
690 return;
691 }
692 break;
693 }
694 tcg_gen_andi_i32(ret, arg, (1u << len) - 1);
695 tcg_gen_shli_i32(ret, ret, ofs);
696 }
697 }
698
699 void tcg_gen_extract_i32(TCGv_i32 ret, TCGv_i32 arg,
700 unsigned int ofs, unsigned int len)
701 {
702 tcg_debug_assert(ofs < 32);
703 tcg_debug_assert(len > 0);
704 tcg_debug_assert(len <= 32);
705 tcg_debug_assert(ofs + len <= 32);
706
707 /* Canonicalize certain special cases, even if extract is supported. */
708 if (ofs + len == 32) {
709 tcg_gen_shri_i32(ret, arg, 32 - len);
710 return;
711 }
712 if (ofs == 0) {
713 tcg_gen_andi_i32(ret, arg, (1u << len) - 1);
714 return;
715 }
716
717 if (TCG_TARGET_HAS_extract_i32
718 && TCG_TARGET_extract_i32_valid(ofs, len)) {
719 tcg_gen_op4ii_i32(INDEX_op_extract_i32, ret, arg, ofs, len);
720 return;
721 }
722
723 /* Assume that zero-extension, if available, is cheaper than a shift. */
724 switch (ofs + len) {
725 case 16:
726 if (TCG_TARGET_HAS_ext16u_i32) {
727 tcg_gen_ext16u_i32(ret, arg);
728 tcg_gen_shri_i32(ret, ret, ofs);
729 return;
730 }
731 break;
732 case 8:
733 if (TCG_TARGET_HAS_ext8u_i32) {
734 tcg_gen_ext8u_i32(ret, arg);
735 tcg_gen_shri_i32(ret, ret, ofs);
736 return;
737 }
738 break;
739 }
740
741 /* ??? Ideally we'd know what values are available for immediate AND.
742 Assume that 8 bits are available, plus the special case of 16,
743 so that we get ext8u, ext16u. */
744 switch (len) {
745 case 1 ... 8: case 16:
746 tcg_gen_shri_i32(ret, arg, ofs);
747 tcg_gen_andi_i32(ret, ret, (1u << len) - 1);
748 break;
749 default:
750 tcg_gen_shli_i32(ret, arg, 32 - len - ofs);
751 tcg_gen_shri_i32(ret, ret, 32 - len);
752 break;
753 }
754 }
755
756 void tcg_gen_sextract_i32(TCGv_i32 ret, TCGv_i32 arg,
757 unsigned int ofs, unsigned int len)
758 {
759 tcg_debug_assert(ofs < 32);
760 tcg_debug_assert(len > 0);
761 tcg_debug_assert(len <= 32);
762 tcg_debug_assert(ofs + len <= 32);
763
764 /* Canonicalize certain special cases, even if extract is supported. */
765 if (ofs + len == 32) {
766 tcg_gen_sari_i32(ret, arg, 32 - len);
767 return;
768 }
769 if (ofs == 0) {
770 switch (len) {
771 case 16:
772 tcg_gen_ext16s_i32(ret, arg);
773 return;
774 case 8:
775 tcg_gen_ext8s_i32(ret, arg);
776 return;
777 }
778 }
779
780 if (TCG_TARGET_HAS_sextract_i32
781 && TCG_TARGET_extract_i32_valid(ofs, len)) {
782 tcg_gen_op4ii_i32(INDEX_op_sextract_i32, ret, arg, ofs, len);
783 return;
784 }
785
786 /* Assume that sign-extension, if available, is cheaper than a shift. */
787 switch (ofs + len) {
788 case 16:
789 if (TCG_TARGET_HAS_ext16s_i32) {
790 tcg_gen_ext16s_i32(ret, arg);
791 tcg_gen_sari_i32(ret, ret, ofs);
792 return;
793 }
794 break;
795 case 8:
796 if (TCG_TARGET_HAS_ext8s_i32) {
797 tcg_gen_ext8s_i32(ret, arg);
798 tcg_gen_sari_i32(ret, ret, ofs);
799 return;
800 }
801 break;
802 }
803 switch (len) {
804 case 16:
805 if (TCG_TARGET_HAS_ext16s_i32) {
806 tcg_gen_shri_i32(ret, arg, ofs);
807 tcg_gen_ext16s_i32(ret, ret);
808 return;
809 }
810 break;
811 case 8:
812 if (TCG_TARGET_HAS_ext8s_i32) {
813 tcg_gen_shri_i32(ret, arg, ofs);
814 tcg_gen_ext8s_i32(ret, ret);
815 return;
816 }
817 break;
818 }
819
820 tcg_gen_shli_i32(ret, arg, 32 - len - ofs);
821 tcg_gen_sari_i32(ret, ret, 32 - len);
822 }
823
824 /*
825 * Extract 32-bits from a 64-bit input, ah:al, starting from ofs.
826 * Unlike tcg_gen_extract_i32 above, len is fixed at 32.
827 */
828 void tcg_gen_extract2_i32(TCGv_i32 ret, TCGv_i32 al, TCGv_i32 ah,
829 unsigned int ofs)
830 {
831 tcg_debug_assert(ofs <= 32);
832 if (ofs == 0) {
833 tcg_gen_mov_i32(ret, al);
834 } else if (ofs == 32) {
835 tcg_gen_mov_i32(ret, ah);
836 } else if (al == ah) {
837 tcg_gen_rotri_i32(ret, al, ofs);
838 } else if (TCG_TARGET_HAS_extract2_i32) {
839 tcg_gen_op4i_i32(INDEX_op_extract2_i32, ret, al, ah, ofs);
840 } else {
841 TCGv_i32 t0 = tcg_temp_new_i32();
842 tcg_gen_shri_i32(t0, al, ofs);
843 tcg_gen_deposit_i32(ret, t0, ah, 32 - ofs, ofs);
844 tcg_temp_free_i32(t0);
845 }
846 }
847
848 void tcg_gen_movcond_i32(TCGCond cond, TCGv_i32 ret, TCGv_i32 c1,
849 TCGv_i32 c2, TCGv_i32 v1, TCGv_i32 v2)
850 {
851 if (cond == TCG_COND_ALWAYS) {
852 tcg_gen_mov_i32(ret, v1);
853 } else if (cond == TCG_COND_NEVER) {
854 tcg_gen_mov_i32(ret, v2);
855 } else if (TCG_TARGET_HAS_movcond_i32) {
856 tcg_gen_op6i_i32(INDEX_op_movcond_i32, ret, c1, c2, v1, v2, cond);
857 } else {
858 TCGv_i32 t0 = tcg_temp_new_i32();
859 TCGv_i32 t1 = tcg_temp_new_i32();
860 tcg_gen_setcond_i32(cond, t0, c1, c2);
861 tcg_gen_neg_i32(t0, t0);
862 tcg_gen_and_i32(t1, v1, t0);
863 tcg_gen_andc_i32(ret, v2, t0);
864 tcg_gen_or_i32(ret, ret, t1);
865 tcg_temp_free_i32(t0);
866 tcg_temp_free_i32(t1);
867 }
868 }
869
870 void tcg_gen_add2_i32(TCGv_i32 rl, TCGv_i32 rh, TCGv_i32 al,
871 TCGv_i32 ah, TCGv_i32 bl, TCGv_i32 bh)
872 {
873 if (TCG_TARGET_HAS_add2_i32) {
874 tcg_gen_op6_i32(INDEX_op_add2_i32, rl, rh, al, ah, bl, bh);
875 } else {
876 TCGv_i64 t0 = tcg_temp_new_i64();
877 TCGv_i64 t1 = tcg_temp_new_i64();
878 tcg_gen_concat_i32_i64(t0, al, ah);
879 tcg_gen_concat_i32_i64(t1, bl, bh);
880 tcg_gen_add_i64(t0, t0, t1);
881 tcg_gen_extr_i64_i32(rl, rh, t0);
882 tcg_temp_free_i64(t0);
883 tcg_temp_free_i64(t1);
884 }
885 }
886
887 void tcg_gen_sub2_i32(TCGv_i32 rl, TCGv_i32 rh, TCGv_i32 al,
888 TCGv_i32 ah, TCGv_i32 bl, TCGv_i32 bh)
889 {
890 if (TCG_TARGET_HAS_sub2_i32) {
891 tcg_gen_op6_i32(INDEX_op_sub2_i32, rl, rh, al, ah, bl, bh);
892 } else {
893 TCGv_i64 t0 = tcg_temp_new_i64();
894 TCGv_i64 t1 = tcg_temp_new_i64();
895 tcg_gen_concat_i32_i64(t0, al, ah);
896 tcg_gen_concat_i32_i64(t1, bl, bh);
897 tcg_gen_sub_i64(t0, t0, t1);
898 tcg_gen_extr_i64_i32(rl, rh, t0);
899 tcg_temp_free_i64(t0);
900 tcg_temp_free_i64(t1);
901 }
902 }
903
904 void tcg_gen_mulu2_i32(TCGv_i32 rl, TCGv_i32 rh, TCGv_i32 arg1, TCGv_i32 arg2)
905 {
906 if (TCG_TARGET_HAS_mulu2_i32) {
907 tcg_gen_op4_i32(INDEX_op_mulu2_i32, rl, rh, arg1, arg2);
908 } else if (TCG_TARGET_HAS_muluh_i32) {
909 TCGv_i32 t = tcg_temp_new_i32();
910 tcg_gen_op3_i32(INDEX_op_mul_i32, t, arg1, arg2);
911 tcg_gen_op3_i32(INDEX_op_muluh_i32, rh, arg1, arg2);
912 tcg_gen_mov_i32(rl, t);
913 tcg_temp_free_i32(t);
914 } else {
915 TCGv_i64 t0 = tcg_temp_new_i64();
916 TCGv_i64 t1 = tcg_temp_new_i64();
917 tcg_gen_extu_i32_i64(t0, arg1);
918 tcg_gen_extu_i32_i64(t1, arg2);
919 tcg_gen_mul_i64(t0, t0, t1);
920 tcg_gen_extr_i64_i32(rl, rh, t0);
921 tcg_temp_free_i64(t0);
922 tcg_temp_free_i64(t1);
923 }
924 }
925
926 void tcg_gen_muls2_i32(TCGv_i32 rl, TCGv_i32 rh, TCGv_i32 arg1, TCGv_i32 arg2)
927 {
928 if (TCG_TARGET_HAS_muls2_i32) {
929 tcg_gen_op4_i32(INDEX_op_muls2_i32, rl, rh, arg1, arg2);
930 } else if (TCG_TARGET_HAS_mulsh_i32) {
931 TCGv_i32 t = tcg_temp_new_i32();
932 tcg_gen_op3_i32(INDEX_op_mul_i32, t, arg1, arg2);
933 tcg_gen_op3_i32(INDEX_op_mulsh_i32, rh, arg1, arg2);
934 tcg_gen_mov_i32(rl, t);
935 tcg_temp_free_i32(t);
936 } else if (TCG_TARGET_REG_BITS == 32) {
937 TCGv_i32 t0 = tcg_temp_new_i32();
938 TCGv_i32 t1 = tcg_temp_new_i32();
939 TCGv_i32 t2 = tcg_temp_new_i32();
940 TCGv_i32 t3 = tcg_temp_new_i32();
941 tcg_gen_mulu2_i32(t0, t1, arg1, arg2);
942 /* Adjust for negative inputs. */
943 tcg_gen_sari_i32(t2, arg1, 31);
944 tcg_gen_sari_i32(t3, arg2, 31);
945 tcg_gen_and_i32(t2, t2, arg2);
946 tcg_gen_and_i32(t3, t3, arg1);
947 tcg_gen_sub_i32(rh, t1, t2);
948 tcg_gen_sub_i32(rh, rh, t3);
949 tcg_gen_mov_i32(rl, t0);
950 tcg_temp_free_i32(t0);
951 tcg_temp_free_i32(t1);
952 tcg_temp_free_i32(t2);
953 tcg_temp_free_i32(t3);
954 } else {
955 TCGv_i64 t0 = tcg_temp_new_i64();
956 TCGv_i64 t1 = tcg_temp_new_i64();
957 tcg_gen_ext_i32_i64(t0, arg1);
958 tcg_gen_ext_i32_i64(t1, arg2);
959 tcg_gen_mul_i64(t0, t0, t1);
960 tcg_gen_extr_i64_i32(rl, rh, t0);
961 tcg_temp_free_i64(t0);
962 tcg_temp_free_i64(t1);
963 }
964 }
965
966 void tcg_gen_mulsu2_i32(TCGv_i32 rl, TCGv_i32 rh, TCGv_i32 arg1, TCGv_i32 arg2)
967 {
968 if (TCG_TARGET_REG_BITS == 32) {
969 TCGv_i32 t0 = tcg_temp_new_i32();
970 TCGv_i32 t1 = tcg_temp_new_i32();
971 TCGv_i32 t2 = tcg_temp_new_i32();
972 tcg_gen_mulu2_i32(t0, t1, arg1, arg2);
973 /* Adjust for negative input for the signed arg1. */
974 tcg_gen_sari_i32(t2, arg1, 31);
975 tcg_gen_and_i32(t2, t2, arg2);
976 tcg_gen_sub_i32(rh, t1, t2);
977 tcg_gen_mov_i32(rl, t0);
978 tcg_temp_free_i32(t0);
979 tcg_temp_free_i32(t1);
980 tcg_temp_free_i32(t2);
981 } else {
982 TCGv_i64 t0 = tcg_temp_new_i64();
983 TCGv_i64 t1 = tcg_temp_new_i64();
984 tcg_gen_ext_i32_i64(t0, arg1);
985 tcg_gen_extu_i32_i64(t1, arg2);
986 tcg_gen_mul_i64(t0, t0, t1);
987 tcg_gen_extr_i64_i32(rl, rh, t0);
988 tcg_temp_free_i64(t0);
989 tcg_temp_free_i64(t1);
990 }
991 }
992
993 void tcg_gen_ext8s_i32(TCGv_i32 ret, TCGv_i32 arg)
994 {
995 if (TCG_TARGET_HAS_ext8s_i32) {
996 tcg_gen_op2_i32(INDEX_op_ext8s_i32, ret, arg);
997 } else {
998 tcg_gen_shli_i32(ret, arg, 24);
999 tcg_gen_sari_i32(ret, ret, 24);
1000 }
1001 }
1002
1003 void tcg_gen_ext16s_i32(TCGv_i32 ret, TCGv_i32 arg)
1004 {
1005 if (TCG_TARGET_HAS_ext16s_i32) {
1006 tcg_gen_op2_i32(INDEX_op_ext16s_i32, ret, arg);
1007 } else {
1008 tcg_gen_shli_i32(ret, arg, 16);
1009 tcg_gen_sari_i32(ret, ret, 16);
1010 }
1011 }
1012
1013 void tcg_gen_ext8u_i32(TCGv_i32 ret, TCGv_i32 arg)
1014 {
1015 if (TCG_TARGET_HAS_ext8u_i32) {
1016 tcg_gen_op2_i32(INDEX_op_ext8u_i32, ret, arg);
1017 } else {
1018 tcg_gen_andi_i32(ret, arg, 0xffu);
1019 }
1020 }
1021
1022 void tcg_gen_ext16u_i32(TCGv_i32 ret, TCGv_i32 arg)
1023 {
1024 if (TCG_TARGET_HAS_ext16u_i32) {
1025 tcg_gen_op2_i32(INDEX_op_ext16u_i32, ret, arg);
1026 } else {
1027 tcg_gen_andi_i32(ret, arg, 0xffffu);
1028 }
1029 }
1030
1031 /* Note: we assume the two high bytes are set to zero */
1032 void tcg_gen_bswap16_i32(TCGv_i32 ret, TCGv_i32 arg)
1033 {
1034 if (TCG_TARGET_HAS_bswap16_i32) {
1035 tcg_gen_op2_i32(INDEX_op_bswap16_i32, ret, arg);
1036 } else {
1037 TCGv_i32 t0 = tcg_temp_new_i32();
1038
1039 tcg_gen_ext8u_i32(t0, arg);
1040 tcg_gen_shli_i32(t0, t0, 8);
1041 tcg_gen_shri_i32(ret, arg, 8);
1042 tcg_gen_or_i32(ret, ret, t0);
1043 tcg_temp_free_i32(t0);
1044 }
1045 }
1046
1047 void tcg_gen_bswap32_i32(TCGv_i32 ret, TCGv_i32 arg)
1048 {
1049 if (TCG_TARGET_HAS_bswap32_i32) {
1050 tcg_gen_op2_i32(INDEX_op_bswap32_i32, ret, arg);
1051 } else {
1052 TCGv_i32 t0 = tcg_temp_new_i32();
1053 TCGv_i32 t1 = tcg_temp_new_i32();
1054 TCGv_i32 t2 = tcg_const_i32(0x00ff00ff);
1055
1056 /* arg = abcd */
1057 tcg_gen_shri_i32(t0, arg, 8); /* t0 = .abc */
1058 tcg_gen_and_i32(t1, arg, t2); /* t1 = .b.d */
1059 tcg_gen_and_i32(t0, t0, t2); /* t0 = .a.c */
1060 tcg_gen_shli_i32(t1, t1, 8); /* t1 = b.d. */
1061 tcg_gen_or_i32(ret, t0, t1); /* ret = badc */
1062
1063 tcg_gen_shri_i32(t0, ret, 16); /* t0 = ..ba */
1064 tcg_gen_shli_i32(t1, ret, 16); /* t1 = dc.. */
1065 tcg_gen_or_i32(ret, t0, t1); /* ret = dcba */
1066
1067 tcg_temp_free_i32(t0);
1068 tcg_temp_free_i32(t1);
1069 tcg_temp_free_i32(t2);
1070 }
1071 }
1072
1073 void tcg_gen_smin_i32(TCGv_i32 ret, TCGv_i32 a, TCGv_i32 b)
1074 {
1075 tcg_gen_movcond_i32(TCG_COND_LT, ret, a, b, a, b);
1076 }
1077
1078 void tcg_gen_umin_i32(TCGv_i32 ret, TCGv_i32 a, TCGv_i32 b)
1079 {
1080 tcg_gen_movcond_i32(TCG_COND_LTU, ret, a, b, a, b);
1081 }
1082
1083 void tcg_gen_smax_i32(TCGv_i32 ret, TCGv_i32 a, TCGv_i32 b)
1084 {
1085 tcg_gen_movcond_i32(TCG_COND_LT, ret, a, b, b, a);
1086 }
1087
1088 void tcg_gen_umax_i32(TCGv_i32 ret, TCGv_i32 a, TCGv_i32 b)
1089 {
1090 tcg_gen_movcond_i32(TCG_COND_LTU, ret, a, b, b, a);
1091 }
1092
1093 void tcg_gen_abs_i32(TCGv_i32 ret, TCGv_i32 a)
1094 {
1095 TCGv_i32 t = tcg_temp_new_i32();
1096
1097 tcg_gen_sari_i32(t, a, 31);
1098 tcg_gen_xor_i32(ret, a, t);
1099 tcg_gen_sub_i32(ret, ret, t);
1100 tcg_temp_free_i32(t);
1101 }
1102
1103 /* 64-bit ops */
1104
1105 #if TCG_TARGET_REG_BITS == 32
1106 /* These are all inline for TCG_TARGET_REG_BITS == 64. */
1107
1108 void tcg_gen_discard_i64(TCGv_i64 arg)
1109 {
1110 tcg_gen_discard_i32(TCGV_LOW(arg));
1111 tcg_gen_discard_i32(TCGV_HIGH(arg));
1112 }
1113
1114 void tcg_gen_mov_i64(TCGv_i64 ret, TCGv_i64 arg)
1115 {
1116 tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1117 tcg_gen_mov_i32(TCGV_HIGH(ret), TCGV_HIGH(arg));
1118 }
1119
1120 void tcg_gen_movi_i64(TCGv_i64 ret, int64_t arg)
1121 {
1122 tcg_gen_movi_i32(TCGV_LOW(ret), arg);
1123 tcg_gen_movi_i32(TCGV_HIGH(ret), arg >> 32);
1124 }
1125
1126 void tcg_gen_ld8u_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
1127 {
1128 tcg_gen_ld8u_i32(TCGV_LOW(ret), arg2, offset);
1129 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1130 }
1131
1132 void tcg_gen_ld8s_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
1133 {
1134 tcg_gen_ld8s_i32(TCGV_LOW(ret), arg2, offset);
1135 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
1136 }
1137
1138 void tcg_gen_ld16u_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
1139 {
1140 tcg_gen_ld16u_i32(TCGV_LOW(ret), arg2, offset);
1141 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1142 }
1143
1144 void tcg_gen_ld16s_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
1145 {
1146 tcg_gen_ld16s_i32(TCGV_LOW(ret), arg2, offset);
1147 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
1148 }
1149
1150 void tcg_gen_ld32u_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
1151 {
1152 tcg_gen_ld_i32(TCGV_LOW(ret), arg2, offset);
1153 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1154 }
1155
1156 void tcg_gen_ld32s_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
1157 {
1158 tcg_gen_ld_i32(TCGV_LOW(ret), arg2, offset);
1159 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
1160 }
1161
1162 void tcg_gen_ld_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
1163 {
1164 /* Since arg2 and ret have different types,
1165 they cannot be the same temporary */
1166 #ifdef HOST_WORDS_BIGENDIAN
1167 tcg_gen_ld_i32(TCGV_HIGH(ret), arg2, offset);
1168 tcg_gen_ld_i32(TCGV_LOW(ret), arg2, offset + 4);
1169 #else
1170 tcg_gen_ld_i32(TCGV_LOW(ret), arg2, offset);
1171 tcg_gen_ld_i32(TCGV_HIGH(ret), arg2, offset + 4);
1172 #endif
1173 }
1174
1175 void tcg_gen_st_i64(TCGv_i64 arg1, TCGv_ptr arg2, tcg_target_long offset)
1176 {
1177 #ifdef HOST_WORDS_BIGENDIAN
1178 tcg_gen_st_i32(TCGV_HIGH(arg1), arg2, offset);
1179 tcg_gen_st_i32(TCGV_LOW(arg1), arg2, offset + 4);
1180 #else
1181 tcg_gen_st_i32(TCGV_LOW(arg1), arg2, offset);
1182 tcg_gen_st_i32(TCGV_HIGH(arg1), arg2, offset + 4);
1183 #endif
1184 }
1185
1186 void tcg_gen_and_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1187 {
1188 tcg_gen_and_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1189 tcg_gen_and_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1190 }
1191
1192 void tcg_gen_or_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1193 {
1194 tcg_gen_or_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1195 tcg_gen_or_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1196 }
1197
1198 void tcg_gen_xor_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1199 {
1200 tcg_gen_xor_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1201 tcg_gen_xor_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1202 }
1203
1204 void tcg_gen_shl_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1205 {
1206 gen_helper_shl_i64(ret, arg1, arg2);
1207 }
1208
1209 void tcg_gen_shr_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1210 {
1211 gen_helper_shr_i64(ret, arg1, arg2);
1212 }
1213
1214 void tcg_gen_sar_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1215 {
1216 gen_helper_sar_i64(ret, arg1, arg2);
1217 }
1218
1219 void tcg_gen_mul_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1220 {
1221 TCGv_i64 t0;
1222 TCGv_i32 t1;
1223
1224 t0 = tcg_temp_new_i64();
1225 t1 = tcg_temp_new_i32();
1226
1227 tcg_gen_mulu2_i32(TCGV_LOW(t0), TCGV_HIGH(t0),
1228 TCGV_LOW(arg1), TCGV_LOW(arg2));
1229
1230 tcg_gen_mul_i32(t1, TCGV_LOW(arg1), TCGV_HIGH(arg2));
1231 tcg_gen_add_i32(TCGV_HIGH(t0), TCGV_HIGH(t0), t1);
1232 tcg_gen_mul_i32(t1, TCGV_HIGH(arg1), TCGV_LOW(arg2));
1233 tcg_gen_add_i32(TCGV_HIGH(t0), TCGV_HIGH(t0), t1);
1234
1235 tcg_gen_mov_i64(ret, t0);
1236 tcg_temp_free_i64(t0);
1237 tcg_temp_free_i32(t1);
1238 }
1239 #endif /* TCG_TARGET_REG_SIZE == 32 */
1240
1241 void tcg_gen_addi_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1242 {
1243 /* some cases can be optimized here */
1244 if (arg2 == 0) {
1245 tcg_gen_mov_i64(ret, arg1);
1246 } else {
1247 TCGv_i64 t0 = tcg_const_i64(arg2);
1248 tcg_gen_add_i64(ret, arg1, t0);
1249 tcg_temp_free_i64(t0);
1250 }
1251 }
1252
1253 void tcg_gen_subfi_i64(TCGv_i64 ret, int64_t arg1, TCGv_i64 arg2)
1254 {
1255 if (arg1 == 0 && TCG_TARGET_HAS_neg_i64) {
1256 /* Don't recurse with tcg_gen_neg_i64. */
1257 tcg_gen_op2_i64(INDEX_op_neg_i64, ret, arg2);
1258 } else {
1259 TCGv_i64 t0 = tcg_const_i64(arg1);
1260 tcg_gen_sub_i64(ret, t0, arg2);
1261 tcg_temp_free_i64(t0);
1262 }
1263 }
1264
1265 void tcg_gen_subi_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1266 {
1267 /* some cases can be optimized here */
1268 if (arg2 == 0) {
1269 tcg_gen_mov_i64(ret, arg1);
1270 } else {
1271 TCGv_i64 t0 = tcg_const_i64(arg2);
1272 tcg_gen_sub_i64(ret, arg1, t0);
1273 tcg_temp_free_i64(t0);
1274 }
1275 }
1276
1277 void tcg_gen_andi_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1278 {
1279 TCGv_i64 t0;
1280
1281 if (TCG_TARGET_REG_BITS == 32) {
1282 tcg_gen_andi_i32(TCGV_LOW(ret), TCGV_LOW(arg1), arg2);
1283 tcg_gen_andi_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), arg2 >> 32);
1284 return;
1285 }
1286
1287 /* Some cases can be optimized here. */
1288 switch (arg2) {
1289 case 0:
1290 tcg_gen_movi_i64(ret, 0);
1291 return;
1292 case -1:
1293 tcg_gen_mov_i64(ret, arg1);
1294 return;
1295 case 0xff:
1296 /* Don't recurse with tcg_gen_ext8u_i64. */
1297 if (TCG_TARGET_HAS_ext8u_i64) {
1298 tcg_gen_op2_i64(INDEX_op_ext8u_i64, ret, arg1);
1299 return;
1300 }
1301 break;
1302 case 0xffff:
1303 if (TCG_TARGET_HAS_ext16u_i64) {
1304 tcg_gen_op2_i64(INDEX_op_ext16u_i64, ret, arg1);
1305 return;
1306 }
1307 break;
1308 case 0xffffffffu:
1309 if (TCG_TARGET_HAS_ext32u_i64) {
1310 tcg_gen_op2_i64(INDEX_op_ext32u_i64, ret, arg1);
1311 return;
1312 }
1313 break;
1314 }
1315 t0 = tcg_const_i64(arg2);
1316 tcg_gen_and_i64(ret, arg1, t0);
1317 tcg_temp_free_i64(t0);
1318 }
1319
1320 void tcg_gen_ori_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1321 {
1322 if (TCG_TARGET_REG_BITS == 32) {
1323 tcg_gen_ori_i32(TCGV_LOW(ret), TCGV_LOW(arg1), arg2);
1324 tcg_gen_ori_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), arg2 >> 32);
1325 return;
1326 }
1327 /* Some cases can be optimized here. */
1328 if (arg2 == -1) {
1329 tcg_gen_movi_i64(ret, -1);
1330 } else if (arg2 == 0) {
1331 tcg_gen_mov_i64(ret, arg1);
1332 } else {
1333 TCGv_i64 t0 = tcg_const_i64(arg2);
1334 tcg_gen_or_i64(ret, arg1, t0);
1335 tcg_temp_free_i64(t0);
1336 }
1337 }
1338
1339 void tcg_gen_xori_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1340 {
1341 if (TCG_TARGET_REG_BITS == 32) {
1342 tcg_gen_xori_i32(TCGV_LOW(ret), TCGV_LOW(arg1), arg2);
1343 tcg_gen_xori_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), arg2 >> 32);
1344 return;
1345 }
1346 /* Some cases can be optimized here. */
1347 if (arg2 == 0) {
1348 tcg_gen_mov_i64(ret, arg1);
1349 } else if (arg2 == -1 && TCG_TARGET_HAS_not_i64) {
1350 /* Don't recurse with tcg_gen_not_i64. */
1351 tcg_gen_op2_i64(INDEX_op_not_i64, ret, arg1);
1352 } else {
1353 TCGv_i64 t0 = tcg_const_i64(arg2);
1354 tcg_gen_xor_i64(ret, arg1, t0);
1355 tcg_temp_free_i64(t0);
1356 }
1357 }
1358
1359 static inline void tcg_gen_shifti_i64(TCGv_i64 ret, TCGv_i64 arg1,
1360 unsigned c, bool right, bool arith)
1361 {
1362 tcg_debug_assert(c < 64);
1363 if (c == 0) {
1364 tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg1));
1365 tcg_gen_mov_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1));
1366 } else if (c >= 32) {
1367 c -= 32;
1368 if (right) {
1369 if (arith) {
1370 tcg_gen_sari_i32(TCGV_LOW(ret), TCGV_HIGH(arg1), c);
1371 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), 31);
1372 } else {
1373 tcg_gen_shri_i32(TCGV_LOW(ret), TCGV_HIGH(arg1), c);
1374 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1375 }
1376 } else {
1377 tcg_gen_shli_i32(TCGV_HIGH(ret), TCGV_LOW(arg1), c);
1378 tcg_gen_movi_i32(TCGV_LOW(ret), 0);
1379 }
1380 } else if (right) {
1381 if (TCG_TARGET_HAS_extract2_i32) {
1382 tcg_gen_extract2_i32(TCGV_LOW(ret),
1383 TCGV_LOW(arg1), TCGV_HIGH(arg1), c);
1384 } else {
1385 tcg_gen_shri_i32(TCGV_LOW(ret), TCGV_LOW(arg1), c);
1386 tcg_gen_deposit_i32(TCGV_LOW(ret), TCGV_LOW(ret),
1387 TCGV_HIGH(arg1), 32 - c, c);
1388 }
1389 if (arith) {
1390 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), c);
1391 } else {
1392 tcg_gen_shri_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), c);
1393 }
1394 } else {
1395 if (TCG_TARGET_HAS_extract2_i32) {
1396 tcg_gen_extract2_i32(TCGV_HIGH(ret),
1397 TCGV_LOW(arg1), TCGV_HIGH(arg1), 32 - c);
1398 } else {
1399 TCGv_i32 t0 = tcg_temp_new_i32();
1400 tcg_gen_shri_i32(t0, TCGV_LOW(arg1), 32 - c);
1401 tcg_gen_deposit_i32(TCGV_HIGH(ret), t0,
1402 TCGV_HIGH(arg1), c, 32 - c);
1403 tcg_temp_free_i32(t0);
1404 }
1405 tcg_gen_shli_i32(TCGV_LOW(ret), TCGV_LOW(arg1), c);
1406 }
1407 }
1408
1409 void tcg_gen_shli_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1410 {
1411 tcg_debug_assert(arg2 >= 0 && arg2 < 64);
1412 if (TCG_TARGET_REG_BITS == 32) {
1413 tcg_gen_shifti_i64(ret, arg1, arg2, 0, 0);
1414 } else if (arg2 == 0) {
1415 tcg_gen_mov_i64(ret, arg1);
1416 } else {
1417 TCGv_i64 t0 = tcg_const_i64(arg2);
1418 tcg_gen_shl_i64(ret, arg1, t0);
1419 tcg_temp_free_i64(t0);
1420 }
1421 }
1422
1423 void tcg_gen_shri_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1424 {
1425 tcg_debug_assert(arg2 >= 0 && arg2 < 64);
1426 if (TCG_TARGET_REG_BITS == 32) {
1427 tcg_gen_shifti_i64(ret, arg1, arg2, 1, 0);
1428 } else if (arg2 == 0) {
1429 tcg_gen_mov_i64(ret, arg1);
1430 } else {
1431 TCGv_i64 t0 = tcg_const_i64(arg2);
1432 tcg_gen_shr_i64(ret, arg1, t0);
1433 tcg_temp_free_i64(t0);
1434 }
1435 }
1436
1437 void tcg_gen_sari_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1438 {
1439 tcg_debug_assert(arg2 >= 0 && arg2 < 64);
1440 if (TCG_TARGET_REG_BITS == 32) {
1441 tcg_gen_shifti_i64(ret, arg1, arg2, 1, 1);
1442 } else if (arg2 == 0) {
1443 tcg_gen_mov_i64(ret, arg1);
1444 } else {
1445 TCGv_i64 t0 = tcg_const_i64(arg2);
1446 tcg_gen_sar_i64(ret, arg1, t0);
1447 tcg_temp_free_i64(t0);
1448 }
1449 }
1450
1451 void tcg_gen_brcond_i64(TCGCond cond, TCGv_i64 arg1, TCGv_i64 arg2, TCGLabel *l)
1452 {
1453 if (cond == TCG_COND_ALWAYS) {
1454 tcg_gen_br(l);
1455 } else if (cond != TCG_COND_NEVER) {
1456 l->refs++;
1457 if (TCG_TARGET_REG_BITS == 32) {
1458 tcg_gen_op6ii_i32(INDEX_op_brcond2_i32, TCGV_LOW(arg1),
1459 TCGV_HIGH(arg1), TCGV_LOW(arg2),
1460 TCGV_HIGH(arg2), cond, label_arg(l));
1461 } else {
1462 tcg_gen_op4ii_i64(INDEX_op_brcond_i64, arg1, arg2, cond,
1463 label_arg(l));
1464 }
1465 }
1466 }
1467
1468 void tcg_gen_brcondi_i64(TCGCond cond, TCGv_i64 arg1, int64_t arg2, TCGLabel *l)
1469 {
1470 if (cond == TCG_COND_ALWAYS) {
1471 tcg_gen_br(l);
1472 } else if (cond != TCG_COND_NEVER) {
1473 TCGv_i64 t0 = tcg_const_i64(arg2);
1474 tcg_gen_brcond_i64(cond, arg1, t0, l);
1475 tcg_temp_free_i64(t0);
1476 }
1477 }
1478
1479 void tcg_gen_setcond_i64(TCGCond cond, TCGv_i64 ret,
1480 TCGv_i64 arg1, TCGv_i64 arg2)
1481 {
1482 if (cond == TCG_COND_ALWAYS) {
1483 tcg_gen_movi_i64(ret, 1);
1484 } else if (cond == TCG_COND_NEVER) {
1485 tcg_gen_movi_i64(ret, 0);
1486 } else {
1487 if (TCG_TARGET_REG_BITS == 32) {
1488 tcg_gen_op6i_i32(INDEX_op_setcond2_i32, TCGV_LOW(ret),
1489 TCGV_LOW(arg1), TCGV_HIGH(arg1),
1490 TCGV_LOW(arg2), TCGV_HIGH(arg2), cond);
1491 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1492 } else {
1493 tcg_gen_op4i_i64(INDEX_op_setcond_i64, ret, arg1, arg2, cond);
1494 }
1495 }
1496 }
1497
1498 void tcg_gen_setcondi_i64(TCGCond cond, TCGv_i64 ret,
1499 TCGv_i64 arg1, int64_t arg2)
1500 {
1501 TCGv_i64 t0 = tcg_const_i64(arg2);
1502 tcg_gen_setcond_i64(cond, ret, arg1, t0);
1503 tcg_temp_free_i64(t0);
1504 }
1505
1506 void tcg_gen_muli_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1507 {
1508 if (arg2 == 0) {
1509 tcg_gen_movi_i64(ret, 0);
1510 } else if (is_power_of_2(arg2)) {
1511 tcg_gen_shli_i64(ret, arg1, ctz64(arg2));
1512 } else {
1513 TCGv_i64 t0 = tcg_const_i64(arg2);
1514 tcg_gen_mul_i64(ret, arg1, t0);
1515 tcg_temp_free_i64(t0);
1516 }
1517 }
1518
1519 void tcg_gen_div_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1520 {
1521 if (TCG_TARGET_HAS_div_i64) {
1522 tcg_gen_op3_i64(INDEX_op_div_i64, ret, arg1, arg2);
1523 } else if (TCG_TARGET_HAS_div2_i64) {
1524 TCGv_i64 t0 = tcg_temp_new_i64();
1525 tcg_gen_sari_i64(t0, arg1, 63);
1526 tcg_gen_op5_i64(INDEX_op_div2_i64, ret, t0, arg1, t0, arg2);
1527 tcg_temp_free_i64(t0);
1528 } else {
1529 gen_helper_div_i64(ret, arg1, arg2);
1530 }
1531 }
1532
1533 void tcg_gen_rem_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1534 {
1535 if (TCG_TARGET_HAS_rem_i64) {
1536 tcg_gen_op3_i64(INDEX_op_rem_i64, ret, arg1, arg2);
1537 } else if (TCG_TARGET_HAS_div_i64) {
1538 TCGv_i64 t0 = tcg_temp_new_i64();
1539 tcg_gen_op3_i64(INDEX_op_div_i64, t0, arg1, arg2);
1540 tcg_gen_mul_i64(t0, t0, arg2);
1541 tcg_gen_sub_i64(ret, arg1, t0);
1542 tcg_temp_free_i64(t0);
1543 } else if (TCG_TARGET_HAS_div2_i64) {
1544 TCGv_i64 t0 = tcg_temp_new_i64();
1545 tcg_gen_sari_i64(t0, arg1, 63);
1546 tcg_gen_op5_i64(INDEX_op_div2_i64, t0, ret, arg1, t0, arg2);
1547 tcg_temp_free_i64(t0);
1548 } else {
1549 gen_helper_rem_i64(ret, arg1, arg2);
1550 }
1551 }
1552
1553 void tcg_gen_divu_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1554 {
1555 if (TCG_TARGET_HAS_div_i64) {
1556 tcg_gen_op3_i64(INDEX_op_divu_i64, ret, arg1, arg2);
1557 } else if (TCG_TARGET_HAS_div2_i64) {
1558 TCGv_i64 t0 = tcg_temp_new_i64();
1559 tcg_gen_movi_i64(t0, 0);
1560 tcg_gen_op5_i64(INDEX_op_divu2_i64, ret, t0, arg1, t0, arg2);
1561 tcg_temp_free_i64(t0);
1562 } else {
1563 gen_helper_divu_i64(ret, arg1, arg2);
1564 }
1565 }
1566
1567 void tcg_gen_remu_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1568 {
1569 if (TCG_TARGET_HAS_rem_i64) {
1570 tcg_gen_op3_i64(INDEX_op_remu_i64, ret, arg1, arg2);
1571 } else if (TCG_TARGET_HAS_div_i64) {
1572 TCGv_i64 t0 = tcg_temp_new_i64();
1573 tcg_gen_op3_i64(INDEX_op_divu_i64, t0, arg1, arg2);
1574 tcg_gen_mul_i64(t0, t0, arg2);
1575 tcg_gen_sub_i64(ret, arg1, t0);
1576 tcg_temp_free_i64(t0);
1577 } else if (TCG_TARGET_HAS_div2_i64) {
1578 TCGv_i64 t0 = tcg_temp_new_i64();
1579 tcg_gen_movi_i64(t0, 0);
1580 tcg_gen_op5_i64(INDEX_op_divu2_i64, t0, ret, arg1, t0, arg2);
1581 tcg_temp_free_i64(t0);
1582 } else {
1583 gen_helper_remu_i64(ret, arg1, arg2);
1584 }
1585 }
1586
1587 void tcg_gen_ext8s_i64(TCGv_i64 ret, TCGv_i64 arg)
1588 {
1589 if (TCG_TARGET_REG_BITS == 32) {
1590 tcg_gen_ext8s_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1591 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
1592 } else if (TCG_TARGET_HAS_ext8s_i64) {
1593 tcg_gen_op2_i64(INDEX_op_ext8s_i64, ret, arg);
1594 } else {
1595 tcg_gen_shli_i64(ret, arg, 56);
1596 tcg_gen_sari_i64(ret, ret, 56);
1597 }
1598 }
1599
1600 void tcg_gen_ext16s_i64(TCGv_i64 ret, TCGv_i64 arg)
1601 {
1602 if (TCG_TARGET_REG_BITS == 32) {
1603 tcg_gen_ext16s_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1604 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
1605 } else if (TCG_TARGET_HAS_ext16s_i64) {
1606 tcg_gen_op2_i64(INDEX_op_ext16s_i64, ret, arg);
1607 } else {
1608 tcg_gen_shli_i64(ret, arg, 48);
1609 tcg_gen_sari_i64(ret, ret, 48);
1610 }
1611 }
1612
1613 void tcg_gen_ext32s_i64(TCGv_i64 ret, TCGv_i64 arg)
1614 {
1615 if (TCG_TARGET_REG_BITS == 32) {
1616 tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1617 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
1618 } else if (TCG_TARGET_HAS_ext32s_i64) {
1619 tcg_gen_op2_i64(INDEX_op_ext32s_i64, ret, arg);
1620 } else {
1621 tcg_gen_shli_i64(ret, arg, 32);
1622 tcg_gen_sari_i64(ret, ret, 32);
1623 }
1624 }
1625
1626 void tcg_gen_ext8u_i64(TCGv_i64 ret, TCGv_i64 arg)
1627 {
1628 if (TCG_TARGET_REG_BITS == 32) {
1629 tcg_gen_ext8u_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1630 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1631 } else if (TCG_TARGET_HAS_ext8u_i64) {
1632 tcg_gen_op2_i64(INDEX_op_ext8u_i64, ret, arg);
1633 } else {
1634 tcg_gen_andi_i64(ret, arg, 0xffu);
1635 }
1636 }
1637
1638 void tcg_gen_ext16u_i64(TCGv_i64 ret, TCGv_i64 arg)
1639 {
1640 if (TCG_TARGET_REG_BITS == 32) {
1641 tcg_gen_ext16u_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1642 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1643 } else if (TCG_TARGET_HAS_ext16u_i64) {
1644 tcg_gen_op2_i64(INDEX_op_ext16u_i64, ret, arg);
1645 } else {
1646 tcg_gen_andi_i64(ret, arg, 0xffffu);
1647 }
1648 }
1649
1650 void tcg_gen_ext32u_i64(TCGv_i64 ret, TCGv_i64 arg)
1651 {
1652 if (TCG_TARGET_REG_BITS == 32) {
1653 tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1654 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1655 } else if (TCG_TARGET_HAS_ext32u_i64) {
1656 tcg_gen_op2_i64(INDEX_op_ext32u_i64, ret, arg);
1657 } else {
1658 tcg_gen_andi_i64(ret, arg, 0xffffffffu);
1659 }
1660 }
1661
1662 /* Note: we assume the six high bytes are set to zero */
1663 void tcg_gen_bswap16_i64(TCGv_i64 ret, TCGv_i64 arg)
1664 {
1665 if (TCG_TARGET_REG_BITS == 32) {
1666 tcg_gen_bswap16_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1667 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1668 } else if (TCG_TARGET_HAS_bswap16_i64) {
1669 tcg_gen_op2_i64(INDEX_op_bswap16_i64, ret, arg);
1670 } else {
1671 TCGv_i64 t0 = tcg_temp_new_i64();
1672
1673 tcg_gen_ext8u_i64(t0, arg);
1674 tcg_gen_shli_i64(t0, t0, 8);
1675 tcg_gen_shri_i64(ret, arg, 8);
1676 tcg_gen_or_i64(ret, ret, t0);
1677 tcg_temp_free_i64(t0);
1678 }
1679 }
1680
1681 /* Note: we assume the four high bytes are set to zero */
1682 void tcg_gen_bswap32_i64(TCGv_i64 ret, TCGv_i64 arg)
1683 {
1684 if (TCG_TARGET_REG_BITS == 32) {
1685 tcg_gen_bswap32_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1686 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1687 } else if (TCG_TARGET_HAS_bswap32_i64) {
1688 tcg_gen_op2_i64(INDEX_op_bswap32_i64, ret, arg);
1689 } else {
1690 TCGv_i64 t0 = tcg_temp_new_i64();
1691 TCGv_i64 t1 = tcg_temp_new_i64();
1692 TCGv_i64 t2 = tcg_const_i64(0x00ff00ff);
1693
1694 /* arg = ....abcd */
1695 tcg_gen_shri_i64(t0, arg, 8); /* t0 = .....abc */
1696 tcg_gen_and_i64(t1, arg, t2); /* t1 = .....b.d */
1697 tcg_gen_and_i64(t0, t0, t2); /* t0 = .....a.c */
1698 tcg_gen_shli_i64(t1, t1, 8); /* t1 = ....b.d. */
1699 tcg_gen_or_i64(ret, t0, t1); /* ret = ....badc */
1700
1701 tcg_gen_shli_i64(t1, ret, 48); /* t1 = dc...... */
1702 tcg_gen_shri_i64(t0, ret, 16); /* t0 = ......ba */
1703 tcg_gen_shri_i64(t1, t1, 32); /* t1 = ....dc.. */
1704 tcg_gen_or_i64(ret, t0, t1); /* ret = ....dcba */
1705
1706 tcg_temp_free_i64(t0);
1707 tcg_temp_free_i64(t1);
1708 tcg_temp_free_i64(t2);
1709 }
1710 }
1711
1712 void tcg_gen_bswap64_i64(TCGv_i64 ret, TCGv_i64 arg)
1713 {
1714 if (TCG_TARGET_REG_BITS == 32) {
1715 TCGv_i32 t0, t1;
1716 t0 = tcg_temp_new_i32();
1717 t1 = tcg_temp_new_i32();
1718
1719 tcg_gen_bswap32_i32(t0, TCGV_LOW(arg));
1720 tcg_gen_bswap32_i32(t1, TCGV_HIGH(arg));
1721 tcg_gen_mov_i32(TCGV_LOW(ret), t1);
1722 tcg_gen_mov_i32(TCGV_HIGH(ret), t0);
1723 tcg_temp_free_i32(t0);
1724 tcg_temp_free_i32(t1);
1725 } else if (TCG_TARGET_HAS_bswap64_i64) {
1726 tcg_gen_op2_i64(INDEX_op_bswap64_i64, ret, arg);
1727 } else {
1728 TCGv_i64 t0 = tcg_temp_new_i64();
1729 TCGv_i64 t1 = tcg_temp_new_i64();
1730 TCGv_i64 t2 = tcg_temp_new_i64();
1731
1732 /* arg = abcdefgh */
1733 tcg_gen_movi_i64(t2, 0x00ff00ff00ff00ffull);
1734 tcg_gen_shri_i64(t0, arg, 8); /* t0 = .abcdefg */
1735 tcg_gen_and_i64(t1, arg, t2); /* t1 = .b.d.f.h */
1736 tcg_gen_and_i64(t0, t0, t2); /* t0 = .a.c.e.g */
1737 tcg_gen_shli_i64(t1, t1, 8); /* t1 = b.d.f.h. */
1738 tcg_gen_or_i64(ret, t0, t1); /* ret = badcfehg */
1739
1740 tcg_gen_movi_i64(t2, 0x0000ffff0000ffffull);
1741 tcg_gen_shri_i64(t0, ret, 16); /* t0 = ..badcfe */
1742 tcg_gen_and_i64(t1, ret, t2); /* t1 = ..dc..hg */
1743 tcg_gen_and_i64(t0, t0, t2); /* t0 = ..ba..fe */
1744 tcg_gen_shli_i64(t1, t1, 16); /* t1 = dc..hg.. */
1745 tcg_gen_or_i64(ret, t0, t1); /* ret = dcbahgfe */
1746
1747 tcg_gen_shri_i64(t0, ret, 32); /* t0 = ....dcba */
1748 tcg_gen_shli_i64(t1, ret, 32); /* t1 = hgfe.... */
1749 tcg_gen_or_i64(ret, t0, t1); /* ret = hgfedcba */
1750
1751 tcg_temp_free_i64(t0);
1752 tcg_temp_free_i64(t1);
1753 tcg_temp_free_i64(t2);
1754 }
1755 }
1756
1757 void tcg_gen_not_i64(TCGv_i64 ret, TCGv_i64 arg)
1758 {
1759 if (TCG_TARGET_REG_BITS == 32) {
1760 tcg_gen_not_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1761 tcg_gen_not_i32(TCGV_HIGH(ret), TCGV_HIGH(arg));
1762 } else if (TCG_TARGET_HAS_not_i64) {
1763 tcg_gen_op2_i64(INDEX_op_not_i64, ret, arg);
1764 } else {
1765 tcg_gen_xori_i64(ret, arg, -1);
1766 }
1767 }
1768
1769 void tcg_gen_andc_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1770 {
1771 if (TCG_TARGET_REG_BITS == 32) {
1772 tcg_gen_andc_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1773 tcg_gen_andc_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1774 } else if (TCG_TARGET_HAS_andc_i64) {
1775 tcg_gen_op3_i64(INDEX_op_andc_i64, ret, arg1, arg2);
1776 } else {
1777 TCGv_i64 t0 = tcg_temp_new_i64();
1778 tcg_gen_not_i64(t0, arg2);
1779 tcg_gen_and_i64(ret, arg1, t0);
1780 tcg_temp_free_i64(t0);
1781 }
1782 }
1783
1784 void tcg_gen_eqv_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1785 {
1786 if (TCG_TARGET_REG_BITS == 32) {
1787 tcg_gen_eqv_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1788 tcg_gen_eqv_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1789 } else if (TCG_TARGET_HAS_eqv_i64) {
1790 tcg_gen_op3_i64(INDEX_op_eqv_i64, ret, arg1, arg2);
1791 } else {
1792 tcg_gen_xor_i64(ret, arg1, arg2);
1793 tcg_gen_not_i64(ret, ret);
1794 }
1795 }
1796
1797 void tcg_gen_nand_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1798 {
1799 if (TCG_TARGET_REG_BITS == 32) {
1800 tcg_gen_nand_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1801 tcg_gen_nand_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1802 } else if (TCG_TARGET_HAS_nand_i64) {
1803 tcg_gen_op3_i64(INDEX_op_nand_i64, ret, arg1, arg2);
1804 } else {
1805 tcg_gen_and_i64(ret, arg1, arg2);
1806 tcg_gen_not_i64(ret, ret);
1807 }
1808 }
1809
1810 void tcg_gen_nor_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1811 {
1812 if (TCG_TARGET_REG_BITS == 32) {
1813 tcg_gen_nor_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1814 tcg_gen_nor_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1815 } else if (TCG_TARGET_HAS_nor_i64) {
1816 tcg_gen_op3_i64(INDEX_op_nor_i64, ret, arg1, arg2);
1817 } else {
1818 tcg_gen_or_i64(ret, arg1, arg2);
1819 tcg_gen_not_i64(ret, ret);
1820 }
1821 }
1822
1823 void tcg_gen_orc_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1824 {
1825 if (TCG_TARGET_REG_BITS == 32) {
1826 tcg_gen_orc_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1827 tcg_gen_orc_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1828 } else if (TCG_TARGET_HAS_orc_i64) {
1829 tcg_gen_op3_i64(INDEX_op_orc_i64, ret, arg1, arg2);
1830 } else {
1831 TCGv_i64 t0 = tcg_temp_new_i64();
1832 tcg_gen_not_i64(t0, arg2);
1833 tcg_gen_or_i64(ret, arg1, t0);
1834 tcg_temp_free_i64(t0);
1835 }
1836 }
1837
1838 void tcg_gen_clz_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1839 {
1840 if (TCG_TARGET_HAS_clz_i64) {
1841 tcg_gen_op3_i64(INDEX_op_clz_i64, ret, arg1, arg2);
1842 } else {
1843 gen_helper_clz_i64(ret, arg1, arg2);
1844 }
1845 }
1846
1847 void tcg_gen_clzi_i64(TCGv_i64 ret, TCGv_i64 arg1, uint64_t arg2)
1848 {
1849 if (TCG_TARGET_REG_BITS == 32
1850 && TCG_TARGET_HAS_clz_i32
1851 && arg2 <= 0xffffffffu) {
1852 TCGv_i32 t = tcg_const_i32((uint32_t)arg2 - 32);
1853 tcg_gen_clz_i32(t, TCGV_LOW(arg1), t);
1854 tcg_gen_addi_i32(t, t, 32);
1855 tcg_gen_clz_i32(TCGV_LOW(ret), TCGV_HIGH(arg1), t);
1856 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1857 tcg_temp_free_i32(t);
1858 } else {
1859 TCGv_i64 t = tcg_const_i64(arg2);
1860 tcg_gen_clz_i64(ret, arg1, t);
1861 tcg_temp_free_i64(t);
1862 }
1863 }
1864
1865 void tcg_gen_ctz_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1866 {
1867 if (TCG_TARGET_HAS_ctz_i64) {
1868 tcg_gen_op3_i64(INDEX_op_ctz_i64, ret, arg1, arg2);
1869 } else if (TCG_TARGET_HAS_ctpop_i64 || TCG_TARGET_HAS_clz_i64) {
1870 TCGv_i64 z, t = tcg_temp_new_i64();
1871
1872 if (TCG_TARGET_HAS_ctpop_i64) {
1873 tcg_gen_subi_i64(t, arg1, 1);
1874 tcg_gen_andc_i64(t, t, arg1);
1875 tcg_gen_ctpop_i64(t, t);
1876 } else {
1877 /* Since all non-x86 hosts have clz(0) == 64, don't fight it. */
1878 tcg_gen_neg_i64(t, arg1);
1879 tcg_gen_and_i64(t, t, arg1);
1880 tcg_gen_clzi_i64(t, t, 64);
1881 tcg_gen_xori_i64(t, t, 63);
1882 }
1883 z = tcg_const_i64(0);
1884 tcg_gen_movcond_i64(TCG_COND_EQ, ret, arg1, z, arg2, t);
1885 tcg_temp_free_i64(t);
1886 tcg_temp_free_i64(z);
1887 } else {
1888 gen_helper_ctz_i64(ret, arg1, arg2);
1889 }
1890 }
1891
1892 void tcg_gen_ctzi_i64(TCGv_i64 ret, TCGv_i64 arg1, uint64_t arg2)
1893 {
1894 if (TCG_TARGET_REG_BITS == 32
1895 && TCG_TARGET_HAS_ctz_i32
1896 && arg2 <= 0xffffffffu) {
1897 TCGv_i32 t32 = tcg_const_i32((uint32_t)arg2 - 32);
1898 tcg_gen_ctz_i32(t32, TCGV_HIGH(arg1), t32);
1899 tcg_gen_addi_i32(t32, t32, 32);
1900 tcg_gen_ctz_i32(TCGV_LOW(ret), TCGV_LOW(arg1), t32);
1901 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1902 tcg_temp_free_i32(t32);
1903 } else if (!TCG_TARGET_HAS_ctz_i64
1904 && TCG_TARGET_HAS_ctpop_i64
1905 && arg2 == 64) {
1906 /* This equivalence has the advantage of not requiring a fixup. */
1907 TCGv_i64 t = tcg_temp_new_i64();
1908 tcg_gen_subi_i64(t, arg1, 1);
1909 tcg_gen_andc_i64(t, t, arg1);
1910 tcg_gen_ctpop_i64(ret, t);
1911 tcg_temp_free_i64(t);
1912 } else {
1913 TCGv_i64 t64 = tcg_const_i64(arg2);
1914 tcg_gen_ctz_i64(ret, arg1, t64);
1915 tcg_temp_free_i64(t64);
1916 }
1917 }
1918
1919 void tcg_gen_clrsb_i64(TCGv_i64 ret, TCGv_i64 arg)
1920 {
1921 if (TCG_TARGET_HAS_clz_i64 || TCG_TARGET_HAS_clz_i32) {
1922 TCGv_i64 t = tcg_temp_new_i64();
1923 tcg_gen_sari_i64(t, arg, 63);
1924 tcg_gen_xor_i64(t, t, arg);
1925 tcg_gen_clzi_i64(t, t, 64);
1926 tcg_gen_subi_i64(ret, t, 1);
1927 tcg_temp_free_i64(t);
1928 } else {
1929 gen_helper_clrsb_i64(ret, arg);
1930 }
1931 }
1932
1933 void tcg_gen_ctpop_i64(TCGv_i64 ret, TCGv_i64 arg1)
1934 {
1935 if (TCG_TARGET_HAS_ctpop_i64) {
1936 tcg_gen_op2_i64(INDEX_op_ctpop_i64, ret, arg1);
1937 } else if (TCG_TARGET_REG_BITS == 32 && TCG_TARGET_HAS_ctpop_i32) {
1938 tcg_gen_ctpop_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1));
1939 tcg_gen_ctpop_i32(TCGV_LOW(ret), TCGV_LOW(arg1));
1940 tcg_gen_add_i32(TCGV_LOW(ret), TCGV_LOW(ret), TCGV_HIGH(ret));
1941 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1942 } else {
1943 gen_helper_ctpop_i64(ret, arg1);
1944 }
1945 }
1946
1947 void tcg_gen_rotl_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1948 {
1949 if (TCG_TARGET_HAS_rot_i64) {
1950 tcg_gen_op3_i64(INDEX_op_rotl_i64, ret, arg1, arg2);
1951 } else {
1952 TCGv_i64 t0, t1;
1953 t0 = tcg_temp_new_i64();
1954 t1 = tcg_temp_new_i64();
1955 tcg_gen_shl_i64(t0, arg1, arg2);
1956 tcg_gen_subfi_i64(t1, 64, arg2);
1957 tcg_gen_shr_i64(t1, arg1, t1);
1958 tcg_gen_or_i64(ret, t0, t1);
1959 tcg_temp_free_i64(t0);
1960 tcg_temp_free_i64(t1);
1961 }
1962 }
1963
1964 void tcg_gen_rotli_i64(TCGv_i64 ret, TCGv_i64 arg1, unsigned arg2)
1965 {
1966 tcg_debug_assert(arg2 < 64);
1967 /* some cases can be optimized here */
1968 if (arg2 == 0) {
1969 tcg_gen_mov_i64(ret, arg1);
1970 } else if (TCG_TARGET_HAS_rot_i64) {
1971 TCGv_i64 t0 = tcg_const_i64(arg2);
1972 tcg_gen_rotl_i64(ret, arg1, t0);
1973 tcg_temp_free_i64(t0);
1974 } else {
1975 TCGv_i64 t0, t1;
1976 t0 = tcg_temp_new_i64();
1977 t1 = tcg_temp_new_i64();
1978 tcg_gen_shli_i64(t0, arg1, arg2);
1979 tcg_gen_shri_i64(t1, arg1, 64 - arg2);
1980 tcg_gen_or_i64(ret, t0, t1);
1981 tcg_temp_free_i64(t0);
1982 tcg_temp_free_i64(t1);
1983 }
1984 }
1985
1986 void tcg_gen_rotr_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1987 {
1988 if (TCG_TARGET_HAS_rot_i64) {
1989 tcg_gen_op3_i64(INDEX_op_rotr_i64, ret, arg1, arg2);
1990 } else {
1991 TCGv_i64 t0, t1;
1992 t0 = tcg_temp_new_i64();
1993 t1 = tcg_temp_new_i64();
1994 tcg_gen_shr_i64(t0, arg1, arg2);
1995 tcg_gen_subfi_i64(t1, 64, arg2);
1996 tcg_gen_shl_i64(t1, arg1, t1);
1997 tcg_gen_or_i64(ret, t0, t1);
1998 tcg_temp_free_i64(t0);
1999 tcg_temp_free_i64(t1);
2000 }
2001 }
2002
2003 void tcg_gen_rotri_i64(TCGv_i64 ret, TCGv_i64 arg1, unsigned arg2)
2004 {
2005 tcg_debug_assert(arg2 < 64);
2006 /* some cases can be optimized here */
2007 if (arg2 == 0) {
2008 tcg_gen_mov_i64(ret, arg1);
2009 } else {
2010 tcg_gen_rotli_i64(ret, arg1, 64 - arg2);
2011 }
2012 }
2013
2014 void tcg_gen_deposit_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2,
2015 unsigned int ofs, unsigned int len)
2016 {
2017 uint64_t mask;
2018 TCGv_i64 t1;
2019
2020 tcg_debug_assert(ofs < 64);
2021 tcg_debug_assert(len > 0);
2022 tcg_debug_assert(len <= 64);
2023 tcg_debug_assert(ofs + len <= 64);
2024
2025 if (len == 64) {
2026 tcg_gen_mov_i64(ret, arg2);
2027 return;
2028 }
2029 if (TCG_TARGET_HAS_deposit_i64 && TCG_TARGET_deposit_i64_valid(ofs, len)) {
2030 tcg_gen_op5ii_i64(INDEX_op_deposit_i64, ret, arg1, arg2, ofs, len);
2031 return;
2032 }
2033
2034 if (TCG_TARGET_REG_BITS == 32) {
2035 if (ofs >= 32) {
2036 tcg_gen_deposit_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1),
2037 TCGV_LOW(arg2), ofs - 32, len);
2038 tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg1));
2039 return;
2040 }
2041 if (ofs + len <= 32) {
2042 tcg_gen_deposit_i32(TCGV_LOW(ret), TCGV_LOW(arg1),
2043 TCGV_LOW(arg2), ofs, len);
2044 tcg_gen_mov_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1));
2045 return;
2046 }
2047 }
2048
2049 t1 = tcg_temp_new_i64();
2050
2051 if (TCG_TARGET_HAS_extract2_i64) {
2052 if (ofs + len == 64) {
2053 tcg_gen_shli_i64(t1, arg1, len);
2054 tcg_gen_extract2_i64(ret, t1, arg2, len);
2055 goto done;
2056 }
2057 if (ofs == 0) {
2058 tcg_gen_extract2_i64(ret, arg1, arg2, len);
2059 tcg_gen_rotli_i64(ret, ret, len);
2060 goto done;
2061 }
2062 }
2063
2064 mask = (1ull << len) - 1;
2065 if (ofs + len < 64) {
2066 tcg_gen_andi_i64(t1, arg2, mask);
2067 tcg_gen_shli_i64(t1, t1, ofs);
2068 } else {
2069 tcg_gen_shli_i64(t1, arg2, ofs);
2070 }
2071 tcg_gen_andi_i64(ret, arg1, ~(mask << ofs));
2072 tcg_gen_or_i64(ret, ret, t1);
2073 done:
2074 tcg_temp_free_i64(t1);
2075 }
2076
2077 void tcg_gen_deposit_z_i64(TCGv_i64 ret, TCGv_i64 arg,
2078 unsigned int ofs, unsigned int len)
2079 {
2080 tcg_debug_assert(ofs < 64);
2081 tcg_debug_assert(len > 0);
2082 tcg_debug_assert(len <= 64);
2083 tcg_debug_assert(ofs + len <= 64);
2084
2085 if (ofs + len == 64) {
2086 tcg_gen_shli_i64(ret, arg, ofs);
2087 } else if (ofs == 0) {
2088 tcg_gen_andi_i64(ret, arg, (1ull << len) - 1);
2089 } else if (TCG_TARGET_HAS_deposit_i64
2090 && TCG_TARGET_deposit_i64_valid(ofs, len)) {
2091 TCGv_i64 zero = tcg_const_i64(0);
2092 tcg_gen_op5ii_i64(INDEX_op_deposit_i64, ret, zero, arg, ofs, len);
2093 tcg_temp_free_i64(zero);
2094 } else {
2095 if (TCG_TARGET_REG_BITS == 32) {
2096 if (ofs >= 32) {
2097 tcg_gen_deposit_z_i32(TCGV_HIGH(ret), TCGV_LOW(arg),
2098 ofs - 32, len);
2099 tcg_gen_movi_i32(TCGV_LOW(ret), 0);
2100 return;
2101 }
2102 if (ofs + len <= 32) {
2103 tcg_gen_deposit_z_i32(TCGV_LOW(ret), TCGV_LOW(arg), ofs, len);
2104 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
2105 return;
2106 }
2107 }
2108 /* To help two-operand hosts we prefer to zero-extend first,
2109 which allows ARG to stay live. */
2110 switch (len) {
2111 case 32:
2112 if (TCG_TARGET_HAS_ext32u_i64) {
2113 tcg_gen_ext32u_i64(ret, arg);
2114 tcg_gen_shli_i64(ret, ret, ofs);
2115 return;
2116 }
2117 break;
2118 case 16:
2119 if (TCG_TARGET_HAS_ext16u_i64) {
2120 tcg_gen_ext16u_i64(ret, arg);
2121 tcg_gen_shli_i64(ret, ret, ofs);
2122 return;
2123 }
2124 break;
2125 case 8:
2126 if (TCG_TARGET_HAS_ext8u_i64) {
2127 tcg_gen_ext8u_i64(ret, arg);
2128 tcg_gen_shli_i64(ret, ret, ofs);
2129 return;
2130 }
2131 break;
2132 }
2133 /* Otherwise prefer zero-extension over AND for code size. */
2134 switch (ofs + len) {
2135 case 32:
2136 if (TCG_TARGET_HAS_ext32u_i64) {
2137 tcg_gen_shli_i64(ret, arg, ofs);
2138 tcg_gen_ext32u_i64(ret, ret);
2139 return;
2140 }
2141 break;
2142 case 16:
2143 if (TCG_TARGET_HAS_ext16u_i64) {
2144 tcg_gen_shli_i64(ret, arg, ofs);
2145 tcg_gen_ext16u_i64(ret, ret);
2146 return;
2147 }
2148 break;
2149 case 8:
2150 if (TCG_TARGET_HAS_ext8u_i64) {
2151 tcg_gen_shli_i64(ret, arg, ofs);
2152 tcg_gen_ext8u_i64(ret, ret);
2153 return;
2154 }
2155 break;
2156 }
2157 tcg_gen_andi_i64(ret, arg, (1ull << len) - 1);
2158 tcg_gen_shli_i64(ret, ret, ofs);
2159 }
2160 }
2161
2162 void tcg_gen_extract_i64(TCGv_i64 ret, TCGv_i64 arg,
2163 unsigned int ofs, unsigned int len)
2164 {
2165 tcg_debug_assert(ofs < 64);
2166 tcg_debug_assert(len > 0);
2167 tcg_debug_assert(len <= 64);
2168 tcg_debug_assert(ofs + len <= 64);
2169
2170 /* Canonicalize certain special cases, even if extract is supported. */
2171 if (ofs + len == 64) {
2172 tcg_gen_shri_i64(ret, arg, 64 - len);
2173 return;
2174 }
2175 if (ofs == 0) {
2176 tcg_gen_andi_i64(ret, arg, (1ull << len) - 1);
2177 return;
2178 }
2179
2180 if (TCG_TARGET_REG_BITS == 32) {
2181 /* Look for a 32-bit extract within one of the two words. */
2182 if (ofs >= 32) {
2183 tcg_gen_extract_i32(TCGV_LOW(ret), TCGV_HIGH(arg), ofs - 32, len);
2184 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
2185 return;
2186 }
2187 if (ofs + len <= 32) {
2188 tcg_gen_extract_i32(TCGV_LOW(ret), TCGV_LOW(arg), ofs, len);
2189 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
2190 return;
2191 }
2192 /* The field is split across two words. One double-word
2193 shift is better than two double-word shifts. */
2194 goto do_shift_and;
2195 }
2196
2197 if (TCG_TARGET_HAS_extract_i64
2198 && TCG_TARGET_extract_i64_valid(ofs, len)) {
2199 tcg_gen_op4ii_i64(INDEX_op_extract_i64, ret, arg, ofs, len);
2200 return;
2201 }
2202
2203 /* Assume that zero-extension, if available, is cheaper than a shift. */
2204 switch (ofs + len) {
2205 case 32:
2206 if (TCG_TARGET_HAS_ext32u_i64) {
2207 tcg_gen_ext32u_i64(ret, arg);
2208 tcg_gen_shri_i64(ret, ret, ofs);
2209 return;
2210 }
2211 break;
2212 case 16:
2213 if (TCG_TARGET_HAS_ext16u_i64) {
2214 tcg_gen_ext16u_i64(ret, arg);
2215 tcg_gen_shri_i64(ret, ret, ofs);
2216 return;
2217 }
2218 break;
2219 case 8:
2220 if (TCG_TARGET_HAS_ext8u_i64) {
2221 tcg_gen_ext8u_i64(ret, arg);
2222 tcg_gen_shri_i64(ret, ret, ofs);
2223 return;
2224 }
2225 break;
2226 }
2227
2228 /* ??? Ideally we'd know what values are available for immediate AND.
2229 Assume that 8 bits are available, plus the special cases of 16 and 32,
2230 so that we get ext8u, ext16u, and ext32u. */
2231 switch (len) {
2232 case 1 ... 8: case 16: case 32:
2233 do_shift_and:
2234 tcg_gen_shri_i64(ret, arg, ofs);
2235 tcg_gen_andi_i64(ret, ret, (1ull << len) - 1);
2236 break;
2237 default:
2238 tcg_gen_shli_i64(ret, arg, 64 - len - ofs);
2239 tcg_gen_shri_i64(ret, ret, 64 - len);
2240 break;
2241 }
2242 }
2243
2244 void tcg_gen_sextract_i64(TCGv_i64 ret, TCGv_i64 arg,
2245 unsigned int ofs, unsigned int len)
2246 {
2247 tcg_debug_assert(ofs < 64);
2248 tcg_debug_assert(len > 0);
2249 tcg_debug_assert(len <= 64);
2250 tcg_debug_assert(ofs + len <= 64);
2251
2252 /* Canonicalize certain special cases, even if sextract is supported. */
2253 if (ofs + len == 64) {
2254 tcg_gen_sari_i64(ret, arg, 64 - len);
2255 return;
2256 }
2257 if (ofs == 0) {
2258 switch (len) {
2259 case 32:
2260 tcg_gen_ext32s_i64(ret, arg);
2261 return;
2262 case 16:
2263 tcg_gen_ext16s_i64(ret, arg);
2264 return;
2265 case 8:
2266 tcg_gen_ext8s_i64(ret, arg);
2267 return;
2268 }
2269 }
2270
2271 if (TCG_TARGET_REG_BITS == 32) {
2272 /* Look for a 32-bit extract within one of the two words. */
2273 if (ofs >= 32) {
2274 tcg_gen_sextract_i32(TCGV_LOW(ret), TCGV_HIGH(arg), ofs - 32, len);
2275 } else if (ofs + len <= 32) {
2276 tcg_gen_sextract_i32(TCGV_LOW(ret), TCGV_LOW(arg), ofs, len);
2277 } else if (ofs == 0) {
2278 tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg));
2279 tcg_gen_sextract_i32(TCGV_HIGH(ret), TCGV_HIGH(arg), 0, len - 32);
2280 return;
2281 } else if (len > 32) {
2282 TCGv_i32 t = tcg_temp_new_i32();
2283 /* Extract the bits for the high word normally. */
2284 tcg_gen_sextract_i32(t, TCGV_HIGH(arg), ofs + 32, len - 32);
2285 /* Shift the field down for the low part. */
2286 tcg_gen_shri_i64(ret, arg, ofs);
2287 /* Overwrite the shift into the high part. */
2288 tcg_gen_mov_i32(TCGV_HIGH(ret), t);
2289 tcg_temp_free_i32(t);
2290 return;
2291 } else {
2292 /* Shift the field down for the low part, such that the
2293 field sits at the MSB. */
2294 tcg_gen_shri_i64(ret, arg, ofs + len - 32);
2295 /* Shift the field down from the MSB, sign extending. */
2296 tcg_gen_sari_i32(TCGV_LOW(ret), TCGV_LOW(ret), 32 - len);
2297 }
2298 /* Sign-extend the field from 32 bits. */
2299 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
2300 return;
2301 }
2302
2303 if (TCG_TARGET_HAS_sextract_i64
2304 && TCG_TARGET_extract_i64_valid(ofs, len)) {
2305 tcg_gen_op4ii_i64(INDEX_op_sextract_i64, ret, arg, ofs, len);
2306 return;
2307 }
2308
2309 /* Assume that sign-extension, if available, is cheaper than a shift. */
2310 switch (ofs + len) {
2311 case 32:
2312 if (TCG_TARGET_HAS_ext32s_i64) {
2313 tcg_gen_ext32s_i64(ret, arg);
2314 tcg_gen_sari_i64(ret, ret, ofs);
2315 return;
2316 }
2317 break;
2318 case 16:
2319 if (TCG_TARGET_HAS_ext16s_i64) {
2320 tcg_gen_ext16s_i64(ret, arg);
2321 tcg_gen_sari_i64(ret, ret, ofs);
2322 return;
2323 }
2324 break;
2325 case 8:
2326 if (TCG_TARGET_HAS_ext8s_i64) {
2327 tcg_gen_ext8s_i64(ret, arg);
2328 tcg_gen_sari_i64(ret, ret, ofs);
2329 return;
2330 }
2331 break;
2332 }
2333 switch (len) {
2334 case 32:
2335 if (TCG_TARGET_HAS_ext32s_i64) {
2336 tcg_gen_shri_i64(ret, arg, ofs);
2337 tcg_gen_ext32s_i64(ret, ret);
2338 return;
2339 }
2340 break;
2341 case 16:
2342 if (TCG_TARGET_HAS_ext16s_i64) {
2343 tcg_gen_shri_i64(ret, arg, ofs);
2344 tcg_gen_ext16s_i64(ret, ret);
2345 return;
2346 }
2347 break;
2348 case 8:
2349 if (TCG_TARGET_HAS_ext8s_i64) {
2350 tcg_gen_shri_i64(ret, arg, ofs);
2351 tcg_gen_ext8s_i64(ret, ret);
2352 return;
2353 }
2354 break;
2355 }
2356 tcg_gen_shli_i64(ret, arg, 64 - len - ofs);
2357 tcg_gen_sari_i64(ret, ret, 64 - len);
2358 }
2359
2360 /*
2361 * Extract 64 bits from a 128-bit input, ah:al, starting from ofs.
2362 * Unlike tcg_gen_extract_i64 above, len is fixed at 64.
2363 */
2364 void tcg_gen_extract2_i64(TCGv_i64 ret, TCGv_i64 al, TCGv_i64 ah,
2365 unsigned int ofs)
2366 {
2367 tcg_debug_assert(ofs <= 64);
2368 if (ofs == 0) {
2369 tcg_gen_mov_i64(ret, al);
2370 } else if (ofs == 64) {
2371 tcg_gen_mov_i64(ret, ah);
2372 } else if (al == ah) {
2373 tcg_gen_rotri_i64(ret, al, ofs);
2374 } else if (TCG_TARGET_HAS_extract2_i64) {
2375 tcg_gen_op4i_i64(INDEX_op_extract2_i64, ret, al, ah, ofs);
2376 } else {
2377 TCGv_i64 t0 = tcg_temp_new_i64();
2378 tcg_gen_shri_i64(t0, al, ofs);
2379 tcg_gen_deposit_i64(ret, t0, ah, 64 - ofs, ofs);
2380 tcg_temp_free_i64(t0);
2381 }
2382 }
2383
2384 void tcg_gen_movcond_i64(TCGCond cond, TCGv_i64 ret, TCGv_i64 c1,
2385 TCGv_i64 c2, TCGv_i64 v1, TCGv_i64 v2)
2386 {
2387 if (cond == TCG_COND_ALWAYS) {
2388 tcg_gen_mov_i64(ret, v1);
2389 } else if (cond == TCG_COND_NEVER) {
2390 tcg_gen_mov_i64(ret, v2);
2391 } else if (TCG_TARGET_REG_BITS == 32) {
2392 TCGv_i32 t0 = tcg_temp_new_i32();
2393 TCGv_i32 t1 = tcg_temp_new_i32();
2394 tcg_gen_op6i_i32(INDEX_op_setcond2_i32, t0,
2395 TCGV_LOW(c1), TCGV_HIGH(c1),
2396 TCGV_LOW(c2), TCGV_HIGH(c2), cond);
2397
2398 if (TCG_TARGET_HAS_movcond_i32) {
2399 tcg_gen_movi_i32(t1, 0);
2400 tcg_gen_movcond_i32(TCG_COND_NE, TCGV_LOW(ret), t0, t1,
2401 TCGV_LOW(v1), TCGV_LOW(v2));
2402 tcg_gen_movcond_i32(TCG_COND_NE, TCGV_HIGH(ret), t0, t1,
2403 TCGV_HIGH(v1), TCGV_HIGH(v2));
2404 } else {
2405 tcg_gen_neg_i32(t0, t0);
2406
2407 tcg_gen_and_i32(t1, TCGV_LOW(v1), t0);
2408 tcg_gen_andc_i32(TCGV_LOW(ret), TCGV_LOW(v2), t0);
2409 tcg_gen_or_i32(TCGV_LOW(ret), TCGV_LOW(ret), t1);
2410
2411 tcg_gen_and_i32(t1, TCGV_HIGH(v1), t0);
2412 tcg_gen_andc_i32(TCGV_HIGH(ret), TCGV_HIGH(v2), t0);
2413 tcg_gen_or_i32(TCGV_HIGH(ret), TCGV_HIGH(ret), t1);
2414 }
2415 tcg_temp_free_i32(t0);
2416 tcg_temp_free_i32(t1);
2417 } else if (TCG_TARGET_HAS_movcond_i64) {
2418 tcg_gen_op6i_i64(INDEX_op_movcond_i64, ret, c1, c2, v1, v2, cond);
2419 } else {
2420 TCGv_i64 t0 = tcg_temp_new_i64();
2421 TCGv_i64 t1 = tcg_temp_new_i64();
2422 tcg_gen_setcond_i64(cond, t0, c1, c2);
2423 tcg_gen_neg_i64(t0, t0);
2424 tcg_gen_and_i64(t1, v1, t0);
2425 tcg_gen_andc_i64(ret, v2, t0);
2426 tcg_gen_or_i64(ret, ret, t1);
2427 tcg_temp_free_i64(t0);
2428 tcg_temp_free_i64(t1);
2429 }
2430 }
2431
2432 void tcg_gen_add2_i64(TCGv_i64 rl, TCGv_i64 rh, TCGv_i64 al,
2433 TCGv_i64 ah, TCGv_i64 bl, TCGv_i64 bh)
2434 {
2435 if (TCG_TARGET_HAS_add2_i64) {
2436 tcg_gen_op6_i64(INDEX_op_add2_i64, rl, rh, al, ah, bl, bh);
2437 } else {
2438 TCGv_i64 t0 = tcg_temp_new_i64();
2439 TCGv_i64 t1 = tcg_temp_new_i64();
2440 tcg_gen_add_i64(t0, al, bl);
2441 tcg_gen_setcond_i64(TCG_COND_LTU, t1, t0, al);
2442 tcg_gen_add_i64(rh, ah, bh);
2443 tcg_gen_add_i64(rh, rh, t1);
2444 tcg_gen_mov_i64(rl, t0);
2445 tcg_temp_free_i64(t0);
2446 tcg_temp_free_i64(t1);
2447 }
2448 }
2449
2450 void tcg_gen_sub2_i64(TCGv_i64 rl, TCGv_i64 rh, TCGv_i64 al,
2451 TCGv_i64 ah, TCGv_i64 bl, TCGv_i64 bh)
2452 {
2453 if (TCG_TARGET_HAS_sub2_i64) {
2454 tcg_gen_op6_i64(INDEX_op_sub2_i64, rl, rh, al, ah, bl, bh);
2455 } else {
2456 TCGv_i64 t0 = tcg_temp_new_i64();
2457 TCGv_i64 t1 = tcg_temp_new_i64();
2458 tcg_gen_sub_i64(t0, al, bl);
2459 tcg_gen_setcond_i64(TCG_COND_LTU, t1, al, bl);
2460 tcg_gen_sub_i64(rh, ah, bh);
2461 tcg_gen_sub_i64(rh, rh, t1);
2462 tcg_gen_mov_i64(rl, t0);
2463 tcg_temp_free_i64(t0);
2464 tcg_temp_free_i64(t1);
2465 }
2466 }
2467
2468 void tcg_gen_mulu2_i64(TCGv_i64 rl, TCGv_i64 rh, TCGv_i64 arg1, TCGv_i64 arg2)
2469 {
2470 if (TCG_TARGET_HAS_mulu2_i64) {
2471 tcg_gen_op4_i64(INDEX_op_mulu2_i64, rl, rh, arg1, arg2);
2472 } else if (TCG_TARGET_HAS_muluh_i64) {
2473 TCGv_i64 t = tcg_temp_new_i64();
2474 tcg_gen_op3_i64(INDEX_op_mul_i64, t, arg1, arg2);
2475 tcg_gen_op3_i64(INDEX_op_muluh_i64, rh, arg1, arg2);
2476 tcg_gen_mov_i64(rl, t);
2477 tcg_temp_free_i64(t);
2478 } else {
2479 TCGv_i64 t0 = tcg_temp_new_i64();
2480 tcg_gen_mul_i64(t0, arg1, arg2);
2481 gen_helper_muluh_i64(rh, arg1, arg2);
2482 tcg_gen_mov_i64(rl, t0);
2483 tcg_temp_free_i64(t0);
2484 }
2485 }
2486
2487 void tcg_gen_muls2_i64(TCGv_i64 rl, TCGv_i64 rh, TCGv_i64 arg1, TCGv_i64 arg2)
2488 {
2489 if (TCG_TARGET_HAS_muls2_i64) {
2490 tcg_gen_op4_i64(INDEX_op_muls2_i64, rl, rh, arg1, arg2);
2491 } else if (TCG_TARGET_HAS_mulsh_i64) {
2492 TCGv_i64 t = tcg_temp_new_i64();
2493 tcg_gen_op3_i64(INDEX_op_mul_i64, t, arg1, arg2);
2494 tcg_gen_op3_i64(INDEX_op_mulsh_i64, rh, arg1, arg2);
2495 tcg_gen_mov_i64(rl, t);
2496 tcg_temp_free_i64(t);
2497 } else if (TCG_TARGET_HAS_mulu2_i64 || TCG_TARGET_HAS_muluh_i64) {
2498 TCGv_i64 t0 = tcg_temp_new_i64();
2499 TCGv_i64 t1 = tcg_temp_new_i64();
2500 TCGv_i64 t2 = tcg_temp_new_i64();
2501 TCGv_i64 t3 = tcg_temp_new_i64();
2502 tcg_gen_mulu2_i64(t0, t1, arg1, arg2);
2503 /* Adjust for negative inputs. */
2504 tcg_gen_sari_i64(t2, arg1, 63);
2505 tcg_gen_sari_i64(t3, arg2, 63);
2506 tcg_gen_and_i64(t2, t2, arg2);
2507 tcg_gen_and_i64(t3, t3, arg1);
2508 tcg_gen_sub_i64(rh, t1, t2);
2509 tcg_gen_sub_i64(rh, rh, t3);
2510 tcg_gen_mov_i64(rl, t0);
2511 tcg_temp_free_i64(t0);
2512 tcg_temp_free_i64(t1);
2513 tcg_temp_free_i64(t2);
2514 tcg_temp_free_i64(t3);
2515 } else {
2516 TCGv_i64 t0 = tcg_temp_new_i64();
2517 tcg_gen_mul_i64(t0, arg1, arg2);
2518 gen_helper_mulsh_i64(rh, arg1, arg2);
2519 tcg_gen_mov_i64(rl, t0);
2520 tcg_temp_free_i64(t0);
2521 }
2522 }
2523
2524 void tcg_gen_mulsu2_i64(TCGv_i64 rl, TCGv_i64 rh, TCGv_i64 arg1, TCGv_i64 arg2)
2525 {
2526 TCGv_i64 t0 = tcg_temp_new_i64();
2527 TCGv_i64 t1 = tcg_temp_new_i64();
2528 TCGv_i64 t2 = tcg_temp_new_i64();
2529 tcg_gen_mulu2_i64(t0, t1, arg1, arg2);
2530 /* Adjust for negative input for the signed arg1. */
2531 tcg_gen_sari_i64(t2, arg1, 63);
2532 tcg_gen_and_i64(t2, t2, arg2);
2533 tcg_gen_sub_i64(rh, t1, t2);
2534 tcg_gen_mov_i64(rl, t0);
2535 tcg_temp_free_i64(t0);
2536 tcg_temp_free_i64(t1);
2537 tcg_temp_free_i64(t2);
2538 }
2539
2540 void tcg_gen_smin_i64(TCGv_i64 ret, TCGv_i64 a, TCGv_i64 b)
2541 {
2542 tcg_gen_movcond_i64(TCG_COND_LT, ret, a, b, a, b);
2543 }
2544
2545 void tcg_gen_umin_i64(TCGv_i64 ret, TCGv_i64 a, TCGv_i64 b)
2546 {
2547 tcg_gen_movcond_i64(TCG_COND_LTU, ret, a, b, a, b);
2548 }
2549
2550 void tcg_gen_smax_i64(TCGv_i64 ret, TCGv_i64 a, TCGv_i64 b)
2551 {
2552 tcg_gen_movcond_i64(TCG_COND_LT, ret, a, b, b, a);
2553 }
2554
2555 void tcg_gen_umax_i64(TCGv_i64 ret, TCGv_i64 a, TCGv_i64 b)
2556 {
2557 tcg_gen_movcond_i64(TCG_COND_LTU, ret, a, b, b, a);
2558 }
2559
2560 void tcg_gen_abs_i64(TCGv_i64 ret, TCGv_i64 a)
2561 {
2562 TCGv_i64 t = tcg_temp_new_i64();
2563
2564 tcg_gen_sari_i64(t, a, 63);
2565 tcg_gen_xor_i64(ret, a, t);
2566 tcg_gen_sub_i64(ret, ret, t);
2567 tcg_temp_free_i64(t);
2568 }
2569
2570 /* Size changing operations. */
2571
2572 void tcg_gen_extrl_i64_i32(TCGv_i32 ret, TCGv_i64 arg)
2573 {
2574 if (TCG_TARGET_REG_BITS == 32) {
2575 tcg_gen_mov_i32(ret, TCGV_LOW(arg));
2576 } else if (TCG_TARGET_HAS_extrl_i64_i32) {
2577 tcg_gen_op2(INDEX_op_extrl_i64_i32,
2578 tcgv_i32_arg(ret), tcgv_i64_arg(arg));
2579 } else {
2580 tcg_gen_mov_i32(ret, (TCGv_i32)arg);
2581 }
2582 }
2583
2584 void tcg_gen_extrh_i64_i32(TCGv_i32 ret, TCGv_i64 arg)
2585 {
2586 if (TCG_TARGET_REG_BITS == 32) {
2587 tcg_gen_mov_i32(ret, TCGV_HIGH(arg));
2588 } else if (TCG_TARGET_HAS_extrh_i64_i32) {
2589 tcg_gen_op2(INDEX_op_extrh_i64_i32,
2590 tcgv_i32_arg(ret), tcgv_i64_arg(arg));
2591 } else {
2592 TCGv_i64 t = tcg_temp_new_i64();
2593 tcg_gen_shri_i64(t, arg, 32);
2594 tcg_gen_mov_i32(ret, (TCGv_i32)t);
2595 tcg_temp_free_i64(t);
2596 }
2597 }
2598
2599 void tcg_gen_extu_i32_i64(TCGv_i64 ret, TCGv_i32 arg)
2600 {
2601 if (TCG_TARGET_REG_BITS == 32) {
2602 tcg_gen_mov_i32(TCGV_LOW(ret), arg);
2603 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
2604 } else {
2605 tcg_gen_op2(INDEX_op_extu_i32_i64,
2606 tcgv_i64_arg(ret), tcgv_i32_arg(arg));
2607 }
2608 }
2609
2610 void tcg_gen_ext_i32_i64(TCGv_i64 ret, TCGv_i32 arg)
2611 {
2612 if (TCG_TARGET_REG_BITS == 32) {
2613 tcg_gen_mov_i32(TCGV_LOW(ret), arg);
2614 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
2615 } else {
2616 tcg_gen_op2(INDEX_op_ext_i32_i64,
2617 tcgv_i64_arg(ret), tcgv_i32_arg(arg));
2618 }
2619 }
2620
2621 void tcg_gen_concat_i32_i64(TCGv_i64 dest, TCGv_i32 low, TCGv_i32 high)
2622 {
2623 TCGv_i64 tmp;
2624
2625 if (TCG_TARGET_REG_BITS == 32) {
2626 tcg_gen_mov_i32(TCGV_LOW(dest), low);
2627 tcg_gen_mov_i32(TCGV_HIGH(dest), high);
2628 return;
2629 }
2630
2631 tmp = tcg_temp_new_i64();
2632 /* These extensions are only needed for type correctness.
2633 We may be able to do better given target specific information. */
2634 tcg_gen_extu_i32_i64(tmp, high);
2635 tcg_gen_extu_i32_i64(dest, low);
2636 /* If deposit is available, use it. Otherwise use the extra
2637 knowledge that we have of the zero-extensions above. */
2638 if (TCG_TARGET_HAS_deposit_i64 && TCG_TARGET_deposit_i64_valid(32, 32)) {
2639 tcg_gen_deposit_i64(dest, dest, tmp, 32, 32);
2640 } else {
2641 tcg_gen_shli_i64(tmp, tmp, 32);
2642 tcg_gen_or_i64(dest, dest, tmp);
2643 }
2644 tcg_temp_free_i64(tmp);
2645 }
2646
2647 void tcg_gen_extr_i64_i32(TCGv_i32 lo, TCGv_i32 hi, TCGv_i64 arg)
2648 {
2649 if (TCG_TARGET_REG_BITS == 32) {
2650 tcg_gen_mov_i32(lo, TCGV_LOW(arg));
2651 tcg_gen_mov_i32(hi, TCGV_HIGH(arg));
2652 } else {
2653 tcg_gen_extrl_i64_i32(lo, arg);
2654 tcg_gen_extrh_i64_i32(hi, arg);
2655 }
2656 }
2657
2658 void tcg_gen_extr32_i64(TCGv_i64 lo, TCGv_i64 hi, TCGv_i64 arg)
2659 {
2660 tcg_gen_ext32u_i64(lo, arg);
2661 tcg_gen_shri_i64(hi, arg, 32);
2662 }
2663
2664 /* QEMU specific operations. */
2665
2666 void tcg_gen_exit_tb(TranslationBlock *tb, unsigned idx)
2667 {
2668 uintptr_t val = (uintptr_t)tb + idx;
2669
2670 if (tb == NULL) {
2671 tcg_debug_assert(idx == 0);
2672 } else if (idx <= TB_EXIT_IDXMAX) {
2673 #ifdef CONFIG_DEBUG_TCG
2674 /* This is an exit following a goto_tb. Verify that we have
2675 seen this numbered exit before, via tcg_gen_goto_tb. */
2676 tcg_debug_assert(tcg_ctx->goto_tb_issue_mask & (1 << idx));
2677 #endif
2678 /* When not chaining, exit without indicating a link. */
2679 if (qemu_loglevel_mask(CPU_LOG_TB_NOCHAIN)) {
2680 val = 0;
2681 }
2682 } else {
2683 /* This is an exit via the exitreq label. */
2684 tcg_debug_assert(idx == TB_EXIT_REQUESTED);
2685 }
2686
2687 tcg_gen_op1i(INDEX_op_exit_tb, val);
2688 }
2689
2690 void tcg_gen_goto_tb(unsigned idx)
2691 {
2692 /* We only support two chained exits. */
2693 tcg_debug_assert(idx <= TB_EXIT_IDXMAX);
2694 #ifdef CONFIG_DEBUG_TCG
2695 /* Verify that we havn't seen this numbered exit before. */
2696 tcg_debug_assert((tcg_ctx->goto_tb_issue_mask & (1 << idx)) == 0);
2697 tcg_ctx->goto_tb_issue_mask |= 1 << idx;
2698 #endif
2699 /* When not chaining, we simply fall through to the "fallback" exit. */
2700 if (!qemu_loglevel_mask(CPU_LOG_TB_NOCHAIN)) {
2701 tcg_gen_op1i(INDEX_op_goto_tb, idx);
2702 }
2703 }
2704
2705 void tcg_gen_lookup_and_goto_ptr(void)
2706 {
2707 if (TCG_TARGET_HAS_goto_ptr && !qemu_loglevel_mask(CPU_LOG_TB_NOCHAIN)) {
2708 TCGv_ptr ptr = tcg_temp_new_ptr();
2709 gen_helper_lookup_tb_ptr(ptr, cpu_env);
2710 tcg_gen_op1i(INDEX_op_goto_ptr, tcgv_ptr_arg(ptr));
2711 tcg_temp_free_ptr(ptr);
2712 } else {
2713 tcg_gen_exit_tb(NULL, 0);
2714 }
2715 }
2716
2717 static inline TCGMemOp tcg_canonicalize_memop(TCGMemOp op, bool is64, bool st)
2718 {
2719 /* Trigger the asserts within as early as possible. */
2720 (void)get_alignment_bits(op);
2721
2722 switch (op & MO_SIZE) {
2723 case MO_8:
2724 op &= ~MO_BSWAP;
2725 break;
2726 case MO_16:
2727 break;
2728 case MO_32:
2729 if (!is64) {
2730 op &= ~MO_SIGN;
2731 }
2732 break;
2733 case MO_64:
2734 if (!is64) {
2735 tcg_abort();
2736 }
2737 break;
2738 }
2739 if (st) {
2740 op &= ~MO_SIGN;
2741 }
2742 return op;
2743 }
2744
2745 static void gen_ldst_i32(TCGOpcode opc, TCGv_i32 val, TCGv addr,
2746 TCGMemOp memop, TCGArg idx)
2747 {
2748 TCGMemOpIdx oi = make_memop_idx(memop, idx);
2749 #if TARGET_LONG_BITS == 32
2750 tcg_gen_op3i_i32(opc, val, addr, oi);
2751 #else
2752 if (TCG_TARGET_REG_BITS == 32) {
2753 tcg_gen_op4i_i32(opc, val, TCGV_LOW(addr), TCGV_HIGH(addr), oi);
2754 } else {
2755 tcg_gen_op3(opc, tcgv_i32_arg(val), tcgv_i64_arg(addr), oi);
2756 }
2757 #endif
2758 }
2759
2760 static void gen_ldst_i64(TCGOpcode opc, TCGv_i64 val, TCGv addr,
2761 TCGMemOp memop, TCGArg idx)
2762 {
2763 TCGMemOpIdx oi = make_memop_idx(memop, idx);
2764 #if TARGET_LONG_BITS == 32
2765 if (TCG_TARGET_REG_BITS == 32) {
2766 tcg_gen_op4i_i32(opc, TCGV_LOW(val), TCGV_HIGH(val), addr, oi);
2767 } else {
2768 tcg_gen_op3(opc, tcgv_i64_arg(val), tcgv_i32_arg(addr), oi);
2769 }
2770 #else
2771 if (TCG_TARGET_REG_BITS == 32) {
2772 tcg_gen_op5i_i32(opc, TCGV_LOW(val), TCGV_HIGH(val),
2773 TCGV_LOW(addr), TCGV_HIGH(addr), oi);
2774 } else {
2775 tcg_gen_op3i_i64(opc, val, addr, oi);
2776 }
2777 #endif
2778 }
2779
2780 static void tcg_gen_req_mo(TCGBar type)
2781 {
2782 #ifdef TCG_GUEST_DEFAULT_MO
2783 type &= TCG_GUEST_DEFAULT_MO;
2784 #endif
2785 type &= ~TCG_TARGET_DEFAULT_MO;
2786 if (type) {
2787 tcg_gen_mb(type | TCG_BAR_SC);
2788 }
2789 }
2790
2791 void tcg_gen_qemu_ld_i32(TCGv_i32 val, TCGv addr, TCGArg idx, TCGMemOp memop)
2792 {
2793 TCGMemOp orig_memop;
2794
2795 tcg_gen_req_mo(TCG_MO_LD_LD | TCG_MO_ST_LD);
2796 memop = tcg_canonicalize_memop(memop, 0, 0);
2797 trace_guest_mem_before_tcg(tcg_ctx->cpu, cpu_env,
2798 addr, trace_mem_get_info(memop, 0));
2799
2800 orig_memop = memop;
2801 if (!TCG_TARGET_HAS_MEMORY_BSWAP && (memop & MO_BSWAP)) {
2802 memop &= ~MO_BSWAP;
2803 /* The bswap primitive requires zero-extended input. */
2804 if ((memop & MO_SSIZE) == MO_SW) {
2805 memop &= ~MO_SIGN;
2806 }
2807 }
2808
2809 gen_ldst_i32(INDEX_op_qemu_ld_i32, val, addr, memop, idx);
2810
2811 if ((orig_memop ^ memop) & MO_BSWAP) {
2812 switch (orig_memop & MO_SIZE) {
2813 case MO_16:
2814 tcg_gen_bswap16_i32(val, val);
2815 if (orig_memop & MO_SIGN) {
2816 tcg_gen_ext16s_i32(val, val);
2817 }
2818 break;
2819 case MO_32:
2820 tcg_gen_bswap32_i32(val, val);
2821 break;
2822 default:
2823 g_assert_not_reached();
2824 }
2825 }
2826 }
2827
2828 void tcg_gen_qemu_st_i32(TCGv_i32 val, TCGv addr, TCGArg idx, TCGMemOp memop)
2829 {
2830 TCGv_i32 swap = NULL;
2831
2832 tcg_gen_req_mo(TCG_MO_LD_ST | TCG_MO_ST_ST);
2833 memop = tcg_canonicalize_memop(memop, 0, 1);
2834 trace_guest_mem_before_tcg(tcg_ctx->cpu, cpu_env,
2835 addr, trace_mem_get_info(memop, 1));
2836
2837 if (!TCG_TARGET_HAS_MEMORY_BSWAP && (memop & MO_BSWAP)) {
2838 swap = tcg_temp_new_i32();
2839 switch (memop & MO_SIZE) {
2840 case MO_16:
2841 tcg_gen_ext16u_i32(swap, val);
2842 tcg_gen_bswap16_i32(swap, swap);
2843 break;
2844 case MO_32:
2845 tcg_gen_bswap32_i32(swap, val);
2846 break;
2847 default:
2848 g_assert_not_reached();
2849 }
2850 val = swap;
2851 memop &= ~MO_BSWAP;
2852 }
2853
2854 gen_ldst_i32(INDEX_op_qemu_st_i32, val, addr, memop, idx);
2855
2856 if (swap) {
2857 tcg_temp_free_i32(swap);
2858 }
2859 }
2860
2861 void tcg_gen_qemu_ld_i64(TCGv_i64 val, TCGv addr, TCGArg idx, TCGMemOp memop)
2862 {
2863 TCGMemOp orig_memop;
2864
2865 if (TCG_TARGET_REG_BITS == 32 && (memop & MO_SIZE) < MO_64) {
2866 tcg_gen_qemu_ld_i32(TCGV_LOW(val), addr, idx, memop);
2867 if (memop & MO_SIGN) {
2868 tcg_gen_sari_i32(TCGV_HIGH(val), TCGV_LOW(val), 31);
2869 } else {
2870 tcg_gen_movi_i32(TCGV_HIGH(val), 0);
2871 }
2872 return;
2873 }
2874
2875 tcg_gen_req_mo(TCG_MO_LD_LD | TCG_MO_ST_LD);
2876 memop = tcg_canonicalize_memop(memop, 1, 0);
2877 trace_guest_mem_before_tcg(tcg_ctx->cpu, cpu_env,
2878 addr, trace_mem_get_info(memop, 0));
2879
2880 orig_memop = memop;
2881 if (!TCG_TARGET_HAS_MEMORY_BSWAP && (memop & MO_BSWAP)) {
2882 memop &= ~MO_BSWAP;
2883 /* The bswap primitive requires zero-extended input. */
2884 if ((memop & MO_SIGN) && (memop & MO_SIZE) < MO_64) {
2885 memop &= ~MO_SIGN;
2886 }
2887 }
2888
2889 gen_ldst_i64(INDEX_op_qemu_ld_i64, val, addr, memop, idx);
2890
2891 if ((orig_memop ^ memop) & MO_BSWAP) {
2892 switch (orig_memop & MO_SIZE) {
2893 case MO_16:
2894 tcg_gen_bswap16_i64(val, val);
2895 if (orig_memop & MO_SIGN) {
2896 tcg_gen_ext16s_i64(val, val);
2897 }
2898 break;
2899 case MO_32:
2900 tcg_gen_bswap32_i64(val, val);
2901 if (orig_memop & MO_SIGN) {
2902 tcg_gen_ext32s_i64(val, val);
2903 }
2904 break;
2905 case MO_64:
2906 tcg_gen_bswap64_i64(val, val);
2907 break;
2908 default:
2909 g_assert_not_reached();
2910 }
2911 }
2912 }
2913
2914 void tcg_gen_qemu_st_i64(TCGv_i64 val, TCGv addr, TCGArg idx, TCGMemOp memop)
2915 {
2916 TCGv_i64 swap = NULL;
2917
2918 if (TCG_TARGET_REG_BITS == 32 && (memop & MO_SIZE) < MO_64) {
2919 tcg_gen_qemu_st_i32(TCGV_LOW(val), addr, idx, memop);
2920 return;
2921 }
2922
2923 tcg_gen_req_mo(TCG_MO_LD_ST | TCG_MO_ST_ST);
2924 memop = tcg_canonicalize_memop(memop, 1, 1);
2925 trace_guest_mem_before_tcg(tcg_ctx->cpu, cpu_env,
2926 addr, trace_mem_get_info(memop, 1));
2927
2928 if (!TCG_TARGET_HAS_MEMORY_BSWAP && (memop & MO_BSWAP)) {
2929 swap = tcg_temp_new_i64();
2930 switch (memop & MO_SIZE) {
2931 case MO_16:
2932 tcg_gen_ext16u_i64(swap, val);
2933 tcg_gen_bswap16_i64(swap, swap);
2934 break;
2935 case MO_32:
2936 tcg_gen_ext32u_i64(swap, val);
2937 tcg_gen_bswap32_i64(swap, swap);
2938 break;
2939 case MO_64:
2940 tcg_gen_bswap64_i64(swap, val);
2941 break;
2942 default:
2943 g_assert_not_reached();
2944 }
2945 val = swap;
2946 memop &= ~MO_BSWAP;
2947 }
2948
2949 gen_ldst_i64(INDEX_op_qemu_st_i64, val, addr, memop, idx);
2950
2951 if (swap) {
2952 tcg_temp_free_i64(swap);
2953 }
2954 }
2955
2956 static void tcg_gen_ext_i32(TCGv_i32 ret, TCGv_i32 val, TCGMemOp opc)
2957 {
2958 switch (opc & MO_SSIZE) {
2959 case MO_SB:
2960 tcg_gen_ext8s_i32(ret, val);
2961 break;
2962 case MO_UB:
2963 tcg_gen_ext8u_i32(ret, val);
2964 break;
2965 case MO_SW:
2966 tcg_gen_ext16s_i32(ret, val);
2967 break;
2968 case MO_UW:
2969 tcg_gen_ext16u_i32(ret, val);
2970 break;
2971 default:
2972 tcg_gen_mov_i32(ret, val);
2973 break;
2974 }
2975 }
2976
2977 static void tcg_gen_ext_i64(TCGv_i64 ret, TCGv_i64 val, TCGMemOp opc)
2978 {
2979 switch (opc & MO_SSIZE) {
2980 case MO_SB:
2981 tcg_gen_ext8s_i64(ret, val);
2982 break;
2983 case MO_UB:
2984 tcg_gen_ext8u_i64(ret, val);
2985 break;
2986 case MO_SW:
2987 tcg_gen_ext16s_i64(ret, val);
2988 break;
2989 case MO_UW:
2990 tcg_gen_ext16u_i64(ret, val);
2991 break;
2992 case MO_SL:
2993 tcg_gen_ext32s_i64(ret, val);
2994 break;
2995 case MO_UL:
2996 tcg_gen_ext32u_i64(ret, val);
2997 break;
2998 default:
2999 tcg_gen_mov_i64(ret, val);
3000 break;
3001 }
3002 }
3003
3004 #ifdef CONFIG_SOFTMMU
3005 typedef void (*gen_atomic_cx_i32)(TCGv_i32, TCGv_env, TCGv,
3006 TCGv_i32, TCGv_i32, TCGv_i32);
3007 typedef void (*gen_atomic_cx_i64)(TCGv_i64, TCGv_env, TCGv,
3008 TCGv_i64, TCGv_i64, TCGv_i32);
3009 typedef void (*gen_atomic_op_i32)(TCGv_i32, TCGv_env, TCGv,
3010 TCGv_i32, TCGv_i32);
3011 typedef void (*gen_atomic_op_i64)(TCGv_i64, TCGv_env, TCGv,
3012 TCGv_i64, TCGv_i32);
3013 #else
3014 typedef void (*gen_atomic_cx_i32)(TCGv_i32, TCGv_env, TCGv, TCGv_i32, TCGv_i32);
3015 typedef void (*gen_atomic_cx_i64)(TCGv_i64, TCGv_env, TCGv, TCGv_i64, TCGv_i64);
3016 typedef void (*gen_atomic_op_i32)(TCGv_i32, TCGv_env, TCGv, TCGv_i32);
3017 typedef void (*gen_atomic_op_i64)(TCGv_i64, TCGv_env, TCGv, TCGv_i64);
3018 #endif
3019
3020 #ifdef CONFIG_ATOMIC64
3021 # define WITH_ATOMIC64(X) X,
3022 #else
3023 # define WITH_ATOMIC64(X)
3024 #endif
3025
3026 static void * const table_cmpxchg[16] = {
3027 [MO_8] = gen_helper_atomic_cmpxchgb,
3028 [MO_16 | MO_LE] = gen_helper_atomic_cmpxchgw_le,
3029 [MO_16 | MO_BE] = gen_helper_atomic_cmpxchgw_be,
3030 [MO_32 | MO_LE] = gen_helper_atomic_cmpxchgl_le,
3031 [MO_32 | MO_BE] = gen_helper_atomic_cmpxchgl_be,
3032 WITH_ATOMIC64([MO_64 | MO_LE] = gen_helper_atomic_cmpxchgq_le)
3033 WITH_ATOMIC64([MO_64 | MO_BE] = gen_helper_atomic_cmpxchgq_be)
3034 };
3035
3036 void tcg_gen_atomic_cmpxchg_i32(TCGv_i32 retv, TCGv addr, TCGv_i32 cmpv,
3037 TCGv_i32 newv, TCGArg idx, TCGMemOp memop)
3038 {
3039 memop = tcg_canonicalize_memop(memop, 0, 0);
3040
3041 if (!(tcg_ctx->tb_cflags & CF_PARALLEL)) {
3042 TCGv_i32 t1 = tcg_temp_new_i32();
3043 TCGv_i32 t2 = tcg_temp_new_i32();
3044
3045 tcg_gen_ext_i32(t2, cmpv, memop & MO_SIZE);
3046
3047 tcg_gen_qemu_ld_i32(t1, addr, idx, memop & ~MO_SIGN);
3048 tcg_gen_movcond_i32(TCG_COND_EQ, t2, t1, t2, newv, t1);
3049 tcg_gen_qemu_st_i32(t2, addr, idx, memop);
3050 tcg_temp_free_i32(t2);
3051
3052 if (memop & MO_SIGN) {
3053 tcg_gen_ext_i32(retv, t1, memop);
3054 } else {
3055 tcg_gen_mov_i32(retv, t1);
3056 }
3057 tcg_temp_free_i32(t1);
3058 } else {
3059 gen_atomic_cx_i32 gen;
3060
3061 gen = table_cmpxchg[memop & (MO_SIZE | MO_BSWAP)];
3062 tcg_debug_assert(gen != NULL);
3063
3064 #ifdef CONFIG_SOFTMMU
3065 {
3066 TCGv_i32 oi = tcg_const_i32(make_memop_idx(memop & ~MO_SIGN, idx));
3067 gen(retv, cpu_env, addr, cmpv, newv, oi);
3068 tcg_temp_free_i32(oi);
3069 }
3070 #else
3071 gen(retv, cpu_env, addr, cmpv, newv);
3072 #endif
3073
3074 if (memop & MO_SIGN) {
3075 tcg_gen_ext_i32(retv, retv, memop);
3076 }
3077 }
3078 }
3079
3080 void tcg_gen_atomic_cmpxchg_i64(TCGv_i64 retv, TCGv addr, TCGv_i64 cmpv,
3081 TCGv_i64 newv, TCGArg idx, TCGMemOp memop)
3082 {
3083 memop = tcg_canonicalize_memop(memop, 1, 0);
3084
3085 if (!(tcg_ctx->tb_cflags & CF_PARALLEL)) {
3086 TCGv_i64 t1 = tcg_temp_new_i64();
3087 TCGv_i64 t2 = tcg_temp_new_i64();
3088
3089 tcg_gen_ext_i64(t2, cmpv, memop & MO_SIZE);
3090
3091 tcg_gen_qemu_ld_i64(t1, addr, idx, memop & ~MO_SIGN);
3092 tcg_gen_movcond_i64(TCG_COND_EQ, t2, t1, t2, newv, t1);
3093 tcg_gen_qemu_st_i64(t2, addr, idx, memop);
3094 tcg_temp_free_i64(t2);
3095
3096 if (memop & MO_SIGN) {
3097 tcg_gen_ext_i64(retv, t1, memop);
3098 } else {
3099 tcg_gen_mov_i64(retv, t1);
3100 }
3101 tcg_temp_free_i64(t1);
3102 } else if ((memop & MO_SIZE) == MO_64) {
3103 #ifdef CONFIG_ATOMIC64
3104 gen_atomic_cx_i64 gen;
3105
3106 gen = table_cmpxchg[memop & (MO_SIZE | MO_BSWAP)];
3107 tcg_debug_assert(gen != NULL);
3108
3109 #ifdef CONFIG_SOFTMMU
3110 {
3111 TCGv_i32 oi = tcg_const_i32(make_memop_idx(memop, idx));
3112 gen(retv, cpu_env, addr, cmpv, newv, oi);
3113 tcg_temp_free_i32(oi);
3114 }
3115 #else
3116 gen(retv, cpu_env, addr, cmpv, newv);
3117 #endif
3118 #else
3119 gen_helper_exit_atomic(cpu_env);
3120 /* Produce a result, so that we have a well-formed opcode stream
3121 with respect to uses of the result in the (dead) code following. */
3122 tcg_gen_movi_i64(retv, 0);
3123 #endif /* CONFIG_ATOMIC64 */
3124 } else {
3125 TCGv_i32 c32 = tcg_temp_new_i32();
3126 TCGv_i32 n32 = tcg_temp_new_i32();
3127 TCGv_i32 r32 = tcg_temp_new_i32();
3128
3129 tcg_gen_extrl_i64_i32(c32, cmpv);
3130 tcg_gen_extrl_i64_i32(n32, newv);
3131 tcg_gen_atomic_cmpxchg_i32(r32, addr, c32, n32, idx, memop & ~MO_SIGN);
3132 tcg_temp_free_i32(c32);
3133 tcg_temp_free_i32(n32);
3134
3135 tcg_gen_extu_i32_i64(retv, r32);
3136 tcg_temp_free_i32(r32);
3137
3138 if (memop & MO_SIGN) {
3139 tcg_gen_ext_i64(retv, retv, memop);
3140 }
3141 }
3142 }
3143
3144 static void do_nonatomic_op_i32(TCGv_i32 ret, TCGv addr, TCGv_i32 val,
3145 TCGArg idx, TCGMemOp memop, bool new_val,
3146 void (*gen)(TCGv_i32, TCGv_i32, TCGv_i32))
3147 {
3148 TCGv_i32 t1 = tcg_temp_new_i32();
3149 TCGv_i32 t2 = tcg_temp_new_i32();
3150
3151 memop = tcg_canonicalize_memop(memop, 0, 0);
3152
3153 tcg_gen_qemu_ld_i32(t1, addr, idx, memop & ~MO_SIGN);
3154 gen(t2, t1, val);
3155 tcg_gen_qemu_st_i32(t2, addr, idx, memop);
3156
3157 tcg_gen_ext_i32(ret, (new_val ? t2 : t1), memop);
3158 tcg_temp_free_i32(t1);
3159 tcg_temp_free_i32(t2);
3160 }
3161
3162 static void do_atomic_op_i32(TCGv_i32 ret, TCGv addr, TCGv_i32 val,
3163 TCGArg idx, TCGMemOp memop, void * const table[])
3164 {
3165 gen_atomic_op_i32 gen;
3166
3167 memop = tcg_canonicalize_memop(memop, 0, 0);
3168
3169 gen = table[memop & (MO_SIZE | MO_BSWAP)];
3170 tcg_debug_assert(gen != NULL);
3171
3172 #ifdef CONFIG_SOFTMMU
3173 {
3174 TCGv_i32 oi = tcg_const_i32(make_memop_idx(memop & ~MO_SIGN, idx));
3175 gen(ret, cpu_env, addr, val, oi);
3176 tcg_temp_free_i32(oi);
3177 }
3178 #else
3179 gen(ret, cpu_env, addr, val);
3180 #endif
3181
3182 if (memop & MO_SIGN) {
3183 tcg_gen_ext_i32(ret, ret, memop);
3184 }
3185 }
3186
3187 static void do_nonatomic_op_i64(TCGv_i64 ret, TCGv addr, TCGv_i64 val,
3188 TCGArg idx, TCGMemOp memop, bool new_val,
3189 void (*gen)(TCGv_i64, TCGv_i64, TCGv_i64))
3190 {
3191 TCGv_i64 t1 = tcg_temp_new_i64();
3192 TCGv_i64 t2 = tcg_temp_new_i64();
3193
3194 memop = tcg_canonicalize_memop(memop, 1, 0);
3195
3196 tcg_gen_qemu_ld_i64(t1, addr, idx, memop & ~MO_SIGN);
3197 gen(t2, t1, val);
3198 tcg_gen_qemu_st_i64(t2, addr, idx, memop);
3199
3200 tcg_gen_ext_i64(ret, (new_val ? t2 : t1), memop);
3201 tcg_temp_free_i64(t1);
3202 tcg_temp_free_i64(t2);
3203 }
3204
3205 static void do_atomic_op_i64(TCGv_i64 ret, TCGv addr, TCGv_i64 val,
3206 TCGArg idx, TCGMemOp memop, void * const table[])
3207 {
3208 memop = tcg_canonicalize_memop(memop, 1, 0);
3209
3210 if ((memop & MO_SIZE) == MO_64) {
3211 #ifdef CONFIG_ATOMIC64
3212 gen_atomic_op_i64 gen;
3213
3214 gen = table[memop & (MO_SIZE | MO_BSWAP)];
3215 tcg_debug_assert(gen != NULL);
3216
3217 #ifdef CONFIG_SOFTMMU
3218 {
3219 TCGv_i32 oi = tcg_const_i32(make_memop_idx(memop & ~MO_SIGN, idx));
3220 gen(ret, cpu_env, addr, val, oi);
3221 tcg_temp_free_i32(oi);
3222 }
3223 #else
3224 gen(ret, cpu_env, addr, val);
3225 #endif
3226 #else
3227 gen_helper_exit_atomic(cpu_env);
3228 /* Produce a result, so that we have a well-formed opcode stream
3229 with respect to uses of the result in the (dead) code following. */
3230 tcg_gen_movi_i64(ret, 0);
3231 #endif /* CONFIG_ATOMIC64 */
3232 } else {
3233 TCGv_i32 v32 = tcg_temp_new_i32();
3234 TCGv_i32 r32 = tcg_temp_new_i32();
3235
3236 tcg_gen_extrl_i64_i32(v32, val);
3237 do_atomic_op_i32(r32, addr, v32, idx, memop & ~MO_SIGN, table);
3238 tcg_temp_free_i32(v32);
3239
3240 tcg_gen_extu_i32_i64(ret, r32);
3241 tcg_temp_free_i32(r32);
3242
3243 if (memop & MO_SIGN) {
3244 tcg_gen_ext_i64(ret, ret, memop);
3245 }
3246 }
3247 }
3248
3249 #define GEN_ATOMIC_HELPER(NAME, OP, NEW) \
3250 static void * const table_##NAME[16] = { \
3251 [MO_8] = gen_helper_atomic_##NAME##b, \
3252 [MO_16 | MO_LE] = gen_helper_atomic_##NAME##w_le, \
3253 [MO_16 | MO_BE] = gen_helper_atomic_##NAME##w_be, \
3254 [MO_32 | MO_LE] = gen_helper_atomic_##NAME##l_le, \
3255 [MO_32 | MO_BE] = gen_helper_atomic_##NAME##l_be, \
3256 WITH_ATOMIC64([MO_64 | MO_LE] = gen_helper_atomic_##NAME##q_le) \
3257 WITH_ATOMIC64([MO_64 | MO_BE] = gen_helper_atomic_##NAME##q_be) \
3258 }; \
3259 void tcg_gen_atomic_##NAME##_i32 \
3260 (TCGv_i32 ret, TCGv addr, TCGv_i32 val, TCGArg idx, TCGMemOp memop) \
3261 { \
3262 if (tcg_ctx->tb_cflags & CF_PARALLEL) { \
3263 do_atomic_op_i32(ret, addr, val, idx, memop, table_##NAME); \
3264 } else { \
3265 do_nonatomic_op_i32(ret, addr, val, idx, memop, NEW, \
3266 tcg_gen_##OP##_i32); \
3267 } \
3268 } \
3269 void tcg_gen_atomic_##NAME##_i64 \
3270 (TCGv_i64 ret, TCGv addr, TCGv_i64 val, TCGArg idx, TCGMemOp memop) \
3271 { \
3272 if (tcg_ctx->tb_cflags & CF_PARALLEL) { \
3273 do_atomic_op_i64(ret, addr, val, idx, memop, table_##NAME); \
3274 } else { \
3275 do_nonatomic_op_i64(ret, addr, val, idx, memop, NEW, \
3276 tcg_gen_##OP##_i64); \
3277 } \
3278 }
3279
3280 GEN_ATOMIC_HELPER(fetch_add, add, 0)
3281 GEN_ATOMIC_HELPER(fetch_and, and, 0)
3282 GEN_ATOMIC_HELPER(fetch_or, or, 0)
3283 GEN_ATOMIC_HELPER(fetch_xor, xor, 0)
3284 GEN_ATOMIC_HELPER(fetch_smin, smin, 0)
3285 GEN_ATOMIC_HELPER(fetch_umin, umin, 0)
3286 GEN_ATOMIC_HELPER(fetch_smax, smax, 0)
3287 GEN_ATOMIC_HELPER(fetch_umax, umax, 0)
3288
3289 GEN_ATOMIC_HELPER(add_fetch, add, 1)
3290 GEN_ATOMIC_HELPER(and_fetch, and, 1)
3291 GEN_ATOMIC_HELPER(or_fetch, or, 1)
3292 GEN_ATOMIC_HELPER(xor_fetch, xor, 1)
3293 GEN_ATOMIC_HELPER(smin_fetch, smin, 1)
3294 GEN_ATOMIC_HELPER(umin_fetch, umin, 1)
3295 GEN_ATOMIC_HELPER(smax_fetch, smax, 1)
3296 GEN_ATOMIC_HELPER(umax_fetch, umax, 1)
3297
3298 static void tcg_gen_mov2_i32(TCGv_i32 r, TCGv_i32 a, TCGv_i32 b)
3299 {
3300 tcg_gen_mov_i32(r, b);
3301 }
3302
3303 static void tcg_gen_mov2_i64(TCGv_i64 r, TCGv_i64 a, TCGv_i64 b)
3304 {
3305 tcg_gen_mov_i64(r, b);
3306 }
3307
3308 GEN_ATOMIC_HELPER(xchg, mov2, 0)
3309
3310 #undef GEN_ATOMIC_HELPER
QEMU mirror