?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
block/block-copy: add block_copy_cancel
[qemu.git] / tcg / tcg-op.c
blob70475773f457417a854b5bc29f12046e89053188
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/tcg.h"
29 #include "tcg/tcg-op.h"
30 #include "tcg/tcg-mo.h"
31 #include "trace-tcg.h"
32 #include "trace/mem.h"
33 #include "exec/plugin-gen.h"
34
35 /* Reduce the number of ifdefs below. This assumes that all uses of
36 TCGV_HIGH and TCGV_LOW are properly protected by a conditional that
37 the compiler can eliminate. */
38 #if TCG_TARGET_REG_BITS == 64
39 extern TCGv_i32 TCGV_LOW_link_error(TCGv_i64);
40 extern TCGv_i32 TCGV_HIGH_link_error(TCGv_i64);
41 #define TCGV_LOW TCGV_LOW_link_error
42 #define TCGV_HIGH TCGV_HIGH_link_error
43 #endif
44
45 void tcg_gen_op1(TCGOpcode opc, TCGArg a1)
46 {
47 TCGOp *op = tcg_emit_op(opc);
48 op->args[0] = a1;
49 }
50
51 void tcg_gen_op2(TCGOpcode opc, TCGArg a1, TCGArg a2)
52 {
53 TCGOp *op = tcg_emit_op(opc);
54 op->args[0] = a1;
55 op->args[1] = a2;
56 }
57
58 void tcg_gen_op3(TCGOpcode opc, TCGArg a1, TCGArg a2, TCGArg a3)
59 {
60 TCGOp *op = tcg_emit_op(opc);
61 op->args[0] = a1;
62 op->args[1] = a2;
63 op->args[2] = a3;
64 }
65
66 void tcg_gen_op4(TCGOpcode opc, TCGArg a1, TCGArg a2, TCGArg a3, TCGArg a4)
67 {
68 TCGOp *op = tcg_emit_op(opc);
69 op->args[0] = a1;
70 op->args[1] = a2;
71 op->args[2] = a3;
72 op->args[3] = a4;
73 }
74
75 void tcg_gen_op5(TCGOpcode opc, TCGArg a1, TCGArg a2, TCGArg a3,
76 TCGArg a4, TCGArg a5)
77 {
78 TCGOp *op = tcg_emit_op(opc);
79 op->args[0] = a1;
80 op->args[1] = a2;
81 op->args[2] = a3;
82 op->args[3] = a4;
83 op->args[4] = a5;
84 }
85
86 void tcg_gen_op6(TCGOpcode opc, TCGArg a1, TCGArg a2, TCGArg a3,
87 TCGArg a4, TCGArg a5, TCGArg a6)
88 {
89 TCGOp *op = tcg_emit_op(opc);
90 op->args[0] = a1;
91 op->args[1] = a2;
92 op->args[2] = a3;
93 op->args[3] = a4;
94 op->args[4] = a5;
95 op->args[5] = a6;
96 }
97
98 void tcg_gen_mb(TCGBar mb_type)
99 {
100 if (tcg_ctx->tb_cflags & CF_PARALLEL) {
101 tcg_gen_op1(INDEX_op_mb, mb_type);
102 }
103 }
104
105 /* 32 bit ops */
106
107 void tcg_gen_movi_i32(TCGv_i32 ret, int32_t arg)
108 {
109 tcg_gen_mov_i32(ret, tcg_constant_i32(arg));
110 }
111
112 void tcg_gen_addi_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
113 {
114 /* some cases can be optimized here */
115 if (arg2 == 0) {
116 tcg_gen_mov_i32(ret, arg1);
117 } else {
118 tcg_gen_add_i32(ret, arg1, tcg_constant_i32(arg2));
119 }
120 }
121
122 void tcg_gen_subfi_i32(TCGv_i32 ret, int32_t arg1, TCGv_i32 arg2)
123 {
124 if (arg1 == 0 && TCG_TARGET_HAS_neg_i32) {
125 /* Don't recurse with tcg_gen_neg_i32. */
126 tcg_gen_op2_i32(INDEX_op_neg_i32, ret, arg2);
127 } else {
128 tcg_gen_sub_i32(ret, tcg_constant_i32(arg1), arg2);
129 }
130 }
131
132 void tcg_gen_subi_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
133 {
134 /* some cases can be optimized here */
135 if (arg2 == 0) {
136 tcg_gen_mov_i32(ret, arg1);
137 } else {
138 tcg_gen_sub_i32(ret, arg1, tcg_constant_i32(arg2));
139 }
140 }
141
142 void tcg_gen_andi_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
143 {
144 /* Some cases can be optimized here. */
145 switch (arg2) {
146 case 0:
147 tcg_gen_movi_i32(ret, 0);
148 return;
149 case -1:
150 tcg_gen_mov_i32(ret, arg1);
151 return;
152 case 0xff:
153 /* Don't recurse with tcg_gen_ext8u_i32. */
154 if (TCG_TARGET_HAS_ext8u_i32) {
155 tcg_gen_op2_i32(INDEX_op_ext8u_i32, ret, arg1);
156 return;
157 }
158 break;
159 case 0xffff:
160 if (TCG_TARGET_HAS_ext16u_i32) {
161 tcg_gen_op2_i32(INDEX_op_ext16u_i32, ret, arg1);
162 return;
163 }
164 break;
165 }
166
167 tcg_gen_and_i32(ret, arg1, tcg_constant_i32(arg2));
168 }
169
170 void tcg_gen_ori_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
171 {
172 /* Some cases can be optimized here. */
173 if (arg2 == -1) {
174 tcg_gen_movi_i32(ret, -1);
175 } else if (arg2 == 0) {
176 tcg_gen_mov_i32(ret, arg1);
177 } else {
178 tcg_gen_or_i32(ret, arg1, tcg_constant_i32(arg2));
179 }
180 }
181
182 void tcg_gen_xori_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
183 {
184 /* Some cases can be optimized here. */
185 if (arg2 == 0) {
186 tcg_gen_mov_i32(ret, arg1);
187 } else if (arg2 == -1 && TCG_TARGET_HAS_not_i32) {
188 /* Don't recurse with tcg_gen_not_i32. */
189 tcg_gen_op2_i32(INDEX_op_not_i32, ret, arg1);
190 } else {
191 tcg_gen_xor_i32(ret, arg1, tcg_constant_i32(arg2));
192 }
193 }
194
195 void tcg_gen_shli_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
196 {
197 tcg_debug_assert(arg2 >= 0 && arg2 < 32);
198 if (arg2 == 0) {
199 tcg_gen_mov_i32(ret, arg1);
200 } else {
201 tcg_gen_shl_i32(ret, arg1, tcg_constant_i32(arg2));
202 }
203 }
204
205 void tcg_gen_shri_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
206 {
207 tcg_debug_assert(arg2 >= 0 && arg2 < 32);
208 if (arg2 == 0) {
209 tcg_gen_mov_i32(ret, arg1);
210 } else {
211 tcg_gen_shr_i32(ret, arg1, tcg_constant_i32(arg2));
212 }
213 }
214
215 void tcg_gen_sari_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
216 {
217 tcg_debug_assert(arg2 >= 0 && arg2 < 32);
218 if (arg2 == 0) {
219 tcg_gen_mov_i32(ret, arg1);
220 } else {
221 tcg_gen_sar_i32(ret, arg1, tcg_constant_i32(arg2));
222 }
223 }
224
225 void tcg_gen_brcond_i32(TCGCond cond, TCGv_i32 arg1, TCGv_i32 arg2, TCGLabel *l)
226 {
227 if (cond == TCG_COND_ALWAYS) {
228 tcg_gen_br(l);
229 } else if (cond != TCG_COND_NEVER) {
230 l->refs++;
231 tcg_gen_op4ii_i32(INDEX_op_brcond_i32, arg1, arg2, cond, label_arg(l));
232 }
233 }
234
235 void tcg_gen_brcondi_i32(TCGCond cond, TCGv_i32 arg1, int32_t arg2, TCGLabel *l)
236 {
237 if (cond == TCG_COND_ALWAYS) {
238 tcg_gen_br(l);
239 } else if (cond != TCG_COND_NEVER) {
240 tcg_gen_brcond_i32(cond, arg1, tcg_constant_i32(arg2), l);
241 }
242 }
243
244 void tcg_gen_setcond_i32(TCGCond cond, TCGv_i32 ret,
245 TCGv_i32 arg1, TCGv_i32 arg2)
246 {
247 if (cond == TCG_COND_ALWAYS) {
248 tcg_gen_movi_i32(ret, 1);
249 } else if (cond == TCG_COND_NEVER) {
250 tcg_gen_movi_i32(ret, 0);
251 } else {
252 tcg_gen_op4i_i32(INDEX_op_setcond_i32, ret, arg1, arg2, cond);
253 }
254 }
255
256 void tcg_gen_setcondi_i32(TCGCond cond, TCGv_i32 ret,
257 TCGv_i32 arg1, int32_t arg2)
258 {
259 tcg_gen_setcond_i32(cond, ret, arg1, tcg_constant_i32(arg2));
260 }
261
262 void tcg_gen_muli_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
263 {
264 if (arg2 == 0) {
265 tcg_gen_movi_i32(ret, 0);
266 } else if (is_power_of_2(arg2)) {
267 tcg_gen_shli_i32(ret, arg1, ctz32(arg2));
268 } else {
269 tcg_gen_mul_i32(ret, arg1, tcg_constant_i32(arg2));
270 }
271 }
272
273 void tcg_gen_div_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
274 {
275 if (TCG_TARGET_HAS_div_i32) {
276 tcg_gen_op3_i32(INDEX_op_div_i32, ret, arg1, arg2);
277 } else if (TCG_TARGET_HAS_div2_i32) {
278 TCGv_i32 t0 = tcg_temp_new_i32();
279 tcg_gen_sari_i32(t0, arg1, 31);
280 tcg_gen_op5_i32(INDEX_op_div2_i32, ret, t0, arg1, t0, arg2);
281 tcg_temp_free_i32(t0);
282 } else {
283 gen_helper_div_i32(ret, arg1, arg2);
284 }
285 }
286
287 void tcg_gen_rem_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
288 {
289 if (TCG_TARGET_HAS_rem_i32) {
290 tcg_gen_op3_i32(INDEX_op_rem_i32, ret, arg1, arg2);
291 } else if (TCG_TARGET_HAS_div_i32) {
292 TCGv_i32 t0 = tcg_temp_new_i32();
293 tcg_gen_op3_i32(INDEX_op_div_i32, t0, arg1, arg2);
294 tcg_gen_mul_i32(t0, t0, arg2);
295 tcg_gen_sub_i32(ret, arg1, t0);
296 tcg_temp_free_i32(t0);
297 } else if (TCG_TARGET_HAS_div2_i32) {
298 TCGv_i32 t0 = tcg_temp_new_i32();
299 tcg_gen_sari_i32(t0, arg1, 31);
300 tcg_gen_op5_i32(INDEX_op_div2_i32, t0, ret, arg1, t0, arg2);
301 tcg_temp_free_i32(t0);
302 } else {
303 gen_helper_rem_i32(ret, arg1, arg2);
304 }
305 }
306
307 void tcg_gen_divu_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
308 {
309 if (TCG_TARGET_HAS_div_i32) {
310 tcg_gen_op3_i32(INDEX_op_divu_i32, ret, arg1, arg2);
311 } else if (TCG_TARGET_HAS_div2_i32) {
312 TCGv_i32 t0 = tcg_temp_new_i32();
313 tcg_gen_movi_i32(t0, 0);
314 tcg_gen_op5_i32(INDEX_op_divu2_i32, ret, t0, arg1, t0, arg2);
315 tcg_temp_free_i32(t0);
316 } else {
317 gen_helper_divu_i32(ret, arg1, arg2);
318 }
319 }
320
321 void tcg_gen_remu_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
322 {
323 if (TCG_TARGET_HAS_rem_i32) {
324 tcg_gen_op3_i32(INDEX_op_remu_i32, ret, arg1, arg2);
325 } else if (TCG_TARGET_HAS_div_i32) {
326 TCGv_i32 t0 = tcg_temp_new_i32();
327 tcg_gen_op3_i32(INDEX_op_divu_i32, t0, arg1, arg2);
328 tcg_gen_mul_i32(t0, t0, arg2);
329 tcg_gen_sub_i32(ret, arg1, t0);
330 tcg_temp_free_i32(t0);
331 } else if (TCG_TARGET_HAS_div2_i32) {
332 TCGv_i32 t0 = tcg_temp_new_i32();
333 tcg_gen_movi_i32(t0, 0);
334 tcg_gen_op5_i32(INDEX_op_divu2_i32, t0, ret, arg1, t0, arg2);
335 tcg_temp_free_i32(t0);
336 } else {
337 gen_helper_remu_i32(ret, arg1, arg2);
338 }
339 }
340
341 void tcg_gen_andc_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
342 {
343 if (TCG_TARGET_HAS_andc_i32) {
344 tcg_gen_op3_i32(INDEX_op_andc_i32, ret, arg1, arg2);
345 } else {
346 TCGv_i32 t0 = tcg_temp_new_i32();
347 tcg_gen_not_i32(t0, arg2);
348 tcg_gen_and_i32(ret, arg1, t0);
349 tcg_temp_free_i32(t0);
350 }
351 }
352
353 void tcg_gen_eqv_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
354 {
355 if (TCG_TARGET_HAS_eqv_i32) {
356 tcg_gen_op3_i32(INDEX_op_eqv_i32, ret, arg1, arg2);
357 } else {
358 tcg_gen_xor_i32(ret, arg1, arg2);
359 tcg_gen_not_i32(ret, ret);
360 }
361 }
362
363 void tcg_gen_nand_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
364 {
365 if (TCG_TARGET_HAS_nand_i32) {
366 tcg_gen_op3_i32(INDEX_op_nand_i32, ret, arg1, arg2);
367 } else {
368 tcg_gen_and_i32(ret, arg1, arg2);
369 tcg_gen_not_i32(ret, ret);
370 }
371 }
372
373 void tcg_gen_nor_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
374 {
375 if (TCG_TARGET_HAS_nor_i32) {
376 tcg_gen_op3_i32(INDEX_op_nor_i32, ret, arg1, arg2);
377 } else {
378 tcg_gen_or_i32(ret, arg1, arg2);
379 tcg_gen_not_i32(ret, ret);
380 }
381 }
382
383 void tcg_gen_orc_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
384 {
385 if (TCG_TARGET_HAS_orc_i32) {
386 tcg_gen_op3_i32(INDEX_op_orc_i32, ret, arg1, arg2);
387 } else {
388 TCGv_i32 t0 = tcg_temp_new_i32();
389 tcg_gen_not_i32(t0, arg2);
390 tcg_gen_or_i32(ret, arg1, t0);
391 tcg_temp_free_i32(t0);
392 }
393 }
394
395 void tcg_gen_clz_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
396 {
397 if (TCG_TARGET_HAS_clz_i32) {
398 tcg_gen_op3_i32(INDEX_op_clz_i32, ret, arg1, arg2);
399 } else if (TCG_TARGET_HAS_clz_i64) {
400 TCGv_i64 t1 = tcg_temp_new_i64();
401 TCGv_i64 t2 = tcg_temp_new_i64();
402 tcg_gen_extu_i32_i64(t1, arg1);
403 tcg_gen_extu_i32_i64(t2, arg2);
404 tcg_gen_addi_i64(t2, t2, 32);
405 tcg_gen_clz_i64(t1, t1, t2);
406 tcg_gen_extrl_i64_i32(ret, t1);
407 tcg_temp_free_i64(t1);
408 tcg_temp_free_i64(t2);
409 tcg_gen_subi_i32(ret, ret, 32);
410 } else {
411 gen_helper_clz_i32(ret, arg1, arg2);
412 }
413 }
414
415 void tcg_gen_clzi_i32(TCGv_i32 ret, TCGv_i32 arg1, uint32_t arg2)
416 {
417 tcg_gen_clz_i32(ret, arg1, tcg_constant_i32(arg2));
418 }
419
420 void tcg_gen_ctz_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
421 {
422 if (TCG_TARGET_HAS_ctz_i32) {
423 tcg_gen_op3_i32(INDEX_op_ctz_i32, ret, arg1, arg2);
424 } else if (TCG_TARGET_HAS_ctz_i64) {
425 TCGv_i64 t1 = tcg_temp_new_i64();
426 TCGv_i64 t2 = tcg_temp_new_i64();
427 tcg_gen_extu_i32_i64(t1, arg1);
428 tcg_gen_extu_i32_i64(t2, arg2);
429 tcg_gen_ctz_i64(t1, t1, t2);
430 tcg_gen_extrl_i64_i32(ret, t1);
431 tcg_temp_free_i64(t1);
432 tcg_temp_free_i64(t2);
433 } else if (TCG_TARGET_HAS_ctpop_i32
434 || TCG_TARGET_HAS_ctpop_i64
435 || TCG_TARGET_HAS_clz_i32
436 || TCG_TARGET_HAS_clz_i64) {
437 TCGv_i32 z, t = tcg_temp_new_i32();
438
439 if (TCG_TARGET_HAS_ctpop_i32 || TCG_TARGET_HAS_ctpop_i64) {
440 tcg_gen_subi_i32(t, arg1, 1);
441 tcg_gen_andc_i32(t, t, arg1);
442 tcg_gen_ctpop_i32(t, t);
443 } else {
444 /* Since all non-x86 hosts have clz(0) == 32, don't fight it. */
445 tcg_gen_neg_i32(t, arg1);
446 tcg_gen_and_i32(t, t, arg1);
447 tcg_gen_clzi_i32(t, t, 32);
448 tcg_gen_xori_i32(t, t, 31);
449 }
450 z = tcg_constant_i32(0);
451 tcg_gen_movcond_i32(TCG_COND_EQ, ret, arg1, z, arg2, t);
452 tcg_temp_free_i32(t);
453 } else {
454 gen_helper_ctz_i32(ret, arg1, arg2);
455 }
456 }
457
458 void tcg_gen_ctzi_i32(TCGv_i32 ret, TCGv_i32 arg1, uint32_t arg2)
459 {
460 if (!TCG_TARGET_HAS_ctz_i32 && TCG_TARGET_HAS_ctpop_i32 && arg2 == 32) {
461 /* This equivalence has the advantage of not requiring a fixup. */
462 TCGv_i32 t = tcg_temp_new_i32();
463 tcg_gen_subi_i32(t, arg1, 1);
464 tcg_gen_andc_i32(t, t, arg1);
465 tcg_gen_ctpop_i32(ret, t);
466 tcg_temp_free_i32(t);
467 } else {
468 tcg_gen_ctz_i32(ret, arg1, tcg_constant_i32(arg2));
469 }
470 }
471
472 void tcg_gen_clrsb_i32(TCGv_i32 ret, TCGv_i32 arg)
473 {
474 if (TCG_TARGET_HAS_clz_i32) {
475 TCGv_i32 t = tcg_temp_new_i32();
476 tcg_gen_sari_i32(t, arg, 31);
477 tcg_gen_xor_i32(t, t, arg);
478 tcg_gen_clzi_i32(t, t, 32);
479 tcg_gen_subi_i32(ret, t, 1);
480 tcg_temp_free_i32(t);
481 } else {
482 gen_helper_clrsb_i32(ret, arg);
483 }
484 }
485
486 void tcg_gen_ctpop_i32(TCGv_i32 ret, TCGv_i32 arg1)
487 {
488 if (TCG_TARGET_HAS_ctpop_i32) {
489 tcg_gen_op2_i32(INDEX_op_ctpop_i32, ret, arg1);
490 } else if (TCG_TARGET_HAS_ctpop_i64) {
491 TCGv_i64 t = tcg_temp_new_i64();
492 tcg_gen_extu_i32_i64(t, arg1);
493 tcg_gen_ctpop_i64(t, t);
494 tcg_gen_extrl_i64_i32(ret, t);
495 tcg_temp_free_i64(t);
496 } else {
497 gen_helper_ctpop_i32(ret, arg1);
498 }
499 }
500
501 void tcg_gen_rotl_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
502 {
503 if (TCG_TARGET_HAS_rot_i32) {
504 tcg_gen_op3_i32(INDEX_op_rotl_i32, ret, arg1, arg2);
505 } else {
506 TCGv_i32 t0, t1;
507
508 t0 = tcg_temp_new_i32();
509 t1 = tcg_temp_new_i32();
510 tcg_gen_shl_i32(t0, arg1, arg2);
511 tcg_gen_subfi_i32(t1, 32, arg2);
512 tcg_gen_shr_i32(t1, arg1, t1);
513 tcg_gen_or_i32(ret, t0, t1);
514 tcg_temp_free_i32(t0);
515 tcg_temp_free_i32(t1);
516 }
517 }
518
519 void tcg_gen_rotli_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
520 {
521 tcg_debug_assert(arg2 >= 0 && arg2 < 32);
522 /* some cases can be optimized here */
523 if (arg2 == 0) {
524 tcg_gen_mov_i32(ret, arg1);
525 } else if (TCG_TARGET_HAS_rot_i32) {
526 tcg_gen_rotl_i32(ret, arg1, tcg_constant_i32(arg2));
527 } else {
528 TCGv_i32 t0, t1;
529 t0 = tcg_temp_new_i32();
530 t1 = tcg_temp_new_i32();
531 tcg_gen_shli_i32(t0, arg1, arg2);
532 tcg_gen_shri_i32(t1, arg1, 32 - arg2);
533 tcg_gen_or_i32(ret, t0, t1);
534 tcg_temp_free_i32(t0);
535 tcg_temp_free_i32(t1);
536 }
537 }
538
539 void tcg_gen_rotr_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2)
540 {
541 if (TCG_TARGET_HAS_rot_i32) {
542 tcg_gen_op3_i32(INDEX_op_rotr_i32, ret, arg1, arg2);
543 } else {
544 TCGv_i32 t0, t1;
545
546 t0 = tcg_temp_new_i32();
547 t1 = tcg_temp_new_i32();
548 tcg_gen_shr_i32(t0, arg1, arg2);
549 tcg_gen_subfi_i32(t1, 32, arg2);
550 tcg_gen_shl_i32(t1, arg1, t1);
551 tcg_gen_or_i32(ret, t0, t1);
552 tcg_temp_free_i32(t0);
553 tcg_temp_free_i32(t1);
554 }
555 }
556
557 void tcg_gen_rotri_i32(TCGv_i32 ret, TCGv_i32 arg1, int32_t arg2)
558 {
559 tcg_debug_assert(arg2 >= 0 && arg2 < 32);
560 /* some cases can be optimized here */
561 if (arg2 == 0) {
562 tcg_gen_mov_i32(ret, arg1);
563 } else {
564 tcg_gen_rotli_i32(ret, arg1, 32 - arg2);
565 }
566 }
567
568 void tcg_gen_deposit_i32(TCGv_i32 ret, TCGv_i32 arg1, TCGv_i32 arg2,
569 unsigned int ofs, unsigned int len)
570 {
571 uint32_t mask;
572 TCGv_i32 t1;
573
574 tcg_debug_assert(ofs < 32);
575 tcg_debug_assert(len > 0);
576 tcg_debug_assert(len <= 32);
577 tcg_debug_assert(ofs + len <= 32);
578
579 if (len == 32) {
580 tcg_gen_mov_i32(ret, arg2);
581 return;
582 }
583 if (TCG_TARGET_HAS_deposit_i32 && TCG_TARGET_deposit_i32_valid(ofs, len)) {
584 tcg_gen_op5ii_i32(INDEX_op_deposit_i32, ret, arg1, arg2, ofs, len);
585 return;
586 }
587
588 t1 = tcg_temp_new_i32();
589
590 if (TCG_TARGET_HAS_extract2_i32) {
591 if (ofs + len == 32) {
592 tcg_gen_shli_i32(t1, arg1, len);
593 tcg_gen_extract2_i32(ret, t1, arg2, len);
594 goto done;
595 }
596 if (ofs == 0) {
597 tcg_gen_extract2_i32(ret, arg1, arg2, len);
598 tcg_gen_rotli_i32(ret, ret, len);
599 goto done;
600 }
601 }
602
603 mask = (1u << len) - 1;
604 if (ofs + len < 32) {
605 tcg_gen_andi_i32(t1, arg2, mask);
606 tcg_gen_shli_i32(t1, t1, ofs);
607 } else {
608 tcg_gen_shli_i32(t1, arg2, ofs);
609 }
610 tcg_gen_andi_i32(ret, arg1, ~(mask << ofs));
611 tcg_gen_or_i32(ret, ret, t1);
612 done:
613 tcg_temp_free_i32(t1);
614 }
615
616 void tcg_gen_deposit_z_i32(TCGv_i32 ret, TCGv_i32 arg,
617 unsigned int ofs, unsigned int len)
618 {
619 tcg_debug_assert(ofs < 32);
620 tcg_debug_assert(len > 0);
621 tcg_debug_assert(len <= 32);
622 tcg_debug_assert(ofs + len <= 32);
623
624 if (ofs + len == 32) {
625 tcg_gen_shli_i32(ret, arg, ofs);
626 } else if (ofs == 0) {
627 tcg_gen_andi_i32(ret, arg, (1u << len) - 1);
628 } else if (TCG_TARGET_HAS_deposit_i32
629 && TCG_TARGET_deposit_i32_valid(ofs, len)) {
630 TCGv_i32 zero = tcg_constant_i32(0);
631 tcg_gen_op5ii_i32(INDEX_op_deposit_i32, ret, zero, arg, ofs, len);
632 } else {
633 /* To help two-operand hosts we prefer to zero-extend first,
634 which allows ARG to stay live. */
635 switch (len) {
636 case 16:
637 if (TCG_TARGET_HAS_ext16u_i32) {
638 tcg_gen_ext16u_i32(ret, arg);
639 tcg_gen_shli_i32(ret, ret, ofs);
640 return;
641 }
642 break;
643 case 8:
644 if (TCG_TARGET_HAS_ext8u_i32) {
645 tcg_gen_ext8u_i32(ret, arg);
646 tcg_gen_shli_i32(ret, ret, ofs);
647 return;
648 }
649 break;
650 }
651 /* Otherwise prefer zero-extension over AND for code size. */
652 switch (ofs + len) {
653 case 16:
654 if (TCG_TARGET_HAS_ext16u_i32) {
655 tcg_gen_shli_i32(ret, arg, ofs);
656 tcg_gen_ext16u_i32(ret, ret);
657 return;
658 }
659 break;
660 case 8:
661 if (TCG_TARGET_HAS_ext8u_i32) {
662 tcg_gen_shli_i32(ret, arg, ofs);
663 tcg_gen_ext8u_i32(ret, ret);
664 return;
665 }
666 break;
667 }
668 tcg_gen_andi_i32(ret, arg, (1u << len) - 1);
669 tcg_gen_shli_i32(ret, ret, ofs);
670 }
671 }
672
673 void tcg_gen_extract_i32(TCGv_i32 ret, TCGv_i32 arg,
674 unsigned int ofs, unsigned int len)
675 {
676 tcg_debug_assert(ofs < 32);
677 tcg_debug_assert(len > 0);
678 tcg_debug_assert(len <= 32);
679 tcg_debug_assert(ofs + len <= 32);
680
681 /* Canonicalize certain special cases, even if extract is supported. */
682 if (ofs + len == 32) {
683 tcg_gen_shri_i32(ret, arg, 32 - len);
684 return;
685 }
686 if (ofs == 0) {
687 tcg_gen_andi_i32(ret, arg, (1u << len) - 1);
688 return;
689 }
690
691 if (TCG_TARGET_HAS_extract_i32
692 && TCG_TARGET_extract_i32_valid(ofs, len)) {
693 tcg_gen_op4ii_i32(INDEX_op_extract_i32, ret, arg, ofs, len);
694 return;
695 }
696
697 /* Assume that zero-extension, if available, is cheaper than a shift. */
698 switch (ofs + len) {
699 case 16:
700 if (TCG_TARGET_HAS_ext16u_i32) {
701 tcg_gen_ext16u_i32(ret, arg);
702 tcg_gen_shri_i32(ret, ret, ofs);
703 return;
704 }
705 break;
706 case 8:
707 if (TCG_TARGET_HAS_ext8u_i32) {
708 tcg_gen_ext8u_i32(ret, arg);
709 tcg_gen_shri_i32(ret, ret, ofs);
710 return;
711 }
712 break;
713 }
714
715 /* ??? Ideally we'd know what values are available for immediate AND.
716 Assume that 8 bits are available, plus the special case of 16,
717 so that we get ext8u, ext16u. */
718 switch (len) {
719 case 1 ... 8: case 16:
720 tcg_gen_shri_i32(ret, arg, ofs);
721 tcg_gen_andi_i32(ret, ret, (1u << len) - 1);
722 break;
723 default:
724 tcg_gen_shli_i32(ret, arg, 32 - len - ofs);
725 tcg_gen_shri_i32(ret, ret, 32 - len);
726 break;
727 }
728 }
729
730 void tcg_gen_sextract_i32(TCGv_i32 ret, TCGv_i32 arg,
731 unsigned int ofs, unsigned int len)
732 {
733 tcg_debug_assert(ofs < 32);
734 tcg_debug_assert(len > 0);
735 tcg_debug_assert(len <= 32);
736 tcg_debug_assert(ofs + len <= 32);
737
738 /* Canonicalize certain special cases, even if extract is supported. */
739 if (ofs + len == 32) {
740 tcg_gen_sari_i32(ret, arg, 32 - len);
741 return;
742 }
743 if (ofs == 0) {
744 switch (len) {
745 case 16:
746 tcg_gen_ext16s_i32(ret, arg);
747 return;
748 case 8:
749 tcg_gen_ext8s_i32(ret, arg);
750 return;
751 }
752 }
753
754 if (TCG_TARGET_HAS_sextract_i32
755 && TCG_TARGET_extract_i32_valid(ofs, len)) {
756 tcg_gen_op4ii_i32(INDEX_op_sextract_i32, ret, arg, ofs, len);
757 return;
758 }
759
760 /* Assume that sign-extension, if available, is cheaper than a shift. */
761 switch (ofs + len) {
762 case 16:
763 if (TCG_TARGET_HAS_ext16s_i32) {
764 tcg_gen_ext16s_i32(ret, arg);
765 tcg_gen_sari_i32(ret, ret, ofs);
766 return;
767 }
768 break;
769 case 8:
770 if (TCG_TARGET_HAS_ext8s_i32) {
771 tcg_gen_ext8s_i32(ret, arg);
772 tcg_gen_sari_i32(ret, ret, ofs);
773 return;
774 }
775 break;
776 }
777 switch (len) {
778 case 16:
779 if (TCG_TARGET_HAS_ext16s_i32) {
780 tcg_gen_shri_i32(ret, arg, ofs);
781 tcg_gen_ext16s_i32(ret, ret);
782 return;
783 }
784 break;
785 case 8:
786 if (TCG_TARGET_HAS_ext8s_i32) {
787 tcg_gen_shri_i32(ret, arg, ofs);
788 tcg_gen_ext8s_i32(ret, ret);
789 return;
790 }
791 break;
792 }
793
794 tcg_gen_shli_i32(ret, arg, 32 - len - ofs);
795 tcg_gen_sari_i32(ret, ret, 32 - len);
796 }
797
798 /*
799 * Extract 32-bits from a 64-bit input, ah:al, starting from ofs.
800 * Unlike tcg_gen_extract_i32 above, len is fixed at 32.
801 */
802 void tcg_gen_extract2_i32(TCGv_i32 ret, TCGv_i32 al, TCGv_i32 ah,
803 unsigned int ofs)
804 {
805 tcg_debug_assert(ofs <= 32);
806 if (ofs == 0) {
807 tcg_gen_mov_i32(ret, al);
808 } else if (ofs == 32) {
809 tcg_gen_mov_i32(ret, ah);
810 } else if (al == ah) {
811 tcg_gen_rotri_i32(ret, al, ofs);
812 } else if (TCG_TARGET_HAS_extract2_i32) {
813 tcg_gen_op4i_i32(INDEX_op_extract2_i32, ret, al, ah, ofs);
814 } else {
815 TCGv_i32 t0 = tcg_temp_new_i32();
816 tcg_gen_shri_i32(t0, al, ofs);
817 tcg_gen_deposit_i32(ret, t0, ah, 32 - ofs, ofs);
818 tcg_temp_free_i32(t0);
819 }
820 }
821
822 void tcg_gen_movcond_i32(TCGCond cond, TCGv_i32 ret, TCGv_i32 c1,
823 TCGv_i32 c2, TCGv_i32 v1, TCGv_i32 v2)
824 {
825 if (cond == TCG_COND_ALWAYS) {
826 tcg_gen_mov_i32(ret, v1);
827 } else if (cond == TCG_COND_NEVER) {
828 tcg_gen_mov_i32(ret, v2);
829 } else if (TCG_TARGET_HAS_movcond_i32) {
830 tcg_gen_op6i_i32(INDEX_op_movcond_i32, ret, c1, c2, v1, v2, cond);
831 } else {
832 TCGv_i32 t0 = tcg_temp_new_i32();
833 TCGv_i32 t1 = tcg_temp_new_i32();
834 tcg_gen_setcond_i32(cond, t0, c1, c2);
835 tcg_gen_neg_i32(t0, t0);
836 tcg_gen_and_i32(t1, v1, t0);
837 tcg_gen_andc_i32(ret, v2, t0);
838 tcg_gen_or_i32(ret, ret, t1);
839 tcg_temp_free_i32(t0);
840 tcg_temp_free_i32(t1);
841 }
842 }
843
844 void tcg_gen_add2_i32(TCGv_i32 rl, TCGv_i32 rh, TCGv_i32 al,
845 TCGv_i32 ah, TCGv_i32 bl, TCGv_i32 bh)
846 {
847 if (TCG_TARGET_HAS_add2_i32) {
848 tcg_gen_op6_i32(INDEX_op_add2_i32, rl, rh, al, ah, bl, bh);
849 } else {
850 TCGv_i64 t0 = tcg_temp_new_i64();
851 TCGv_i64 t1 = tcg_temp_new_i64();
852 tcg_gen_concat_i32_i64(t0, al, ah);
853 tcg_gen_concat_i32_i64(t1, bl, bh);
854 tcg_gen_add_i64(t0, t0, t1);
855 tcg_gen_extr_i64_i32(rl, rh, t0);
856 tcg_temp_free_i64(t0);
857 tcg_temp_free_i64(t1);
858 }
859 }
860
861 void tcg_gen_sub2_i32(TCGv_i32 rl, TCGv_i32 rh, TCGv_i32 al,
862 TCGv_i32 ah, TCGv_i32 bl, TCGv_i32 bh)
863 {
864 if (TCG_TARGET_HAS_sub2_i32) {
865 tcg_gen_op6_i32(INDEX_op_sub2_i32, rl, rh, al, ah, bl, bh);
866 } else {
867 TCGv_i64 t0 = tcg_temp_new_i64();
868 TCGv_i64 t1 = tcg_temp_new_i64();
869 tcg_gen_concat_i32_i64(t0, al, ah);
870 tcg_gen_concat_i32_i64(t1, bl, bh);
871 tcg_gen_sub_i64(t0, t0, t1);
872 tcg_gen_extr_i64_i32(rl, rh, t0);
873 tcg_temp_free_i64(t0);
874 tcg_temp_free_i64(t1);
875 }
876 }
877
878 void tcg_gen_mulu2_i32(TCGv_i32 rl, TCGv_i32 rh, TCGv_i32 arg1, TCGv_i32 arg2)
879 {
880 if (TCG_TARGET_HAS_mulu2_i32) {
881 tcg_gen_op4_i32(INDEX_op_mulu2_i32, rl, rh, arg1, arg2);
882 } else if (TCG_TARGET_HAS_muluh_i32) {
883 TCGv_i32 t = tcg_temp_new_i32();
884 tcg_gen_op3_i32(INDEX_op_mul_i32, t, arg1, arg2);
885 tcg_gen_op3_i32(INDEX_op_muluh_i32, rh, arg1, arg2);
886 tcg_gen_mov_i32(rl, t);
887 tcg_temp_free_i32(t);
888 } else {
889 TCGv_i64 t0 = tcg_temp_new_i64();
890 TCGv_i64 t1 = tcg_temp_new_i64();
891 tcg_gen_extu_i32_i64(t0, arg1);
892 tcg_gen_extu_i32_i64(t1, arg2);
893 tcg_gen_mul_i64(t0, t0, t1);
894 tcg_gen_extr_i64_i32(rl, rh, t0);
895 tcg_temp_free_i64(t0);
896 tcg_temp_free_i64(t1);
897 }
898 }
899
900 void tcg_gen_muls2_i32(TCGv_i32 rl, TCGv_i32 rh, TCGv_i32 arg1, TCGv_i32 arg2)
901 {
902 if (TCG_TARGET_HAS_muls2_i32) {
903 tcg_gen_op4_i32(INDEX_op_muls2_i32, rl, rh, arg1, arg2);
904 } else if (TCG_TARGET_HAS_mulsh_i32) {
905 TCGv_i32 t = tcg_temp_new_i32();
906 tcg_gen_op3_i32(INDEX_op_mul_i32, t, arg1, arg2);
907 tcg_gen_op3_i32(INDEX_op_mulsh_i32, rh, arg1, arg2);
908 tcg_gen_mov_i32(rl, t);
909 tcg_temp_free_i32(t);
910 } else if (TCG_TARGET_REG_BITS == 32) {
911 TCGv_i32 t0 = tcg_temp_new_i32();
912 TCGv_i32 t1 = tcg_temp_new_i32();
913 TCGv_i32 t2 = tcg_temp_new_i32();
914 TCGv_i32 t3 = tcg_temp_new_i32();
915 tcg_gen_mulu2_i32(t0, t1, arg1, arg2);
916 /* Adjust for negative inputs. */
917 tcg_gen_sari_i32(t2, arg1, 31);
918 tcg_gen_sari_i32(t3, arg2, 31);
919 tcg_gen_and_i32(t2, t2, arg2);
920 tcg_gen_and_i32(t3, t3, arg1);
921 tcg_gen_sub_i32(rh, t1, t2);
922 tcg_gen_sub_i32(rh, rh, t3);
923 tcg_gen_mov_i32(rl, t0);
924 tcg_temp_free_i32(t0);
925 tcg_temp_free_i32(t1);
926 tcg_temp_free_i32(t2);
927 tcg_temp_free_i32(t3);
928 } else {
929 TCGv_i64 t0 = tcg_temp_new_i64();
930 TCGv_i64 t1 = tcg_temp_new_i64();
931 tcg_gen_ext_i32_i64(t0, arg1);
932 tcg_gen_ext_i32_i64(t1, arg2);
933 tcg_gen_mul_i64(t0, t0, t1);
934 tcg_gen_extr_i64_i32(rl, rh, t0);
935 tcg_temp_free_i64(t0);
936 tcg_temp_free_i64(t1);
937 }
938 }
939
940 void tcg_gen_mulsu2_i32(TCGv_i32 rl, TCGv_i32 rh, TCGv_i32 arg1, TCGv_i32 arg2)
941 {
942 if (TCG_TARGET_REG_BITS == 32) {
943 TCGv_i32 t0 = tcg_temp_new_i32();
944 TCGv_i32 t1 = tcg_temp_new_i32();
945 TCGv_i32 t2 = tcg_temp_new_i32();
946 tcg_gen_mulu2_i32(t0, t1, arg1, arg2);
947 /* Adjust for negative input for the signed arg1. */
948 tcg_gen_sari_i32(t2, arg1, 31);
949 tcg_gen_and_i32(t2, t2, arg2);
950 tcg_gen_sub_i32(rh, t1, t2);
951 tcg_gen_mov_i32(rl, t0);
952 tcg_temp_free_i32(t0);
953 tcg_temp_free_i32(t1);
954 tcg_temp_free_i32(t2);
955 } else {
956 TCGv_i64 t0 = tcg_temp_new_i64();
957 TCGv_i64 t1 = tcg_temp_new_i64();
958 tcg_gen_ext_i32_i64(t0, arg1);
959 tcg_gen_extu_i32_i64(t1, arg2);
960 tcg_gen_mul_i64(t0, t0, t1);
961 tcg_gen_extr_i64_i32(rl, rh, t0);
962 tcg_temp_free_i64(t0);
963 tcg_temp_free_i64(t1);
964 }
965 }
966
967 void tcg_gen_ext8s_i32(TCGv_i32 ret, TCGv_i32 arg)
968 {
969 if (TCG_TARGET_HAS_ext8s_i32) {
970 tcg_gen_op2_i32(INDEX_op_ext8s_i32, ret, arg);
971 } else {
972 tcg_gen_shli_i32(ret, arg, 24);
973 tcg_gen_sari_i32(ret, ret, 24);
974 }
975 }
976
977 void tcg_gen_ext16s_i32(TCGv_i32 ret, TCGv_i32 arg)
978 {
979 if (TCG_TARGET_HAS_ext16s_i32) {
980 tcg_gen_op2_i32(INDEX_op_ext16s_i32, ret, arg);
981 } else {
982 tcg_gen_shli_i32(ret, arg, 16);
983 tcg_gen_sari_i32(ret, ret, 16);
984 }
985 }
986
987 void tcg_gen_ext8u_i32(TCGv_i32 ret, TCGv_i32 arg)
988 {
989 if (TCG_TARGET_HAS_ext8u_i32) {
990 tcg_gen_op2_i32(INDEX_op_ext8u_i32, ret, arg);
991 } else {
992 tcg_gen_andi_i32(ret, arg, 0xffu);
993 }
994 }
995
996 void tcg_gen_ext16u_i32(TCGv_i32 ret, TCGv_i32 arg)
997 {
998 if (TCG_TARGET_HAS_ext16u_i32) {
999 tcg_gen_op2_i32(INDEX_op_ext16u_i32, ret, arg);
1000 } else {
1001 tcg_gen_andi_i32(ret, arg, 0xffffu);
1002 }
1003 }
1004
1005 /* Note: we assume the two high bytes are set to zero */
1006 void tcg_gen_bswap16_i32(TCGv_i32 ret, TCGv_i32 arg)
1007 {
1008 if (TCG_TARGET_HAS_bswap16_i32) {
1009 tcg_gen_op2_i32(INDEX_op_bswap16_i32, ret, arg);
1010 } else {
1011 TCGv_i32 t0 = tcg_temp_new_i32();
1012
1013 tcg_gen_ext8u_i32(t0, arg);
1014 tcg_gen_shli_i32(t0, t0, 8);
1015 tcg_gen_shri_i32(ret, arg, 8);
1016 tcg_gen_or_i32(ret, ret, t0);
1017 tcg_temp_free_i32(t0);
1018 }
1019 }
1020
1021 void tcg_gen_bswap32_i32(TCGv_i32 ret, TCGv_i32 arg)
1022 {
1023 if (TCG_TARGET_HAS_bswap32_i32) {
1024 tcg_gen_op2_i32(INDEX_op_bswap32_i32, ret, arg);
1025 } else {
1026 TCGv_i32 t0 = tcg_temp_new_i32();
1027 TCGv_i32 t1 = tcg_temp_new_i32();
1028 TCGv_i32 t2 = tcg_constant_i32(0x00ff00ff);
1029
1030 /* arg = abcd */
1031 tcg_gen_shri_i32(t0, arg, 8); /* t0 = .abc */
1032 tcg_gen_and_i32(t1, arg, t2); /* t1 = .b.d */
1033 tcg_gen_and_i32(t0, t0, t2); /* t0 = .a.c */
1034 tcg_gen_shli_i32(t1, t1, 8); /* t1 = b.d. */
1035 tcg_gen_or_i32(ret, t0, t1); /* ret = badc */
1036
1037 tcg_gen_shri_i32(t0, ret, 16); /* t0 = ..ba */
1038 tcg_gen_shli_i32(t1, ret, 16); /* t1 = dc.. */
1039 tcg_gen_or_i32(ret, t0, t1); /* ret = dcba */
1040
1041 tcg_temp_free_i32(t0);
1042 tcg_temp_free_i32(t1);
1043 }
1044 }
1045
1046 void tcg_gen_smin_i32(TCGv_i32 ret, TCGv_i32 a, TCGv_i32 b)
1047 {
1048 tcg_gen_movcond_i32(TCG_COND_LT, ret, a, b, a, b);
1049 }
1050
1051 void tcg_gen_umin_i32(TCGv_i32 ret, TCGv_i32 a, TCGv_i32 b)
1052 {
1053 tcg_gen_movcond_i32(TCG_COND_LTU, ret, a, b, a, b);
1054 }
1055
1056 void tcg_gen_smax_i32(TCGv_i32 ret, TCGv_i32 a, TCGv_i32 b)
1057 {
1058 tcg_gen_movcond_i32(TCG_COND_LT, ret, a, b, b, a);
1059 }
1060
1061 void tcg_gen_umax_i32(TCGv_i32 ret, TCGv_i32 a, TCGv_i32 b)
1062 {
1063 tcg_gen_movcond_i32(TCG_COND_LTU, ret, a, b, b, a);
1064 }
1065
1066 void tcg_gen_abs_i32(TCGv_i32 ret, TCGv_i32 a)
1067 {
1068 TCGv_i32 t = tcg_temp_new_i32();
1069
1070 tcg_gen_sari_i32(t, a, 31);
1071 tcg_gen_xor_i32(ret, a, t);
1072 tcg_gen_sub_i32(ret, ret, t);
1073 tcg_temp_free_i32(t);
1074 }
1075
1076 /* 64-bit ops */
1077
1078 #if TCG_TARGET_REG_BITS == 32
1079 /* These are all inline for TCG_TARGET_REG_BITS == 64. */
1080
1081 void tcg_gen_discard_i64(TCGv_i64 arg)
1082 {
1083 tcg_gen_discard_i32(TCGV_LOW(arg));
1084 tcg_gen_discard_i32(TCGV_HIGH(arg));
1085 }
1086
1087 void tcg_gen_mov_i64(TCGv_i64 ret, TCGv_i64 arg)
1088 {
1089 TCGTemp *ts = tcgv_i64_temp(arg);
1090
1091 /* Canonicalize TCGv_i64 TEMP_CONST into TCGv_i32 TEMP_CONST. */
1092 if (ts->kind == TEMP_CONST) {
1093 tcg_gen_movi_i64(ret, ts->val);
1094 } else {
1095 tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1096 tcg_gen_mov_i32(TCGV_HIGH(ret), TCGV_HIGH(arg));
1097 }
1098 }
1099
1100 void tcg_gen_movi_i64(TCGv_i64 ret, int64_t arg)
1101 {
1102 tcg_gen_movi_i32(TCGV_LOW(ret), arg);
1103 tcg_gen_movi_i32(TCGV_HIGH(ret), arg >> 32);
1104 }
1105
1106 void tcg_gen_ld8u_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
1107 {
1108 tcg_gen_ld8u_i32(TCGV_LOW(ret), arg2, offset);
1109 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1110 }
1111
1112 void tcg_gen_ld8s_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
1113 {
1114 tcg_gen_ld8s_i32(TCGV_LOW(ret), arg2, offset);
1115 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
1116 }
1117
1118 void tcg_gen_ld16u_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
1119 {
1120 tcg_gen_ld16u_i32(TCGV_LOW(ret), arg2, offset);
1121 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1122 }
1123
1124 void tcg_gen_ld16s_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
1125 {
1126 tcg_gen_ld16s_i32(TCGV_LOW(ret), arg2, offset);
1127 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
1128 }
1129
1130 void tcg_gen_ld32u_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
1131 {
1132 tcg_gen_ld_i32(TCGV_LOW(ret), arg2, offset);
1133 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1134 }
1135
1136 void tcg_gen_ld32s_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
1137 {
1138 tcg_gen_ld_i32(TCGV_LOW(ret), arg2, offset);
1139 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
1140 }
1141
1142 void tcg_gen_ld_i64(TCGv_i64 ret, TCGv_ptr arg2, tcg_target_long offset)
1143 {
1144 /* Since arg2 and ret have different types,
1145 they cannot be the same temporary */
1146 #ifdef HOST_WORDS_BIGENDIAN
1147 tcg_gen_ld_i32(TCGV_HIGH(ret), arg2, offset);
1148 tcg_gen_ld_i32(TCGV_LOW(ret), arg2, offset + 4);
1149 #else
1150 tcg_gen_ld_i32(TCGV_LOW(ret), arg2, offset);
1151 tcg_gen_ld_i32(TCGV_HIGH(ret), arg2, offset + 4);
1152 #endif
1153 }
1154
1155 void tcg_gen_st_i64(TCGv_i64 arg1, TCGv_ptr arg2, tcg_target_long offset)
1156 {
1157 #ifdef HOST_WORDS_BIGENDIAN
1158 tcg_gen_st_i32(TCGV_HIGH(arg1), arg2, offset);
1159 tcg_gen_st_i32(TCGV_LOW(arg1), arg2, offset + 4);
1160 #else
1161 tcg_gen_st_i32(TCGV_LOW(arg1), arg2, offset);
1162 tcg_gen_st_i32(TCGV_HIGH(arg1), arg2, offset + 4);
1163 #endif
1164 }
1165
1166 void tcg_gen_and_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1167 {
1168 tcg_gen_and_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1169 tcg_gen_and_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1170 }
1171
1172 void tcg_gen_or_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1173 {
1174 tcg_gen_or_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1175 tcg_gen_or_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1176 }
1177
1178 void tcg_gen_xor_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1179 {
1180 tcg_gen_xor_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1181 tcg_gen_xor_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1182 }
1183
1184 void tcg_gen_shl_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1185 {
1186 gen_helper_shl_i64(ret, arg1, arg2);
1187 }
1188
1189 void tcg_gen_shr_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1190 {
1191 gen_helper_shr_i64(ret, arg1, arg2);
1192 }
1193
1194 void tcg_gen_sar_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1195 {
1196 gen_helper_sar_i64(ret, arg1, arg2);
1197 }
1198
1199 void tcg_gen_mul_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1200 {
1201 TCGv_i64 t0;
1202 TCGv_i32 t1;
1203
1204 t0 = tcg_temp_new_i64();
1205 t1 = tcg_temp_new_i32();
1206
1207 tcg_gen_mulu2_i32(TCGV_LOW(t0), TCGV_HIGH(t0),
1208 TCGV_LOW(arg1), TCGV_LOW(arg2));
1209
1210 tcg_gen_mul_i32(t1, TCGV_LOW(arg1), TCGV_HIGH(arg2));
1211 tcg_gen_add_i32(TCGV_HIGH(t0), TCGV_HIGH(t0), t1);
1212 tcg_gen_mul_i32(t1, TCGV_HIGH(arg1), TCGV_LOW(arg2));
1213 tcg_gen_add_i32(TCGV_HIGH(t0), TCGV_HIGH(t0), t1);
1214
1215 tcg_gen_mov_i64(ret, t0);
1216 tcg_temp_free_i64(t0);
1217 tcg_temp_free_i32(t1);
1218 }
1219
1220 #else
1221
1222 void tcg_gen_movi_i64(TCGv_i64 ret, int64_t arg)
1223 {
1224 tcg_gen_mov_i64(ret, tcg_constant_i64(arg));
1225 }
1226
1227 #endif /* TCG_TARGET_REG_SIZE == 32 */
1228
1229 void tcg_gen_addi_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1230 {
1231 /* some cases can be optimized here */
1232 if (arg2 == 0) {
1233 tcg_gen_mov_i64(ret, arg1);
1234 } else if (TCG_TARGET_REG_BITS == 64) {
1235 tcg_gen_add_i64(ret, arg1, tcg_constant_i64(arg2));
1236 } else {
1237 tcg_gen_add2_i32(TCGV_LOW(ret), TCGV_HIGH(ret),
1238 TCGV_LOW(arg1), TCGV_HIGH(arg1),
1239 tcg_constant_i32(arg2), tcg_constant_i32(arg2 >> 32));
1240 }
1241 }
1242
1243 void tcg_gen_subfi_i64(TCGv_i64 ret, int64_t arg1, TCGv_i64 arg2)
1244 {
1245 if (arg1 == 0 && TCG_TARGET_HAS_neg_i64) {
1246 /* Don't recurse with tcg_gen_neg_i64. */
1247 tcg_gen_op2_i64(INDEX_op_neg_i64, ret, arg2);
1248 } else if (TCG_TARGET_REG_BITS == 64) {
1249 tcg_gen_sub_i64(ret, tcg_constant_i64(arg1), arg2);
1250 } else {
1251 tcg_gen_sub2_i32(TCGV_LOW(ret), TCGV_HIGH(ret),
1252 tcg_constant_i32(arg1), tcg_constant_i32(arg1 >> 32),
1253 TCGV_LOW(arg2), TCGV_HIGH(arg2));
1254 }
1255 }
1256
1257 void tcg_gen_subi_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1258 {
1259 /* some cases can be optimized here */
1260 if (arg2 == 0) {
1261 tcg_gen_mov_i64(ret, arg1);
1262 } else if (TCG_TARGET_REG_BITS == 64) {
1263 tcg_gen_sub_i64(ret, arg1, tcg_constant_i64(arg2));
1264 } else {
1265 tcg_gen_sub2_i32(TCGV_LOW(ret), TCGV_HIGH(ret),
1266 TCGV_LOW(arg1), TCGV_HIGH(arg1),
1267 tcg_constant_i32(arg2), tcg_constant_i32(arg2 >> 32));
1268 }
1269 }
1270
1271 void tcg_gen_andi_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1272 {
1273 if (TCG_TARGET_REG_BITS == 32) {
1274 tcg_gen_andi_i32(TCGV_LOW(ret), TCGV_LOW(arg1), arg2);
1275 tcg_gen_andi_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), arg2 >> 32);
1276 return;
1277 }
1278
1279 /* Some cases can be optimized here. */
1280 switch (arg2) {
1281 case 0:
1282 tcg_gen_movi_i64(ret, 0);
1283 return;
1284 case -1:
1285 tcg_gen_mov_i64(ret, arg1);
1286 return;
1287 case 0xff:
1288 /* Don't recurse with tcg_gen_ext8u_i64. */
1289 if (TCG_TARGET_HAS_ext8u_i64) {
1290 tcg_gen_op2_i64(INDEX_op_ext8u_i64, ret, arg1);
1291 return;
1292 }
1293 break;
1294 case 0xffff:
1295 if (TCG_TARGET_HAS_ext16u_i64) {
1296 tcg_gen_op2_i64(INDEX_op_ext16u_i64, ret, arg1);
1297 return;
1298 }
1299 break;
1300 case 0xffffffffu:
1301 if (TCG_TARGET_HAS_ext32u_i64) {
1302 tcg_gen_op2_i64(INDEX_op_ext32u_i64, ret, arg1);
1303 return;
1304 }
1305 break;
1306 }
1307
1308 tcg_gen_and_i64(ret, arg1, tcg_constant_i64(arg2));
1309 }
1310
1311 void tcg_gen_ori_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1312 {
1313 if (TCG_TARGET_REG_BITS == 32) {
1314 tcg_gen_ori_i32(TCGV_LOW(ret), TCGV_LOW(arg1), arg2);
1315 tcg_gen_ori_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), arg2 >> 32);
1316 return;
1317 }
1318 /* Some cases can be optimized here. */
1319 if (arg2 == -1) {
1320 tcg_gen_movi_i64(ret, -1);
1321 } else if (arg2 == 0) {
1322 tcg_gen_mov_i64(ret, arg1);
1323 } else {
1324 tcg_gen_or_i64(ret, arg1, tcg_constant_i64(arg2));
1325 }
1326 }
1327
1328 void tcg_gen_xori_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1329 {
1330 if (TCG_TARGET_REG_BITS == 32) {
1331 tcg_gen_xori_i32(TCGV_LOW(ret), TCGV_LOW(arg1), arg2);
1332 tcg_gen_xori_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), arg2 >> 32);
1333 return;
1334 }
1335 /* Some cases can be optimized here. */
1336 if (arg2 == 0) {
1337 tcg_gen_mov_i64(ret, arg1);
1338 } else if (arg2 == -1 && TCG_TARGET_HAS_not_i64) {
1339 /* Don't recurse with tcg_gen_not_i64. */
1340 tcg_gen_op2_i64(INDEX_op_not_i64, ret, arg1);
1341 } else {
1342 tcg_gen_xor_i64(ret, arg1, tcg_constant_i64(arg2));
1343 }
1344 }
1345
1346 static inline void tcg_gen_shifti_i64(TCGv_i64 ret, TCGv_i64 arg1,
1347 unsigned c, bool right, bool arith)
1348 {
1349 tcg_debug_assert(c < 64);
1350 if (c == 0) {
1351 tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg1));
1352 tcg_gen_mov_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1));
1353 } else if (c >= 32) {
1354 c -= 32;
1355 if (right) {
1356 if (arith) {
1357 tcg_gen_sari_i32(TCGV_LOW(ret), TCGV_HIGH(arg1), c);
1358 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), 31);
1359 } else {
1360 tcg_gen_shri_i32(TCGV_LOW(ret), TCGV_HIGH(arg1), c);
1361 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1362 }
1363 } else {
1364 tcg_gen_shli_i32(TCGV_HIGH(ret), TCGV_LOW(arg1), c);
1365 tcg_gen_movi_i32(TCGV_LOW(ret), 0);
1366 }
1367 } else if (right) {
1368 if (TCG_TARGET_HAS_extract2_i32) {
1369 tcg_gen_extract2_i32(TCGV_LOW(ret),
1370 TCGV_LOW(arg1), TCGV_HIGH(arg1), c);
1371 } else {
1372 tcg_gen_shri_i32(TCGV_LOW(ret), TCGV_LOW(arg1), c);
1373 tcg_gen_deposit_i32(TCGV_LOW(ret), TCGV_LOW(ret),
1374 TCGV_HIGH(arg1), 32 - c, c);
1375 }
1376 if (arith) {
1377 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), c);
1378 } else {
1379 tcg_gen_shri_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), c);
1380 }
1381 } else {
1382 if (TCG_TARGET_HAS_extract2_i32) {
1383 tcg_gen_extract2_i32(TCGV_HIGH(ret),
1384 TCGV_LOW(arg1), TCGV_HIGH(arg1), 32 - c);
1385 } else {
1386 TCGv_i32 t0 = tcg_temp_new_i32();
1387 tcg_gen_shri_i32(t0, TCGV_LOW(arg1), 32 - c);
1388 tcg_gen_deposit_i32(TCGV_HIGH(ret), t0,
1389 TCGV_HIGH(arg1), c, 32 - c);
1390 tcg_temp_free_i32(t0);
1391 }
1392 tcg_gen_shli_i32(TCGV_LOW(ret), TCGV_LOW(arg1), c);
1393 }
1394 }
1395
1396 void tcg_gen_shli_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1397 {
1398 tcg_debug_assert(arg2 >= 0 && arg2 < 64);
1399 if (TCG_TARGET_REG_BITS == 32) {
1400 tcg_gen_shifti_i64(ret, arg1, arg2, 0, 0);
1401 } else if (arg2 == 0) {
1402 tcg_gen_mov_i64(ret, arg1);
1403 } else {
1404 tcg_gen_shl_i64(ret, arg1, tcg_constant_i64(arg2));
1405 }
1406 }
1407
1408 void tcg_gen_shri_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1409 {
1410 tcg_debug_assert(arg2 >= 0 && arg2 < 64);
1411 if (TCG_TARGET_REG_BITS == 32) {
1412 tcg_gen_shifti_i64(ret, arg1, arg2, 1, 0);
1413 } else if (arg2 == 0) {
1414 tcg_gen_mov_i64(ret, arg1);
1415 } else {
1416 tcg_gen_shr_i64(ret, arg1, tcg_constant_i64(arg2));
1417 }
1418 }
1419
1420 void tcg_gen_sari_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1421 {
1422 tcg_debug_assert(arg2 >= 0 && arg2 < 64);
1423 if (TCG_TARGET_REG_BITS == 32) {
1424 tcg_gen_shifti_i64(ret, arg1, arg2, 1, 1);
1425 } else if (arg2 == 0) {
1426 tcg_gen_mov_i64(ret, arg1);
1427 } else {
1428 tcg_gen_sar_i64(ret, arg1, tcg_constant_i64(arg2));
1429 }
1430 }
1431
1432 void tcg_gen_brcond_i64(TCGCond cond, TCGv_i64 arg1, TCGv_i64 arg2, TCGLabel *l)
1433 {
1434 if (cond == TCG_COND_ALWAYS) {
1435 tcg_gen_br(l);
1436 } else if (cond != TCG_COND_NEVER) {
1437 l->refs++;
1438 if (TCG_TARGET_REG_BITS == 32) {
1439 tcg_gen_op6ii_i32(INDEX_op_brcond2_i32, TCGV_LOW(arg1),
1440 TCGV_HIGH(arg1), TCGV_LOW(arg2),
1441 TCGV_HIGH(arg2), cond, label_arg(l));
1442 } else {
1443 tcg_gen_op4ii_i64(INDEX_op_brcond_i64, arg1, arg2, cond,
1444 label_arg(l));
1445 }
1446 }
1447 }
1448
1449 void tcg_gen_brcondi_i64(TCGCond cond, TCGv_i64 arg1, int64_t arg2, TCGLabel *l)
1450 {
1451 if (TCG_TARGET_REG_BITS == 64) {
1452 tcg_gen_brcond_i64(cond, arg1, tcg_constant_i64(arg2), l);
1453 } else if (cond == TCG_COND_ALWAYS) {
1454 tcg_gen_br(l);
1455 } else if (cond != TCG_COND_NEVER) {
1456 l->refs++;
1457 tcg_gen_op6ii_i32(INDEX_op_brcond2_i32,
1458 TCGV_LOW(arg1), TCGV_HIGH(arg1),
1459 tcg_constant_i32(arg2),
1460 tcg_constant_i32(arg2 >> 32),
1461 cond, label_arg(l));
1462 }
1463 }
1464
1465 void tcg_gen_setcond_i64(TCGCond cond, TCGv_i64 ret,
1466 TCGv_i64 arg1, TCGv_i64 arg2)
1467 {
1468 if (cond == TCG_COND_ALWAYS) {
1469 tcg_gen_movi_i64(ret, 1);
1470 } else if (cond == TCG_COND_NEVER) {
1471 tcg_gen_movi_i64(ret, 0);
1472 } else {
1473 if (TCG_TARGET_REG_BITS == 32) {
1474 tcg_gen_op6i_i32(INDEX_op_setcond2_i32, TCGV_LOW(ret),
1475 TCGV_LOW(arg1), TCGV_HIGH(arg1),
1476 TCGV_LOW(arg2), TCGV_HIGH(arg2), cond);
1477 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1478 } else {
1479 tcg_gen_op4i_i64(INDEX_op_setcond_i64, ret, arg1, arg2, cond);
1480 }
1481 }
1482 }
1483
1484 void tcg_gen_setcondi_i64(TCGCond cond, TCGv_i64 ret,
1485 TCGv_i64 arg1, int64_t arg2)
1486 {
1487 if (TCG_TARGET_REG_BITS == 64) {
1488 tcg_gen_setcond_i64(cond, ret, arg1, tcg_constant_i64(arg2));
1489 } else if (cond == TCG_COND_ALWAYS) {
1490 tcg_gen_movi_i64(ret, 1);
1491 } else if (cond == TCG_COND_NEVER) {
1492 tcg_gen_movi_i64(ret, 0);
1493 } else {
1494 tcg_gen_op6i_i32(INDEX_op_setcond2_i32, TCGV_LOW(ret),
1495 TCGV_LOW(arg1), TCGV_HIGH(arg1),
1496 tcg_constant_i32(arg2),
1497 tcg_constant_i32(arg2 >> 32), cond);
1498 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1499 }
1500 }
1501
1502 void tcg_gen_muli_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1503 {
1504 if (arg2 == 0) {
1505 tcg_gen_movi_i64(ret, 0);
1506 } else if (is_power_of_2(arg2)) {
1507 tcg_gen_shli_i64(ret, arg1, ctz64(arg2));
1508 } else {
1509 TCGv_i64 t0 = tcg_const_i64(arg2);
1510 tcg_gen_mul_i64(ret, arg1, t0);
1511 tcg_temp_free_i64(t0);
1512 }
1513 }
1514
1515 void tcg_gen_div_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1516 {
1517 if (TCG_TARGET_HAS_div_i64) {
1518 tcg_gen_op3_i64(INDEX_op_div_i64, ret, arg1, arg2);
1519 } else if (TCG_TARGET_HAS_div2_i64) {
1520 TCGv_i64 t0 = tcg_temp_new_i64();
1521 tcg_gen_sari_i64(t0, arg1, 63);
1522 tcg_gen_op5_i64(INDEX_op_div2_i64, ret, t0, arg1, t0, arg2);
1523 tcg_temp_free_i64(t0);
1524 } else {
1525 gen_helper_div_i64(ret, arg1, arg2);
1526 }
1527 }
1528
1529 void tcg_gen_rem_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1530 {
1531 if (TCG_TARGET_HAS_rem_i64) {
1532 tcg_gen_op3_i64(INDEX_op_rem_i64, ret, arg1, arg2);
1533 } else if (TCG_TARGET_HAS_div_i64) {
1534 TCGv_i64 t0 = tcg_temp_new_i64();
1535 tcg_gen_op3_i64(INDEX_op_div_i64, t0, arg1, arg2);
1536 tcg_gen_mul_i64(t0, t0, arg2);
1537 tcg_gen_sub_i64(ret, arg1, t0);
1538 tcg_temp_free_i64(t0);
1539 } else if (TCG_TARGET_HAS_div2_i64) {
1540 TCGv_i64 t0 = tcg_temp_new_i64();
1541 tcg_gen_sari_i64(t0, arg1, 63);
1542 tcg_gen_op5_i64(INDEX_op_div2_i64, t0, ret, arg1, t0, arg2);
1543 tcg_temp_free_i64(t0);
1544 } else {
1545 gen_helper_rem_i64(ret, arg1, arg2);
1546 }
1547 }
1548
1549 void tcg_gen_divu_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1550 {
1551 if (TCG_TARGET_HAS_div_i64) {
1552 tcg_gen_op3_i64(INDEX_op_divu_i64, ret, arg1, arg2);
1553 } else if (TCG_TARGET_HAS_div2_i64) {
1554 TCGv_i64 t0 = tcg_temp_new_i64();
1555 tcg_gen_movi_i64(t0, 0);
1556 tcg_gen_op5_i64(INDEX_op_divu2_i64, ret, t0, arg1, t0, arg2);
1557 tcg_temp_free_i64(t0);
1558 } else {
1559 gen_helper_divu_i64(ret, arg1, arg2);
1560 }
1561 }
1562
1563 void tcg_gen_remu_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1564 {
1565 if (TCG_TARGET_HAS_rem_i64) {
1566 tcg_gen_op3_i64(INDEX_op_remu_i64, ret, arg1, arg2);
1567 } else if (TCG_TARGET_HAS_div_i64) {
1568 TCGv_i64 t0 = tcg_temp_new_i64();
1569 tcg_gen_op3_i64(INDEX_op_divu_i64, t0, arg1, arg2);
1570 tcg_gen_mul_i64(t0, t0, arg2);
1571 tcg_gen_sub_i64(ret, arg1, t0);
1572 tcg_temp_free_i64(t0);
1573 } else if (TCG_TARGET_HAS_div2_i64) {
1574 TCGv_i64 t0 = tcg_temp_new_i64();
1575 tcg_gen_movi_i64(t0, 0);
1576 tcg_gen_op5_i64(INDEX_op_divu2_i64, t0, ret, arg1, t0, arg2);
1577 tcg_temp_free_i64(t0);
1578 } else {
1579 gen_helper_remu_i64(ret, arg1, arg2);
1580 }
1581 }
1582
1583 void tcg_gen_ext8s_i64(TCGv_i64 ret, TCGv_i64 arg)
1584 {
1585 if (TCG_TARGET_REG_BITS == 32) {
1586 tcg_gen_ext8s_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1587 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
1588 } else if (TCG_TARGET_HAS_ext8s_i64) {
1589 tcg_gen_op2_i64(INDEX_op_ext8s_i64, ret, arg);
1590 } else {
1591 tcg_gen_shli_i64(ret, arg, 56);
1592 tcg_gen_sari_i64(ret, ret, 56);
1593 }
1594 }
1595
1596 void tcg_gen_ext16s_i64(TCGv_i64 ret, TCGv_i64 arg)
1597 {
1598 if (TCG_TARGET_REG_BITS == 32) {
1599 tcg_gen_ext16s_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1600 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
1601 } else if (TCG_TARGET_HAS_ext16s_i64) {
1602 tcg_gen_op2_i64(INDEX_op_ext16s_i64, ret, arg);
1603 } else {
1604 tcg_gen_shli_i64(ret, arg, 48);
1605 tcg_gen_sari_i64(ret, ret, 48);
1606 }
1607 }
1608
1609 void tcg_gen_ext32s_i64(TCGv_i64 ret, TCGv_i64 arg)
1610 {
1611 if (TCG_TARGET_REG_BITS == 32) {
1612 tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1613 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
1614 } else if (TCG_TARGET_HAS_ext32s_i64) {
1615 tcg_gen_op2_i64(INDEX_op_ext32s_i64, ret, arg);
1616 } else {
1617 tcg_gen_shli_i64(ret, arg, 32);
1618 tcg_gen_sari_i64(ret, ret, 32);
1619 }
1620 }
1621
1622 void tcg_gen_ext8u_i64(TCGv_i64 ret, TCGv_i64 arg)
1623 {
1624 if (TCG_TARGET_REG_BITS == 32) {
1625 tcg_gen_ext8u_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1626 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1627 } else if (TCG_TARGET_HAS_ext8u_i64) {
1628 tcg_gen_op2_i64(INDEX_op_ext8u_i64, ret, arg);
1629 } else {
1630 tcg_gen_andi_i64(ret, arg, 0xffu);
1631 }
1632 }
1633
1634 void tcg_gen_ext16u_i64(TCGv_i64 ret, TCGv_i64 arg)
1635 {
1636 if (TCG_TARGET_REG_BITS == 32) {
1637 tcg_gen_ext16u_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1638 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1639 } else if (TCG_TARGET_HAS_ext16u_i64) {
1640 tcg_gen_op2_i64(INDEX_op_ext16u_i64, ret, arg);
1641 } else {
1642 tcg_gen_andi_i64(ret, arg, 0xffffu);
1643 }
1644 }
1645
1646 void tcg_gen_ext32u_i64(TCGv_i64 ret, TCGv_i64 arg)
1647 {
1648 if (TCG_TARGET_REG_BITS == 32) {
1649 tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1650 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1651 } else if (TCG_TARGET_HAS_ext32u_i64) {
1652 tcg_gen_op2_i64(INDEX_op_ext32u_i64, ret, arg);
1653 } else {
1654 tcg_gen_andi_i64(ret, arg, 0xffffffffu);
1655 }
1656 }
1657
1658 /* Note: we assume the six high bytes are set to zero */
1659 void tcg_gen_bswap16_i64(TCGv_i64 ret, TCGv_i64 arg)
1660 {
1661 if (TCG_TARGET_REG_BITS == 32) {
1662 tcg_gen_bswap16_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1663 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1664 } else if (TCG_TARGET_HAS_bswap16_i64) {
1665 tcg_gen_op2_i64(INDEX_op_bswap16_i64, ret, arg);
1666 } else {
1667 TCGv_i64 t0 = tcg_temp_new_i64();
1668
1669 tcg_gen_ext8u_i64(t0, arg);
1670 tcg_gen_shli_i64(t0, t0, 8);
1671 tcg_gen_shri_i64(ret, arg, 8);
1672 tcg_gen_or_i64(ret, ret, t0);
1673 tcg_temp_free_i64(t0);
1674 }
1675 }
1676
1677 /* Note: we assume the four high bytes are set to zero */
1678 void tcg_gen_bswap32_i64(TCGv_i64 ret, TCGv_i64 arg)
1679 {
1680 if (TCG_TARGET_REG_BITS == 32) {
1681 tcg_gen_bswap32_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1682 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1683 } else if (TCG_TARGET_HAS_bswap32_i64) {
1684 tcg_gen_op2_i64(INDEX_op_bswap32_i64, ret, arg);
1685 } else {
1686 TCGv_i64 t0 = tcg_temp_new_i64();
1687 TCGv_i64 t1 = tcg_temp_new_i64();
1688 TCGv_i64 t2 = tcg_constant_i64(0x00ff00ff);
1689
1690 /* arg = ....abcd */
1691 tcg_gen_shri_i64(t0, arg, 8); /* t0 = .....abc */
1692 tcg_gen_and_i64(t1, arg, t2); /* t1 = .....b.d */
1693 tcg_gen_and_i64(t0, t0, t2); /* t0 = .....a.c */
1694 tcg_gen_shli_i64(t1, t1, 8); /* t1 = ....b.d. */
1695 tcg_gen_or_i64(ret, t0, t1); /* ret = ....badc */
1696
1697 tcg_gen_shli_i64(t1, ret, 48); /* t1 = dc...... */
1698 tcg_gen_shri_i64(t0, ret, 16); /* t0 = ......ba */
1699 tcg_gen_shri_i64(t1, t1, 32); /* t1 = ....dc.. */
1700 tcg_gen_or_i64(ret, t0, t1); /* ret = ....dcba */
1701
1702 tcg_temp_free_i64(t0);
1703 tcg_temp_free_i64(t1);
1704 }
1705 }
1706
1707 void tcg_gen_bswap64_i64(TCGv_i64 ret, TCGv_i64 arg)
1708 {
1709 if (TCG_TARGET_REG_BITS == 32) {
1710 TCGv_i32 t0, t1;
1711 t0 = tcg_temp_new_i32();
1712 t1 = tcg_temp_new_i32();
1713
1714 tcg_gen_bswap32_i32(t0, TCGV_LOW(arg));
1715 tcg_gen_bswap32_i32(t1, TCGV_HIGH(arg));
1716 tcg_gen_mov_i32(TCGV_LOW(ret), t1);
1717 tcg_gen_mov_i32(TCGV_HIGH(ret), t0);
1718 tcg_temp_free_i32(t0);
1719 tcg_temp_free_i32(t1);
1720 } else if (TCG_TARGET_HAS_bswap64_i64) {
1721 tcg_gen_op2_i64(INDEX_op_bswap64_i64, ret, arg);
1722 } else {
1723 TCGv_i64 t0 = tcg_temp_new_i64();
1724 TCGv_i64 t1 = tcg_temp_new_i64();
1725 TCGv_i64 t2 = tcg_temp_new_i64();
1726
1727 /* arg = abcdefgh */
1728 tcg_gen_movi_i64(t2, 0x00ff00ff00ff00ffull);
1729 tcg_gen_shri_i64(t0, arg, 8); /* t0 = .abcdefg */
1730 tcg_gen_and_i64(t1, arg, t2); /* t1 = .b.d.f.h */
1731 tcg_gen_and_i64(t0, t0, t2); /* t0 = .a.c.e.g */
1732 tcg_gen_shli_i64(t1, t1, 8); /* t1 = b.d.f.h. */
1733 tcg_gen_or_i64(ret, t0, t1); /* ret = badcfehg */
1734
1735 tcg_gen_movi_i64(t2, 0x0000ffff0000ffffull);
1736 tcg_gen_shri_i64(t0, ret, 16); /* t0 = ..badcfe */
1737 tcg_gen_and_i64(t1, ret, t2); /* t1 = ..dc..hg */
1738 tcg_gen_and_i64(t0, t0, t2); /* t0 = ..ba..fe */
1739 tcg_gen_shli_i64(t1, t1, 16); /* t1 = dc..hg.. */
1740 tcg_gen_or_i64(ret, t0, t1); /* ret = dcbahgfe */
1741
1742 tcg_gen_shri_i64(t0, ret, 32); /* t0 = ....dcba */
1743 tcg_gen_shli_i64(t1, ret, 32); /* t1 = hgfe.... */
1744 tcg_gen_or_i64(ret, t0, t1); /* ret = hgfedcba */
1745
1746 tcg_temp_free_i64(t0);
1747 tcg_temp_free_i64(t1);
1748 tcg_temp_free_i64(t2);
1749 }
1750 }
1751
1752 void tcg_gen_not_i64(TCGv_i64 ret, TCGv_i64 arg)
1753 {
1754 if (TCG_TARGET_REG_BITS == 32) {
1755 tcg_gen_not_i32(TCGV_LOW(ret), TCGV_LOW(arg));
1756 tcg_gen_not_i32(TCGV_HIGH(ret), TCGV_HIGH(arg));
1757 } else if (TCG_TARGET_HAS_not_i64) {
1758 tcg_gen_op2_i64(INDEX_op_not_i64, ret, arg);
1759 } else {
1760 tcg_gen_xori_i64(ret, arg, -1);
1761 }
1762 }
1763
1764 void tcg_gen_andc_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1765 {
1766 if (TCG_TARGET_REG_BITS == 32) {
1767 tcg_gen_andc_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1768 tcg_gen_andc_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1769 } else if (TCG_TARGET_HAS_andc_i64) {
1770 tcg_gen_op3_i64(INDEX_op_andc_i64, ret, arg1, arg2);
1771 } else {
1772 TCGv_i64 t0 = tcg_temp_new_i64();
1773 tcg_gen_not_i64(t0, arg2);
1774 tcg_gen_and_i64(ret, arg1, t0);
1775 tcg_temp_free_i64(t0);
1776 }
1777 }
1778
1779 void tcg_gen_eqv_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1780 {
1781 if (TCG_TARGET_REG_BITS == 32) {
1782 tcg_gen_eqv_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1783 tcg_gen_eqv_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1784 } else if (TCG_TARGET_HAS_eqv_i64) {
1785 tcg_gen_op3_i64(INDEX_op_eqv_i64, ret, arg1, arg2);
1786 } else {
1787 tcg_gen_xor_i64(ret, arg1, arg2);
1788 tcg_gen_not_i64(ret, ret);
1789 }
1790 }
1791
1792 void tcg_gen_nand_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1793 {
1794 if (TCG_TARGET_REG_BITS == 32) {
1795 tcg_gen_nand_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1796 tcg_gen_nand_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1797 } else if (TCG_TARGET_HAS_nand_i64) {
1798 tcg_gen_op3_i64(INDEX_op_nand_i64, ret, arg1, arg2);
1799 } else {
1800 tcg_gen_and_i64(ret, arg1, arg2);
1801 tcg_gen_not_i64(ret, ret);
1802 }
1803 }
1804
1805 void tcg_gen_nor_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1806 {
1807 if (TCG_TARGET_REG_BITS == 32) {
1808 tcg_gen_nor_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1809 tcg_gen_nor_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1810 } else if (TCG_TARGET_HAS_nor_i64) {
1811 tcg_gen_op3_i64(INDEX_op_nor_i64, ret, arg1, arg2);
1812 } else {
1813 tcg_gen_or_i64(ret, arg1, arg2);
1814 tcg_gen_not_i64(ret, ret);
1815 }
1816 }
1817
1818 void tcg_gen_orc_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1819 {
1820 if (TCG_TARGET_REG_BITS == 32) {
1821 tcg_gen_orc_i32(TCGV_LOW(ret), TCGV_LOW(arg1), TCGV_LOW(arg2));
1822 tcg_gen_orc_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1), TCGV_HIGH(arg2));
1823 } else if (TCG_TARGET_HAS_orc_i64) {
1824 tcg_gen_op3_i64(INDEX_op_orc_i64, ret, arg1, arg2);
1825 } else {
1826 TCGv_i64 t0 = tcg_temp_new_i64();
1827 tcg_gen_not_i64(t0, arg2);
1828 tcg_gen_or_i64(ret, arg1, t0);
1829 tcg_temp_free_i64(t0);
1830 }
1831 }
1832
1833 void tcg_gen_clz_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1834 {
1835 if (TCG_TARGET_HAS_clz_i64) {
1836 tcg_gen_op3_i64(INDEX_op_clz_i64, ret, arg1, arg2);
1837 } else {
1838 gen_helper_clz_i64(ret, arg1, arg2);
1839 }
1840 }
1841
1842 void tcg_gen_clzi_i64(TCGv_i64 ret, TCGv_i64 arg1, uint64_t arg2)
1843 {
1844 if (TCG_TARGET_REG_BITS == 32
1845 && TCG_TARGET_HAS_clz_i32
1846 && arg2 <= 0xffffffffu) {
1847 TCGv_i32 t = tcg_temp_new_i32();
1848 tcg_gen_clzi_i32(t, TCGV_LOW(arg1), arg2 - 32);
1849 tcg_gen_addi_i32(t, t, 32);
1850 tcg_gen_clz_i32(TCGV_LOW(ret), TCGV_HIGH(arg1), t);
1851 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1852 tcg_temp_free_i32(t);
1853 } else {
1854 TCGv_i64 t0 = tcg_const_i64(arg2);
1855 tcg_gen_clz_i64(ret, arg1, t0);
1856 tcg_temp_free_i64(t0);
1857 }
1858 }
1859
1860 void tcg_gen_ctz_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1861 {
1862 if (TCG_TARGET_HAS_ctz_i64) {
1863 tcg_gen_op3_i64(INDEX_op_ctz_i64, ret, arg1, arg2);
1864 } else if (TCG_TARGET_HAS_ctpop_i64 || TCG_TARGET_HAS_clz_i64) {
1865 TCGv_i64 z, t = tcg_temp_new_i64();
1866
1867 if (TCG_TARGET_HAS_ctpop_i64) {
1868 tcg_gen_subi_i64(t, arg1, 1);
1869 tcg_gen_andc_i64(t, t, arg1);
1870 tcg_gen_ctpop_i64(t, t);
1871 } else {
1872 /* Since all non-x86 hosts have clz(0) == 64, don't fight it. */
1873 tcg_gen_neg_i64(t, arg1);
1874 tcg_gen_and_i64(t, t, arg1);
1875 tcg_gen_clzi_i64(t, t, 64);
1876 tcg_gen_xori_i64(t, t, 63);
1877 }
1878 z = tcg_constant_i64(0);
1879 tcg_gen_movcond_i64(TCG_COND_EQ, ret, arg1, z, arg2, t);
1880 tcg_temp_free_i64(t);
1881 tcg_temp_free_i64(z);
1882 } else {
1883 gen_helper_ctz_i64(ret, arg1, arg2);
1884 }
1885 }
1886
1887 void tcg_gen_ctzi_i64(TCGv_i64 ret, TCGv_i64 arg1, uint64_t arg2)
1888 {
1889 if (TCG_TARGET_REG_BITS == 32
1890 && TCG_TARGET_HAS_ctz_i32
1891 && arg2 <= 0xffffffffu) {
1892 TCGv_i32 t32 = tcg_temp_new_i32();
1893 tcg_gen_ctzi_i32(t32, TCGV_HIGH(arg1), arg2 - 32);
1894 tcg_gen_addi_i32(t32, t32, 32);
1895 tcg_gen_ctz_i32(TCGV_LOW(ret), TCGV_LOW(arg1), t32);
1896 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1897 tcg_temp_free_i32(t32);
1898 } else if (!TCG_TARGET_HAS_ctz_i64
1899 && TCG_TARGET_HAS_ctpop_i64
1900 && arg2 == 64) {
1901 /* This equivalence has the advantage of not requiring a fixup. */
1902 TCGv_i64 t = tcg_temp_new_i64();
1903 tcg_gen_subi_i64(t, arg1, 1);
1904 tcg_gen_andc_i64(t, t, arg1);
1905 tcg_gen_ctpop_i64(ret, t);
1906 tcg_temp_free_i64(t);
1907 } else {
1908 TCGv_i64 t0 = tcg_const_i64(arg2);
1909 tcg_gen_ctz_i64(ret, arg1, t0);
1910 tcg_temp_free_i64(t0);
1911 }
1912 }
1913
1914 void tcg_gen_clrsb_i64(TCGv_i64 ret, TCGv_i64 arg)
1915 {
1916 if (TCG_TARGET_HAS_clz_i64 || TCG_TARGET_HAS_clz_i32) {
1917 TCGv_i64 t = tcg_temp_new_i64();
1918 tcg_gen_sari_i64(t, arg, 63);
1919 tcg_gen_xor_i64(t, t, arg);
1920 tcg_gen_clzi_i64(t, t, 64);
1921 tcg_gen_subi_i64(ret, t, 1);
1922 tcg_temp_free_i64(t);
1923 } else {
1924 gen_helper_clrsb_i64(ret, arg);
1925 }
1926 }
1927
1928 void tcg_gen_ctpop_i64(TCGv_i64 ret, TCGv_i64 arg1)
1929 {
1930 if (TCG_TARGET_HAS_ctpop_i64) {
1931 tcg_gen_op2_i64(INDEX_op_ctpop_i64, ret, arg1);
1932 } else if (TCG_TARGET_REG_BITS == 32 && TCG_TARGET_HAS_ctpop_i32) {
1933 tcg_gen_ctpop_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1));
1934 tcg_gen_ctpop_i32(TCGV_LOW(ret), TCGV_LOW(arg1));
1935 tcg_gen_add_i32(TCGV_LOW(ret), TCGV_LOW(ret), TCGV_HIGH(ret));
1936 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
1937 } else {
1938 gen_helper_ctpop_i64(ret, arg1);
1939 }
1940 }
1941
1942 void tcg_gen_rotl_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1943 {
1944 if (TCG_TARGET_HAS_rot_i64) {
1945 tcg_gen_op3_i64(INDEX_op_rotl_i64, ret, arg1, arg2);
1946 } else {
1947 TCGv_i64 t0, t1;
1948 t0 = tcg_temp_new_i64();
1949 t1 = tcg_temp_new_i64();
1950 tcg_gen_shl_i64(t0, arg1, arg2);
1951 tcg_gen_subfi_i64(t1, 64, arg2);
1952 tcg_gen_shr_i64(t1, arg1, t1);
1953 tcg_gen_or_i64(ret, t0, t1);
1954 tcg_temp_free_i64(t0);
1955 tcg_temp_free_i64(t1);
1956 }
1957 }
1958
1959 void tcg_gen_rotli_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1960 {
1961 tcg_debug_assert(arg2 >= 0 && arg2 < 64);
1962 /* some cases can be optimized here */
1963 if (arg2 == 0) {
1964 tcg_gen_mov_i64(ret, arg1);
1965 } else if (TCG_TARGET_HAS_rot_i64) {
1966 tcg_gen_rotl_i64(ret, arg1, tcg_constant_i64(arg2));
1967 } else {
1968 TCGv_i64 t0, t1;
1969 t0 = tcg_temp_new_i64();
1970 t1 = tcg_temp_new_i64();
1971 tcg_gen_shli_i64(t0, arg1, arg2);
1972 tcg_gen_shri_i64(t1, arg1, 64 - arg2);
1973 tcg_gen_or_i64(ret, t0, t1);
1974 tcg_temp_free_i64(t0);
1975 tcg_temp_free_i64(t1);
1976 }
1977 }
1978
1979 void tcg_gen_rotr_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2)
1980 {
1981 if (TCG_TARGET_HAS_rot_i64) {
1982 tcg_gen_op3_i64(INDEX_op_rotr_i64, ret, arg1, arg2);
1983 } else {
1984 TCGv_i64 t0, t1;
1985 t0 = tcg_temp_new_i64();
1986 t1 = tcg_temp_new_i64();
1987 tcg_gen_shr_i64(t0, arg1, arg2);
1988 tcg_gen_subfi_i64(t1, 64, arg2);
1989 tcg_gen_shl_i64(t1, arg1, t1);
1990 tcg_gen_or_i64(ret, t0, t1);
1991 tcg_temp_free_i64(t0);
1992 tcg_temp_free_i64(t1);
1993 }
1994 }
1995
1996 void tcg_gen_rotri_i64(TCGv_i64 ret, TCGv_i64 arg1, int64_t arg2)
1997 {
1998 tcg_debug_assert(arg2 >= 0 && arg2 < 64);
1999 /* some cases can be optimized here */
2000 if (arg2 == 0) {
2001 tcg_gen_mov_i64(ret, arg1);
2002 } else {
2003 tcg_gen_rotli_i64(ret, arg1, 64 - arg2);
2004 }
2005 }
2006
2007 void tcg_gen_deposit_i64(TCGv_i64 ret, TCGv_i64 arg1, TCGv_i64 arg2,
2008 unsigned int ofs, unsigned int len)
2009 {
2010 uint64_t mask;
2011 TCGv_i64 t1;
2012
2013 tcg_debug_assert(ofs < 64);
2014 tcg_debug_assert(len > 0);
2015 tcg_debug_assert(len <= 64);
2016 tcg_debug_assert(ofs + len <= 64);
2017
2018 if (len == 64) {
2019 tcg_gen_mov_i64(ret, arg2);
2020 return;
2021 }
2022 if (TCG_TARGET_HAS_deposit_i64 && TCG_TARGET_deposit_i64_valid(ofs, len)) {
2023 tcg_gen_op5ii_i64(INDEX_op_deposit_i64, ret, arg1, arg2, ofs, len);
2024 return;
2025 }
2026
2027 if (TCG_TARGET_REG_BITS == 32) {
2028 if (ofs >= 32) {
2029 tcg_gen_deposit_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1),
2030 TCGV_LOW(arg2), ofs - 32, len);
2031 tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg1));
2032 return;
2033 }
2034 if (ofs + len <= 32) {
2035 tcg_gen_deposit_i32(TCGV_LOW(ret), TCGV_LOW(arg1),
2036 TCGV_LOW(arg2), ofs, len);
2037 tcg_gen_mov_i32(TCGV_HIGH(ret), TCGV_HIGH(arg1));
2038 return;
2039 }
2040 }
2041
2042 t1 = tcg_temp_new_i64();
2043
2044 if (TCG_TARGET_HAS_extract2_i64) {
2045 if (ofs + len == 64) {
2046 tcg_gen_shli_i64(t1, arg1, len);
2047 tcg_gen_extract2_i64(ret, t1, arg2, len);
2048 goto done;
2049 }
2050 if (ofs == 0) {
2051 tcg_gen_extract2_i64(ret, arg1, arg2, len);
2052 tcg_gen_rotli_i64(ret, ret, len);
2053 goto done;
2054 }
2055 }
2056
2057 mask = (1ull << len) - 1;
2058 if (ofs + len < 64) {
2059 tcg_gen_andi_i64(t1, arg2, mask);
2060 tcg_gen_shli_i64(t1, t1, ofs);
2061 } else {
2062 tcg_gen_shli_i64(t1, arg2, ofs);
2063 }
2064 tcg_gen_andi_i64(ret, arg1, ~(mask << ofs));
2065 tcg_gen_or_i64(ret, ret, t1);
2066 done:
2067 tcg_temp_free_i64(t1);
2068 }
2069
2070 void tcg_gen_deposit_z_i64(TCGv_i64 ret, TCGv_i64 arg,
2071 unsigned int ofs, unsigned int len)
2072 {
2073 tcg_debug_assert(ofs < 64);
2074 tcg_debug_assert(len > 0);
2075 tcg_debug_assert(len <= 64);
2076 tcg_debug_assert(ofs + len <= 64);
2077
2078 if (ofs + len == 64) {
2079 tcg_gen_shli_i64(ret, arg, ofs);
2080 } else if (ofs == 0) {
2081 tcg_gen_andi_i64(ret, arg, (1ull << len) - 1);
2082 } else if (TCG_TARGET_HAS_deposit_i64
2083 && TCG_TARGET_deposit_i64_valid(ofs, len)) {
2084 TCGv_i64 zero = tcg_constant_i64(0);
2085 tcg_gen_op5ii_i64(INDEX_op_deposit_i64, ret, zero, arg, ofs, len);
2086 } else {
2087 if (TCG_TARGET_REG_BITS == 32) {
2088 if (ofs >= 32) {
2089 tcg_gen_deposit_z_i32(TCGV_HIGH(ret), TCGV_LOW(arg),
2090 ofs - 32, len);
2091 tcg_gen_movi_i32(TCGV_LOW(ret), 0);
2092 return;
2093 }
2094 if (ofs + len <= 32) {
2095 tcg_gen_deposit_z_i32(TCGV_LOW(ret), TCGV_LOW(arg), ofs, len);
2096 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
2097 return;
2098 }
2099 }
2100 /* To help two-operand hosts we prefer to zero-extend first,
2101 which allows ARG to stay live. */
2102 switch (len) {
2103 case 32:
2104 if (TCG_TARGET_HAS_ext32u_i64) {
2105 tcg_gen_ext32u_i64(ret, arg);
2106 tcg_gen_shli_i64(ret, ret, ofs);
2107 return;
2108 }
2109 break;
2110 case 16:
2111 if (TCG_TARGET_HAS_ext16u_i64) {
2112 tcg_gen_ext16u_i64(ret, arg);
2113 tcg_gen_shli_i64(ret, ret, ofs);
2114 return;
2115 }
2116 break;
2117 case 8:
2118 if (TCG_TARGET_HAS_ext8u_i64) {
2119 tcg_gen_ext8u_i64(ret, arg);
2120 tcg_gen_shli_i64(ret, ret, ofs);
2121 return;
2122 }
2123 break;
2124 }
2125 /* Otherwise prefer zero-extension over AND for code size. */
2126 switch (ofs + len) {
2127 case 32:
2128 if (TCG_TARGET_HAS_ext32u_i64) {
2129 tcg_gen_shli_i64(ret, arg, ofs);
2130 tcg_gen_ext32u_i64(ret, ret);
2131 return;
2132 }
2133 break;
2134 case 16:
2135 if (TCG_TARGET_HAS_ext16u_i64) {
2136 tcg_gen_shli_i64(ret, arg, ofs);
2137 tcg_gen_ext16u_i64(ret, ret);
2138 return;
2139 }
2140 break;
2141 case 8:
2142 if (TCG_TARGET_HAS_ext8u_i64) {
2143 tcg_gen_shli_i64(ret, arg, ofs);
2144 tcg_gen_ext8u_i64(ret, ret);
2145 return;
2146 }
2147 break;
2148 }
2149 tcg_gen_andi_i64(ret, arg, (1ull << len) - 1);
2150 tcg_gen_shli_i64(ret, ret, ofs);
2151 }
2152 }
2153
2154 void tcg_gen_extract_i64(TCGv_i64 ret, TCGv_i64 arg,
2155 unsigned int ofs, unsigned int len)
2156 {
2157 tcg_debug_assert(ofs < 64);
2158 tcg_debug_assert(len > 0);
2159 tcg_debug_assert(len <= 64);
2160 tcg_debug_assert(ofs + len <= 64);
2161
2162 /* Canonicalize certain special cases, even if extract is supported. */
2163 if (ofs + len == 64) {
2164 tcg_gen_shri_i64(ret, arg, 64 - len);
2165 return;
2166 }
2167 if (ofs == 0) {
2168 tcg_gen_andi_i64(ret, arg, (1ull << len) - 1);
2169 return;
2170 }
2171
2172 if (TCG_TARGET_REG_BITS == 32) {
2173 /* Look for a 32-bit extract within one of the two words. */
2174 if (ofs >= 32) {
2175 tcg_gen_extract_i32(TCGV_LOW(ret), TCGV_HIGH(arg), ofs - 32, len);
2176 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
2177 return;
2178 }
2179 if (ofs + len <= 32) {
2180 tcg_gen_extract_i32(TCGV_LOW(ret), TCGV_LOW(arg), ofs, len);
2181 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
2182 return;
2183 }
2184 /* The field is split across two words. One double-word
2185 shift is better than two double-word shifts. */
2186 goto do_shift_and;
2187 }
2188
2189 if (TCG_TARGET_HAS_extract_i64
2190 && TCG_TARGET_extract_i64_valid(ofs, len)) {
2191 tcg_gen_op4ii_i64(INDEX_op_extract_i64, ret, arg, ofs, len);
2192 return;
2193 }
2194
2195 /* Assume that zero-extension, if available, is cheaper than a shift. */
2196 switch (ofs + len) {
2197 case 32:
2198 if (TCG_TARGET_HAS_ext32u_i64) {
2199 tcg_gen_ext32u_i64(ret, arg);
2200 tcg_gen_shri_i64(ret, ret, ofs);
2201 return;
2202 }
2203 break;
2204 case 16:
2205 if (TCG_TARGET_HAS_ext16u_i64) {
2206 tcg_gen_ext16u_i64(ret, arg);
2207 tcg_gen_shri_i64(ret, ret, ofs);
2208 return;
2209 }
2210 break;
2211 case 8:
2212 if (TCG_TARGET_HAS_ext8u_i64) {
2213 tcg_gen_ext8u_i64(ret, arg);
2214 tcg_gen_shri_i64(ret, ret, ofs);
2215 return;
2216 }
2217 break;
2218 }
2219
2220 /* ??? Ideally we'd know what values are available for immediate AND.
2221 Assume that 8 bits are available, plus the special cases of 16 and 32,
2222 so that we get ext8u, ext16u, and ext32u. */
2223 switch (len) {
2224 case 1 ... 8: case 16: case 32:
2225 do_shift_and:
2226 tcg_gen_shri_i64(ret, arg, ofs);
2227 tcg_gen_andi_i64(ret, ret, (1ull << len) - 1);
2228 break;
2229 default:
2230 tcg_gen_shli_i64(ret, arg, 64 - len - ofs);
2231 tcg_gen_shri_i64(ret, ret, 64 - len);
2232 break;
2233 }
2234 }
2235
2236 void tcg_gen_sextract_i64(TCGv_i64 ret, TCGv_i64 arg,
2237 unsigned int ofs, unsigned int len)
2238 {
2239 tcg_debug_assert(ofs < 64);
2240 tcg_debug_assert(len > 0);
2241 tcg_debug_assert(len <= 64);
2242 tcg_debug_assert(ofs + len <= 64);
2243
2244 /* Canonicalize certain special cases, even if sextract is supported. */
2245 if (ofs + len == 64) {
2246 tcg_gen_sari_i64(ret, arg, 64 - len);
2247 return;
2248 }
2249 if (ofs == 0) {
2250 switch (len) {
2251 case 32:
2252 tcg_gen_ext32s_i64(ret, arg);
2253 return;
2254 case 16:
2255 tcg_gen_ext16s_i64(ret, arg);
2256 return;
2257 case 8:
2258 tcg_gen_ext8s_i64(ret, arg);
2259 return;
2260 }
2261 }
2262
2263 if (TCG_TARGET_REG_BITS == 32) {
2264 /* Look for a 32-bit extract within one of the two words. */
2265 if (ofs >= 32) {
2266 tcg_gen_sextract_i32(TCGV_LOW(ret), TCGV_HIGH(arg), ofs - 32, len);
2267 } else if (ofs + len <= 32) {
2268 tcg_gen_sextract_i32(TCGV_LOW(ret), TCGV_LOW(arg), ofs, len);
2269 } else if (ofs == 0) {
2270 tcg_gen_mov_i32(TCGV_LOW(ret), TCGV_LOW(arg));
2271 tcg_gen_sextract_i32(TCGV_HIGH(ret), TCGV_HIGH(arg), 0, len - 32);
2272 return;
2273 } else if (len > 32) {
2274 TCGv_i32 t = tcg_temp_new_i32();
2275 /* Extract the bits for the high word normally. */
2276 tcg_gen_sextract_i32(t, TCGV_HIGH(arg), ofs + 32, len - 32);
2277 /* Shift the field down for the low part. */
2278 tcg_gen_shri_i64(ret, arg, ofs);
2279 /* Overwrite the shift into the high part. */
2280 tcg_gen_mov_i32(TCGV_HIGH(ret), t);
2281 tcg_temp_free_i32(t);
2282 return;
2283 } else {
2284 /* Shift the field down for the low part, such that the
2285 field sits at the MSB. */
2286 tcg_gen_shri_i64(ret, arg, ofs + len - 32);
2287 /* Shift the field down from the MSB, sign extending. */
2288 tcg_gen_sari_i32(TCGV_LOW(ret), TCGV_LOW(ret), 32 - len);
2289 }
2290 /* Sign-extend the field from 32 bits. */
2291 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
2292 return;
2293 }
2294
2295 if (TCG_TARGET_HAS_sextract_i64
2296 && TCG_TARGET_extract_i64_valid(ofs, len)) {
2297 tcg_gen_op4ii_i64(INDEX_op_sextract_i64, ret, arg, ofs, len);
2298 return;
2299 }
2300
2301 /* Assume that sign-extension, if available, is cheaper than a shift. */
2302 switch (ofs + len) {
2303 case 32:
2304 if (TCG_TARGET_HAS_ext32s_i64) {
2305 tcg_gen_ext32s_i64(ret, arg);
2306 tcg_gen_sari_i64(ret, ret, ofs);
2307 return;
2308 }
2309 break;
2310 case 16:
2311 if (TCG_TARGET_HAS_ext16s_i64) {
2312 tcg_gen_ext16s_i64(ret, arg);
2313 tcg_gen_sari_i64(ret, ret, ofs);
2314 return;
2315 }
2316 break;
2317 case 8:
2318 if (TCG_TARGET_HAS_ext8s_i64) {
2319 tcg_gen_ext8s_i64(ret, arg);
2320 tcg_gen_sari_i64(ret, ret, ofs);
2321 return;
2322 }
2323 break;
2324 }
2325 switch (len) {
2326 case 32:
2327 if (TCG_TARGET_HAS_ext32s_i64) {
2328 tcg_gen_shri_i64(ret, arg, ofs);
2329 tcg_gen_ext32s_i64(ret, ret);
2330 return;
2331 }
2332 break;
2333 case 16:
2334 if (TCG_TARGET_HAS_ext16s_i64) {
2335 tcg_gen_shri_i64(ret, arg, ofs);
2336 tcg_gen_ext16s_i64(ret, ret);
2337 return;
2338 }
2339 break;
2340 case 8:
2341 if (TCG_TARGET_HAS_ext8s_i64) {
2342 tcg_gen_shri_i64(ret, arg, ofs);
2343 tcg_gen_ext8s_i64(ret, ret);
2344 return;
2345 }
2346 break;
2347 }
2348 tcg_gen_shli_i64(ret, arg, 64 - len - ofs);
2349 tcg_gen_sari_i64(ret, ret, 64 - len);
2350 }
2351
2352 /*
2353 * Extract 64 bits from a 128-bit input, ah:al, starting from ofs.
2354 * Unlike tcg_gen_extract_i64 above, len is fixed at 64.
2355 */
2356 void tcg_gen_extract2_i64(TCGv_i64 ret, TCGv_i64 al, TCGv_i64 ah,
2357 unsigned int ofs)
2358 {
2359 tcg_debug_assert(ofs <= 64);
2360 if (ofs == 0) {
2361 tcg_gen_mov_i64(ret, al);
2362 } else if (ofs == 64) {
2363 tcg_gen_mov_i64(ret, ah);
2364 } else if (al == ah) {
2365 tcg_gen_rotri_i64(ret, al, ofs);
2366 } else if (TCG_TARGET_HAS_extract2_i64) {
2367 tcg_gen_op4i_i64(INDEX_op_extract2_i64, ret, al, ah, ofs);
2368 } else {
2369 TCGv_i64 t0 = tcg_temp_new_i64();
2370 tcg_gen_shri_i64(t0, al, ofs);
2371 tcg_gen_deposit_i64(ret, t0, ah, 64 - ofs, ofs);
2372 tcg_temp_free_i64(t0);
2373 }
2374 }
2375
2376 void tcg_gen_movcond_i64(TCGCond cond, TCGv_i64 ret, TCGv_i64 c1,
2377 TCGv_i64 c2, TCGv_i64 v1, TCGv_i64 v2)
2378 {
2379 if (cond == TCG_COND_ALWAYS) {
2380 tcg_gen_mov_i64(ret, v1);
2381 } else if (cond == TCG_COND_NEVER) {
2382 tcg_gen_mov_i64(ret, v2);
2383 } else if (TCG_TARGET_REG_BITS == 32) {
2384 TCGv_i32 t0 = tcg_temp_new_i32();
2385 TCGv_i32 t1 = tcg_temp_new_i32();
2386 tcg_gen_op6i_i32(INDEX_op_setcond2_i32, t0,
2387 TCGV_LOW(c1), TCGV_HIGH(c1),
2388 TCGV_LOW(c2), TCGV_HIGH(c2), cond);
2389
2390 if (TCG_TARGET_HAS_movcond_i32) {
2391 tcg_gen_movi_i32(t1, 0);
2392 tcg_gen_movcond_i32(TCG_COND_NE, TCGV_LOW(ret), t0, t1,
2393 TCGV_LOW(v1), TCGV_LOW(v2));
2394 tcg_gen_movcond_i32(TCG_COND_NE, TCGV_HIGH(ret), t0, t1,
2395 TCGV_HIGH(v1), TCGV_HIGH(v2));
2396 } else {
2397 tcg_gen_neg_i32(t0, t0);
2398
2399 tcg_gen_and_i32(t1, TCGV_LOW(v1), t0);
2400 tcg_gen_andc_i32(TCGV_LOW(ret), TCGV_LOW(v2), t0);
2401 tcg_gen_or_i32(TCGV_LOW(ret), TCGV_LOW(ret), t1);
2402
2403 tcg_gen_and_i32(t1, TCGV_HIGH(v1), t0);
2404 tcg_gen_andc_i32(TCGV_HIGH(ret), TCGV_HIGH(v2), t0);
2405 tcg_gen_or_i32(TCGV_HIGH(ret), TCGV_HIGH(ret), t1);
2406 }
2407 tcg_temp_free_i32(t0);
2408 tcg_temp_free_i32(t1);
2409 } else if (TCG_TARGET_HAS_movcond_i64) {
2410 tcg_gen_op6i_i64(INDEX_op_movcond_i64, ret, c1, c2, v1, v2, cond);
2411 } else {
2412 TCGv_i64 t0 = tcg_temp_new_i64();
2413 TCGv_i64 t1 = tcg_temp_new_i64();
2414 tcg_gen_setcond_i64(cond, t0, c1, c2);
2415 tcg_gen_neg_i64(t0, t0);
2416 tcg_gen_and_i64(t1, v1, t0);
2417 tcg_gen_andc_i64(ret, v2, t0);
2418 tcg_gen_or_i64(ret, ret, t1);
2419 tcg_temp_free_i64(t0);
2420 tcg_temp_free_i64(t1);
2421 }
2422 }
2423
2424 void tcg_gen_add2_i64(TCGv_i64 rl, TCGv_i64 rh, TCGv_i64 al,
2425 TCGv_i64 ah, TCGv_i64 bl, TCGv_i64 bh)
2426 {
2427 if (TCG_TARGET_HAS_add2_i64) {
2428 tcg_gen_op6_i64(INDEX_op_add2_i64, rl, rh, al, ah, bl, bh);
2429 } else {
2430 TCGv_i64 t0 = tcg_temp_new_i64();
2431 TCGv_i64 t1 = tcg_temp_new_i64();
2432 tcg_gen_add_i64(t0, al, bl);
2433 tcg_gen_setcond_i64(TCG_COND_LTU, t1, t0, al);
2434 tcg_gen_add_i64(rh, ah, bh);
2435 tcg_gen_add_i64(rh, rh, t1);
2436 tcg_gen_mov_i64(rl, t0);
2437 tcg_temp_free_i64(t0);
2438 tcg_temp_free_i64(t1);
2439 }
2440 }
2441
2442 void tcg_gen_sub2_i64(TCGv_i64 rl, TCGv_i64 rh, TCGv_i64 al,
2443 TCGv_i64 ah, TCGv_i64 bl, TCGv_i64 bh)
2444 {
2445 if (TCG_TARGET_HAS_sub2_i64) {
2446 tcg_gen_op6_i64(INDEX_op_sub2_i64, rl, rh, al, ah, bl, bh);
2447 } else {
2448 TCGv_i64 t0 = tcg_temp_new_i64();
2449 TCGv_i64 t1 = tcg_temp_new_i64();
2450 tcg_gen_sub_i64(t0, al, bl);
2451 tcg_gen_setcond_i64(TCG_COND_LTU, t1, al, bl);
2452 tcg_gen_sub_i64(rh, ah, bh);
2453 tcg_gen_sub_i64(rh, rh, t1);
2454 tcg_gen_mov_i64(rl, t0);
2455 tcg_temp_free_i64(t0);
2456 tcg_temp_free_i64(t1);
2457 }
2458 }
2459
2460 void tcg_gen_mulu2_i64(TCGv_i64 rl, TCGv_i64 rh, TCGv_i64 arg1, TCGv_i64 arg2)
2461 {
2462 if (TCG_TARGET_HAS_mulu2_i64) {
2463 tcg_gen_op4_i64(INDEX_op_mulu2_i64, rl, rh, arg1, arg2);
2464 } else if (TCG_TARGET_HAS_muluh_i64) {
2465 TCGv_i64 t = tcg_temp_new_i64();
2466 tcg_gen_op3_i64(INDEX_op_mul_i64, t, arg1, arg2);
2467 tcg_gen_op3_i64(INDEX_op_muluh_i64, rh, arg1, arg2);
2468 tcg_gen_mov_i64(rl, t);
2469 tcg_temp_free_i64(t);
2470 } else {
2471 TCGv_i64 t0 = tcg_temp_new_i64();
2472 tcg_gen_mul_i64(t0, arg1, arg2);
2473 gen_helper_muluh_i64(rh, arg1, arg2);
2474 tcg_gen_mov_i64(rl, t0);
2475 tcg_temp_free_i64(t0);
2476 }
2477 }
2478
2479 void tcg_gen_muls2_i64(TCGv_i64 rl, TCGv_i64 rh, TCGv_i64 arg1, TCGv_i64 arg2)
2480 {
2481 if (TCG_TARGET_HAS_muls2_i64) {
2482 tcg_gen_op4_i64(INDEX_op_muls2_i64, rl, rh, arg1, arg2);
2483 } else if (TCG_TARGET_HAS_mulsh_i64) {
2484 TCGv_i64 t = tcg_temp_new_i64();
2485 tcg_gen_op3_i64(INDEX_op_mul_i64, t, arg1, arg2);
2486 tcg_gen_op3_i64(INDEX_op_mulsh_i64, rh, arg1, arg2);
2487 tcg_gen_mov_i64(rl, t);
2488 tcg_temp_free_i64(t);
2489 } else if (TCG_TARGET_HAS_mulu2_i64 || TCG_TARGET_HAS_muluh_i64) {
2490 TCGv_i64 t0 = tcg_temp_new_i64();
2491 TCGv_i64 t1 = tcg_temp_new_i64();
2492 TCGv_i64 t2 = tcg_temp_new_i64();
2493 TCGv_i64 t3 = tcg_temp_new_i64();
2494 tcg_gen_mulu2_i64(t0, t1, arg1, arg2);
2495 /* Adjust for negative inputs. */
2496 tcg_gen_sari_i64(t2, arg1, 63);
2497 tcg_gen_sari_i64(t3, arg2, 63);
2498 tcg_gen_and_i64(t2, t2, arg2);
2499 tcg_gen_and_i64(t3, t3, arg1);
2500 tcg_gen_sub_i64(rh, t1, t2);
2501 tcg_gen_sub_i64(rh, rh, t3);
2502 tcg_gen_mov_i64(rl, t0);
2503 tcg_temp_free_i64(t0);
2504 tcg_temp_free_i64(t1);
2505 tcg_temp_free_i64(t2);
2506 tcg_temp_free_i64(t3);
2507 } else {
2508 TCGv_i64 t0 = tcg_temp_new_i64();
2509 tcg_gen_mul_i64(t0, arg1, arg2);
2510 gen_helper_mulsh_i64(rh, arg1, arg2);
2511 tcg_gen_mov_i64(rl, t0);
2512 tcg_temp_free_i64(t0);
2513 }
2514 }
2515
2516 void tcg_gen_mulsu2_i64(TCGv_i64 rl, TCGv_i64 rh, TCGv_i64 arg1, TCGv_i64 arg2)
2517 {
2518 TCGv_i64 t0 = tcg_temp_new_i64();
2519 TCGv_i64 t1 = tcg_temp_new_i64();
2520 TCGv_i64 t2 = tcg_temp_new_i64();
2521 tcg_gen_mulu2_i64(t0, t1, arg1, arg2);
2522 /* Adjust for negative input for the signed arg1. */
2523 tcg_gen_sari_i64(t2, arg1, 63);
2524 tcg_gen_and_i64(t2, t2, arg2);
2525 tcg_gen_sub_i64(rh, t1, t2);
2526 tcg_gen_mov_i64(rl, t0);
2527 tcg_temp_free_i64(t0);
2528 tcg_temp_free_i64(t1);
2529 tcg_temp_free_i64(t2);
2530 }
2531
2532 void tcg_gen_smin_i64(TCGv_i64 ret, TCGv_i64 a, TCGv_i64 b)
2533 {
2534 tcg_gen_movcond_i64(TCG_COND_LT, ret, a, b, a, b);
2535 }
2536
2537 void tcg_gen_umin_i64(TCGv_i64 ret, TCGv_i64 a, TCGv_i64 b)
2538 {
2539 tcg_gen_movcond_i64(TCG_COND_LTU, ret, a, b, a, b);
2540 }
2541
2542 void tcg_gen_smax_i64(TCGv_i64 ret, TCGv_i64 a, TCGv_i64 b)
2543 {
2544 tcg_gen_movcond_i64(TCG_COND_LT, ret, a, b, b, a);
2545 }
2546
2547 void tcg_gen_umax_i64(TCGv_i64 ret, TCGv_i64 a, TCGv_i64 b)
2548 {
2549 tcg_gen_movcond_i64(TCG_COND_LTU, ret, a, b, b, a);
2550 }
2551
2552 void tcg_gen_abs_i64(TCGv_i64 ret, TCGv_i64 a)
2553 {
2554 TCGv_i64 t = tcg_temp_new_i64();
2555
2556 tcg_gen_sari_i64(t, a, 63);
2557 tcg_gen_xor_i64(ret, a, t);
2558 tcg_gen_sub_i64(ret, ret, t);
2559 tcg_temp_free_i64(t);
2560 }
2561
2562 /* Size changing operations. */
2563
2564 void tcg_gen_extrl_i64_i32(TCGv_i32 ret, TCGv_i64 arg)
2565 {
2566 if (TCG_TARGET_REG_BITS == 32) {
2567 tcg_gen_mov_i32(ret, TCGV_LOW(arg));
2568 } else if (TCG_TARGET_HAS_extrl_i64_i32) {
2569 tcg_gen_op2(INDEX_op_extrl_i64_i32,
2570 tcgv_i32_arg(ret), tcgv_i64_arg(arg));
2571 } else {
2572 tcg_gen_mov_i32(ret, (TCGv_i32)arg);
2573 }
2574 }
2575
2576 void tcg_gen_extrh_i64_i32(TCGv_i32 ret, TCGv_i64 arg)
2577 {
2578 if (TCG_TARGET_REG_BITS == 32) {
2579 tcg_gen_mov_i32(ret, TCGV_HIGH(arg));
2580 } else if (TCG_TARGET_HAS_extrh_i64_i32) {
2581 tcg_gen_op2(INDEX_op_extrh_i64_i32,
2582 tcgv_i32_arg(ret), tcgv_i64_arg(arg));
2583 } else {
2584 TCGv_i64 t = tcg_temp_new_i64();
2585 tcg_gen_shri_i64(t, arg, 32);
2586 tcg_gen_mov_i32(ret, (TCGv_i32)t);
2587 tcg_temp_free_i64(t);
2588 }
2589 }
2590
2591 void tcg_gen_extu_i32_i64(TCGv_i64 ret, TCGv_i32 arg)
2592 {
2593 if (TCG_TARGET_REG_BITS == 32) {
2594 tcg_gen_mov_i32(TCGV_LOW(ret), arg);
2595 tcg_gen_movi_i32(TCGV_HIGH(ret), 0);
2596 } else {
2597 tcg_gen_op2(INDEX_op_extu_i32_i64,
2598 tcgv_i64_arg(ret), tcgv_i32_arg(arg));
2599 }
2600 }
2601
2602 void tcg_gen_ext_i32_i64(TCGv_i64 ret, TCGv_i32 arg)
2603 {
2604 if (TCG_TARGET_REG_BITS == 32) {
2605 tcg_gen_mov_i32(TCGV_LOW(ret), arg);
2606 tcg_gen_sari_i32(TCGV_HIGH(ret), TCGV_LOW(ret), 31);
2607 } else {
2608 tcg_gen_op2(INDEX_op_ext_i32_i64,
2609 tcgv_i64_arg(ret), tcgv_i32_arg(arg));
2610 }
2611 }
2612
2613 void tcg_gen_concat_i32_i64(TCGv_i64 dest, TCGv_i32 low, TCGv_i32 high)
2614 {
2615 TCGv_i64 tmp;
2616
2617 if (TCG_TARGET_REG_BITS == 32) {
2618 tcg_gen_mov_i32(TCGV_LOW(dest), low);
2619 tcg_gen_mov_i32(TCGV_HIGH(dest), high);
2620 return;
2621 }
2622
2623 tmp = tcg_temp_new_i64();
2624 /* These extensions are only needed for type correctness.
2625 We may be able to do better given target specific information. */
2626 tcg_gen_extu_i32_i64(tmp, high);
2627 tcg_gen_extu_i32_i64(dest, low);
2628 /* If deposit is available, use it. Otherwise use the extra
2629 knowledge that we have of the zero-extensions above. */
2630 if (TCG_TARGET_HAS_deposit_i64 && TCG_TARGET_deposit_i64_valid(32, 32)) {
2631 tcg_gen_deposit_i64(dest, dest, tmp, 32, 32);
2632 } else {
2633 tcg_gen_shli_i64(tmp, tmp, 32);
2634 tcg_gen_or_i64(dest, dest, tmp);
2635 }
2636 tcg_temp_free_i64(tmp);
2637 }
2638
2639 void tcg_gen_extr_i64_i32(TCGv_i32 lo, TCGv_i32 hi, TCGv_i64 arg)
2640 {
2641 if (TCG_TARGET_REG_BITS == 32) {
2642 tcg_gen_mov_i32(lo, TCGV_LOW(arg));
2643 tcg_gen_mov_i32(hi, TCGV_HIGH(arg));
2644 } else {
2645 tcg_gen_extrl_i64_i32(lo, arg);
2646 tcg_gen_extrh_i64_i32(hi, arg);
2647 }
2648 }
2649
2650 void tcg_gen_extr32_i64(TCGv_i64 lo, TCGv_i64 hi, TCGv_i64 arg)
2651 {
2652 tcg_gen_ext32u_i64(lo, arg);
2653 tcg_gen_shri_i64(hi, arg, 32);
2654 }
2655
2656 /* QEMU specific operations. */
2657
2658 void tcg_gen_exit_tb(const TranslationBlock *tb, unsigned idx)
2659 {
2660 /*
2661 * Let the jit code return the read-only version of the
2662 * TranslationBlock, so that we minimize the pc-relative
2663 * distance of the address of the exit_tb code to TB.
2664 * This will improve utilization of pc-relative address loads.
2665 *
2666 * TODO: Move this to translator_loop, so that all const
2667 * TranslationBlock pointers refer to read-only memory.
2668 * This requires coordination with targets that do not use
2669 * the translator_loop.
2670 */
2671 uintptr_t val = (uintptr_t)tcg_splitwx_to_rx((void *)tb) + idx;
2672
2673 if (tb == NULL) {
2674 tcg_debug_assert(idx == 0);
2675 } else if (idx <= TB_EXIT_IDXMAX) {
2676 #ifdef CONFIG_DEBUG_TCG
2677 /* This is an exit following a goto_tb. Verify that we have
2678 seen this numbered exit before, via tcg_gen_goto_tb. */
2679 tcg_debug_assert(tcg_ctx->goto_tb_issue_mask & (1 << idx));
2680 #endif
2681 /* When not chaining, exit without indicating a link. */
2682 if (qemu_loglevel_mask(CPU_LOG_TB_NOCHAIN)) {
2683 val = 0;
2684 }
2685 } else {
2686 /* This is an exit via the exitreq label. */
2687 tcg_debug_assert(idx == TB_EXIT_REQUESTED);
2688 }
2689
2690 plugin_gen_disable_mem_helpers();
2691 tcg_gen_op1i(INDEX_op_exit_tb, val);
2692 }
2693
2694 void tcg_gen_goto_tb(unsigned idx)
2695 {
2696 /* We only support two chained exits. */
2697 tcg_debug_assert(idx <= TB_EXIT_IDXMAX);
2698 #ifdef CONFIG_DEBUG_TCG
2699 /* Verify that we havn't seen this numbered exit before. */
2700 tcg_debug_assert((tcg_ctx->goto_tb_issue_mask & (1 << idx)) == 0);
2701 tcg_ctx->goto_tb_issue_mask |= 1 << idx;
2702 #endif
2703 plugin_gen_disable_mem_helpers();
2704 /* When not chaining, we simply fall through to the "fallback" exit. */
2705 if (!qemu_loglevel_mask(CPU_LOG_TB_NOCHAIN)) {
2706 tcg_gen_op1i(INDEX_op_goto_tb, idx);
2707 }
2708 }
2709
2710 void tcg_gen_lookup_and_goto_ptr(void)
2711 {
2712 if (TCG_TARGET_HAS_goto_ptr && !qemu_loglevel_mask(CPU_LOG_TB_NOCHAIN)) {
2713 TCGv_ptr ptr;
2714
2715 plugin_gen_disable_mem_helpers();
2716 ptr = tcg_temp_new_ptr();
2717 gen_helper_lookup_tb_ptr(ptr, cpu_env);
2718 tcg_gen_op1i(INDEX_op_goto_ptr, tcgv_ptr_arg(ptr));
2719 tcg_temp_free_ptr(ptr);
2720 } else {
2721 tcg_gen_exit_tb(NULL, 0);
2722 }
2723 }
2724
2725 static inline MemOp tcg_canonicalize_memop(MemOp op, bool is64, bool st)
2726 {
2727 /* Trigger the asserts within as early as possible. */
2728 (void)get_alignment_bits(op);
2729
2730 switch (op & MO_SIZE) {
2731 case MO_8:
2732 op &= ~MO_BSWAP;
2733 break;
2734 case MO_16:
2735 break;
2736 case MO_32:
2737 if (!is64) {
2738 op &= ~MO_SIGN;
2739 }
2740 break;
2741 case MO_64:
2742 if (!is64) {
2743 tcg_abort();
2744 }
2745 break;
2746 }
2747 if (st) {
2748 op &= ~MO_SIGN;
2749 }
2750 return op;
2751 }
2752
2753 static void gen_ldst_i32(TCGOpcode opc, TCGv_i32 val, TCGv addr,
2754 MemOp memop, TCGArg idx)
2755 {
2756 TCGMemOpIdx oi = make_memop_idx(memop, idx);
2757 #if TARGET_LONG_BITS == 32
2758 tcg_gen_op3i_i32(opc, val, addr, oi);
2759 #else
2760 if (TCG_TARGET_REG_BITS == 32) {
2761 tcg_gen_op4i_i32(opc, val, TCGV_LOW(addr), TCGV_HIGH(addr), oi);
2762 } else {
2763 tcg_gen_op3(opc, tcgv_i32_arg(val), tcgv_i64_arg(addr), oi);
2764 }
2765 #endif
2766 }
2767
2768 static void gen_ldst_i64(TCGOpcode opc, TCGv_i64 val, TCGv addr,
2769 MemOp memop, TCGArg idx)
2770 {
2771 TCGMemOpIdx oi = make_memop_idx(memop, idx);
2772 #if TARGET_LONG_BITS == 32
2773 if (TCG_TARGET_REG_BITS == 32) {
2774 tcg_gen_op4i_i32(opc, TCGV_LOW(val), TCGV_HIGH(val), addr, oi);
2775 } else {
2776 tcg_gen_op3(opc, tcgv_i64_arg(val), tcgv_i32_arg(addr), oi);
2777 }
2778 #else
2779 if (TCG_TARGET_REG_BITS == 32) {
2780 tcg_gen_op5i_i32(opc, TCGV_LOW(val), TCGV_HIGH(val),
2781 TCGV_LOW(addr), TCGV_HIGH(addr), oi);
2782 } else {
2783 tcg_gen_op3i_i64(opc, val, addr, oi);
2784 }
2785 #endif
2786 }
2787
2788 static void tcg_gen_req_mo(TCGBar type)
2789 {
2790 #ifdef TCG_GUEST_DEFAULT_MO
2791 type &= TCG_GUEST_DEFAULT_MO;
2792 #endif
2793 type &= ~TCG_TARGET_DEFAULT_MO;
2794 if (type) {
2795 tcg_gen_mb(type | TCG_BAR_SC);
2796 }
2797 }
2798
2799 static inline TCGv plugin_prep_mem_callbacks(TCGv vaddr)
2800 {
2801 #ifdef CONFIG_PLUGIN
2802 if (tcg_ctx->plugin_insn != NULL) {
2803 /* Save a copy of the vaddr for use after a load. */
2804 TCGv temp = tcg_temp_new();
2805 tcg_gen_mov_tl(temp, vaddr);
2806 return temp;
2807 }
2808 #endif
2809 return vaddr;
2810 }
2811
2812 static inline void plugin_gen_mem_callbacks(TCGv vaddr, uint16_t info)
2813 {
2814 #ifdef CONFIG_PLUGIN
2815 if (tcg_ctx->plugin_insn != NULL) {
2816 plugin_gen_empty_mem_callback(vaddr, info);
2817 tcg_temp_free(vaddr);
2818 }
2819 #endif
2820 }
2821
2822 void tcg_gen_qemu_ld_i32(TCGv_i32 val, TCGv addr, TCGArg idx, MemOp memop)
2823 {
2824 MemOp orig_memop;
2825 uint16_t info = trace_mem_get_info(memop, idx, 0);
2826
2827 tcg_gen_req_mo(TCG_MO_LD_LD | TCG_MO_ST_LD);
2828 memop = tcg_canonicalize_memop(memop, 0, 0);
2829 trace_guest_mem_before_tcg(tcg_ctx->cpu, cpu_env, addr, info);
2830
2831 orig_memop = memop;
2832 if (!TCG_TARGET_HAS_MEMORY_BSWAP && (memop & MO_BSWAP)) {
2833 memop &= ~MO_BSWAP;
2834 /* The bswap primitive requires zero-extended input. */
2835 if ((memop & MO_SSIZE) == MO_SW) {
2836 memop &= ~MO_SIGN;
2837 }
2838 }
2839
2840 addr = plugin_prep_mem_callbacks(addr);
2841 gen_ldst_i32(INDEX_op_qemu_ld_i32, val, addr, memop, idx);
2842 plugin_gen_mem_callbacks(addr, info);
2843
2844 if ((orig_memop ^ memop) & MO_BSWAP) {
2845 switch (orig_memop & MO_SIZE) {
2846 case MO_16:
2847 tcg_gen_bswap16_i32(val, val);
2848 if (orig_memop & MO_SIGN) {
2849 tcg_gen_ext16s_i32(val, val);
2850 }
2851 break;
2852 case MO_32:
2853 tcg_gen_bswap32_i32(val, val);
2854 break;
2855 default:
2856 g_assert_not_reached();
2857 }
2858 }
2859 }
2860
2861 void tcg_gen_qemu_st_i32(TCGv_i32 val, TCGv addr, TCGArg idx, MemOp memop)
2862 {
2863 TCGv_i32 swap = NULL;
2864 uint16_t info = trace_mem_get_info(memop, idx, 1);
2865
2866 tcg_gen_req_mo(TCG_MO_LD_ST | TCG_MO_ST_ST);
2867 memop = tcg_canonicalize_memop(memop, 0, 1);
2868 trace_guest_mem_before_tcg(tcg_ctx->cpu, cpu_env, addr, info);
2869
2870 if (!TCG_TARGET_HAS_MEMORY_BSWAP && (memop & MO_BSWAP)) {
2871 swap = tcg_temp_new_i32();
2872 switch (memop & MO_SIZE) {
2873 case MO_16:
2874 tcg_gen_ext16u_i32(swap, val);
2875 tcg_gen_bswap16_i32(swap, swap);
2876 break;
2877 case MO_32:
2878 tcg_gen_bswap32_i32(swap, val);
2879 break;
2880 default:
2881 g_assert_not_reached();
2882 }
2883 val = swap;
2884 memop &= ~MO_BSWAP;
2885 }
2886
2887 addr = plugin_prep_mem_callbacks(addr);
2888 if (TCG_TARGET_HAS_qemu_st8_i32 && (memop & MO_SIZE) == MO_8) {
2889 gen_ldst_i32(INDEX_op_qemu_st8_i32, val, addr, memop, idx);
2890 } else {
2891 gen_ldst_i32(INDEX_op_qemu_st_i32, val, addr, memop, idx);
2892 }
2893 plugin_gen_mem_callbacks(addr, info);
2894
2895 if (swap) {
2896 tcg_temp_free_i32(swap);
2897 }
2898 }
2899
2900 void tcg_gen_qemu_ld_i64(TCGv_i64 val, TCGv addr, TCGArg idx, MemOp memop)
2901 {
2902 MemOp orig_memop;
2903 uint16_t info;
2904
2905 if (TCG_TARGET_REG_BITS == 32 && (memop & MO_SIZE) < MO_64) {
2906 tcg_gen_qemu_ld_i32(TCGV_LOW(val), addr, idx, memop);
2907 if (memop & MO_SIGN) {
2908 tcg_gen_sari_i32(TCGV_HIGH(val), TCGV_LOW(val), 31);
2909 } else {
2910 tcg_gen_movi_i32(TCGV_HIGH(val), 0);
2911 }
2912 return;
2913 }
2914
2915 tcg_gen_req_mo(TCG_MO_LD_LD | TCG_MO_ST_LD);
2916 memop = tcg_canonicalize_memop(memop, 1, 0);
2917 info = trace_mem_get_info(memop, idx, 0);
2918 trace_guest_mem_before_tcg(tcg_ctx->cpu, cpu_env, addr, info);
2919
2920 orig_memop = memop;
2921 if (!TCG_TARGET_HAS_MEMORY_BSWAP && (memop & MO_BSWAP)) {
2922 memop &= ~MO_BSWAP;
2923 /* The bswap primitive requires zero-extended input. */
2924 if ((memop & MO_SIGN) && (memop & MO_SIZE) < MO_64) {
2925 memop &= ~MO_SIGN;
2926 }
2927 }
2928
2929 addr = plugin_prep_mem_callbacks(addr);
2930 gen_ldst_i64(INDEX_op_qemu_ld_i64, val, addr, memop, idx);
2931 plugin_gen_mem_callbacks(addr, info);
2932
2933 if ((orig_memop ^ memop) & MO_BSWAP) {
2934 switch (orig_memop & MO_SIZE) {
2935 case MO_16:
2936 tcg_gen_bswap16_i64(val, val);
2937 if (orig_memop & MO_SIGN) {
2938 tcg_gen_ext16s_i64(val, val);
2939 }
2940 break;
2941 case MO_32:
2942 tcg_gen_bswap32_i64(val, val);
2943 if (orig_memop & MO_SIGN) {
2944 tcg_gen_ext32s_i64(val, val);
2945 }
2946 break;
2947 case MO_64:
2948 tcg_gen_bswap64_i64(val, val);
2949 break;
2950 default:
2951 g_assert_not_reached();
2952 }
2953 }
2954 }
2955
2956 void tcg_gen_qemu_st_i64(TCGv_i64 val, TCGv addr, TCGArg idx, MemOp memop)
2957 {
2958 TCGv_i64 swap = NULL;
2959 uint16_t info;
2960
2961 if (TCG_TARGET_REG_BITS == 32 && (memop & MO_SIZE) < MO_64) {
2962 tcg_gen_qemu_st_i32(TCGV_LOW(val), addr, idx, memop);
2963 return;
2964 }
2965
2966 tcg_gen_req_mo(TCG_MO_LD_ST | TCG_MO_ST_ST);
2967 memop = tcg_canonicalize_memop(memop, 1, 1);
2968 info = trace_mem_get_info(memop, idx, 1);
2969 trace_guest_mem_before_tcg(tcg_ctx->cpu, cpu_env, addr, info);
2970
2971 if (!TCG_TARGET_HAS_MEMORY_BSWAP && (memop & MO_BSWAP)) {
2972 swap = tcg_temp_new_i64();
2973 switch (memop & MO_SIZE) {
2974 case MO_16:
2975 tcg_gen_ext16u_i64(swap, val);
2976 tcg_gen_bswap16_i64(swap, swap);
2977 break;
2978 case MO_32:
2979 tcg_gen_ext32u_i64(swap, val);
2980 tcg_gen_bswap32_i64(swap, swap);
2981 break;
2982 case MO_64:
2983 tcg_gen_bswap64_i64(swap, val);
2984 break;
2985 default:
2986 g_assert_not_reached();
2987 }
2988 val = swap;
2989 memop &= ~MO_BSWAP;
2990 }
2991
2992 addr = plugin_prep_mem_callbacks(addr);
2993 gen_ldst_i64(INDEX_op_qemu_st_i64, val, addr, memop, idx);
2994 plugin_gen_mem_callbacks(addr, info);
2995
2996 if (swap) {
2997 tcg_temp_free_i64(swap);
2998 }
2999 }
3000
3001 static void tcg_gen_ext_i32(TCGv_i32 ret, TCGv_i32 val, MemOp opc)
3002 {
3003 switch (opc & MO_SSIZE) {
3004 case MO_SB:
3005 tcg_gen_ext8s_i32(ret, val);
3006 break;
3007 case MO_UB:
3008 tcg_gen_ext8u_i32(ret, val);
3009 break;
3010 case MO_SW:
3011 tcg_gen_ext16s_i32(ret, val);
3012 break;
3013 case MO_UW:
3014 tcg_gen_ext16u_i32(ret, val);
3015 break;
3016 default:
3017 tcg_gen_mov_i32(ret, val);
3018 break;
3019 }
3020 }
3021
3022 static void tcg_gen_ext_i64(TCGv_i64 ret, TCGv_i64 val, MemOp opc)
3023 {
3024 switch (opc & MO_SSIZE) {
3025 case MO_SB:
3026 tcg_gen_ext8s_i64(ret, val);
3027 break;
3028 case MO_UB:
3029 tcg_gen_ext8u_i64(ret, val);
3030 break;
3031 case MO_SW:
3032 tcg_gen_ext16s_i64(ret, val);
3033 break;
3034 case MO_UW:
3035 tcg_gen_ext16u_i64(ret, val);
3036 break;
3037 case MO_SL:
3038 tcg_gen_ext32s_i64(ret, val);
3039 break;
3040 case MO_UL:
3041 tcg_gen_ext32u_i64(ret, val);
3042 break;
3043 default:
3044 tcg_gen_mov_i64(ret, val);
3045 break;
3046 }
3047 }
3048
3049 #ifdef CONFIG_SOFTMMU
3050 typedef void (*gen_atomic_cx_i32)(TCGv_i32, TCGv_env, TCGv,
3051 TCGv_i32, TCGv_i32, TCGv_i32);
3052 typedef void (*gen_atomic_cx_i64)(TCGv_i64, TCGv_env, TCGv,
3053 TCGv_i64, TCGv_i64, TCGv_i32);
3054 typedef void (*gen_atomic_op_i32)(TCGv_i32, TCGv_env, TCGv,
3055 TCGv_i32, TCGv_i32);
3056 typedef void (*gen_atomic_op_i64)(TCGv_i64, TCGv_env, TCGv,
3057 TCGv_i64, TCGv_i32);
3058 #else
3059 typedef void (*gen_atomic_cx_i32)(TCGv_i32, TCGv_env, TCGv, TCGv_i32, TCGv_i32);
3060 typedef void (*gen_atomic_cx_i64)(TCGv_i64, TCGv_env, TCGv, TCGv_i64, TCGv_i64);
3061 typedef void (*gen_atomic_op_i32)(TCGv_i32, TCGv_env, TCGv, TCGv_i32);
3062 typedef void (*gen_atomic_op_i64)(TCGv_i64, TCGv_env, TCGv, TCGv_i64);
3063 #endif
3064
3065 #ifdef CONFIG_ATOMIC64
3066 # define WITH_ATOMIC64(X) X,
3067 #else
3068 # define WITH_ATOMIC64(X)
3069 #endif
3070
3071 static void * const table_cmpxchg[16] = {
3072 [MO_8] = gen_helper_atomic_cmpxchgb,
3073 [MO_16 | MO_LE] = gen_helper_atomic_cmpxchgw_le,
3074 [MO_16 | MO_BE] = gen_helper_atomic_cmpxchgw_be,
3075 [MO_32 | MO_LE] = gen_helper_atomic_cmpxchgl_le,
3076 [MO_32 | MO_BE] = gen_helper_atomic_cmpxchgl_be,
3077 WITH_ATOMIC64([MO_64 | MO_LE] = gen_helper_atomic_cmpxchgq_le)
3078 WITH_ATOMIC64([MO_64 | MO_BE] = gen_helper_atomic_cmpxchgq_be)
3079 };
3080
3081 void tcg_gen_atomic_cmpxchg_i32(TCGv_i32 retv, TCGv addr, TCGv_i32 cmpv,
3082 TCGv_i32 newv, TCGArg idx, MemOp memop)
3083 {
3084 memop = tcg_canonicalize_memop(memop, 0, 0);
3085
3086 if (!(tcg_ctx->tb_cflags & CF_PARALLEL)) {
3087 TCGv_i32 t1 = tcg_temp_new_i32();
3088 TCGv_i32 t2 = tcg_temp_new_i32();
3089
3090 tcg_gen_ext_i32(t2, cmpv, memop & MO_SIZE);
3091
3092 tcg_gen_qemu_ld_i32(t1, addr, idx, memop & ~MO_SIGN);
3093 tcg_gen_movcond_i32(TCG_COND_EQ, t2, t1, t2, newv, t1);
3094 tcg_gen_qemu_st_i32(t2, addr, idx, memop);
3095 tcg_temp_free_i32(t2);
3096
3097 if (memop & MO_SIGN) {
3098 tcg_gen_ext_i32(retv, t1, memop);
3099 } else {
3100 tcg_gen_mov_i32(retv, t1);
3101 }
3102 tcg_temp_free_i32(t1);
3103 } else {
3104 gen_atomic_cx_i32 gen;
3105
3106 gen = table_cmpxchg[memop & (MO_SIZE | MO_BSWAP)];
3107 tcg_debug_assert(gen != NULL);
3108
3109 #ifdef CONFIG_SOFTMMU
3110 {
3111 TCGMemOpIdx oi = make_memop_idx(memop & ~MO_SIGN, idx);
3112 gen(retv, cpu_env, addr, cmpv, newv, tcg_constant_i32(oi));
3113 }
3114 #else
3115 gen(retv, cpu_env, addr, cmpv, newv);
3116 #endif
3117
3118 if (memop & MO_SIGN) {
3119 tcg_gen_ext_i32(retv, retv, memop);
3120 }
3121 }
3122 }
3123
3124 void tcg_gen_atomic_cmpxchg_i64(TCGv_i64 retv, TCGv addr, TCGv_i64 cmpv,
3125 TCGv_i64 newv, TCGArg idx, MemOp memop)
3126 {
3127 memop = tcg_canonicalize_memop(memop, 1, 0);
3128
3129 if (!(tcg_ctx->tb_cflags & CF_PARALLEL)) {
3130 TCGv_i64 t1 = tcg_temp_new_i64();
3131 TCGv_i64 t2 = tcg_temp_new_i64();
3132
3133 tcg_gen_ext_i64(t2, cmpv, memop & MO_SIZE);
3134
3135 tcg_gen_qemu_ld_i64(t1, addr, idx, memop & ~MO_SIGN);
3136 tcg_gen_movcond_i64(TCG_COND_EQ, t2, t1, t2, newv, t1);
3137 tcg_gen_qemu_st_i64(t2, addr, idx, memop);
3138 tcg_temp_free_i64(t2);
3139
3140 if (memop & MO_SIGN) {
3141 tcg_gen_ext_i64(retv, t1, memop);
3142 } else {
3143 tcg_gen_mov_i64(retv, t1);
3144 }
3145 tcg_temp_free_i64(t1);
3146 } else if ((memop & MO_SIZE) == MO_64) {
3147 #ifdef CONFIG_ATOMIC64
3148 gen_atomic_cx_i64 gen;
3149
3150 gen = table_cmpxchg[memop & (MO_SIZE | MO_BSWAP)];
3151 tcg_debug_assert(gen != NULL);
3152
3153 #ifdef CONFIG_SOFTMMU
3154 {
3155 TCGMemOpIdx oi = make_memop_idx(memop, idx);
3156 gen(retv, cpu_env, addr, cmpv, newv, tcg_constant_i32(oi));
3157 }
3158 #else
3159 gen(retv, cpu_env, addr, cmpv, newv);
3160 #endif
3161 #else
3162 gen_helper_exit_atomic(cpu_env);
3163 /* Produce a result, so that we have a well-formed opcode stream
3164 with respect to uses of the result in the (dead) code following. */
3165 tcg_gen_movi_i64(retv, 0);
3166 #endif /* CONFIG_ATOMIC64 */
3167 } else {
3168 TCGv_i32 c32 = tcg_temp_new_i32();
3169 TCGv_i32 n32 = tcg_temp_new_i32();
3170 TCGv_i32 r32 = tcg_temp_new_i32();
3171
3172 tcg_gen_extrl_i64_i32(c32, cmpv);
3173 tcg_gen_extrl_i64_i32(n32, newv);
3174 tcg_gen_atomic_cmpxchg_i32(r32, addr, c32, n32, idx, memop & ~MO_SIGN);
3175 tcg_temp_free_i32(c32);
3176 tcg_temp_free_i32(n32);
3177
3178 tcg_gen_extu_i32_i64(retv, r32);
3179 tcg_temp_free_i32(r32);
3180
3181 if (memop & MO_SIGN) {
3182 tcg_gen_ext_i64(retv, retv, memop);
3183 }
3184 }
3185 }
3186
3187 static void do_nonatomic_op_i32(TCGv_i32 ret, TCGv addr, TCGv_i32 val,
3188 TCGArg idx, MemOp memop, bool new_val,
3189 void (*gen)(TCGv_i32, TCGv_i32, TCGv_i32))
3190 {
3191 TCGv_i32 t1 = tcg_temp_new_i32();
3192 TCGv_i32 t2 = tcg_temp_new_i32();
3193
3194 memop = tcg_canonicalize_memop(memop, 0, 0);
3195
3196 tcg_gen_qemu_ld_i32(t1, addr, idx, memop);
3197 tcg_gen_ext_i32(t2, val, memop);
3198 gen(t2, t1, t2);
3199 tcg_gen_qemu_st_i32(t2, addr, idx, memop);
3200
3201 tcg_gen_ext_i32(ret, (new_val ? t2 : t1), memop);
3202 tcg_temp_free_i32(t1);
3203 tcg_temp_free_i32(t2);
3204 }
3205
3206 static void do_atomic_op_i32(TCGv_i32 ret, TCGv addr, TCGv_i32 val,
3207 TCGArg idx, MemOp memop, void * const table[])
3208 {
3209 gen_atomic_op_i32 gen;
3210
3211 memop = tcg_canonicalize_memop(memop, 0, 0);
3212
3213 gen = table[memop & (MO_SIZE | MO_BSWAP)];
3214 tcg_debug_assert(gen != NULL);
3215
3216 #ifdef CONFIG_SOFTMMU
3217 {
3218 TCGMemOpIdx oi = make_memop_idx(memop & ~MO_SIGN, idx);
3219 gen(ret, cpu_env, addr, val, tcg_constant_i32(oi));
3220 }
3221 #else
3222 gen(ret, cpu_env, addr, val);
3223 #endif
3224
3225 if (memop & MO_SIGN) {
3226 tcg_gen_ext_i32(ret, ret, memop);
3227 }
3228 }
3229
3230 static void do_nonatomic_op_i64(TCGv_i64 ret, TCGv addr, TCGv_i64 val,
3231 TCGArg idx, MemOp memop, bool new_val,
3232 void (*gen)(TCGv_i64, TCGv_i64, TCGv_i64))
3233 {
3234 TCGv_i64 t1 = tcg_temp_new_i64();
3235 TCGv_i64 t2 = tcg_temp_new_i64();
3236
3237 memop = tcg_canonicalize_memop(memop, 1, 0);
3238
3239 tcg_gen_qemu_ld_i64(t1, addr, idx, memop);
3240 tcg_gen_ext_i64(t2, val, memop);
3241 gen(t2, t1, t2);
3242 tcg_gen_qemu_st_i64(t2, addr, idx, memop);
3243
3244 tcg_gen_ext_i64(ret, (new_val ? t2 : t1), memop);
3245 tcg_temp_free_i64(t1);
3246 tcg_temp_free_i64(t2);
3247 }
3248
3249 static void do_atomic_op_i64(TCGv_i64 ret, TCGv addr, TCGv_i64 val,
3250 TCGArg idx, MemOp memop, void * const table[])
3251 {
3252 memop = tcg_canonicalize_memop(memop, 1, 0);
3253
3254 if ((memop & MO_SIZE) == MO_64) {
3255 #ifdef CONFIG_ATOMIC64
3256 gen_atomic_op_i64 gen;
3257
3258 gen = table[memop & (MO_SIZE | MO_BSWAP)];
3259 tcg_debug_assert(gen != NULL);
3260
3261 #ifdef CONFIG_SOFTMMU
3262 {
3263 TCGMemOpIdx oi = make_memop_idx(memop & ~MO_SIGN, idx);
3264 gen(ret, cpu_env, addr, val, tcg_constant_i32(oi));
3265 }
3266 #else
3267 gen(ret, cpu_env, addr, val);
3268 #endif
3269 #else
3270 gen_helper_exit_atomic(cpu_env);
3271 /* Produce a result, so that we have a well-formed opcode stream
3272 with respect to uses of the result in the (dead) code following. */
3273 tcg_gen_movi_i64(ret, 0);
3274 #endif /* CONFIG_ATOMIC64 */
3275 } else {
3276 TCGv_i32 v32 = tcg_temp_new_i32();
3277 TCGv_i32 r32 = tcg_temp_new_i32();
3278
3279 tcg_gen_extrl_i64_i32(v32, val);
3280 do_atomic_op_i32(r32, addr, v32, idx, memop & ~MO_SIGN, table);
3281 tcg_temp_free_i32(v32);
3282
3283 tcg_gen_extu_i32_i64(ret, r32);
3284 tcg_temp_free_i32(r32);
3285
3286 if (memop & MO_SIGN) {
3287 tcg_gen_ext_i64(ret, ret, memop);
3288 }
3289 }
3290 }
3291
3292 #define GEN_ATOMIC_HELPER(NAME, OP, NEW) \
3293 static void * const table_##NAME[16] = { \
3294 [MO_8] = gen_helper_atomic_##NAME##b, \
3295 [MO_16 | MO_LE] = gen_helper_atomic_##NAME##w_le, \
3296 [MO_16 | MO_BE] = gen_helper_atomic_##NAME##w_be, \
3297 [MO_32 | MO_LE] = gen_helper_atomic_##NAME##l_le, \
3298 [MO_32 | MO_BE] = gen_helper_atomic_##NAME##l_be, \
3299 WITH_ATOMIC64([MO_64 | MO_LE] = gen_helper_atomic_##NAME##q_le) \
3300 WITH_ATOMIC64([MO_64 | MO_BE] = gen_helper_atomic_##NAME##q_be) \
3301 }; \
3302 void tcg_gen_atomic_##NAME##_i32 \
3303 (TCGv_i32 ret, TCGv addr, TCGv_i32 val, TCGArg idx, MemOp memop) \
3304 { \
3305 if (tcg_ctx->tb_cflags & CF_PARALLEL) { \
3306 do_atomic_op_i32(ret, addr, val, idx, memop, table_##NAME); \
3307 } else { \
3308 do_nonatomic_op_i32(ret, addr, val, idx, memop, NEW, \
3309 tcg_gen_##OP##_i32); \
3310 } \
3311 } \
3312 void tcg_gen_atomic_##NAME##_i64 \
3313 (TCGv_i64 ret, TCGv addr, TCGv_i64 val, TCGArg idx, MemOp memop) \
3314 { \
3315 if (tcg_ctx->tb_cflags & CF_PARALLEL) { \
3316 do_atomic_op_i64(ret, addr, val, idx, memop, table_##NAME); \
3317 } else { \
3318 do_nonatomic_op_i64(ret, addr, val, idx, memop, NEW, \
3319 tcg_gen_##OP##_i64); \
3320 } \
3321 }
3322
3323 GEN_ATOMIC_HELPER(fetch_add, add, 0)
3324 GEN_ATOMIC_HELPER(fetch_and, and, 0)
3325 GEN_ATOMIC_HELPER(fetch_or, or, 0)
3326 GEN_ATOMIC_HELPER(fetch_xor, xor, 0)
3327 GEN_ATOMIC_HELPER(fetch_smin, smin, 0)
3328 GEN_ATOMIC_HELPER(fetch_umin, umin, 0)
3329 GEN_ATOMIC_HELPER(fetch_smax, smax, 0)
3330 GEN_ATOMIC_HELPER(fetch_umax, umax, 0)
3331
3332 GEN_ATOMIC_HELPER(add_fetch, add, 1)
3333 GEN_ATOMIC_HELPER(and_fetch, and, 1)
3334 GEN_ATOMIC_HELPER(or_fetch, or, 1)
3335 GEN_ATOMIC_HELPER(xor_fetch, xor, 1)
3336 GEN_ATOMIC_HELPER(smin_fetch, smin, 1)
3337 GEN_ATOMIC_HELPER(umin_fetch, umin, 1)
3338 GEN_ATOMIC_HELPER(smax_fetch, smax, 1)
3339 GEN_ATOMIC_HELPER(umax_fetch, umax, 1)
3340
3341 static void tcg_gen_mov2_i32(TCGv_i32 r, TCGv_i32 a, TCGv_i32 b)
3342 {
3343 tcg_gen_mov_i32(r, b);
3344 }
3345
3346 static void tcg_gen_mov2_i64(TCGv_i64 r, TCGv_i64 a, TCGv_i64 b)
3347 {
3348 tcg_gen_mov_i64(r, b);
3349 }
3350
3351 GEN_ATOMIC_HELPER(xchg, mov2, 0)
3352
3353 #undef GEN_ATOMIC_HELPER
QEmu