]> git.proxmox.com Git - qemu.git/blob - tcg/optimize.c
tcg: Do constant folding on double-word comparisons
[qemu.git] / tcg / optimize.c
1 /*
2 * Optimizations for Tiny Code Generator for QEMU
3 *
4 * Copyright (c) 2010 Samsung Electronics.
5 * Contributed by Kirill Batuzov <batuzovk@ispras.ru>
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a copy
8 * of this software and associated documentation files (the "Software"), to deal
9 * in the Software without restriction, including without limitation the rights
10 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
11 * copies of the Software, and to permit persons to whom the Software is
12 * furnished to do so, subject to the following conditions:
13 *
14 * The above copyright notice and this permission notice shall be included in
15 * all copies or substantial portions of the Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
23 * THE SOFTWARE.
24 */
25
26 #include "config.h"
27
28 #include <stdlib.h>
29 #include <stdio.h>
30
31 #include "qemu-common.h"
32 #include "tcg-op.h"
33
34 #define CASE_OP_32_64(x) \
35 glue(glue(case INDEX_op_, x), _i32): \
36 glue(glue(case INDEX_op_, x), _i64)
37
38 typedef enum {
39 TCG_TEMP_UNDEF = 0,
40 TCG_TEMP_CONST,
41 TCG_TEMP_COPY,
42 } tcg_temp_state;
43
44 struct tcg_temp_info {
45 tcg_temp_state state;
46 uint16_t prev_copy;
47 uint16_t next_copy;
48 tcg_target_ulong val;
49 };
50
51 static struct tcg_temp_info temps[TCG_MAX_TEMPS];
52
53 /* Reset TEMP's state to TCG_TEMP_UNDEF. If TEMP only had one copy, remove
54 the copy flag from the left temp. */
55 static void reset_temp(TCGArg temp)
56 {
57 if (temps[temp].state == TCG_TEMP_COPY) {
58 if (temps[temp].prev_copy == temps[temp].next_copy) {
59 temps[temps[temp].next_copy].state = TCG_TEMP_UNDEF;
60 } else {
61 temps[temps[temp].next_copy].prev_copy = temps[temp].prev_copy;
62 temps[temps[temp].prev_copy].next_copy = temps[temp].next_copy;
63 }
64 }
65 temps[temp].state = TCG_TEMP_UNDEF;
66 }
67
68 static int op_bits(TCGOpcode op)
69 {
70 const TCGOpDef *def = &tcg_op_defs[op];
71 return def->flags & TCG_OPF_64BIT ? 64 : 32;
72 }
73
74 static TCGOpcode op_to_movi(TCGOpcode op)
75 {
76 switch (op_bits(op)) {
77 case 32:
78 return INDEX_op_movi_i32;
79 case 64:
80 return INDEX_op_movi_i64;
81 default:
82 fprintf(stderr, "op_to_movi: unexpected return value of "
83 "function op_bits.\n");
84 tcg_abort();
85 }
86 }
87
88 static TCGArg find_better_copy(TCGContext *s, TCGArg temp)
89 {
90 TCGArg i;
91
92 /* If this is already a global, we can't do better. */
93 if (temp < s->nb_globals) {
94 return temp;
95 }
96
97 /* Search for a global first. */
98 for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) {
99 if (i < s->nb_globals) {
100 return i;
101 }
102 }
103
104 /* If it is a temp, search for a temp local. */
105 if (!s->temps[temp].temp_local) {
106 for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) {
107 if (s->temps[i].temp_local) {
108 return i;
109 }
110 }
111 }
112
113 /* Failure to find a better representation, return the same temp. */
114 return temp;
115 }
116
117 static bool temps_are_copies(TCGArg arg1, TCGArg arg2)
118 {
119 TCGArg i;
120
121 if (arg1 == arg2) {
122 return true;
123 }
124
125 if (temps[arg1].state != TCG_TEMP_COPY
126 || temps[arg2].state != TCG_TEMP_COPY) {
127 return false;
128 }
129
130 for (i = temps[arg1].next_copy ; i != arg1 ; i = temps[i].next_copy) {
131 if (i == arg2) {
132 return true;
133 }
134 }
135
136 return false;
137 }
138
139 static void tcg_opt_gen_mov(TCGContext *s, TCGArg *gen_args,
140 TCGArg dst, TCGArg src)
141 {
142 reset_temp(dst);
143 assert(temps[src].state != TCG_TEMP_CONST);
144
145 if (s->temps[src].type == s->temps[dst].type) {
146 if (temps[src].state != TCG_TEMP_COPY) {
147 temps[src].state = TCG_TEMP_COPY;
148 temps[src].next_copy = src;
149 temps[src].prev_copy = src;
150 }
151 temps[dst].state = TCG_TEMP_COPY;
152 temps[dst].next_copy = temps[src].next_copy;
153 temps[dst].prev_copy = src;
154 temps[temps[dst].next_copy].prev_copy = dst;
155 temps[src].next_copy = dst;
156 }
157
158 gen_args[0] = dst;
159 gen_args[1] = src;
160 }
161
162 static void tcg_opt_gen_movi(TCGArg *gen_args, TCGArg dst, TCGArg val)
163 {
164 reset_temp(dst);
165 temps[dst].state = TCG_TEMP_CONST;
166 temps[dst].val = val;
167 gen_args[0] = dst;
168 gen_args[1] = val;
169 }
170
171 static TCGOpcode op_to_mov(TCGOpcode op)
172 {
173 switch (op_bits(op)) {
174 case 32:
175 return INDEX_op_mov_i32;
176 case 64:
177 return INDEX_op_mov_i64;
178 default:
179 fprintf(stderr, "op_to_mov: unexpected return value of "
180 "function op_bits.\n");
181 tcg_abort();
182 }
183 }
184
185 static TCGArg do_constant_folding_2(TCGOpcode op, TCGArg x, TCGArg y)
186 {
187 switch (op) {
188 CASE_OP_32_64(add):
189 return x + y;
190
191 CASE_OP_32_64(sub):
192 return x - y;
193
194 CASE_OP_32_64(mul):
195 return x * y;
196
197 CASE_OP_32_64(and):
198 return x & y;
199
200 CASE_OP_32_64(or):
201 return x | y;
202
203 CASE_OP_32_64(xor):
204 return x ^ y;
205
206 case INDEX_op_shl_i32:
207 return (uint32_t)x << (uint32_t)y;
208
209 case INDEX_op_shl_i64:
210 return (uint64_t)x << (uint64_t)y;
211
212 case INDEX_op_shr_i32:
213 return (uint32_t)x >> (uint32_t)y;
214
215 case INDEX_op_shr_i64:
216 return (uint64_t)x >> (uint64_t)y;
217
218 case INDEX_op_sar_i32:
219 return (int32_t)x >> (int32_t)y;
220
221 case INDEX_op_sar_i64:
222 return (int64_t)x >> (int64_t)y;
223
224 case INDEX_op_rotr_i32:
225 x = ((uint32_t)x << (32 - y)) | ((uint32_t)x >> y);
226 return x;
227
228 case INDEX_op_rotr_i64:
229 x = ((uint64_t)x << (64 - y)) | ((uint64_t)x >> y);
230 return x;
231
232 case INDEX_op_rotl_i32:
233 x = ((uint32_t)x << y) | ((uint32_t)x >> (32 - y));
234 return x;
235
236 case INDEX_op_rotl_i64:
237 x = ((uint64_t)x << y) | ((uint64_t)x >> (64 - y));
238 return x;
239
240 CASE_OP_32_64(not):
241 return ~x;
242
243 CASE_OP_32_64(neg):
244 return -x;
245
246 CASE_OP_32_64(andc):
247 return x & ~y;
248
249 CASE_OP_32_64(orc):
250 return x | ~y;
251
252 CASE_OP_32_64(eqv):
253 return ~(x ^ y);
254
255 CASE_OP_32_64(nand):
256 return ~(x & y);
257
258 CASE_OP_32_64(nor):
259 return ~(x | y);
260
261 CASE_OP_32_64(ext8s):
262 return (int8_t)x;
263
264 CASE_OP_32_64(ext16s):
265 return (int16_t)x;
266
267 CASE_OP_32_64(ext8u):
268 return (uint8_t)x;
269
270 CASE_OP_32_64(ext16u):
271 return (uint16_t)x;
272
273 case INDEX_op_ext32s_i64:
274 return (int32_t)x;
275
276 case INDEX_op_ext32u_i64:
277 return (uint32_t)x;
278
279 default:
280 fprintf(stderr,
281 "Unrecognized operation %d in do_constant_folding.\n", op);
282 tcg_abort();
283 }
284 }
285
286 static TCGArg do_constant_folding(TCGOpcode op, TCGArg x, TCGArg y)
287 {
288 TCGArg res = do_constant_folding_2(op, x, y);
289 if (op_bits(op) == 32) {
290 res &= 0xffffffff;
291 }
292 return res;
293 }
294
295 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c)
296 {
297 switch (c) {
298 case TCG_COND_EQ:
299 return x == y;
300 case TCG_COND_NE:
301 return x != y;
302 case TCG_COND_LT:
303 return (int32_t)x < (int32_t)y;
304 case TCG_COND_GE:
305 return (int32_t)x >= (int32_t)y;
306 case TCG_COND_LE:
307 return (int32_t)x <= (int32_t)y;
308 case TCG_COND_GT:
309 return (int32_t)x > (int32_t)y;
310 case TCG_COND_LTU:
311 return x < y;
312 case TCG_COND_GEU:
313 return x >= y;
314 case TCG_COND_LEU:
315 return x <= y;
316 case TCG_COND_GTU:
317 return x > y;
318 default:
319 tcg_abort();
320 }
321 }
322
323 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c)
324 {
325 switch (c) {
326 case TCG_COND_EQ:
327 return x == y;
328 case TCG_COND_NE:
329 return x != y;
330 case TCG_COND_LT:
331 return (int64_t)x < (int64_t)y;
332 case TCG_COND_GE:
333 return (int64_t)x >= (int64_t)y;
334 case TCG_COND_LE:
335 return (int64_t)x <= (int64_t)y;
336 case TCG_COND_GT:
337 return (int64_t)x > (int64_t)y;
338 case TCG_COND_LTU:
339 return x < y;
340 case TCG_COND_GEU:
341 return x >= y;
342 case TCG_COND_LEU:
343 return x <= y;
344 case TCG_COND_GTU:
345 return x > y;
346 default:
347 tcg_abort();
348 }
349 }
350
351 static bool do_constant_folding_cond_eq(TCGCond c)
352 {
353 switch (c) {
354 case TCG_COND_GT:
355 case TCG_COND_LTU:
356 case TCG_COND_LT:
357 case TCG_COND_GTU:
358 case TCG_COND_NE:
359 return 0;
360 case TCG_COND_GE:
361 case TCG_COND_GEU:
362 case TCG_COND_LE:
363 case TCG_COND_LEU:
364 case TCG_COND_EQ:
365 return 1;
366 default:
367 tcg_abort();
368 }
369 }
370
371 /* Return 2 if the condition can't be simplified, and the result
372 of the condition (0 or 1) if it can */
373 static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x,
374 TCGArg y, TCGCond c)
375 {
376 if (temps[x].state == TCG_TEMP_CONST && temps[y].state == TCG_TEMP_CONST) {
377 switch (op_bits(op)) {
378 case 32:
379 return do_constant_folding_cond_32(temps[x].val, temps[y].val, c);
380 case 64:
381 return do_constant_folding_cond_64(temps[x].val, temps[y].val, c);
382 default:
383 tcg_abort();
384 }
385 } else if (temps_are_copies(x, y)) {
386 return do_constant_folding_cond_eq(c);
387 } else if (temps[y].state == TCG_TEMP_CONST && temps[y].val == 0) {
388 switch (c) {
389 case TCG_COND_LTU:
390 return 0;
391 case TCG_COND_GEU:
392 return 1;
393 default:
394 return 2;
395 }
396 } else {
397 return 2;
398 }
399 }
400
401 /* Return 2 if the condition can't be simplified, and the result
402 of the condition (0 or 1) if it can */
403 static TCGArg do_constant_folding_cond2(TCGArg *p1, TCGArg *p2, TCGCond c)
404 {
405 TCGArg al = p1[0], ah = p1[1];
406 TCGArg bl = p2[0], bh = p2[1];
407
408 if (temps[bl].state == TCG_TEMP_CONST
409 && temps[bh].state == TCG_TEMP_CONST) {
410 uint64_t b = ((uint64_t)temps[bh].val << 32) | (uint32_t)temps[bl].val;
411
412 if (temps[al].state == TCG_TEMP_CONST
413 && temps[ah].state == TCG_TEMP_CONST) {
414 uint64_t a;
415 a = ((uint64_t)temps[ah].val << 32) | (uint32_t)temps[al].val;
416 return do_constant_folding_cond_64(a, b, c);
417 }
418 if (b == 0) {
419 switch (c) {
420 case TCG_COND_LTU:
421 return 0;
422 case TCG_COND_GEU:
423 return 1;
424 default:
425 break;
426 }
427 }
428 }
429 if (temps_are_copies(al, bl) && temps_are_copies(ah, bh)) {
430 return do_constant_folding_cond_eq(c);
431 }
432 return 2;
433 }
434
435 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
436 {
437 TCGArg a1 = *p1, a2 = *p2;
438 int sum = 0;
439 sum += temps[a1].state == TCG_TEMP_CONST;
440 sum -= temps[a2].state == TCG_TEMP_CONST;
441
442 /* Prefer the constant in second argument, and then the form
443 op a, a, b, which is better handled on non-RISC hosts. */
444 if (sum > 0 || (sum == 0 && dest == a2)) {
445 *p1 = a2;
446 *p2 = a1;
447 return true;
448 }
449 return false;
450 }
451
452 static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
453 {
454 int sum = 0;
455 sum += temps[p1[0]].state == TCG_TEMP_CONST;
456 sum += temps[p1[1]].state == TCG_TEMP_CONST;
457 sum -= temps[p2[0]].state == TCG_TEMP_CONST;
458 sum -= temps[p2[1]].state == TCG_TEMP_CONST;
459 if (sum > 0) {
460 TCGArg t;
461 t = p1[0], p1[0] = p2[0], p2[0] = t;
462 t = p1[1], p1[1] = p2[1], p2[1] = t;
463 return true;
464 }
465 return false;
466 }
467
468 /* Propagate constants and copies, fold constant expressions. */
469 static TCGArg *tcg_constant_folding(TCGContext *s, uint16_t *tcg_opc_ptr,
470 TCGArg *args, TCGOpDef *tcg_op_defs)
471 {
472 int i, nb_ops, op_index, nb_temps, nb_globals, nb_call_args;
473 TCGOpcode op;
474 const TCGOpDef *def;
475 TCGArg *gen_args;
476 TCGArg tmp;
477
478 /* Array VALS has an element for each temp.
479 If this temp holds a constant then its value is kept in VALS' element.
480 If this temp is a copy of other ones then the other copies are
481 available through the doubly linked circular list. */
482
483 nb_temps = s->nb_temps;
484 nb_globals = s->nb_globals;
485 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
486
487 nb_ops = tcg_opc_ptr - gen_opc_buf;
488 gen_args = args;
489 for (op_index = 0; op_index < nb_ops; op_index++) {
490 op = gen_opc_buf[op_index];
491 def = &tcg_op_defs[op];
492 /* Do copy propagation */
493 if (op == INDEX_op_call) {
494 int nb_oargs = args[0] >> 16;
495 int nb_iargs = args[0] & 0xffff;
496 for (i = nb_oargs + 1; i < nb_oargs + nb_iargs + 1; i++) {
497 if (temps[args[i]].state == TCG_TEMP_COPY) {
498 args[i] = find_better_copy(s, args[i]);
499 }
500 }
501 } else {
502 for (i = def->nb_oargs; i < def->nb_oargs + def->nb_iargs; i++) {
503 if (temps[args[i]].state == TCG_TEMP_COPY) {
504 args[i] = find_better_copy(s, args[i]);
505 }
506 }
507 }
508
509 /* For commutative operations make constant second argument */
510 switch (op) {
511 CASE_OP_32_64(add):
512 CASE_OP_32_64(mul):
513 CASE_OP_32_64(and):
514 CASE_OP_32_64(or):
515 CASE_OP_32_64(xor):
516 CASE_OP_32_64(eqv):
517 CASE_OP_32_64(nand):
518 CASE_OP_32_64(nor):
519 swap_commutative(args[0], &args[1], &args[2]);
520 break;
521 CASE_OP_32_64(brcond):
522 if (swap_commutative(-1, &args[0], &args[1])) {
523 args[2] = tcg_swap_cond(args[2]);
524 }
525 break;
526 CASE_OP_32_64(setcond):
527 if (swap_commutative(args[0], &args[1], &args[2])) {
528 args[3] = tcg_swap_cond(args[3]);
529 }
530 break;
531 CASE_OP_32_64(movcond):
532 if (swap_commutative(-1, &args[1], &args[2])) {
533 args[5] = tcg_swap_cond(args[5]);
534 }
535 /* For movcond, we canonicalize the "false" input reg to match
536 the destination reg so that the tcg backend can implement
537 a "move if true" operation. */
538 if (swap_commutative(args[0], &args[4], &args[3])) {
539 args[5] = tcg_invert_cond(args[5]);
540 }
541 break;
542 case INDEX_op_add2_i32:
543 swap_commutative(args[0], &args[2], &args[4]);
544 swap_commutative(args[1], &args[3], &args[5]);
545 break;
546 case INDEX_op_brcond2_i32:
547 if (swap_commutative2(&args[0], &args[2])) {
548 args[4] = tcg_swap_cond(args[4]);
549 }
550 break;
551 case INDEX_op_setcond2_i32:
552 if (swap_commutative2(&args[1], &args[3])) {
553 args[5] = tcg_swap_cond(args[5]);
554 }
555 break;
556 default:
557 break;
558 }
559
560 /* Simplify expressions for "shift/rot r, 0, a => movi r, 0" */
561 switch (op) {
562 CASE_OP_32_64(shl):
563 CASE_OP_32_64(shr):
564 CASE_OP_32_64(sar):
565 CASE_OP_32_64(rotl):
566 CASE_OP_32_64(rotr):
567 if (temps[args[1]].state == TCG_TEMP_CONST
568 && temps[args[1]].val == 0) {
569 gen_opc_buf[op_index] = op_to_movi(op);
570 tcg_opt_gen_movi(gen_args, args[0], 0);
571 args += 3;
572 gen_args += 2;
573 continue;
574 }
575 break;
576 default:
577 break;
578 }
579
580 /* Simplify expression for "op r, a, 0 => mov r, a" cases */
581 switch (op) {
582 CASE_OP_32_64(add):
583 CASE_OP_32_64(sub):
584 CASE_OP_32_64(shl):
585 CASE_OP_32_64(shr):
586 CASE_OP_32_64(sar):
587 CASE_OP_32_64(rotl):
588 CASE_OP_32_64(rotr):
589 CASE_OP_32_64(or):
590 CASE_OP_32_64(xor):
591 if (temps[args[1]].state == TCG_TEMP_CONST) {
592 /* Proceed with possible constant folding. */
593 break;
594 }
595 if (temps[args[2]].state == TCG_TEMP_CONST
596 && temps[args[2]].val == 0) {
597 if (temps_are_copies(args[0], args[1])) {
598 gen_opc_buf[op_index] = INDEX_op_nop;
599 } else {
600 gen_opc_buf[op_index] = op_to_mov(op);
601 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
602 gen_args += 2;
603 }
604 args += 3;
605 continue;
606 }
607 break;
608 default:
609 break;
610 }
611
612 /* Simplify expression for "op r, a, 0 => movi r, 0" cases */
613 switch (op) {
614 CASE_OP_32_64(and):
615 CASE_OP_32_64(mul):
616 if ((temps[args[2]].state == TCG_TEMP_CONST
617 && temps[args[2]].val == 0)) {
618 gen_opc_buf[op_index] = op_to_movi(op);
619 tcg_opt_gen_movi(gen_args, args[0], 0);
620 args += 3;
621 gen_args += 2;
622 continue;
623 }
624 break;
625 default:
626 break;
627 }
628
629 /* Simplify expression for "op r, a, a => mov r, a" cases */
630 switch (op) {
631 CASE_OP_32_64(or):
632 CASE_OP_32_64(and):
633 if (temps_are_copies(args[1], args[2])) {
634 if (temps_are_copies(args[0], args[1])) {
635 gen_opc_buf[op_index] = INDEX_op_nop;
636 } else {
637 gen_opc_buf[op_index] = op_to_mov(op);
638 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
639 gen_args += 2;
640 }
641 args += 3;
642 continue;
643 }
644 break;
645 default:
646 break;
647 }
648
649 /* Simplify expression for "op r, a, a => movi r, 0" cases */
650 switch (op) {
651 CASE_OP_32_64(sub):
652 CASE_OP_32_64(xor):
653 if (temps_are_copies(args[1], args[2])) {
654 gen_opc_buf[op_index] = op_to_movi(op);
655 tcg_opt_gen_movi(gen_args, args[0], 0);
656 gen_args += 2;
657 args += 3;
658 continue;
659 }
660 break;
661 default:
662 break;
663 }
664
665 /* Propagate constants through copy operations and do constant
666 folding. Constants will be substituted to arguments by register
667 allocator where needed and possible. Also detect copies. */
668 switch (op) {
669 CASE_OP_32_64(mov):
670 if (temps_are_copies(args[0], args[1])) {
671 args += 2;
672 gen_opc_buf[op_index] = INDEX_op_nop;
673 break;
674 }
675 if (temps[args[1]].state != TCG_TEMP_CONST) {
676 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
677 gen_args += 2;
678 args += 2;
679 break;
680 }
681 /* Source argument is constant. Rewrite the operation and
682 let movi case handle it. */
683 op = op_to_movi(op);
684 gen_opc_buf[op_index] = op;
685 args[1] = temps[args[1]].val;
686 /* fallthrough */
687 CASE_OP_32_64(movi):
688 tcg_opt_gen_movi(gen_args, args[0], args[1]);
689 gen_args += 2;
690 args += 2;
691 break;
692
693 CASE_OP_32_64(not):
694 CASE_OP_32_64(neg):
695 CASE_OP_32_64(ext8s):
696 CASE_OP_32_64(ext8u):
697 CASE_OP_32_64(ext16s):
698 CASE_OP_32_64(ext16u):
699 case INDEX_op_ext32s_i64:
700 case INDEX_op_ext32u_i64:
701 if (temps[args[1]].state == TCG_TEMP_CONST) {
702 gen_opc_buf[op_index] = op_to_movi(op);
703 tmp = do_constant_folding(op, temps[args[1]].val, 0);
704 tcg_opt_gen_movi(gen_args, args[0], tmp);
705 gen_args += 2;
706 args += 2;
707 break;
708 }
709 goto do_default;
710
711 CASE_OP_32_64(add):
712 CASE_OP_32_64(sub):
713 CASE_OP_32_64(mul):
714 CASE_OP_32_64(or):
715 CASE_OP_32_64(and):
716 CASE_OP_32_64(xor):
717 CASE_OP_32_64(shl):
718 CASE_OP_32_64(shr):
719 CASE_OP_32_64(sar):
720 CASE_OP_32_64(rotl):
721 CASE_OP_32_64(rotr):
722 CASE_OP_32_64(andc):
723 CASE_OP_32_64(orc):
724 CASE_OP_32_64(eqv):
725 CASE_OP_32_64(nand):
726 CASE_OP_32_64(nor):
727 if (temps[args[1]].state == TCG_TEMP_CONST
728 && temps[args[2]].state == TCG_TEMP_CONST) {
729 gen_opc_buf[op_index] = op_to_movi(op);
730 tmp = do_constant_folding(op, temps[args[1]].val,
731 temps[args[2]].val);
732 tcg_opt_gen_movi(gen_args, args[0], tmp);
733 gen_args += 2;
734 args += 3;
735 break;
736 }
737 goto do_default;
738
739 CASE_OP_32_64(deposit):
740 if (temps[args[1]].state == TCG_TEMP_CONST
741 && temps[args[2]].state == TCG_TEMP_CONST) {
742 gen_opc_buf[op_index] = op_to_movi(op);
743 tmp = ((1ull << args[4]) - 1);
744 tmp = (temps[args[1]].val & ~(tmp << args[3]))
745 | ((temps[args[2]].val & tmp) << args[3]);
746 tcg_opt_gen_movi(gen_args, args[0], tmp);
747 gen_args += 2;
748 args += 5;
749 break;
750 }
751 goto do_default;
752
753 CASE_OP_32_64(setcond):
754 tmp = do_constant_folding_cond(op, args[1], args[2], args[3]);
755 if (tmp != 2) {
756 gen_opc_buf[op_index] = op_to_movi(op);
757 tcg_opt_gen_movi(gen_args, args[0], tmp);
758 gen_args += 2;
759 args += 4;
760 break;
761 }
762 goto do_default;
763
764 CASE_OP_32_64(brcond):
765 tmp = do_constant_folding_cond(op, args[0], args[1], args[2]);
766 if (tmp != 2) {
767 if (tmp) {
768 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
769 gen_opc_buf[op_index] = INDEX_op_br;
770 gen_args[0] = args[3];
771 gen_args += 1;
772 } else {
773 gen_opc_buf[op_index] = INDEX_op_nop;
774 }
775 args += 4;
776 break;
777 }
778 goto do_default;
779
780 CASE_OP_32_64(movcond):
781 tmp = do_constant_folding_cond(op, args[1], args[2], args[5]);
782 if (tmp != 2) {
783 if (temps_are_copies(args[0], args[4-tmp])) {
784 gen_opc_buf[op_index] = INDEX_op_nop;
785 } else if (temps[args[4-tmp]].state == TCG_TEMP_CONST) {
786 gen_opc_buf[op_index] = op_to_movi(op);
787 tcg_opt_gen_movi(gen_args, args[0], temps[args[4-tmp]].val);
788 gen_args += 2;
789 } else {
790 gen_opc_buf[op_index] = op_to_mov(op);
791 tcg_opt_gen_mov(s, gen_args, args[0], args[4-tmp]);
792 gen_args += 2;
793 }
794 args += 6;
795 break;
796 }
797 goto do_default;
798
799 case INDEX_op_brcond2_i32:
800 tmp = do_constant_folding_cond2(&args[0], &args[2], args[4]);
801 if (tmp != 2) {
802 if (tmp) {
803 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
804 gen_opc_buf[op_index] = INDEX_op_br;
805 gen_args[0] = args[5];
806 gen_args += 1;
807 } else {
808 gen_opc_buf[op_index] = INDEX_op_nop;
809 }
810 } else if ((args[4] == TCG_COND_LT || args[4] == TCG_COND_GE)
811 && temps[args[2]].state == TCG_TEMP_CONST
812 && temps[args[3]].state == TCG_TEMP_CONST
813 && temps[args[2]].val == 0
814 && temps[args[3]].val == 0) {
815 /* Simplify LT/GE comparisons vs zero to a single compare
816 vs the high word of the input. */
817 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
818 gen_opc_buf[op_index] = INDEX_op_brcond_i32;
819 gen_args[0] = args[1];
820 gen_args[1] = args[3];
821 gen_args[2] = args[4];
822 gen_args[3] = args[5];
823 gen_args += 4;
824 } else {
825 goto do_default;
826 }
827 args += 6;
828 break;
829
830 case INDEX_op_setcond2_i32:
831 tmp = do_constant_folding_cond2(&args[1], &args[3], args[5]);
832 if (tmp != 2) {
833 gen_opc_buf[op_index] = INDEX_op_movi_i32;
834 tcg_opt_gen_movi(gen_args, args[0], tmp);
835 gen_args += 2;
836 } else if ((args[5] == TCG_COND_LT || args[5] == TCG_COND_GE)
837 && temps[args[3]].state == TCG_TEMP_CONST
838 && temps[args[4]].state == TCG_TEMP_CONST
839 && temps[args[3]].val == 0
840 && temps[args[4]].val == 0) {
841 /* Simplify LT/GE comparisons vs zero to a single compare
842 vs the high word of the input. */
843 gen_opc_buf[op_index] = INDEX_op_setcond_i32;
844 gen_args[0] = args[0];
845 gen_args[1] = args[2];
846 gen_args[2] = args[4];
847 gen_args[3] = args[5];
848 gen_args += 4;
849 } else {
850 goto do_default;
851 }
852 args += 6;
853 break;
854
855 case INDEX_op_call:
856 nb_call_args = (args[0] >> 16) + (args[0] & 0xffff);
857 if (!(args[nb_call_args + 1] & (TCG_CALL_CONST | TCG_CALL_PURE))) {
858 for (i = 0; i < nb_globals; i++) {
859 reset_temp(i);
860 }
861 }
862 for (i = 0; i < (args[0] >> 16); i++) {
863 reset_temp(args[i + 1]);
864 }
865 i = nb_call_args + 3;
866 while (i) {
867 *gen_args = *args;
868 args++;
869 gen_args++;
870 i--;
871 }
872 break;
873
874 default:
875 do_default:
876 /* Default case: we know nothing about operation (or were unable
877 to compute the operation result) so no propagation is done.
878 We trash everything if the operation is the end of a basic
879 block, otherwise we only trash the output args. */
880 if (def->flags & TCG_OPF_BB_END) {
881 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
882 } else {
883 for (i = 0; i < def->nb_oargs; i++) {
884 reset_temp(args[i]);
885 }
886 }
887 for (i = 0; i < def->nb_args; i++) {
888 gen_args[i] = args[i];
889 }
890 args += def->nb_args;
891 gen_args += def->nb_args;
892 break;
893 }
894 }
895
896 return gen_args;
897 }
898
899 TCGArg *tcg_optimize(TCGContext *s, uint16_t *tcg_opc_ptr,
900 TCGArg *args, TCGOpDef *tcg_op_defs)
901 {
902 TCGArg *res;
903 res = tcg_constant_folding(s, tcg_opc_ptr, args, tcg_op_defs);
904 return res;
905 }