]> git.proxmox.com Git - mirror_qemu.git/blob - tcg/optimize.c
tcg: Optimize double-word comparisons against zero
[mirror_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 /* Return 2 if the condition can't be simplified, and the result
296 of the condition (0 or 1) if it can */
297 static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x,
298 TCGArg y, TCGCond c)
299 {
300 if (temps[x].state == TCG_TEMP_CONST && temps[y].state == TCG_TEMP_CONST) {
301 switch (op_bits(op)) {
302 case 32:
303 switch (c) {
304 case TCG_COND_EQ:
305 return (uint32_t)temps[x].val == (uint32_t)temps[y].val;
306 case TCG_COND_NE:
307 return (uint32_t)temps[x].val != (uint32_t)temps[y].val;
308 case TCG_COND_LT:
309 return (int32_t)temps[x].val < (int32_t)temps[y].val;
310 case TCG_COND_GE:
311 return (int32_t)temps[x].val >= (int32_t)temps[y].val;
312 case TCG_COND_LE:
313 return (int32_t)temps[x].val <= (int32_t)temps[y].val;
314 case TCG_COND_GT:
315 return (int32_t)temps[x].val > (int32_t)temps[y].val;
316 case TCG_COND_LTU:
317 return (uint32_t)temps[x].val < (uint32_t)temps[y].val;
318 case TCG_COND_GEU:
319 return (uint32_t)temps[x].val >= (uint32_t)temps[y].val;
320 case TCG_COND_LEU:
321 return (uint32_t)temps[x].val <= (uint32_t)temps[y].val;
322 case TCG_COND_GTU:
323 return (uint32_t)temps[x].val > (uint32_t)temps[y].val;
324 default:
325 break;
326 }
327 break;
328 case 64:
329 switch (c) {
330 case TCG_COND_EQ:
331 return (uint64_t)temps[x].val == (uint64_t)temps[y].val;
332 case TCG_COND_NE:
333 return (uint64_t)temps[x].val != (uint64_t)temps[y].val;
334 case TCG_COND_LT:
335 return (int64_t)temps[x].val < (int64_t)temps[y].val;
336 case TCG_COND_GE:
337 return (int64_t)temps[x].val >= (int64_t)temps[y].val;
338 case TCG_COND_LE:
339 return (int64_t)temps[x].val <= (int64_t)temps[y].val;
340 case TCG_COND_GT:
341 return (int64_t)temps[x].val > (int64_t)temps[y].val;
342 case TCG_COND_LTU:
343 return (uint64_t)temps[x].val < (uint64_t)temps[y].val;
344 case TCG_COND_GEU:
345 return (uint64_t)temps[x].val >= (uint64_t)temps[y].val;
346 case TCG_COND_LEU:
347 return (uint64_t)temps[x].val <= (uint64_t)temps[y].val;
348 case TCG_COND_GTU:
349 return (uint64_t)temps[x].val > (uint64_t)temps[y].val;
350 default:
351 break;
352 }
353 break;
354 }
355 } else if (temps_are_copies(x, y)) {
356 switch (c) {
357 case TCG_COND_GT:
358 case TCG_COND_LTU:
359 case TCG_COND_LT:
360 case TCG_COND_GTU:
361 case TCG_COND_NE:
362 return 0;
363 case TCG_COND_GE:
364 case TCG_COND_GEU:
365 case TCG_COND_LE:
366 case TCG_COND_LEU:
367 case TCG_COND_EQ:
368 return 1;
369 default:
370 break;
371 }
372 } else if (temps[y].state == TCG_TEMP_CONST && temps[y].val == 0) {
373 switch (c) {
374 case TCG_COND_LTU:
375 return 0;
376 case TCG_COND_GEU:
377 return 1;
378 default:
379 return 2;
380 }
381 } else {
382 return 2;
383 }
384
385 fprintf(stderr,
386 "Unrecognized bitness %d or condition %d in "
387 "do_constant_folding_cond.\n", op_bits(op), c);
388 tcg_abort();
389 }
390
391 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
392 {
393 TCGArg a1 = *p1, a2 = *p2;
394 int sum = 0;
395 sum += temps[a1].state == TCG_TEMP_CONST;
396 sum -= temps[a2].state == TCG_TEMP_CONST;
397
398 /* Prefer the constant in second argument, and then the form
399 op a, a, b, which is better handled on non-RISC hosts. */
400 if (sum > 0 || (sum == 0 && dest == a2)) {
401 *p1 = a2;
402 *p2 = a1;
403 return true;
404 }
405 return false;
406 }
407
408 static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
409 {
410 int sum = 0;
411 sum += temps[p1[0]].state == TCG_TEMP_CONST;
412 sum += temps[p1[1]].state == TCG_TEMP_CONST;
413 sum -= temps[p2[0]].state == TCG_TEMP_CONST;
414 sum -= temps[p2[1]].state == TCG_TEMP_CONST;
415 if (sum > 0) {
416 TCGArg t;
417 t = p1[0], p1[0] = p2[0], p2[0] = t;
418 t = p1[1], p1[1] = p2[1], p2[1] = t;
419 return true;
420 }
421 return false;
422 }
423
424 /* Propagate constants and copies, fold constant expressions. */
425 static TCGArg *tcg_constant_folding(TCGContext *s, uint16_t *tcg_opc_ptr,
426 TCGArg *args, TCGOpDef *tcg_op_defs)
427 {
428 int i, nb_ops, op_index, nb_temps, nb_globals, nb_call_args;
429 TCGOpcode op;
430 const TCGOpDef *def;
431 TCGArg *gen_args;
432 TCGArg tmp;
433
434 /* Array VALS has an element for each temp.
435 If this temp holds a constant then its value is kept in VALS' element.
436 If this temp is a copy of other ones then the other copies are
437 available through the doubly linked circular list. */
438
439 nb_temps = s->nb_temps;
440 nb_globals = s->nb_globals;
441 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
442
443 nb_ops = tcg_opc_ptr - gen_opc_buf;
444 gen_args = args;
445 for (op_index = 0; op_index < nb_ops; op_index++) {
446 op = gen_opc_buf[op_index];
447 def = &tcg_op_defs[op];
448 /* Do copy propagation */
449 if (op == INDEX_op_call) {
450 int nb_oargs = args[0] >> 16;
451 int nb_iargs = args[0] & 0xffff;
452 for (i = nb_oargs + 1; i < nb_oargs + nb_iargs + 1; i++) {
453 if (temps[args[i]].state == TCG_TEMP_COPY) {
454 args[i] = find_better_copy(s, args[i]);
455 }
456 }
457 } else {
458 for (i = def->nb_oargs; i < def->nb_oargs + def->nb_iargs; i++) {
459 if (temps[args[i]].state == TCG_TEMP_COPY) {
460 args[i] = find_better_copy(s, args[i]);
461 }
462 }
463 }
464
465 /* For commutative operations make constant second argument */
466 switch (op) {
467 CASE_OP_32_64(add):
468 CASE_OP_32_64(mul):
469 CASE_OP_32_64(and):
470 CASE_OP_32_64(or):
471 CASE_OP_32_64(xor):
472 CASE_OP_32_64(eqv):
473 CASE_OP_32_64(nand):
474 CASE_OP_32_64(nor):
475 swap_commutative(args[0], &args[1], &args[2]);
476 break;
477 CASE_OP_32_64(brcond):
478 if (swap_commutative(-1, &args[0], &args[1])) {
479 args[2] = tcg_swap_cond(args[2]);
480 }
481 break;
482 CASE_OP_32_64(setcond):
483 if (swap_commutative(args[0], &args[1], &args[2])) {
484 args[3] = tcg_swap_cond(args[3]);
485 }
486 break;
487 CASE_OP_32_64(movcond):
488 if (swap_commutative(-1, &args[1], &args[2])) {
489 args[5] = tcg_swap_cond(args[5]);
490 }
491 /* For movcond, we canonicalize the "false" input reg to match
492 the destination reg so that the tcg backend can implement
493 a "move if true" operation. */
494 if (swap_commutative(args[0], &args[4], &args[3])) {
495 args[5] = tcg_invert_cond(args[5]);
496 }
497 break;
498 case INDEX_op_add2_i32:
499 swap_commutative(args[0], &args[2], &args[4]);
500 swap_commutative(args[1], &args[3], &args[5]);
501 break;
502 case INDEX_op_brcond2_i32:
503 if (swap_commutative2(&args[0], &args[2])) {
504 args[4] = tcg_swap_cond(args[4]);
505 }
506 break;
507 case INDEX_op_setcond2_i32:
508 if (swap_commutative2(&args[1], &args[3])) {
509 args[5] = tcg_swap_cond(args[5]);
510 }
511 break;
512 default:
513 break;
514 }
515
516 /* Simplify expressions for "shift/rot r, 0, a => movi r, 0" */
517 switch (op) {
518 CASE_OP_32_64(shl):
519 CASE_OP_32_64(shr):
520 CASE_OP_32_64(sar):
521 CASE_OP_32_64(rotl):
522 CASE_OP_32_64(rotr):
523 if (temps[args[1]].state == TCG_TEMP_CONST
524 && temps[args[1]].val == 0) {
525 gen_opc_buf[op_index] = op_to_movi(op);
526 tcg_opt_gen_movi(gen_args, args[0], 0);
527 args += 3;
528 gen_args += 2;
529 continue;
530 }
531 break;
532 default:
533 break;
534 }
535
536 /* Simplify expression for "op r, a, 0 => mov r, a" cases */
537 switch (op) {
538 CASE_OP_32_64(add):
539 CASE_OP_32_64(sub):
540 CASE_OP_32_64(shl):
541 CASE_OP_32_64(shr):
542 CASE_OP_32_64(sar):
543 CASE_OP_32_64(rotl):
544 CASE_OP_32_64(rotr):
545 CASE_OP_32_64(or):
546 CASE_OP_32_64(xor):
547 if (temps[args[1]].state == TCG_TEMP_CONST) {
548 /* Proceed with possible constant folding. */
549 break;
550 }
551 if (temps[args[2]].state == TCG_TEMP_CONST
552 && temps[args[2]].val == 0) {
553 if (temps_are_copies(args[0], args[1])) {
554 gen_opc_buf[op_index] = INDEX_op_nop;
555 } else {
556 gen_opc_buf[op_index] = op_to_mov(op);
557 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
558 gen_args += 2;
559 }
560 args += 3;
561 continue;
562 }
563 break;
564 default:
565 break;
566 }
567
568 /* Simplify expression for "op r, a, 0 => movi r, 0" cases */
569 switch (op) {
570 CASE_OP_32_64(and):
571 CASE_OP_32_64(mul):
572 if ((temps[args[2]].state == TCG_TEMP_CONST
573 && temps[args[2]].val == 0)) {
574 gen_opc_buf[op_index] = op_to_movi(op);
575 tcg_opt_gen_movi(gen_args, args[0], 0);
576 args += 3;
577 gen_args += 2;
578 continue;
579 }
580 break;
581 default:
582 break;
583 }
584
585 /* Simplify expression for "op r, a, a => mov r, a" cases */
586 switch (op) {
587 CASE_OP_32_64(or):
588 CASE_OP_32_64(and):
589 if (temps_are_copies(args[1], args[2])) {
590 if (temps_are_copies(args[0], args[1])) {
591 gen_opc_buf[op_index] = INDEX_op_nop;
592 } else {
593 gen_opc_buf[op_index] = op_to_mov(op);
594 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
595 gen_args += 2;
596 }
597 args += 3;
598 continue;
599 }
600 break;
601 default:
602 break;
603 }
604
605 /* Simplify expression for "op r, a, a => movi r, 0" cases */
606 switch (op) {
607 CASE_OP_32_64(sub):
608 CASE_OP_32_64(xor):
609 if (temps_are_copies(args[1], args[2])) {
610 gen_opc_buf[op_index] = op_to_movi(op);
611 tcg_opt_gen_movi(gen_args, args[0], 0);
612 gen_args += 2;
613 args += 3;
614 continue;
615 }
616 break;
617 default:
618 break;
619 }
620
621 /* Propagate constants through copy operations and do constant
622 folding. Constants will be substituted to arguments by register
623 allocator where needed and possible. Also detect copies. */
624 switch (op) {
625 CASE_OP_32_64(mov):
626 if (temps_are_copies(args[0], args[1])) {
627 args += 2;
628 gen_opc_buf[op_index] = INDEX_op_nop;
629 break;
630 }
631 if (temps[args[1]].state != TCG_TEMP_CONST) {
632 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
633 gen_args += 2;
634 args += 2;
635 break;
636 }
637 /* Source argument is constant. Rewrite the operation and
638 let movi case handle it. */
639 op = op_to_movi(op);
640 gen_opc_buf[op_index] = op;
641 args[1] = temps[args[1]].val;
642 /* fallthrough */
643 CASE_OP_32_64(movi):
644 tcg_opt_gen_movi(gen_args, args[0], args[1]);
645 gen_args += 2;
646 args += 2;
647 break;
648
649 CASE_OP_32_64(not):
650 CASE_OP_32_64(neg):
651 CASE_OP_32_64(ext8s):
652 CASE_OP_32_64(ext8u):
653 CASE_OP_32_64(ext16s):
654 CASE_OP_32_64(ext16u):
655 case INDEX_op_ext32s_i64:
656 case INDEX_op_ext32u_i64:
657 if (temps[args[1]].state == TCG_TEMP_CONST) {
658 gen_opc_buf[op_index] = op_to_movi(op);
659 tmp = do_constant_folding(op, temps[args[1]].val, 0);
660 tcg_opt_gen_movi(gen_args, args[0], tmp);
661 gen_args += 2;
662 args += 2;
663 break;
664 }
665 goto do_default;
666
667 CASE_OP_32_64(add):
668 CASE_OP_32_64(sub):
669 CASE_OP_32_64(mul):
670 CASE_OP_32_64(or):
671 CASE_OP_32_64(and):
672 CASE_OP_32_64(xor):
673 CASE_OP_32_64(shl):
674 CASE_OP_32_64(shr):
675 CASE_OP_32_64(sar):
676 CASE_OP_32_64(rotl):
677 CASE_OP_32_64(rotr):
678 CASE_OP_32_64(andc):
679 CASE_OP_32_64(orc):
680 CASE_OP_32_64(eqv):
681 CASE_OP_32_64(nand):
682 CASE_OP_32_64(nor):
683 if (temps[args[1]].state == TCG_TEMP_CONST
684 && temps[args[2]].state == TCG_TEMP_CONST) {
685 gen_opc_buf[op_index] = op_to_movi(op);
686 tmp = do_constant_folding(op, temps[args[1]].val,
687 temps[args[2]].val);
688 tcg_opt_gen_movi(gen_args, args[0], tmp);
689 gen_args += 2;
690 args += 3;
691 break;
692 }
693 goto do_default;
694
695 CASE_OP_32_64(deposit):
696 if (temps[args[1]].state == TCG_TEMP_CONST
697 && temps[args[2]].state == TCG_TEMP_CONST) {
698 gen_opc_buf[op_index] = op_to_movi(op);
699 tmp = ((1ull << args[4]) - 1);
700 tmp = (temps[args[1]].val & ~(tmp << args[3]))
701 | ((temps[args[2]].val & tmp) << args[3]);
702 tcg_opt_gen_movi(gen_args, args[0], tmp);
703 gen_args += 2;
704 args += 5;
705 break;
706 }
707 goto do_default;
708
709 CASE_OP_32_64(setcond):
710 tmp = do_constant_folding_cond(op, args[1], args[2], args[3]);
711 if (tmp != 2) {
712 gen_opc_buf[op_index] = op_to_movi(op);
713 tcg_opt_gen_movi(gen_args, args[0], tmp);
714 gen_args += 2;
715 args += 4;
716 break;
717 }
718 goto do_default;
719
720 CASE_OP_32_64(brcond):
721 tmp = do_constant_folding_cond(op, args[0], args[1], args[2]);
722 if (tmp != 2) {
723 if (tmp) {
724 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
725 gen_opc_buf[op_index] = INDEX_op_br;
726 gen_args[0] = args[3];
727 gen_args += 1;
728 } else {
729 gen_opc_buf[op_index] = INDEX_op_nop;
730 }
731 args += 4;
732 break;
733 }
734 goto do_default;
735
736 CASE_OP_32_64(movcond):
737 tmp = do_constant_folding_cond(op, args[1], args[2], args[5]);
738 if (tmp != 2) {
739 if (temps_are_copies(args[0], args[4-tmp])) {
740 gen_opc_buf[op_index] = INDEX_op_nop;
741 } else if (temps[args[4-tmp]].state == TCG_TEMP_CONST) {
742 gen_opc_buf[op_index] = op_to_movi(op);
743 tcg_opt_gen_movi(gen_args, args[0], temps[args[4-tmp]].val);
744 gen_args += 2;
745 } else {
746 gen_opc_buf[op_index] = op_to_mov(op);
747 tcg_opt_gen_mov(s, gen_args, args[0], args[4-tmp]);
748 gen_args += 2;
749 }
750 args += 6;
751 break;
752 }
753 goto do_default;
754
755 case INDEX_op_brcond2_i32:
756 /* Simplify LT/GE comparisons vs zero to a single compare
757 vs the high word of the input. */
758 if ((args[4] == TCG_COND_LT || args[4] == TCG_COND_GE)
759 && temps[args[2]].state == TCG_TEMP_CONST
760 && temps[args[3]].state == TCG_TEMP_CONST
761 && temps[args[2]].val == 0
762 && temps[args[3]].val == 0) {
763 gen_opc_buf[op_index] = INDEX_op_brcond_i32;
764 gen_args[0] = args[1];
765 gen_args[1] = args[3];
766 gen_args[2] = args[4];
767 gen_args[3] = args[5];
768 gen_args += 4;
769 args += 6;
770 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
771 break;
772 }
773 goto do_default;
774
775 case INDEX_op_setcond2_i32:
776 /* Simplify LT/GE comparisons vs zero to a single compare
777 vs the high word of the input. */
778 if ((args[5] == TCG_COND_LT || args[5] == TCG_COND_GE)
779 && temps[args[3]].state == TCG_TEMP_CONST
780 && temps[args[4]].state == TCG_TEMP_CONST
781 && temps[args[3]].val == 0
782 && temps[args[4]].val == 0) {
783 gen_opc_buf[op_index] = INDEX_op_setcond_i32;
784 gen_args[0] = args[0];
785 gen_args[1] = args[2];
786 gen_args[2] = args[4];
787 gen_args[3] = args[5];
788 gen_args += 4;
789 args += 6;
790 break;
791 }
792 goto do_default;
793
794 case INDEX_op_call:
795 nb_call_args = (args[0] >> 16) + (args[0] & 0xffff);
796 if (!(args[nb_call_args + 1] & (TCG_CALL_CONST | TCG_CALL_PURE))) {
797 for (i = 0; i < nb_globals; i++) {
798 reset_temp(i);
799 }
800 }
801 for (i = 0; i < (args[0] >> 16); i++) {
802 reset_temp(args[i + 1]);
803 }
804 i = nb_call_args + 3;
805 while (i) {
806 *gen_args = *args;
807 args++;
808 gen_args++;
809 i--;
810 }
811 break;
812
813 default:
814 do_default:
815 /* Default case: we know nothing about operation (or were unable
816 to compute the operation result) so no propagation is done.
817 We trash everything if the operation is the end of a basic
818 block, otherwise we only trash the output args. */
819 if (def->flags & TCG_OPF_BB_END) {
820 memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
821 } else {
822 for (i = 0; i < def->nb_oargs; i++) {
823 reset_temp(args[i]);
824 }
825 }
826 for (i = 0; i < def->nb_args; i++) {
827 gen_args[i] = args[i];
828 }
829 args += def->nb_args;
830 gen_args += def->nb_args;
831 break;
832 }
833 }
834
835 return gen_args;
836 }
837
838 TCGArg *tcg_optimize(TCGContext *s, uint16_t *tcg_opc_ptr,
839 TCGArg *args, TCGOpDef *tcg_op_defs)
840 {
841 TCGArg *res;
842 res = tcg_constant_folding(s, tcg_opc_ptr, args, tcg_op_defs);
843 return res;
844 }