block migration: fix return value
[qemu/cris-port.git] / tcg / optimize.c
blob34ae3c285752401b7de8633c6c2ba1fa400d3015
1 /*
2 * Optimizations for Tiny Code Generator for QEMU
4 * Copyright (c) 2010 Samsung Electronics.
5 * Contributed by Kirill Batuzov <batuzovk@ispras.ru>
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:
14 * The above copyright notice and this permission notice shall be included in
15 * all copies or substantial portions of the Software.
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.
26 #include "config.h"
28 #include <stdlib.h>
29 #include <stdio.h>
31 #include "qemu-common.h"
32 #include "tcg-op.h"
34 #define CASE_OP_32_64(x) \
35 glue(glue(case INDEX_op_, x), _i32): \
36 glue(glue(case INDEX_op_, x), _i64)
38 typedef enum {
39 TCG_TEMP_UNDEF = 0,
40 TCG_TEMP_CONST,
41 TCG_TEMP_COPY,
42 } tcg_temp_state;
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 tcg_target_ulong mask;
52 static struct tcg_temp_info temps[TCG_MAX_TEMPS];
54 /* Reset TEMP's state to TCG_TEMP_UNDEF. If TEMP only had one copy, remove
55 the copy flag from the left temp. */
56 static void reset_temp(TCGArg temp)
58 if (temps[temp].state == TCG_TEMP_COPY) {
59 if (temps[temp].prev_copy == temps[temp].next_copy) {
60 temps[temps[temp].next_copy].state = TCG_TEMP_UNDEF;
61 } else {
62 temps[temps[temp].next_copy].prev_copy = temps[temp].prev_copy;
63 temps[temps[temp].prev_copy].next_copy = temps[temp].next_copy;
66 temps[temp].state = TCG_TEMP_UNDEF;
67 temps[temp].mask = -1;
70 /* Reset all temporaries, given that there are NB_TEMPS of them. */
71 static void reset_all_temps(int nb_temps)
73 int i;
74 for (i = 0; i < nb_temps; i++) {
75 temps[i].state = TCG_TEMP_UNDEF;
76 temps[i].mask = -1;
80 static int op_bits(TCGOpcode op)
82 const TCGOpDef *def = &tcg_op_defs[op];
83 return def->flags & TCG_OPF_64BIT ? 64 : 32;
86 static TCGOpcode op_to_mov(TCGOpcode op)
88 switch (op_bits(op)) {
89 case 32:
90 return INDEX_op_mov_i32;
91 case 64:
92 return INDEX_op_mov_i64;
93 default:
94 fprintf(stderr, "op_to_mov: unexpected return value of "
95 "function op_bits.\n");
96 tcg_abort();
100 static TCGOpcode op_to_movi(TCGOpcode op)
102 switch (op_bits(op)) {
103 case 32:
104 return INDEX_op_movi_i32;
105 case 64:
106 return INDEX_op_movi_i64;
107 default:
108 fprintf(stderr, "op_to_movi: unexpected return value of "
109 "function op_bits.\n");
110 tcg_abort();
114 static TCGArg find_better_copy(TCGContext *s, TCGArg temp)
116 TCGArg i;
118 /* If this is already a global, we can't do better. */
119 if (temp < s->nb_globals) {
120 return temp;
123 /* Search for a global first. */
124 for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) {
125 if (i < s->nb_globals) {
126 return i;
130 /* If it is a temp, search for a temp local. */
131 if (!s->temps[temp].temp_local) {
132 for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) {
133 if (s->temps[i].temp_local) {
134 return i;
139 /* Failure to find a better representation, return the same temp. */
140 return temp;
143 static bool temps_are_copies(TCGArg arg1, TCGArg arg2)
145 TCGArg i;
147 if (arg1 == arg2) {
148 return true;
151 if (temps[arg1].state != TCG_TEMP_COPY
152 || temps[arg2].state != TCG_TEMP_COPY) {
153 return false;
156 for (i = temps[arg1].next_copy ; i != arg1 ; i = temps[i].next_copy) {
157 if (i == arg2) {
158 return true;
162 return false;
165 static void tcg_opt_gen_mov(TCGContext *s, int op_index, TCGArg *gen_args,
166 TCGOpcode old_op, TCGArg dst, TCGArg src)
168 TCGOpcode new_op = op_to_mov(old_op);
169 tcg_target_ulong mask;
171 s->gen_opc_buf[op_index] = new_op;
173 reset_temp(dst);
174 mask = temps[src].mask;
175 if (TCG_TARGET_REG_BITS > 32 && new_op == INDEX_op_mov_i32) {
176 /* High bits of the destination are now garbage. */
177 mask |= ~0xffffffffull;
179 temps[dst].mask = mask;
181 assert(temps[src].state != TCG_TEMP_CONST);
183 if (s->temps[src].type == s->temps[dst].type) {
184 if (temps[src].state != TCG_TEMP_COPY) {
185 temps[src].state = TCG_TEMP_COPY;
186 temps[src].next_copy = src;
187 temps[src].prev_copy = src;
189 temps[dst].state = TCG_TEMP_COPY;
190 temps[dst].next_copy = temps[src].next_copy;
191 temps[dst].prev_copy = src;
192 temps[temps[dst].next_copy].prev_copy = dst;
193 temps[src].next_copy = dst;
196 gen_args[0] = dst;
197 gen_args[1] = src;
200 static void tcg_opt_gen_movi(TCGContext *s, int op_index, TCGArg *gen_args,
201 TCGOpcode old_op, TCGArg dst, TCGArg val)
203 TCGOpcode new_op = op_to_movi(old_op);
204 tcg_target_ulong mask;
206 s->gen_opc_buf[op_index] = new_op;
208 reset_temp(dst);
209 temps[dst].state = TCG_TEMP_CONST;
210 temps[dst].val = val;
211 mask = val;
212 if (TCG_TARGET_REG_BITS > 32 && new_op == INDEX_op_mov_i32) {
213 /* High bits of the destination are now garbage. */
214 mask |= ~0xffffffffull;
216 temps[dst].mask = mask;
218 gen_args[0] = dst;
219 gen_args[1] = val;
222 static TCGArg do_constant_folding_2(TCGOpcode op, TCGArg x, TCGArg y)
224 uint64_t l64, h64;
226 switch (op) {
227 CASE_OP_32_64(add):
228 return x + y;
230 CASE_OP_32_64(sub):
231 return x - y;
233 CASE_OP_32_64(mul):
234 return x * y;
236 CASE_OP_32_64(and):
237 return x & y;
239 CASE_OP_32_64(or):
240 return x | y;
242 CASE_OP_32_64(xor):
243 return x ^ y;
245 case INDEX_op_shl_i32:
246 return (uint32_t)x << (y & 31);
248 case INDEX_op_shl_i64:
249 return (uint64_t)x << (y & 63);
251 case INDEX_op_shr_i32:
252 return (uint32_t)x >> (y & 31);
254 case INDEX_op_trunc_shr_i32:
255 case INDEX_op_shr_i64:
256 return (uint64_t)x >> (y & 63);
258 case INDEX_op_sar_i32:
259 return (int32_t)x >> (y & 31);
261 case INDEX_op_sar_i64:
262 return (int64_t)x >> (y & 63);
264 case INDEX_op_rotr_i32:
265 return ror32(x, y & 31);
267 case INDEX_op_rotr_i64:
268 return ror64(x, y & 63);
270 case INDEX_op_rotl_i32:
271 return rol32(x, y & 31);
273 case INDEX_op_rotl_i64:
274 return rol64(x, y & 63);
276 CASE_OP_32_64(not):
277 return ~x;
279 CASE_OP_32_64(neg):
280 return -x;
282 CASE_OP_32_64(andc):
283 return x & ~y;
285 CASE_OP_32_64(orc):
286 return x | ~y;
288 CASE_OP_32_64(eqv):
289 return ~(x ^ y);
291 CASE_OP_32_64(nand):
292 return ~(x & y);
294 CASE_OP_32_64(nor):
295 return ~(x | y);
297 CASE_OP_32_64(ext8s):
298 return (int8_t)x;
300 CASE_OP_32_64(ext16s):
301 return (int16_t)x;
303 CASE_OP_32_64(ext8u):
304 return (uint8_t)x;
306 CASE_OP_32_64(ext16u):
307 return (uint16_t)x;
309 case INDEX_op_ext32s_i64:
310 return (int32_t)x;
312 case INDEX_op_ext32u_i64:
313 return (uint32_t)x;
315 case INDEX_op_muluh_i32:
316 return ((uint64_t)(uint32_t)x * (uint32_t)y) >> 32;
317 case INDEX_op_mulsh_i32:
318 return ((int64_t)(int32_t)x * (int32_t)y) >> 32;
320 case INDEX_op_muluh_i64:
321 mulu64(&l64, &h64, x, y);
322 return h64;
323 case INDEX_op_mulsh_i64:
324 muls64(&l64, &h64, x, y);
325 return h64;
327 case INDEX_op_div_i32:
328 /* Avoid crashing on divide by zero, otherwise undefined. */
329 return (int32_t)x / ((int32_t)y ? : 1);
330 case INDEX_op_divu_i32:
331 return (uint32_t)x / ((uint32_t)y ? : 1);
332 case INDEX_op_div_i64:
333 return (int64_t)x / ((int64_t)y ? : 1);
334 case INDEX_op_divu_i64:
335 return (uint64_t)x / ((uint64_t)y ? : 1);
337 case INDEX_op_rem_i32:
338 return (int32_t)x % ((int32_t)y ? : 1);
339 case INDEX_op_remu_i32:
340 return (uint32_t)x % ((uint32_t)y ? : 1);
341 case INDEX_op_rem_i64:
342 return (int64_t)x % ((int64_t)y ? : 1);
343 case INDEX_op_remu_i64:
344 return (uint64_t)x % ((uint64_t)y ? : 1);
346 default:
347 fprintf(stderr,
348 "Unrecognized operation %d in do_constant_folding.\n", op);
349 tcg_abort();
353 static TCGArg do_constant_folding(TCGOpcode op, TCGArg x, TCGArg y)
355 TCGArg res = do_constant_folding_2(op, x, y);
356 if (op_bits(op) == 32) {
357 res &= 0xffffffff;
359 return res;
362 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c)
364 switch (c) {
365 case TCG_COND_EQ:
366 return x == y;
367 case TCG_COND_NE:
368 return x != y;
369 case TCG_COND_LT:
370 return (int32_t)x < (int32_t)y;
371 case TCG_COND_GE:
372 return (int32_t)x >= (int32_t)y;
373 case TCG_COND_LE:
374 return (int32_t)x <= (int32_t)y;
375 case TCG_COND_GT:
376 return (int32_t)x > (int32_t)y;
377 case TCG_COND_LTU:
378 return x < y;
379 case TCG_COND_GEU:
380 return x >= y;
381 case TCG_COND_LEU:
382 return x <= y;
383 case TCG_COND_GTU:
384 return x > y;
385 default:
386 tcg_abort();
390 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c)
392 switch (c) {
393 case TCG_COND_EQ:
394 return x == y;
395 case TCG_COND_NE:
396 return x != y;
397 case TCG_COND_LT:
398 return (int64_t)x < (int64_t)y;
399 case TCG_COND_GE:
400 return (int64_t)x >= (int64_t)y;
401 case TCG_COND_LE:
402 return (int64_t)x <= (int64_t)y;
403 case TCG_COND_GT:
404 return (int64_t)x > (int64_t)y;
405 case TCG_COND_LTU:
406 return x < y;
407 case TCG_COND_GEU:
408 return x >= y;
409 case TCG_COND_LEU:
410 return x <= y;
411 case TCG_COND_GTU:
412 return x > y;
413 default:
414 tcg_abort();
418 static bool do_constant_folding_cond_eq(TCGCond c)
420 switch (c) {
421 case TCG_COND_GT:
422 case TCG_COND_LTU:
423 case TCG_COND_LT:
424 case TCG_COND_GTU:
425 case TCG_COND_NE:
426 return 0;
427 case TCG_COND_GE:
428 case TCG_COND_GEU:
429 case TCG_COND_LE:
430 case TCG_COND_LEU:
431 case TCG_COND_EQ:
432 return 1;
433 default:
434 tcg_abort();
438 /* Return 2 if the condition can't be simplified, and the result
439 of the condition (0 or 1) if it can */
440 static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x,
441 TCGArg y, TCGCond c)
443 if (temps[x].state == TCG_TEMP_CONST && temps[y].state == TCG_TEMP_CONST) {
444 switch (op_bits(op)) {
445 case 32:
446 return do_constant_folding_cond_32(temps[x].val, temps[y].val, c);
447 case 64:
448 return do_constant_folding_cond_64(temps[x].val, temps[y].val, c);
449 default:
450 tcg_abort();
452 } else if (temps_are_copies(x, y)) {
453 return do_constant_folding_cond_eq(c);
454 } else if (temps[y].state == TCG_TEMP_CONST && temps[y].val == 0) {
455 switch (c) {
456 case TCG_COND_LTU:
457 return 0;
458 case TCG_COND_GEU:
459 return 1;
460 default:
461 return 2;
463 } else {
464 return 2;
468 /* Return 2 if the condition can't be simplified, and the result
469 of the condition (0 or 1) if it can */
470 static TCGArg do_constant_folding_cond2(TCGArg *p1, TCGArg *p2, TCGCond c)
472 TCGArg al = p1[0], ah = p1[1];
473 TCGArg bl = p2[0], bh = p2[1];
475 if (temps[bl].state == TCG_TEMP_CONST
476 && temps[bh].state == TCG_TEMP_CONST) {
477 uint64_t b = ((uint64_t)temps[bh].val << 32) | (uint32_t)temps[bl].val;
479 if (temps[al].state == TCG_TEMP_CONST
480 && temps[ah].state == TCG_TEMP_CONST) {
481 uint64_t a;
482 a = ((uint64_t)temps[ah].val << 32) | (uint32_t)temps[al].val;
483 return do_constant_folding_cond_64(a, b, c);
485 if (b == 0) {
486 switch (c) {
487 case TCG_COND_LTU:
488 return 0;
489 case TCG_COND_GEU:
490 return 1;
491 default:
492 break;
496 if (temps_are_copies(al, bl) && temps_are_copies(ah, bh)) {
497 return do_constant_folding_cond_eq(c);
499 return 2;
502 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
504 TCGArg a1 = *p1, a2 = *p2;
505 int sum = 0;
506 sum += temps[a1].state == TCG_TEMP_CONST;
507 sum -= temps[a2].state == TCG_TEMP_CONST;
509 /* Prefer the constant in second argument, and then the form
510 op a, a, b, which is better handled on non-RISC hosts. */
511 if (sum > 0 || (sum == 0 && dest == a2)) {
512 *p1 = a2;
513 *p2 = a1;
514 return true;
516 return false;
519 static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
521 int sum = 0;
522 sum += temps[p1[0]].state == TCG_TEMP_CONST;
523 sum += temps[p1[1]].state == TCG_TEMP_CONST;
524 sum -= temps[p2[0]].state == TCG_TEMP_CONST;
525 sum -= temps[p2[1]].state == TCG_TEMP_CONST;
526 if (sum > 0) {
527 TCGArg t;
528 t = p1[0], p1[0] = p2[0], p2[0] = t;
529 t = p1[1], p1[1] = p2[1], p2[1] = t;
530 return true;
532 return false;
535 /* Propagate constants and copies, fold constant expressions. */
536 static TCGArg *tcg_constant_folding(TCGContext *s, uint16_t *tcg_opc_ptr,
537 TCGArg *args, TCGOpDef *tcg_op_defs)
539 int nb_ops, op_index, nb_temps, nb_globals;
540 TCGArg *gen_args;
542 /* Array VALS has an element for each temp.
543 If this temp holds a constant then its value is kept in VALS' element.
544 If this temp is a copy of other ones then the other copies are
545 available through the doubly linked circular list. */
547 nb_temps = s->nb_temps;
548 nb_globals = s->nb_globals;
549 reset_all_temps(nb_temps);
551 nb_ops = tcg_opc_ptr - s->gen_opc_buf;
552 gen_args = args;
553 for (op_index = 0; op_index < nb_ops; op_index++) {
554 TCGOpcode op = s->gen_opc_buf[op_index];
555 const TCGOpDef *def = &tcg_op_defs[op];
556 tcg_target_ulong mask, partmask, affected;
557 int nb_oargs, nb_iargs, nb_args, i;
558 TCGArg tmp;
560 if (op == INDEX_op_call) {
561 *gen_args++ = tmp = *args++;
562 nb_oargs = tmp >> 16;
563 nb_iargs = tmp & 0xffff;
564 nb_args = nb_oargs + nb_iargs + def->nb_cargs;
565 } else {
566 nb_oargs = def->nb_oargs;
567 nb_iargs = def->nb_iargs;
568 nb_args = def->nb_args;
571 /* Do copy propagation */
572 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
573 if (temps[args[i]].state == TCG_TEMP_COPY) {
574 args[i] = find_better_copy(s, args[i]);
578 /* For commutative operations make constant second argument */
579 switch (op) {
580 CASE_OP_32_64(add):
581 CASE_OP_32_64(mul):
582 CASE_OP_32_64(and):
583 CASE_OP_32_64(or):
584 CASE_OP_32_64(xor):
585 CASE_OP_32_64(eqv):
586 CASE_OP_32_64(nand):
587 CASE_OP_32_64(nor):
588 CASE_OP_32_64(muluh):
589 CASE_OP_32_64(mulsh):
590 swap_commutative(args[0], &args[1], &args[2]);
591 break;
592 CASE_OP_32_64(brcond):
593 if (swap_commutative(-1, &args[0], &args[1])) {
594 args[2] = tcg_swap_cond(args[2]);
596 break;
597 CASE_OP_32_64(setcond):
598 if (swap_commutative(args[0], &args[1], &args[2])) {
599 args[3] = tcg_swap_cond(args[3]);
601 break;
602 CASE_OP_32_64(movcond):
603 if (swap_commutative(-1, &args[1], &args[2])) {
604 args[5] = tcg_swap_cond(args[5]);
606 /* For movcond, we canonicalize the "false" input reg to match
607 the destination reg so that the tcg backend can implement
608 a "move if true" operation. */
609 if (swap_commutative(args[0], &args[4], &args[3])) {
610 args[5] = tcg_invert_cond(args[5]);
612 break;
613 CASE_OP_32_64(add2):
614 swap_commutative(args[0], &args[2], &args[4]);
615 swap_commutative(args[1], &args[3], &args[5]);
616 break;
617 CASE_OP_32_64(mulu2):
618 CASE_OP_32_64(muls2):
619 swap_commutative(args[0], &args[2], &args[3]);
620 break;
621 case INDEX_op_brcond2_i32:
622 if (swap_commutative2(&args[0], &args[2])) {
623 args[4] = tcg_swap_cond(args[4]);
625 break;
626 case INDEX_op_setcond2_i32:
627 if (swap_commutative2(&args[1], &args[3])) {
628 args[5] = tcg_swap_cond(args[5]);
630 break;
631 default:
632 break;
635 /* Simplify expressions for "shift/rot r, 0, a => movi r, 0",
636 and "sub r, 0, a => neg r, a" case. */
637 switch (op) {
638 CASE_OP_32_64(shl):
639 CASE_OP_32_64(shr):
640 CASE_OP_32_64(sar):
641 CASE_OP_32_64(rotl):
642 CASE_OP_32_64(rotr):
643 if (temps[args[1]].state == TCG_TEMP_CONST
644 && temps[args[1]].val == 0) {
645 tcg_opt_gen_movi(s, op_index, gen_args, op, args[0], 0);
646 args += 3;
647 gen_args += 2;
648 continue;
650 break;
651 CASE_OP_32_64(sub):
653 TCGOpcode neg_op;
654 bool have_neg;
656 if (temps[args[2]].state == TCG_TEMP_CONST) {
657 /* Proceed with possible constant folding. */
658 break;
660 if (op == INDEX_op_sub_i32) {
661 neg_op = INDEX_op_neg_i32;
662 have_neg = TCG_TARGET_HAS_neg_i32;
663 } else {
664 neg_op = INDEX_op_neg_i64;
665 have_neg = TCG_TARGET_HAS_neg_i64;
667 if (!have_neg) {
668 break;
670 if (temps[args[1]].state == TCG_TEMP_CONST
671 && temps[args[1]].val == 0) {
672 s->gen_opc_buf[op_index] = neg_op;
673 reset_temp(args[0]);
674 gen_args[0] = args[0];
675 gen_args[1] = args[2];
676 args += 3;
677 gen_args += 2;
678 continue;
681 break;
682 CASE_OP_32_64(xor):
683 CASE_OP_32_64(nand):
684 if (temps[args[1]].state != TCG_TEMP_CONST
685 && temps[args[2]].state == TCG_TEMP_CONST
686 && temps[args[2]].val == -1) {
687 i = 1;
688 goto try_not;
690 break;
691 CASE_OP_32_64(nor):
692 if (temps[args[1]].state != TCG_TEMP_CONST
693 && temps[args[2]].state == TCG_TEMP_CONST
694 && temps[args[2]].val == 0) {
695 i = 1;
696 goto try_not;
698 break;
699 CASE_OP_32_64(andc):
700 if (temps[args[2]].state != TCG_TEMP_CONST
701 && temps[args[1]].state == TCG_TEMP_CONST
702 && temps[args[1]].val == -1) {
703 i = 2;
704 goto try_not;
706 break;
707 CASE_OP_32_64(orc):
708 CASE_OP_32_64(eqv):
709 if (temps[args[2]].state != TCG_TEMP_CONST
710 && temps[args[1]].state == TCG_TEMP_CONST
711 && temps[args[1]].val == 0) {
712 i = 2;
713 goto try_not;
715 break;
716 try_not:
718 TCGOpcode not_op;
719 bool have_not;
721 if (def->flags & TCG_OPF_64BIT) {
722 not_op = INDEX_op_not_i64;
723 have_not = TCG_TARGET_HAS_not_i64;
724 } else {
725 not_op = INDEX_op_not_i32;
726 have_not = TCG_TARGET_HAS_not_i32;
728 if (!have_not) {
729 break;
731 s->gen_opc_buf[op_index] = not_op;
732 reset_temp(args[0]);
733 gen_args[0] = args[0];
734 gen_args[1] = args[i];
735 args += 3;
736 gen_args += 2;
737 continue;
739 default:
740 break;
743 /* Simplify expression for "op r, a, const => mov r, a" cases */
744 switch (op) {
745 CASE_OP_32_64(add):
746 CASE_OP_32_64(sub):
747 CASE_OP_32_64(shl):
748 CASE_OP_32_64(shr):
749 CASE_OP_32_64(sar):
750 CASE_OP_32_64(rotl):
751 CASE_OP_32_64(rotr):
752 CASE_OP_32_64(or):
753 CASE_OP_32_64(xor):
754 CASE_OP_32_64(andc):
755 if (temps[args[1]].state != TCG_TEMP_CONST
756 && temps[args[2]].state == TCG_TEMP_CONST
757 && temps[args[2]].val == 0) {
758 goto do_mov3;
760 break;
761 CASE_OP_32_64(and):
762 CASE_OP_32_64(orc):
763 CASE_OP_32_64(eqv):
764 if (temps[args[1]].state != TCG_TEMP_CONST
765 && temps[args[2]].state == TCG_TEMP_CONST
766 && temps[args[2]].val == -1) {
767 goto do_mov3;
769 break;
770 do_mov3:
771 if (temps_are_copies(args[0], args[1])) {
772 s->gen_opc_buf[op_index] = INDEX_op_nop;
773 } else {
774 tcg_opt_gen_mov(s, op_index, gen_args, op, args[0], args[1]);
775 gen_args += 2;
777 args += 3;
778 continue;
779 default:
780 break;
783 /* Simplify using known-zero bits. Currently only ops with a single
784 output argument is supported. */
785 mask = -1;
786 affected = -1;
787 switch (op) {
788 CASE_OP_32_64(ext8s):
789 if ((temps[args[1]].mask & 0x80) != 0) {
790 break;
792 CASE_OP_32_64(ext8u):
793 mask = 0xff;
794 goto and_const;
795 CASE_OP_32_64(ext16s):
796 if ((temps[args[1]].mask & 0x8000) != 0) {
797 break;
799 CASE_OP_32_64(ext16u):
800 mask = 0xffff;
801 goto and_const;
802 case INDEX_op_ext32s_i64:
803 if ((temps[args[1]].mask & 0x80000000) != 0) {
804 break;
806 case INDEX_op_ext32u_i64:
807 mask = 0xffffffffU;
808 goto and_const;
810 CASE_OP_32_64(and):
811 mask = temps[args[2]].mask;
812 if (temps[args[2]].state == TCG_TEMP_CONST) {
813 and_const:
814 affected = temps[args[1]].mask & ~mask;
816 mask = temps[args[1]].mask & mask;
817 break;
819 CASE_OP_32_64(andc):
820 /* Known-zeros does not imply known-ones. Therefore unless
821 args[2] is constant, we can't infer anything from it. */
822 if (temps[args[2]].state == TCG_TEMP_CONST) {
823 mask = ~temps[args[2]].mask;
824 goto and_const;
826 /* But we certainly know nothing outside args[1] may be set. */
827 mask = temps[args[1]].mask;
828 break;
830 case INDEX_op_sar_i32:
831 if (temps[args[2]].state == TCG_TEMP_CONST) {
832 tmp = temps[args[2]].val & 31;
833 mask = (int32_t)temps[args[1]].mask >> tmp;
835 break;
836 case INDEX_op_sar_i64:
837 if (temps[args[2]].state == TCG_TEMP_CONST) {
838 tmp = temps[args[2]].val & 63;
839 mask = (int64_t)temps[args[1]].mask >> tmp;
841 break;
843 case INDEX_op_shr_i32:
844 if (temps[args[2]].state == TCG_TEMP_CONST) {
845 tmp = temps[args[2]].val & 31;
846 mask = (uint32_t)temps[args[1]].mask >> tmp;
848 break;
849 case INDEX_op_shr_i64:
850 if (temps[args[2]].state == TCG_TEMP_CONST) {
851 tmp = temps[args[2]].val & 63;
852 mask = (uint64_t)temps[args[1]].mask >> tmp;
854 break;
856 case INDEX_op_trunc_shr_i32:
857 mask = (uint64_t)temps[args[1]].mask >> args[2];
858 break;
860 CASE_OP_32_64(shl):
861 if (temps[args[2]].state == TCG_TEMP_CONST) {
862 tmp = temps[args[2]].val & (TCG_TARGET_REG_BITS - 1);
863 mask = temps[args[1]].mask << tmp;
865 break;
867 CASE_OP_32_64(neg):
868 /* Set to 1 all bits to the left of the rightmost. */
869 mask = -(temps[args[1]].mask & -temps[args[1]].mask);
870 break;
872 CASE_OP_32_64(deposit):
873 mask = deposit64(temps[args[1]].mask, args[3], args[4],
874 temps[args[2]].mask);
875 break;
877 CASE_OP_32_64(or):
878 CASE_OP_32_64(xor):
879 mask = temps[args[1]].mask | temps[args[2]].mask;
880 break;
882 CASE_OP_32_64(setcond):
883 case INDEX_op_setcond2_i32:
884 mask = 1;
885 break;
887 CASE_OP_32_64(movcond):
888 mask = temps[args[3]].mask | temps[args[4]].mask;
889 break;
891 CASE_OP_32_64(ld8u):
892 mask = 0xff;
893 break;
894 CASE_OP_32_64(ld16u):
895 mask = 0xffff;
896 break;
897 case INDEX_op_ld32u_i64:
898 mask = 0xffffffffu;
899 break;
901 CASE_OP_32_64(qemu_ld):
903 TCGMemOp mop = args[nb_oargs + nb_iargs];
904 if (!(mop & MO_SIGN)) {
905 mask = (2ULL << ((8 << (mop & MO_SIZE)) - 1)) - 1;
908 break;
910 default:
911 break;
914 /* 32-bit ops generate 32-bit results. For the result is zero test
915 below, we can ignore high bits, but for further optimizations we
916 need to record that the high bits contain garbage. */
917 partmask = mask;
918 if (!(def->flags & TCG_OPF_64BIT)) {
919 mask |= ~(tcg_target_ulong)0xffffffffu;
920 partmask &= 0xffffffffu;
921 affected &= 0xffffffffu;
924 if (partmask == 0) {
925 assert(nb_oargs == 1);
926 tcg_opt_gen_movi(s, op_index, gen_args, op, args[0], 0);
927 args += nb_args;
928 gen_args += 2;
929 continue;
931 if (affected == 0) {
932 assert(nb_oargs == 1);
933 if (temps_are_copies(args[0], args[1])) {
934 s->gen_opc_buf[op_index] = INDEX_op_nop;
935 } else if (temps[args[1]].state != TCG_TEMP_CONST) {
936 tcg_opt_gen_mov(s, op_index, gen_args, op, args[0], args[1]);
937 gen_args += 2;
938 } else {
939 tcg_opt_gen_movi(s, op_index, gen_args, op,
940 args[0], temps[args[1]].val);
941 gen_args += 2;
943 args += nb_args;
944 continue;
947 /* Simplify expression for "op r, a, 0 => movi r, 0" cases */
948 switch (op) {
949 CASE_OP_32_64(and):
950 CASE_OP_32_64(mul):
951 CASE_OP_32_64(muluh):
952 CASE_OP_32_64(mulsh):
953 if ((temps[args[2]].state == TCG_TEMP_CONST
954 && temps[args[2]].val == 0)) {
955 tcg_opt_gen_movi(s, op_index, gen_args, op, args[0], 0);
956 args += 3;
957 gen_args += 2;
958 continue;
960 break;
961 default:
962 break;
965 /* Simplify expression for "op r, a, a => mov r, a" cases */
966 switch (op) {
967 CASE_OP_32_64(or):
968 CASE_OP_32_64(and):
969 if (temps_are_copies(args[1], args[2])) {
970 if (temps_are_copies(args[0], args[1])) {
971 s->gen_opc_buf[op_index] = INDEX_op_nop;
972 } else {
973 tcg_opt_gen_mov(s, op_index, gen_args, op,
974 args[0], args[1]);
975 gen_args += 2;
977 args += 3;
978 continue;
980 break;
981 default:
982 break;
985 /* Simplify expression for "op r, a, a => movi r, 0" cases */
986 switch (op) {
987 CASE_OP_32_64(andc):
988 CASE_OP_32_64(sub):
989 CASE_OP_32_64(xor):
990 if (temps_are_copies(args[1], args[2])) {
991 tcg_opt_gen_movi(s, op_index, gen_args, op, args[0], 0);
992 gen_args += 2;
993 args += 3;
994 continue;
996 break;
997 default:
998 break;
1001 /* Propagate constants through copy operations and do constant
1002 folding. Constants will be substituted to arguments by register
1003 allocator where needed and possible. Also detect copies. */
1004 switch (op) {
1005 CASE_OP_32_64(mov):
1006 if (temps_are_copies(args[0], args[1])) {
1007 args += 2;
1008 s->gen_opc_buf[op_index] = INDEX_op_nop;
1009 break;
1011 if (temps[args[1]].state != TCG_TEMP_CONST) {
1012 tcg_opt_gen_mov(s, op_index, gen_args, op, args[0], args[1]);
1013 gen_args += 2;
1014 args += 2;
1015 break;
1017 /* Source argument is constant. Rewrite the operation and
1018 let movi case handle it. */
1019 args[1] = temps[args[1]].val;
1020 /* fallthrough */
1021 CASE_OP_32_64(movi):
1022 tcg_opt_gen_movi(s, op_index, gen_args, op, args[0], args[1]);
1023 gen_args += 2;
1024 args += 2;
1025 break;
1027 CASE_OP_32_64(not):
1028 CASE_OP_32_64(neg):
1029 CASE_OP_32_64(ext8s):
1030 CASE_OP_32_64(ext8u):
1031 CASE_OP_32_64(ext16s):
1032 CASE_OP_32_64(ext16u):
1033 case INDEX_op_ext32s_i64:
1034 case INDEX_op_ext32u_i64:
1035 if (temps[args[1]].state == TCG_TEMP_CONST) {
1036 tmp = do_constant_folding(op, temps[args[1]].val, 0);
1037 tcg_opt_gen_movi(s, op_index, gen_args, op, args[0], tmp);
1038 gen_args += 2;
1039 args += 2;
1040 break;
1042 goto do_default;
1044 case INDEX_op_trunc_shr_i32:
1045 if (temps[args[1]].state == TCG_TEMP_CONST) {
1046 tmp = do_constant_folding(op, temps[args[1]].val, args[2]);
1047 tcg_opt_gen_movi(s, op_index, gen_args, op, args[0], tmp);
1048 gen_args += 2;
1049 args += 3;
1050 break;
1052 goto do_default;
1054 CASE_OP_32_64(add):
1055 CASE_OP_32_64(sub):
1056 CASE_OP_32_64(mul):
1057 CASE_OP_32_64(or):
1058 CASE_OP_32_64(and):
1059 CASE_OP_32_64(xor):
1060 CASE_OP_32_64(shl):
1061 CASE_OP_32_64(shr):
1062 CASE_OP_32_64(sar):
1063 CASE_OP_32_64(rotl):
1064 CASE_OP_32_64(rotr):
1065 CASE_OP_32_64(andc):
1066 CASE_OP_32_64(orc):
1067 CASE_OP_32_64(eqv):
1068 CASE_OP_32_64(nand):
1069 CASE_OP_32_64(nor):
1070 CASE_OP_32_64(muluh):
1071 CASE_OP_32_64(mulsh):
1072 CASE_OP_32_64(div):
1073 CASE_OP_32_64(divu):
1074 CASE_OP_32_64(rem):
1075 CASE_OP_32_64(remu):
1076 if (temps[args[1]].state == TCG_TEMP_CONST
1077 && temps[args[2]].state == TCG_TEMP_CONST) {
1078 tmp = do_constant_folding(op, temps[args[1]].val,
1079 temps[args[2]].val);
1080 tcg_opt_gen_movi(s, op_index, gen_args, op, args[0], tmp);
1081 gen_args += 2;
1082 args += 3;
1083 break;
1085 goto do_default;
1087 CASE_OP_32_64(deposit):
1088 if (temps[args[1]].state == TCG_TEMP_CONST
1089 && temps[args[2]].state == TCG_TEMP_CONST) {
1090 tmp = deposit64(temps[args[1]].val, args[3], args[4],
1091 temps[args[2]].val);
1092 tcg_opt_gen_movi(s, op_index, gen_args, op, args[0], tmp);
1093 gen_args += 2;
1094 args += 5;
1095 break;
1097 goto do_default;
1099 CASE_OP_32_64(setcond):
1100 tmp = do_constant_folding_cond(op, args[1], args[2], args[3]);
1101 if (tmp != 2) {
1102 tcg_opt_gen_movi(s, op_index, gen_args, op, args[0], tmp);
1103 gen_args += 2;
1104 args += 4;
1105 break;
1107 goto do_default;
1109 CASE_OP_32_64(brcond):
1110 tmp = do_constant_folding_cond(op, args[0], args[1], args[2]);
1111 if (tmp != 2) {
1112 if (tmp) {
1113 reset_all_temps(nb_temps);
1114 s->gen_opc_buf[op_index] = INDEX_op_br;
1115 gen_args[0] = args[3];
1116 gen_args += 1;
1117 } else {
1118 s->gen_opc_buf[op_index] = INDEX_op_nop;
1120 args += 4;
1121 break;
1123 goto do_default;
1125 CASE_OP_32_64(movcond):
1126 tmp = do_constant_folding_cond(op, args[1], args[2], args[5]);
1127 if (tmp != 2) {
1128 if (temps_are_copies(args[0], args[4-tmp])) {
1129 s->gen_opc_buf[op_index] = INDEX_op_nop;
1130 } else if (temps[args[4-tmp]].state == TCG_TEMP_CONST) {
1131 tcg_opt_gen_movi(s, op_index, gen_args, op,
1132 args[0], temps[args[4-tmp]].val);
1133 gen_args += 2;
1134 } else {
1135 tcg_opt_gen_mov(s, op_index, gen_args, op,
1136 args[0], args[4-tmp]);
1137 gen_args += 2;
1139 args += 6;
1140 break;
1142 goto do_default;
1144 case INDEX_op_add2_i32:
1145 case INDEX_op_sub2_i32:
1146 if (temps[args[2]].state == TCG_TEMP_CONST
1147 && temps[args[3]].state == TCG_TEMP_CONST
1148 && temps[args[4]].state == TCG_TEMP_CONST
1149 && temps[args[5]].state == TCG_TEMP_CONST) {
1150 uint32_t al = temps[args[2]].val;
1151 uint32_t ah = temps[args[3]].val;
1152 uint32_t bl = temps[args[4]].val;
1153 uint32_t bh = temps[args[5]].val;
1154 uint64_t a = ((uint64_t)ah << 32) | al;
1155 uint64_t b = ((uint64_t)bh << 32) | bl;
1156 TCGArg rl, rh;
1158 if (op == INDEX_op_add2_i32) {
1159 a += b;
1160 } else {
1161 a -= b;
1164 /* We emit the extra nop when we emit the add2/sub2. */
1165 assert(s->gen_opc_buf[op_index + 1] == INDEX_op_nop);
1167 rl = args[0];
1168 rh = args[1];
1169 tcg_opt_gen_movi(s, op_index, &gen_args[0],
1170 op, rl, (uint32_t)a);
1171 tcg_opt_gen_movi(s, ++op_index, &gen_args[2],
1172 op, rh, (uint32_t)(a >> 32));
1173 gen_args += 4;
1174 args += 6;
1175 break;
1177 goto do_default;
1179 case INDEX_op_mulu2_i32:
1180 if (temps[args[2]].state == TCG_TEMP_CONST
1181 && temps[args[3]].state == TCG_TEMP_CONST) {
1182 uint32_t a = temps[args[2]].val;
1183 uint32_t b = temps[args[3]].val;
1184 uint64_t r = (uint64_t)a * b;
1185 TCGArg rl, rh;
1187 /* We emit the extra nop when we emit the mulu2. */
1188 assert(s->gen_opc_buf[op_index + 1] == INDEX_op_nop);
1190 rl = args[0];
1191 rh = args[1];
1192 tcg_opt_gen_movi(s, op_index, &gen_args[0],
1193 op, rl, (uint32_t)r);
1194 tcg_opt_gen_movi(s, ++op_index, &gen_args[2],
1195 op, rh, (uint32_t)(r >> 32));
1196 gen_args += 4;
1197 args += 4;
1198 break;
1200 goto do_default;
1202 case INDEX_op_brcond2_i32:
1203 tmp = do_constant_folding_cond2(&args[0], &args[2], args[4]);
1204 if (tmp != 2) {
1205 if (tmp) {
1206 do_brcond_true:
1207 reset_all_temps(nb_temps);
1208 s->gen_opc_buf[op_index] = INDEX_op_br;
1209 gen_args[0] = args[5];
1210 gen_args += 1;
1211 } else {
1212 do_brcond_false:
1213 s->gen_opc_buf[op_index] = INDEX_op_nop;
1215 } else if ((args[4] == TCG_COND_LT || args[4] == TCG_COND_GE)
1216 && temps[args[2]].state == TCG_TEMP_CONST
1217 && temps[args[3]].state == TCG_TEMP_CONST
1218 && temps[args[2]].val == 0
1219 && temps[args[3]].val == 0) {
1220 /* Simplify LT/GE comparisons vs zero to a single compare
1221 vs the high word of the input. */
1222 do_brcond_high:
1223 reset_all_temps(nb_temps);
1224 s->gen_opc_buf[op_index] = INDEX_op_brcond_i32;
1225 gen_args[0] = args[1];
1226 gen_args[1] = args[3];
1227 gen_args[2] = args[4];
1228 gen_args[3] = args[5];
1229 gen_args += 4;
1230 } else if (args[4] == TCG_COND_EQ) {
1231 /* Simplify EQ comparisons where one of the pairs
1232 can be simplified. */
1233 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1234 args[0], args[2], TCG_COND_EQ);
1235 if (tmp == 0) {
1236 goto do_brcond_false;
1237 } else if (tmp == 1) {
1238 goto do_brcond_high;
1240 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1241 args[1], args[3], TCG_COND_EQ);
1242 if (tmp == 0) {
1243 goto do_brcond_false;
1244 } else if (tmp != 1) {
1245 goto do_default;
1247 do_brcond_low:
1248 reset_all_temps(nb_temps);
1249 s->gen_opc_buf[op_index] = INDEX_op_brcond_i32;
1250 gen_args[0] = args[0];
1251 gen_args[1] = args[2];
1252 gen_args[2] = args[4];
1253 gen_args[3] = args[5];
1254 gen_args += 4;
1255 } else if (args[4] == TCG_COND_NE) {
1256 /* Simplify NE comparisons where one of the pairs
1257 can be simplified. */
1258 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1259 args[0], args[2], TCG_COND_NE);
1260 if (tmp == 0) {
1261 goto do_brcond_high;
1262 } else if (tmp == 1) {
1263 goto do_brcond_true;
1265 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1266 args[1], args[3], TCG_COND_NE);
1267 if (tmp == 0) {
1268 goto do_brcond_low;
1269 } else if (tmp == 1) {
1270 goto do_brcond_true;
1272 goto do_default;
1273 } else {
1274 goto do_default;
1276 args += 6;
1277 break;
1279 case INDEX_op_setcond2_i32:
1280 tmp = do_constant_folding_cond2(&args[1], &args[3], args[5]);
1281 if (tmp != 2) {
1282 do_setcond_const:
1283 tcg_opt_gen_movi(s, op_index, gen_args, op, args[0], tmp);
1284 gen_args += 2;
1285 } else if ((args[5] == TCG_COND_LT || args[5] == TCG_COND_GE)
1286 && temps[args[3]].state == TCG_TEMP_CONST
1287 && temps[args[4]].state == TCG_TEMP_CONST
1288 && temps[args[3]].val == 0
1289 && temps[args[4]].val == 0) {
1290 /* Simplify LT/GE comparisons vs zero to a single compare
1291 vs the high word of the input. */
1292 do_setcond_high:
1293 s->gen_opc_buf[op_index] = INDEX_op_setcond_i32;
1294 reset_temp(args[0]);
1295 temps[args[0]].mask = 1;
1296 gen_args[0] = args[0];
1297 gen_args[1] = args[2];
1298 gen_args[2] = args[4];
1299 gen_args[3] = args[5];
1300 gen_args += 4;
1301 } else if (args[5] == TCG_COND_EQ) {
1302 /* Simplify EQ comparisons where one of the pairs
1303 can be simplified. */
1304 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1305 args[1], args[3], TCG_COND_EQ);
1306 if (tmp == 0) {
1307 goto do_setcond_const;
1308 } else if (tmp == 1) {
1309 goto do_setcond_high;
1311 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1312 args[2], args[4], TCG_COND_EQ);
1313 if (tmp == 0) {
1314 goto do_setcond_high;
1315 } else if (tmp != 1) {
1316 goto do_default;
1318 do_setcond_low:
1319 reset_temp(args[0]);
1320 temps[args[0]].mask = 1;
1321 s->gen_opc_buf[op_index] = INDEX_op_setcond_i32;
1322 gen_args[0] = args[0];
1323 gen_args[1] = args[1];
1324 gen_args[2] = args[3];
1325 gen_args[3] = args[5];
1326 gen_args += 4;
1327 } else if (args[5] == TCG_COND_NE) {
1328 /* Simplify NE comparisons where one of the pairs
1329 can be simplified. */
1330 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1331 args[1], args[3], TCG_COND_NE);
1332 if (tmp == 0) {
1333 goto do_setcond_high;
1334 } else if (tmp == 1) {
1335 goto do_setcond_const;
1337 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1338 args[2], args[4], TCG_COND_NE);
1339 if (tmp == 0) {
1340 goto do_setcond_low;
1341 } else if (tmp == 1) {
1342 goto do_setcond_const;
1344 goto do_default;
1345 } else {
1346 goto do_default;
1348 args += 6;
1349 break;
1351 case INDEX_op_call:
1352 if (!(args[nb_oargs + nb_iargs + 1]
1353 & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) {
1354 for (i = 0; i < nb_globals; i++) {
1355 reset_temp(i);
1358 goto do_reset_output;
1360 default:
1361 do_default:
1362 /* Default case: we know nothing about operation (or were unable
1363 to compute the operation result) so no propagation is done.
1364 We trash everything if the operation is the end of a basic
1365 block, otherwise we only trash the output args. "mask" is
1366 the non-zero bits mask for the first output arg. */
1367 if (def->flags & TCG_OPF_BB_END) {
1368 reset_all_temps(nb_temps);
1369 } else {
1370 do_reset_output:
1371 for (i = 0; i < nb_oargs; i++) {
1372 reset_temp(args[i]);
1373 /* Save the corresponding known-zero bits mask for the
1374 first output argument (only one supported so far). */
1375 if (i == 0) {
1376 temps[args[i]].mask = mask;
1380 for (i = 0; i < nb_args; i++) {
1381 gen_args[i] = args[i];
1383 args += nb_args;
1384 gen_args += nb_args;
1385 break;
1389 return gen_args;
1392 TCGArg *tcg_optimize(TCGContext *s, uint16_t *tcg_opc_ptr,
1393 TCGArg *args, TCGOpDef *tcg_op_defs)
1395 TCGArg *res;
1396 res = tcg_constant_folding(s, tcg_opc_ptr, args, tcg_op_defs);
1397 return res;