s390x: fix flat file load on 32 bit systems
[qemu/qmp-unstable.git] / tcg / optimize.c
blob89e2d6a3b3ff06275331e502a87b0ed644ad4d47
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_movi(TCGOpcode op)
88 switch (op_bits(op)) {
89 case 32:
90 return INDEX_op_movi_i32;
91 case 64:
92 return INDEX_op_movi_i64;
93 default:
94 fprintf(stderr, "op_to_movi: unexpected return value of "
95 "function op_bits.\n");
96 tcg_abort();
100 static TCGArg find_better_copy(TCGContext *s, TCGArg temp)
102 TCGArg i;
104 /* If this is already a global, we can't do better. */
105 if (temp < s->nb_globals) {
106 return temp;
109 /* Search for a global first. */
110 for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) {
111 if (i < s->nb_globals) {
112 return i;
116 /* If it is a temp, search for a temp local. */
117 if (!s->temps[temp].temp_local) {
118 for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) {
119 if (s->temps[i].temp_local) {
120 return i;
125 /* Failure to find a better representation, return the same temp. */
126 return temp;
129 static bool temps_are_copies(TCGArg arg1, TCGArg arg2)
131 TCGArg i;
133 if (arg1 == arg2) {
134 return true;
137 if (temps[arg1].state != TCG_TEMP_COPY
138 || temps[arg2].state != TCG_TEMP_COPY) {
139 return false;
142 for (i = temps[arg1].next_copy ; i != arg1 ; i = temps[i].next_copy) {
143 if (i == arg2) {
144 return true;
148 return false;
151 static void tcg_opt_gen_mov(TCGContext *s, TCGArg *gen_args,
152 TCGArg dst, TCGArg src)
154 reset_temp(dst);
155 temps[dst].mask = temps[src].mask;
156 assert(temps[src].state != TCG_TEMP_CONST);
158 if (s->temps[src].type == s->temps[dst].type) {
159 if (temps[src].state != TCG_TEMP_COPY) {
160 temps[src].state = TCG_TEMP_COPY;
161 temps[src].next_copy = src;
162 temps[src].prev_copy = src;
164 temps[dst].state = TCG_TEMP_COPY;
165 temps[dst].next_copy = temps[src].next_copy;
166 temps[dst].prev_copy = src;
167 temps[temps[dst].next_copy].prev_copy = dst;
168 temps[src].next_copy = dst;
171 gen_args[0] = dst;
172 gen_args[1] = src;
175 static void tcg_opt_gen_movi(TCGArg *gen_args, TCGArg dst, TCGArg val)
177 reset_temp(dst);
178 temps[dst].state = TCG_TEMP_CONST;
179 temps[dst].val = val;
180 temps[dst].mask = val;
181 gen_args[0] = dst;
182 gen_args[1] = val;
185 static TCGOpcode op_to_mov(TCGOpcode op)
187 switch (op_bits(op)) {
188 case 32:
189 return INDEX_op_mov_i32;
190 case 64:
191 return INDEX_op_mov_i64;
192 default:
193 fprintf(stderr, "op_to_mov: unexpected return value of "
194 "function op_bits.\n");
195 tcg_abort();
199 static TCGArg do_constant_folding_2(TCGOpcode op, TCGArg x, TCGArg y)
201 uint64_t l64, h64;
203 switch (op) {
204 CASE_OP_32_64(add):
205 return x + y;
207 CASE_OP_32_64(sub):
208 return x - y;
210 CASE_OP_32_64(mul):
211 return x * y;
213 CASE_OP_32_64(and):
214 return x & y;
216 CASE_OP_32_64(or):
217 return x | y;
219 CASE_OP_32_64(xor):
220 return x ^ y;
222 case INDEX_op_shl_i32:
223 return (uint32_t)x << (uint32_t)y;
225 case INDEX_op_shl_i64:
226 return (uint64_t)x << (uint64_t)y;
228 case INDEX_op_shr_i32:
229 return (uint32_t)x >> (uint32_t)y;
231 case INDEX_op_shr_i64:
232 return (uint64_t)x >> (uint64_t)y;
234 case INDEX_op_sar_i32:
235 return (int32_t)x >> (int32_t)y;
237 case INDEX_op_sar_i64:
238 return (int64_t)x >> (int64_t)y;
240 case INDEX_op_rotr_i32:
241 return ror32(x, y);
243 case INDEX_op_rotr_i64:
244 return ror64(x, y);
246 case INDEX_op_rotl_i32:
247 return rol32(x, y);
249 case INDEX_op_rotl_i64:
250 return rol64(x, y);
252 CASE_OP_32_64(not):
253 return ~x;
255 CASE_OP_32_64(neg):
256 return -x;
258 CASE_OP_32_64(andc):
259 return x & ~y;
261 CASE_OP_32_64(orc):
262 return x | ~y;
264 CASE_OP_32_64(eqv):
265 return ~(x ^ y);
267 CASE_OP_32_64(nand):
268 return ~(x & y);
270 CASE_OP_32_64(nor):
271 return ~(x | y);
273 CASE_OP_32_64(ext8s):
274 return (int8_t)x;
276 CASE_OP_32_64(ext16s):
277 return (int16_t)x;
279 CASE_OP_32_64(ext8u):
280 return (uint8_t)x;
282 CASE_OP_32_64(ext16u):
283 return (uint16_t)x;
285 case INDEX_op_ext32s_i64:
286 return (int32_t)x;
288 case INDEX_op_ext32u_i64:
289 return (uint32_t)x;
291 case INDEX_op_muluh_i32:
292 return ((uint64_t)(uint32_t)x * (uint32_t)y) >> 32;
293 case INDEX_op_mulsh_i32:
294 return ((int64_t)(int32_t)x * (int32_t)y) >> 32;
296 case INDEX_op_muluh_i64:
297 mulu64(&l64, &h64, x, y);
298 return h64;
299 case INDEX_op_mulsh_i64:
300 muls64(&l64, &h64, x, y);
301 return h64;
303 case INDEX_op_div_i32:
304 /* Avoid crashing on divide by zero, otherwise undefined. */
305 return (int32_t)x / ((int32_t)y ? : 1);
306 case INDEX_op_divu_i32:
307 return (uint32_t)x / ((uint32_t)y ? : 1);
308 case INDEX_op_div_i64:
309 return (int64_t)x / ((int64_t)y ? : 1);
310 case INDEX_op_divu_i64:
311 return (uint64_t)x / ((uint64_t)y ? : 1);
313 case INDEX_op_rem_i32:
314 return (int32_t)x % ((int32_t)y ? : 1);
315 case INDEX_op_remu_i32:
316 return (uint32_t)x % ((uint32_t)y ? : 1);
317 case INDEX_op_rem_i64:
318 return (int64_t)x % ((int64_t)y ? : 1);
319 case INDEX_op_remu_i64:
320 return (uint64_t)x % ((uint64_t)y ? : 1);
322 default:
323 fprintf(stderr,
324 "Unrecognized operation %d in do_constant_folding.\n", op);
325 tcg_abort();
329 static TCGArg do_constant_folding(TCGOpcode op, TCGArg x, TCGArg y)
331 TCGArg res = do_constant_folding_2(op, x, y);
332 if (op_bits(op) == 32) {
333 res &= 0xffffffff;
335 return res;
338 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c)
340 switch (c) {
341 case TCG_COND_EQ:
342 return x == y;
343 case TCG_COND_NE:
344 return x != y;
345 case TCG_COND_LT:
346 return (int32_t)x < (int32_t)y;
347 case TCG_COND_GE:
348 return (int32_t)x >= (int32_t)y;
349 case TCG_COND_LE:
350 return (int32_t)x <= (int32_t)y;
351 case TCG_COND_GT:
352 return (int32_t)x > (int32_t)y;
353 case TCG_COND_LTU:
354 return x < y;
355 case TCG_COND_GEU:
356 return x >= y;
357 case TCG_COND_LEU:
358 return x <= y;
359 case TCG_COND_GTU:
360 return x > y;
361 default:
362 tcg_abort();
366 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c)
368 switch (c) {
369 case TCG_COND_EQ:
370 return x == y;
371 case TCG_COND_NE:
372 return x != y;
373 case TCG_COND_LT:
374 return (int64_t)x < (int64_t)y;
375 case TCG_COND_GE:
376 return (int64_t)x >= (int64_t)y;
377 case TCG_COND_LE:
378 return (int64_t)x <= (int64_t)y;
379 case TCG_COND_GT:
380 return (int64_t)x > (int64_t)y;
381 case TCG_COND_LTU:
382 return x < y;
383 case TCG_COND_GEU:
384 return x >= y;
385 case TCG_COND_LEU:
386 return x <= y;
387 case TCG_COND_GTU:
388 return x > y;
389 default:
390 tcg_abort();
394 static bool do_constant_folding_cond_eq(TCGCond c)
396 switch (c) {
397 case TCG_COND_GT:
398 case TCG_COND_LTU:
399 case TCG_COND_LT:
400 case TCG_COND_GTU:
401 case TCG_COND_NE:
402 return 0;
403 case TCG_COND_GE:
404 case TCG_COND_GEU:
405 case TCG_COND_LE:
406 case TCG_COND_LEU:
407 case TCG_COND_EQ:
408 return 1;
409 default:
410 tcg_abort();
414 /* Return 2 if the condition can't be simplified, and the result
415 of the condition (0 or 1) if it can */
416 static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x,
417 TCGArg y, TCGCond c)
419 if (temps[x].state == TCG_TEMP_CONST && temps[y].state == TCG_TEMP_CONST) {
420 switch (op_bits(op)) {
421 case 32:
422 return do_constant_folding_cond_32(temps[x].val, temps[y].val, c);
423 case 64:
424 return do_constant_folding_cond_64(temps[x].val, temps[y].val, c);
425 default:
426 tcg_abort();
428 } else if (temps_are_copies(x, y)) {
429 return do_constant_folding_cond_eq(c);
430 } else if (temps[y].state == TCG_TEMP_CONST && temps[y].val == 0) {
431 switch (c) {
432 case TCG_COND_LTU:
433 return 0;
434 case TCG_COND_GEU:
435 return 1;
436 default:
437 return 2;
439 } else {
440 return 2;
444 /* Return 2 if the condition can't be simplified, and the result
445 of the condition (0 or 1) if it can */
446 static TCGArg do_constant_folding_cond2(TCGArg *p1, TCGArg *p2, TCGCond c)
448 TCGArg al = p1[0], ah = p1[1];
449 TCGArg bl = p2[0], bh = p2[1];
451 if (temps[bl].state == TCG_TEMP_CONST
452 && temps[bh].state == TCG_TEMP_CONST) {
453 uint64_t b = ((uint64_t)temps[bh].val << 32) | (uint32_t)temps[bl].val;
455 if (temps[al].state == TCG_TEMP_CONST
456 && temps[ah].state == TCG_TEMP_CONST) {
457 uint64_t a;
458 a = ((uint64_t)temps[ah].val << 32) | (uint32_t)temps[al].val;
459 return do_constant_folding_cond_64(a, b, c);
461 if (b == 0) {
462 switch (c) {
463 case TCG_COND_LTU:
464 return 0;
465 case TCG_COND_GEU:
466 return 1;
467 default:
468 break;
472 if (temps_are_copies(al, bl) && temps_are_copies(ah, bh)) {
473 return do_constant_folding_cond_eq(c);
475 return 2;
478 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
480 TCGArg a1 = *p1, a2 = *p2;
481 int sum = 0;
482 sum += temps[a1].state == TCG_TEMP_CONST;
483 sum -= temps[a2].state == TCG_TEMP_CONST;
485 /* Prefer the constant in second argument, and then the form
486 op a, a, b, which is better handled on non-RISC hosts. */
487 if (sum > 0 || (sum == 0 && dest == a2)) {
488 *p1 = a2;
489 *p2 = a1;
490 return true;
492 return false;
495 static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
497 int sum = 0;
498 sum += temps[p1[0]].state == TCG_TEMP_CONST;
499 sum += temps[p1[1]].state == TCG_TEMP_CONST;
500 sum -= temps[p2[0]].state == TCG_TEMP_CONST;
501 sum -= temps[p2[1]].state == TCG_TEMP_CONST;
502 if (sum > 0) {
503 TCGArg t;
504 t = p1[0], p1[0] = p2[0], p2[0] = t;
505 t = p1[1], p1[1] = p2[1], p2[1] = t;
506 return true;
508 return false;
511 /* Propagate constants and copies, fold constant expressions. */
512 static TCGArg *tcg_constant_folding(TCGContext *s, uint16_t *tcg_opc_ptr,
513 TCGArg *args, TCGOpDef *tcg_op_defs)
515 int i, nb_ops, op_index, nb_temps, nb_globals, nb_call_args;
516 tcg_target_ulong mask, affected;
517 TCGOpcode op;
518 const TCGOpDef *def;
519 TCGArg *gen_args;
520 TCGArg tmp;
522 /* Array VALS has an element for each temp.
523 If this temp holds a constant then its value is kept in VALS' element.
524 If this temp is a copy of other ones then the other copies are
525 available through the doubly linked circular list. */
527 nb_temps = s->nb_temps;
528 nb_globals = s->nb_globals;
529 reset_all_temps(nb_temps);
531 nb_ops = tcg_opc_ptr - s->gen_opc_buf;
532 gen_args = args;
533 for (op_index = 0; op_index < nb_ops; op_index++) {
534 op = s->gen_opc_buf[op_index];
535 def = &tcg_op_defs[op];
536 /* Do copy propagation */
537 if (op == INDEX_op_call) {
538 int nb_oargs = args[0] >> 16;
539 int nb_iargs = args[0] & 0xffff;
540 for (i = nb_oargs + 1; i < nb_oargs + nb_iargs + 1; i++) {
541 if (temps[args[i]].state == TCG_TEMP_COPY) {
542 args[i] = find_better_copy(s, args[i]);
545 } else {
546 for (i = def->nb_oargs; i < def->nb_oargs + def->nb_iargs; i++) {
547 if (temps[args[i]].state == TCG_TEMP_COPY) {
548 args[i] = find_better_copy(s, args[i]);
553 /* For commutative operations make constant second argument */
554 switch (op) {
555 CASE_OP_32_64(add):
556 CASE_OP_32_64(mul):
557 CASE_OP_32_64(and):
558 CASE_OP_32_64(or):
559 CASE_OP_32_64(xor):
560 CASE_OP_32_64(eqv):
561 CASE_OP_32_64(nand):
562 CASE_OP_32_64(nor):
563 CASE_OP_32_64(muluh):
564 CASE_OP_32_64(mulsh):
565 swap_commutative(args[0], &args[1], &args[2]);
566 break;
567 CASE_OP_32_64(brcond):
568 if (swap_commutative(-1, &args[0], &args[1])) {
569 args[2] = tcg_swap_cond(args[2]);
571 break;
572 CASE_OP_32_64(setcond):
573 if (swap_commutative(args[0], &args[1], &args[2])) {
574 args[3] = tcg_swap_cond(args[3]);
576 break;
577 CASE_OP_32_64(movcond):
578 if (swap_commutative(-1, &args[1], &args[2])) {
579 args[5] = tcg_swap_cond(args[5]);
581 /* For movcond, we canonicalize the "false" input reg to match
582 the destination reg so that the tcg backend can implement
583 a "move if true" operation. */
584 if (swap_commutative(args[0], &args[4], &args[3])) {
585 args[5] = tcg_invert_cond(args[5]);
587 break;
588 CASE_OP_32_64(add2):
589 swap_commutative(args[0], &args[2], &args[4]);
590 swap_commutative(args[1], &args[3], &args[5]);
591 break;
592 CASE_OP_32_64(mulu2):
593 CASE_OP_32_64(muls2):
594 swap_commutative(args[0], &args[2], &args[3]);
595 break;
596 case INDEX_op_brcond2_i32:
597 if (swap_commutative2(&args[0], &args[2])) {
598 args[4] = tcg_swap_cond(args[4]);
600 break;
601 case INDEX_op_setcond2_i32:
602 if (swap_commutative2(&args[1], &args[3])) {
603 args[5] = tcg_swap_cond(args[5]);
605 break;
606 default:
607 break;
610 /* Simplify expressions for "shift/rot r, 0, a => movi r, 0",
611 and "sub r, 0, a => neg r, a" case. */
612 switch (op) {
613 CASE_OP_32_64(shl):
614 CASE_OP_32_64(shr):
615 CASE_OP_32_64(sar):
616 CASE_OP_32_64(rotl):
617 CASE_OP_32_64(rotr):
618 if (temps[args[1]].state == TCG_TEMP_CONST
619 && temps[args[1]].val == 0) {
620 s->gen_opc_buf[op_index] = op_to_movi(op);
621 tcg_opt_gen_movi(gen_args, args[0], 0);
622 args += 3;
623 gen_args += 2;
624 continue;
626 break;
627 CASE_OP_32_64(sub):
629 TCGOpcode neg_op;
630 bool have_neg;
632 if (temps[args[2]].state == TCG_TEMP_CONST) {
633 /* Proceed with possible constant folding. */
634 break;
636 if (op == INDEX_op_sub_i32) {
637 neg_op = INDEX_op_neg_i32;
638 have_neg = TCG_TARGET_HAS_neg_i32;
639 } else {
640 neg_op = INDEX_op_neg_i64;
641 have_neg = TCG_TARGET_HAS_neg_i64;
643 if (!have_neg) {
644 break;
646 if (temps[args[1]].state == TCG_TEMP_CONST
647 && temps[args[1]].val == 0) {
648 s->gen_opc_buf[op_index] = neg_op;
649 reset_temp(args[0]);
650 gen_args[0] = args[0];
651 gen_args[1] = args[2];
652 args += 3;
653 gen_args += 2;
654 continue;
657 break;
658 default:
659 break;
662 /* Simplify expression for "op r, a, 0 => mov r, a" cases */
663 switch (op) {
664 CASE_OP_32_64(add):
665 CASE_OP_32_64(sub):
666 CASE_OP_32_64(shl):
667 CASE_OP_32_64(shr):
668 CASE_OP_32_64(sar):
669 CASE_OP_32_64(rotl):
670 CASE_OP_32_64(rotr):
671 CASE_OP_32_64(or):
672 CASE_OP_32_64(xor):
673 if (temps[args[1]].state == TCG_TEMP_CONST) {
674 /* Proceed with possible constant folding. */
675 break;
677 if (temps[args[2]].state == TCG_TEMP_CONST
678 && temps[args[2]].val == 0) {
679 if (temps_are_copies(args[0], args[1])) {
680 s->gen_opc_buf[op_index] = INDEX_op_nop;
681 } else {
682 s->gen_opc_buf[op_index] = op_to_mov(op);
683 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
684 gen_args += 2;
686 args += 3;
687 continue;
689 break;
690 default:
691 break;
694 /* Simplify using known-zero bits */
695 mask = -1;
696 affected = -1;
697 switch (op) {
698 CASE_OP_32_64(ext8s):
699 if ((temps[args[1]].mask & 0x80) != 0) {
700 break;
702 CASE_OP_32_64(ext8u):
703 mask = 0xff;
704 goto and_const;
705 CASE_OP_32_64(ext16s):
706 if ((temps[args[1]].mask & 0x8000) != 0) {
707 break;
709 CASE_OP_32_64(ext16u):
710 mask = 0xffff;
711 goto and_const;
712 case INDEX_op_ext32s_i64:
713 if ((temps[args[1]].mask & 0x80000000) != 0) {
714 break;
716 case INDEX_op_ext32u_i64:
717 mask = 0xffffffffU;
718 goto and_const;
720 CASE_OP_32_64(and):
721 mask = temps[args[2]].mask;
722 if (temps[args[2]].state == TCG_TEMP_CONST) {
723 and_const:
724 affected = temps[args[1]].mask & ~mask;
726 mask = temps[args[1]].mask & mask;
727 break;
729 CASE_OP_32_64(sar):
730 if (temps[args[2]].state == TCG_TEMP_CONST) {
731 mask = ((tcg_target_long)temps[args[1]].mask
732 >> temps[args[2]].val);
734 break;
736 CASE_OP_32_64(shr):
737 if (temps[args[2]].state == TCG_TEMP_CONST) {
738 mask = temps[args[1]].mask >> temps[args[2]].val;
740 break;
742 CASE_OP_32_64(shl):
743 if (temps[args[2]].state == TCG_TEMP_CONST) {
744 mask = temps[args[1]].mask << temps[args[2]].val;
746 break;
748 CASE_OP_32_64(neg):
749 /* Set to 1 all bits to the left of the rightmost. */
750 mask = -(temps[args[1]].mask & -temps[args[1]].mask);
751 break;
753 CASE_OP_32_64(deposit):
754 tmp = ((1ull << args[4]) - 1);
755 mask = ((temps[args[1]].mask & ~(tmp << args[3]))
756 | ((temps[args[2]].mask & tmp) << args[3]));
757 break;
759 CASE_OP_32_64(or):
760 CASE_OP_32_64(xor):
761 mask = temps[args[1]].mask | temps[args[2]].mask;
762 break;
764 CASE_OP_32_64(setcond):
765 mask = 1;
766 break;
768 CASE_OP_32_64(movcond):
769 mask = temps[args[3]].mask | temps[args[4]].mask;
770 break;
772 default:
773 break;
776 if (mask == 0) {
777 assert(def->nb_oargs == 1);
778 s->gen_opc_buf[op_index] = op_to_movi(op);
779 tcg_opt_gen_movi(gen_args, args[0], 0);
780 args += def->nb_oargs + def->nb_iargs + def->nb_cargs;
781 gen_args += 2;
782 continue;
784 if (affected == 0) {
785 assert(def->nb_oargs == 1);
786 if (temps_are_copies(args[0], args[1])) {
787 s->gen_opc_buf[op_index] = INDEX_op_nop;
788 } else if (temps[args[1]].state != TCG_TEMP_CONST) {
789 s->gen_opc_buf[op_index] = op_to_mov(op);
790 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
791 gen_args += 2;
792 } else {
793 s->gen_opc_buf[op_index] = op_to_movi(op);
794 tcg_opt_gen_movi(gen_args, args[0], temps[args[1]].val);
795 gen_args += 2;
797 args += def->nb_iargs + 1;
798 continue;
801 /* Simplify expression for "op r, a, 0 => movi r, 0" cases */
802 switch (op) {
803 CASE_OP_32_64(and):
804 CASE_OP_32_64(mul):
805 CASE_OP_32_64(muluh):
806 CASE_OP_32_64(mulsh):
807 if ((temps[args[2]].state == TCG_TEMP_CONST
808 && temps[args[2]].val == 0)) {
809 s->gen_opc_buf[op_index] = op_to_movi(op);
810 tcg_opt_gen_movi(gen_args, args[0], 0);
811 args += 3;
812 gen_args += 2;
813 continue;
815 break;
816 default:
817 break;
820 /* Simplify expression for "op r, a, a => mov r, a" cases */
821 switch (op) {
822 CASE_OP_32_64(or):
823 CASE_OP_32_64(and):
824 if (temps_are_copies(args[1], args[2])) {
825 if (temps_are_copies(args[0], args[1])) {
826 s->gen_opc_buf[op_index] = INDEX_op_nop;
827 } else {
828 s->gen_opc_buf[op_index] = op_to_mov(op);
829 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
830 gen_args += 2;
832 args += 3;
833 continue;
835 break;
836 default:
837 break;
840 /* Simplify expression for "op r, a, a => movi r, 0" cases */
841 switch (op) {
842 CASE_OP_32_64(sub):
843 CASE_OP_32_64(xor):
844 if (temps_are_copies(args[1], args[2])) {
845 s->gen_opc_buf[op_index] = op_to_movi(op);
846 tcg_opt_gen_movi(gen_args, args[0], 0);
847 gen_args += 2;
848 args += 3;
849 continue;
851 break;
852 default:
853 break;
856 /* Propagate constants through copy operations and do constant
857 folding. Constants will be substituted to arguments by register
858 allocator where needed and possible. Also detect copies. */
859 switch (op) {
860 CASE_OP_32_64(mov):
861 if (temps_are_copies(args[0], args[1])) {
862 args += 2;
863 s->gen_opc_buf[op_index] = INDEX_op_nop;
864 break;
866 if (temps[args[1]].state != TCG_TEMP_CONST) {
867 tcg_opt_gen_mov(s, gen_args, args[0], args[1]);
868 gen_args += 2;
869 args += 2;
870 break;
872 /* Source argument is constant. Rewrite the operation and
873 let movi case handle it. */
874 op = op_to_movi(op);
875 s->gen_opc_buf[op_index] = op;
876 args[1] = temps[args[1]].val;
877 /* fallthrough */
878 CASE_OP_32_64(movi):
879 tcg_opt_gen_movi(gen_args, args[0], args[1]);
880 gen_args += 2;
881 args += 2;
882 break;
884 CASE_OP_32_64(not):
885 CASE_OP_32_64(neg):
886 CASE_OP_32_64(ext8s):
887 CASE_OP_32_64(ext8u):
888 CASE_OP_32_64(ext16s):
889 CASE_OP_32_64(ext16u):
890 case INDEX_op_ext32s_i64:
891 case INDEX_op_ext32u_i64:
892 if (temps[args[1]].state == TCG_TEMP_CONST) {
893 s->gen_opc_buf[op_index] = op_to_movi(op);
894 tmp = do_constant_folding(op, temps[args[1]].val, 0);
895 tcg_opt_gen_movi(gen_args, args[0], tmp);
896 gen_args += 2;
897 args += 2;
898 break;
900 goto do_default;
902 CASE_OP_32_64(add):
903 CASE_OP_32_64(sub):
904 CASE_OP_32_64(mul):
905 CASE_OP_32_64(or):
906 CASE_OP_32_64(and):
907 CASE_OP_32_64(xor):
908 CASE_OP_32_64(shl):
909 CASE_OP_32_64(shr):
910 CASE_OP_32_64(sar):
911 CASE_OP_32_64(rotl):
912 CASE_OP_32_64(rotr):
913 CASE_OP_32_64(andc):
914 CASE_OP_32_64(orc):
915 CASE_OP_32_64(eqv):
916 CASE_OP_32_64(nand):
917 CASE_OP_32_64(nor):
918 CASE_OP_32_64(muluh):
919 CASE_OP_32_64(mulsh):
920 CASE_OP_32_64(div):
921 CASE_OP_32_64(divu):
922 CASE_OP_32_64(rem):
923 CASE_OP_32_64(remu):
924 if (temps[args[1]].state == TCG_TEMP_CONST
925 && temps[args[2]].state == TCG_TEMP_CONST) {
926 s->gen_opc_buf[op_index] = op_to_movi(op);
927 tmp = do_constant_folding(op, temps[args[1]].val,
928 temps[args[2]].val);
929 tcg_opt_gen_movi(gen_args, args[0], tmp);
930 gen_args += 2;
931 args += 3;
932 break;
934 goto do_default;
936 CASE_OP_32_64(deposit):
937 if (temps[args[1]].state == TCG_TEMP_CONST
938 && temps[args[2]].state == TCG_TEMP_CONST) {
939 s->gen_opc_buf[op_index] = op_to_movi(op);
940 tmp = ((1ull << args[4]) - 1);
941 tmp = (temps[args[1]].val & ~(tmp << args[3]))
942 | ((temps[args[2]].val & tmp) << args[3]);
943 tcg_opt_gen_movi(gen_args, args[0], tmp);
944 gen_args += 2;
945 args += 5;
946 break;
948 goto do_default;
950 CASE_OP_32_64(setcond):
951 tmp = do_constant_folding_cond(op, args[1], args[2], args[3]);
952 if (tmp != 2) {
953 s->gen_opc_buf[op_index] = op_to_movi(op);
954 tcg_opt_gen_movi(gen_args, args[0], tmp);
955 gen_args += 2;
956 args += 4;
957 break;
959 goto do_default;
961 CASE_OP_32_64(brcond):
962 tmp = do_constant_folding_cond(op, args[0], args[1], args[2]);
963 if (tmp != 2) {
964 if (tmp) {
965 reset_all_temps(nb_temps);
966 s->gen_opc_buf[op_index] = INDEX_op_br;
967 gen_args[0] = args[3];
968 gen_args += 1;
969 } else {
970 s->gen_opc_buf[op_index] = INDEX_op_nop;
972 args += 4;
973 break;
975 goto do_default;
977 CASE_OP_32_64(movcond):
978 tmp = do_constant_folding_cond(op, args[1], args[2], args[5]);
979 if (tmp != 2) {
980 if (temps_are_copies(args[0], args[4-tmp])) {
981 s->gen_opc_buf[op_index] = INDEX_op_nop;
982 } else if (temps[args[4-tmp]].state == TCG_TEMP_CONST) {
983 s->gen_opc_buf[op_index] = op_to_movi(op);
984 tcg_opt_gen_movi(gen_args, args[0], temps[args[4-tmp]].val);
985 gen_args += 2;
986 } else {
987 s->gen_opc_buf[op_index] = op_to_mov(op);
988 tcg_opt_gen_mov(s, gen_args, args[0], args[4-tmp]);
989 gen_args += 2;
991 args += 6;
992 break;
994 goto do_default;
996 case INDEX_op_add2_i32:
997 case INDEX_op_sub2_i32:
998 if (temps[args[2]].state == TCG_TEMP_CONST
999 && temps[args[3]].state == TCG_TEMP_CONST
1000 && temps[args[4]].state == TCG_TEMP_CONST
1001 && temps[args[5]].state == TCG_TEMP_CONST) {
1002 uint32_t al = temps[args[2]].val;
1003 uint32_t ah = temps[args[3]].val;
1004 uint32_t bl = temps[args[4]].val;
1005 uint32_t bh = temps[args[5]].val;
1006 uint64_t a = ((uint64_t)ah << 32) | al;
1007 uint64_t b = ((uint64_t)bh << 32) | bl;
1008 TCGArg rl, rh;
1010 if (op == INDEX_op_add2_i32) {
1011 a += b;
1012 } else {
1013 a -= b;
1016 /* We emit the extra nop when we emit the add2/sub2. */
1017 assert(s->gen_opc_buf[op_index + 1] == INDEX_op_nop);
1019 rl = args[0];
1020 rh = args[1];
1021 s->gen_opc_buf[op_index] = INDEX_op_movi_i32;
1022 s->gen_opc_buf[++op_index] = INDEX_op_movi_i32;
1023 tcg_opt_gen_movi(&gen_args[0], rl, (uint32_t)a);
1024 tcg_opt_gen_movi(&gen_args[2], rh, (uint32_t)(a >> 32));
1025 gen_args += 4;
1026 args += 6;
1027 break;
1029 goto do_default;
1031 case INDEX_op_mulu2_i32:
1032 if (temps[args[2]].state == TCG_TEMP_CONST
1033 && temps[args[3]].state == TCG_TEMP_CONST) {
1034 uint32_t a = temps[args[2]].val;
1035 uint32_t b = temps[args[3]].val;
1036 uint64_t r = (uint64_t)a * b;
1037 TCGArg rl, rh;
1039 /* We emit the extra nop when we emit the mulu2. */
1040 assert(s->gen_opc_buf[op_index + 1] == INDEX_op_nop);
1042 rl = args[0];
1043 rh = args[1];
1044 s->gen_opc_buf[op_index] = INDEX_op_movi_i32;
1045 s->gen_opc_buf[++op_index] = INDEX_op_movi_i32;
1046 tcg_opt_gen_movi(&gen_args[0], rl, (uint32_t)r);
1047 tcg_opt_gen_movi(&gen_args[2], rh, (uint32_t)(r >> 32));
1048 gen_args += 4;
1049 args += 4;
1050 break;
1052 goto do_default;
1054 case INDEX_op_brcond2_i32:
1055 tmp = do_constant_folding_cond2(&args[0], &args[2], args[4]);
1056 if (tmp != 2) {
1057 if (tmp) {
1058 reset_all_temps(nb_temps);
1059 s->gen_opc_buf[op_index] = INDEX_op_br;
1060 gen_args[0] = args[5];
1061 gen_args += 1;
1062 } else {
1063 s->gen_opc_buf[op_index] = INDEX_op_nop;
1065 } else if ((args[4] == TCG_COND_LT || args[4] == TCG_COND_GE)
1066 && temps[args[2]].state == TCG_TEMP_CONST
1067 && temps[args[3]].state == TCG_TEMP_CONST
1068 && temps[args[2]].val == 0
1069 && temps[args[3]].val == 0) {
1070 /* Simplify LT/GE comparisons vs zero to a single compare
1071 vs the high word of the input. */
1072 reset_all_temps(nb_temps);
1073 s->gen_opc_buf[op_index] = INDEX_op_brcond_i32;
1074 gen_args[0] = args[1];
1075 gen_args[1] = args[3];
1076 gen_args[2] = args[4];
1077 gen_args[3] = args[5];
1078 gen_args += 4;
1079 } else {
1080 goto do_default;
1082 args += 6;
1083 break;
1085 case INDEX_op_setcond2_i32:
1086 tmp = do_constant_folding_cond2(&args[1], &args[3], args[5]);
1087 if (tmp != 2) {
1088 s->gen_opc_buf[op_index] = INDEX_op_movi_i32;
1089 tcg_opt_gen_movi(gen_args, args[0], tmp);
1090 gen_args += 2;
1091 } else if ((args[5] == TCG_COND_LT || args[5] == TCG_COND_GE)
1092 && temps[args[3]].state == TCG_TEMP_CONST
1093 && temps[args[4]].state == TCG_TEMP_CONST
1094 && temps[args[3]].val == 0
1095 && temps[args[4]].val == 0) {
1096 /* Simplify LT/GE comparisons vs zero to a single compare
1097 vs the high word of the input. */
1098 s->gen_opc_buf[op_index] = INDEX_op_setcond_i32;
1099 reset_temp(args[0]);
1100 gen_args[0] = args[0];
1101 gen_args[1] = args[2];
1102 gen_args[2] = args[4];
1103 gen_args[3] = args[5];
1104 gen_args += 4;
1105 } else {
1106 goto do_default;
1108 args += 6;
1109 break;
1111 case INDEX_op_call:
1112 nb_call_args = (args[0] >> 16) + (args[0] & 0xffff);
1113 if (!(args[nb_call_args + 1] & (TCG_CALL_NO_READ_GLOBALS |
1114 TCG_CALL_NO_WRITE_GLOBALS))) {
1115 for (i = 0; i < nb_globals; i++) {
1116 reset_temp(i);
1119 for (i = 0; i < (args[0] >> 16); i++) {
1120 reset_temp(args[i + 1]);
1122 i = nb_call_args + 3;
1123 while (i) {
1124 *gen_args = *args;
1125 args++;
1126 gen_args++;
1127 i--;
1129 break;
1131 default:
1132 do_default:
1133 /* Default case: we know nothing about operation (or were unable
1134 to compute the operation result) so no propagation is done.
1135 We trash everything if the operation is the end of a basic
1136 block, otherwise we only trash the output args. "mask" is
1137 the non-zero bits mask for the first output arg. */
1138 if (def->flags & TCG_OPF_BB_END) {
1139 reset_all_temps(nb_temps);
1140 } else {
1141 for (i = 0; i < def->nb_oargs; i++) {
1142 reset_temp(args[i]);
1145 for (i = 0; i < def->nb_args; i++) {
1146 gen_args[i] = args[i];
1148 args += def->nb_args;
1149 gen_args += def->nb_args;
1150 break;
1154 return gen_args;
1157 TCGArg *tcg_optimize(TCGContext *s, uint16_t *tcg_opc_ptr,
1158 TCGArg *args, TCGOpDef *tcg_op_defs)
1160 TCGArg *res;
1161 res = tcg_constant_folding(s, tcg_opc_ptr, args, tcg_op_defs);
1162 return res;