Merge remote-tracking branch 'remotes/stefanberger/tags/pull-tpm-2017-10-04-3' into...
[qemu/kevin.git] / tcg / optimize.c
blobadfc56ce62023ac77a0b3d62fdfd28be6089ae27
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 "qemu/osdep.h"
27 #include "qemu-common.h"
28 #include "exec/cpu-common.h"
29 #include "tcg-op.h"
31 #define CASE_OP_32_64(x) \
32 glue(glue(case INDEX_op_, x), _i32): \
33 glue(glue(case INDEX_op_, x), _i64)
35 struct tcg_temp_info {
36 bool is_const;
37 uint16_t prev_copy;
38 uint16_t next_copy;
39 tcg_target_ulong val;
40 tcg_target_ulong mask;
43 static struct tcg_temp_info temps[TCG_MAX_TEMPS];
44 static TCGTempSet temps_used;
46 static inline bool temp_is_const(TCGArg arg)
48 return temps[arg].is_const;
51 static inline bool temp_is_copy(TCGArg arg)
53 return temps[arg].next_copy != arg;
56 /* Reset TEMP's state, possibly removing the temp for the list of copies. */
57 static void reset_temp(TCGArg temp)
59 temps[temps[temp].next_copy].prev_copy = temps[temp].prev_copy;
60 temps[temps[temp].prev_copy].next_copy = temps[temp].next_copy;
61 temps[temp].next_copy = temp;
62 temps[temp].prev_copy = temp;
63 temps[temp].is_const = false;
64 temps[temp].mask = -1;
67 /* Reset all temporaries, given that there are NB_TEMPS of them. */
68 static void reset_all_temps(int nb_temps)
70 bitmap_zero(temps_used.l, nb_temps);
73 /* Initialize and activate a temporary. */
74 static void init_temp_info(TCGArg temp)
76 if (!test_bit(temp, temps_used.l)) {
77 temps[temp].next_copy = temp;
78 temps[temp].prev_copy = temp;
79 temps[temp].is_const = false;
80 temps[temp].mask = -1;
81 set_bit(temp, temps_used.l);
85 static int op_bits(TCGOpcode op)
87 const TCGOpDef *def = &tcg_op_defs[op];
88 return def->flags & TCG_OPF_64BIT ? 64 : 32;
91 static TCGOpcode op_to_mov(TCGOpcode op)
93 switch (op_bits(op)) {
94 case 32:
95 return INDEX_op_mov_i32;
96 case 64:
97 return INDEX_op_mov_i64;
98 default:
99 fprintf(stderr, "op_to_mov: unexpected return value of "
100 "function op_bits.\n");
101 tcg_abort();
105 static TCGOpcode op_to_movi(TCGOpcode op)
107 switch (op_bits(op)) {
108 case 32:
109 return INDEX_op_movi_i32;
110 case 64:
111 return INDEX_op_movi_i64;
112 default:
113 fprintf(stderr, "op_to_movi: unexpected return value of "
114 "function op_bits.\n");
115 tcg_abort();
119 static TCGArg find_better_copy(TCGContext *s, TCGArg temp)
121 TCGArg i;
123 /* If this is already a global, we can't do better. */
124 if (temp < s->nb_globals) {
125 return temp;
128 /* Search for a global first. */
129 for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) {
130 if (i < s->nb_globals) {
131 return i;
135 /* If it is a temp, search for a temp local. */
136 if (!s->temps[temp].temp_local) {
137 for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) {
138 if (s->temps[i].temp_local) {
139 return i;
144 /* Failure to find a better representation, return the same temp. */
145 return temp;
148 static bool temps_are_copies(TCGArg arg1, TCGArg arg2)
150 TCGArg i;
152 if (arg1 == arg2) {
153 return true;
156 if (!temp_is_copy(arg1) || !temp_is_copy(arg2)) {
157 return false;
160 for (i = temps[arg1].next_copy ; i != arg1 ; i = temps[i].next_copy) {
161 if (i == arg2) {
162 return true;
166 return false;
169 static void tcg_opt_gen_movi(TCGContext *s, TCGOp *op, TCGArg *args,
170 TCGArg dst, TCGArg val)
172 TCGOpcode new_op = op_to_movi(op->opc);
173 tcg_target_ulong mask;
175 op->opc = new_op;
177 reset_temp(dst);
178 temps[dst].is_const = true;
179 temps[dst].val = val;
180 mask = val;
181 if (TCG_TARGET_REG_BITS > 32 && new_op == INDEX_op_movi_i32) {
182 /* High bits of the destination are now garbage. */
183 mask |= ~0xffffffffull;
185 temps[dst].mask = mask;
187 args[0] = dst;
188 args[1] = val;
191 static void tcg_opt_gen_mov(TCGContext *s, TCGOp *op, TCGArg *args,
192 TCGArg dst, TCGArg src)
194 if (temps_are_copies(dst, src)) {
195 tcg_op_remove(s, op);
196 return;
199 TCGOpcode new_op = op_to_mov(op->opc);
200 tcg_target_ulong mask;
202 op->opc = new_op;
204 reset_temp(dst);
205 mask = temps[src].mask;
206 if (TCG_TARGET_REG_BITS > 32 && new_op == INDEX_op_mov_i32) {
207 /* High bits of the destination are now garbage. */
208 mask |= ~0xffffffffull;
210 temps[dst].mask = mask;
212 if (s->temps[src].type == s->temps[dst].type) {
213 temps[dst].next_copy = temps[src].next_copy;
214 temps[dst].prev_copy = src;
215 temps[temps[dst].next_copy].prev_copy = dst;
216 temps[src].next_copy = dst;
217 temps[dst].is_const = temps[src].is_const;
218 temps[dst].val = temps[src].val;
221 args[0] = dst;
222 args[1] = src;
225 static TCGArg do_constant_folding_2(TCGOpcode op, TCGArg x, TCGArg y)
227 uint64_t l64, h64;
229 switch (op) {
230 CASE_OP_32_64(add):
231 return x + y;
233 CASE_OP_32_64(sub):
234 return x - y;
236 CASE_OP_32_64(mul):
237 return x * y;
239 CASE_OP_32_64(and):
240 return x & y;
242 CASE_OP_32_64(or):
243 return x | y;
245 CASE_OP_32_64(xor):
246 return x ^ y;
248 case INDEX_op_shl_i32:
249 return (uint32_t)x << (y & 31);
251 case INDEX_op_shl_i64:
252 return (uint64_t)x << (y & 63);
254 case INDEX_op_shr_i32:
255 return (uint32_t)x >> (y & 31);
257 case INDEX_op_shr_i64:
258 return (uint64_t)x >> (y & 63);
260 case INDEX_op_sar_i32:
261 return (int32_t)x >> (y & 31);
263 case INDEX_op_sar_i64:
264 return (int64_t)x >> (y & 63);
266 case INDEX_op_rotr_i32:
267 return ror32(x, y & 31);
269 case INDEX_op_rotr_i64:
270 return ror64(x, y & 63);
272 case INDEX_op_rotl_i32:
273 return rol32(x, y & 31);
275 case INDEX_op_rotl_i64:
276 return rol64(x, y & 63);
278 CASE_OP_32_64(not):
279 return ~x;
281 CASE_OP_32_64(neg):
282 return -x;
284 CASE_OP_32_64(andc):
285 return x & ~y;
287 CASE_OP_32_64(orc):
288 return x | ~y;
290 CASE_OP_32_64(eqv):
291 return ~(x ^ y);
293 CASE_OP_32_64(nand):
294 return ~(x & y);
296 CASE_OP_32_64(nor):
297 return ~(x | y);
299 case INDEX_op_clz_i32:
300 return (uint32_t)x ? clz32(x) : y;
302 case INDEX_op_clz_i64:
303 return x ? clz64(x) : y;
305 case INDEX_op_ctz_i32:
306 return (uint32_t)x ? ctz32(x) : y;
308 case INDEX_op_ctz_i64:
309 return x ? ctz64(x) : y;
311 case INDEX_op_ctpop_i32:
312 return ctpop32(x);
314 case INDEX_op_ctpop_i64:
315 return ctpop64(x);
317 CASE_OP_32_64(ext8s):
318 return (int8_t)x;
320 CASE_OP_32_64(ext16s):
321 return (int16_t)x;
323 CASE_OP_32_64(ext8u):
324 return (uint8_t)x;
326 CASE_OP_32_64(ext16u):
327 return (uint16_t)x;
329 case INDEX_op_ext_i32_i64:
330 case INDEX_op_ext32s_i64:
331 return (int32_t)x;
333 case INDEX_op_extu_i32_i64:
334 case INDEX_op_extrl_i64_i32:
335 case INDEX_op_ext32u_i64:
336 return (uint32_t)x;
338 case INDEX_op_extrh_i64_i32:
339 return (uint64_t)x >> 32;
341 case INDEX_op_muluh_i32:
342 return ((uint64_t)(uint32_t)x * (uint32_t)y) >> 32;
343 case INDEX_op_mulsh_i32:
344 return ((int64_t)(int32_t)x * (int32_t)y) >> 32;
346 case INDEX_op_muluh_i64:
347 mulu64(&l64, &h64, x, y);
348 return h64;
349 case INDEX_op_mulsh_i64:
350 muls64(&l64, &h64, x, y);
351 return h64;
353 case INDEX_op_div_i32:
354 /* Avoid crashing on divide by zero, otherwise undefined. */
355 return (int32_t)x / ((int32_t)y ? : 1);
356 case INDEX_op_divu_i32:
357 return (uint32_t)x / ((uint32_t)y ? : 1);
358 case INDEX_op_div_i64:
359 return (int64_t)x / ((int64_t)y ? : 1);
360 case INDEX_op_divu_i64:
361 return (uint64_t)x / ((uint64_t)y ? : 1);
363 case INDEX_op_rem_i32:
364 return (int32_t)x % ((int32_t)y ? : 1);
365 case INDEX_op_remu_i32:
366 return (uint32_t)x % ((uint32_t)y ? : 1);
367 case INDEX_op_rem_i64:
368 return (int64_t)x % ((int64_t)y ? : 1);
369 case INDEX_op_remu_i64:
370 return (uint64_t)x % ((uint64_t)y ? : 1);
372 default:
373 fprintf(stderr,
374 "Unrecognized operation %d in do_constant_folding.\n", op);
375 tcg_abort();
379 static TCGArg do_constant_folding(TCGOpcode op, TCGArg x, TCGArg y)
381 TCGArg res = do_constant_folding_2(op, x, y);
382 if (op_bits(op) == 32) {
383 res = (int32_t)res;
385 return res;
388 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c)
390 switch (c) {
391 case TCG_COND_EQ:
392 return x == y;
393 case TCG_COND_NE:
394 return x != y;
395 case TCG_COND_LT:
396 return (int32_t)x < (int32_t)y;
397 case TCG_COND_GE:
398 return (int32_t)x >= (int32_t)y;
399 case TCG_COND_LE:
400 return (int32_t)x <= (int32_t)y;
401 case TCG_COND_GT:
402 return (int32_t)x > (int32_t)y;
403 case TCG_COND_LTU:
404 return x < y;
405 case TCG_COND_GEU:
406 return x >= y;
407 case TCG_COND_LEU:
408 return x <= y;
409 case TCG_COND_GTU:
410 return x > y;
411 default:
412 tcg_abort();
416 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c)
418 switch (c) {
419 case TCG_COND_EQ:
420 return x == y;
421 case TCG_COND_NE:
422 return x != y;
423 case TCG_COND_LT:
424 return (int64_t)x < (int64_t)y;
425 case TCG_COND_GE:
426 return (int64_t)x >= (int64_t)y;
427 case TCG_COND_LE:
428 return (int64_t)x <= (int64_t)y;
429 case TCG_COND_GT:
430 return (int64_t)x > (int64_t)y;
431 case TCG_COND_LTU:
432 return x < y;
433 case TCG_COND_GEU:
434 return x >= y;
435 case TCG_COND_LEU:
436 return x <= y;
437 case TCG_COND_GTU:
438 return x > y;
439 default:
440 tcg_abort();
444 static bool do_constant_folding_cond_eq(TCGCond c)
446 switch (c) {
447 case TCG_COND_GT:
448 case TCG_COND_LTU:
449 case TCG_COND_LT:
450 case TCG_COND_GTU:
451 case TCG_COND_NE:
452 return 0;
453 case TCG_COND_GE:
454 case TCG_COND_GEU:
455 case TCG_COND_LE:
456 case TCG_COND_LEU:
457 case TCG_COND_EQ:
458 return 1;
459 default:
460 tcg_abort();
464 /* Return 2 if the condition can't be simplified, and the result
465 of the condition (0 or 1) if it can */
466 static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x,
467 TCGArg y, TCGCond c)
469 if (temp_is_const(x) && temp_is_const(y)) {
470 switch (op_bits(op)) {
471 case 32:
472 return do_constant_folding_cond_32(temps[x].val, temps[y].val, c);
473 case 64:
474 return do_constant_folding_cond_64(temps[x].val, temps[y].val, c);
475 default:
476 tcg_abort();
478 } else if (temps_are_copies(x, y)) {
479 return do_constant_folding_cond_eq(c);
480 } else if (temp_is_const(y) && temps[y].val == 0) {
481 switch (c) {
482 case TCG_COND_LTU:
483 return 0;
484 case TCG_COND_GEU:
485 return 1;
486 default:
487 return 2;
490 return 2;
493 /* Return 2 if the condition can't be simplified, and the result
494 of the condition (0 or 1) if it can */
495 static TCGArg do_constant_folding_cond2(TCGArg *p1, TCGArg *p2, TCGCond c)
497 TCGArg al = p1[0], ah = p1[1];
498 TCGArg bl = p2[0], bh = p2[1];
500 if (temp_is_const(bl) && temp_is_const(bh)) {
501 uint64_t b = ((uint64_t)temps[bh].val << 32) | (uint32_t)temps[bl].val;
503 if (temp_is_const(al) && temp_is_const(ah)) {
504 uint64_t a;
505 a = ((uint64_t)temps[ah].val << 32) | (uint32_t)temps[al].val;
506 return do_constant_folding_cond_64(a, b, c);
508 if (b == 0) {
509 switch (c) {
510 case TCG_COND_LTU:
511 return 0;
512 case TCG_COND_GEU:
513 return 1;
514 default:
515 break;
519 if (temps_are_copies(al, bl) && temps_are_copies(ah, bh)) {
520 return do_constant_folding_cond_eq(c);
522 return 2;
525 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
527 TCGArg a1 = *p1, a2 = *p2;
528 int sum = 0;
529 sum += temp_is_const(a1);
530 sum -= temp_is_const(a2);
532 /* Prefer the constant in second argument, and then the form
533 op a, a, b, which is better handled on non-RISC hosts. */
534 if (sum > 0 || (sum == 0 && dest == a2)) {
535 *p1 = a2;
536 *p2 = a1;
537 return true;
539 return false;
542 static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
544 int sum = 0;
545 sum += temp_is_const(p1[0]);
546 sum += temp_is_const(p1[1]);
547 sum -= temp_is_const(p2[0]);
548 sum -= temp_is_const(p2[1]);
549 if (sum > 0) {
550 TCGArg t;
551 t = p1[0], p1[0] = p2[0], p2[0] = t;
552 t = p1[1], p1[1] = p2[1], p2[1] = t;
553 return true;
555 return false;
558 /* Propagate constants and copies, fold constant expressions. */
559 void tcg_optimize(TCGContext *s)
561 int oi, oi_next, nb_temps, nb_globals;
562 TCGArg *prev_mb_args = NULL;
564 /* Array VALS has an element for each temp.
565 If this temp holds a constant then its value is kept in VALS' element.
566 If this temp is a copy of other ones then the other copies are
567 available through the doubly linked circular list. */
569 nb_temps = s->nb_temps;
570 nb_globals = s->nb_globals;
571 reset_all_temps(nb_temps);
573 for (oi = s->gen_op_buf[0].next; oi != 0; oi = oi_next) {
574 tcg_target_ulong mask, partmask, affected;
575 int nb_oargs, nb_iargs, i;
576 TCGArg tmp;
578 TCGOp * const op = &s->gen_op_buf[oi];
579 TCGArg * const args = &s->gen_opparam_buf[op->args];
580 TCGOpcode opc = op->opc;
581 const TCGOpDef *def = &tcg_op_defs[opc];
583 oi_next = op->next;
585 /* Count the arguments, and initialize the temps that are
586 going to be used */
587 if (opc == INDEX_op_call) {
588 nb_oargs = op->callo;
589 nb_iargs = op->calli;
590 for (i = 0; i < nb_oargs + nb_iargs; i++) {
591 tmp = args[i];
592 if (tmp != TCG_CALL_DUMMY_ARG) {
593 init_temp_info(tmp);
596 } else {
597 nb_oargs = def->nb_oargs;
598 nb_iargs = def->nb_iargs;
599 for (i = 0; i < nb_oargs + nb_iargs; i++) {
600 init_temp_info(args[i]);
604 /* Do copy propagation */
605 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
606 if (temp_is_copy(args[i])) {
607 args[i] = find_better_copy(s, args[i]);
611 /* For commutative operations make constant second argument */
612 switch (opc) {
613 CASE_OP_32_64(add):
614 CASE_OP_32_64(mul):
615 CASE_OP_32_64(and):
616 CASE_OP_32_64(or):
617 CASE_OP_32_64(xor):
618 CASE_OP_32_64(eqv):
619 CASE_OP_32_64(nand):
620 CASE_OP_32_64(nor):
621 CASE_OP_32_64(muluh):
622 CASE_OP_32_64(mulsh):
623 swap_commutative(args[0], &args[1], &args[2]);
624 break;
625 CASE_OP_32_64(brcond):
626 if (swap_commutative(-1, &args[0], &args[1])) {
627 args[2] = tcg_swap_cond(args[2]);
629 break;
630 CASE_OP_32_64(setcond):
631 if (swap_commutative(args[0], &args[1], &args[2])) {
632 args[3] = tcg_swap_cond(args[3]);
634 break;
635 CASE_OP_32_64(movcond):
636 if (swap_commutative(-1, &args[1], &args[2])) {
637 args[5] = tcg_swap_cond(args[5]);
639 /* For movcond, we canonicalize the "false" input reg to match
640 the destination reg so that the tcg backend can implement
641 a "move if true" operation. */
642 if (swap_commutative(args[0], &args[4], &args[3])) {
643 args[5] = tcg_invert_cond(args[5]);
645 break;
646 CASE_OP_32_64(add2):
647 swap_commutative(args[0], &args[2], &args[4]);
648 swap_commutative(args[1], &args[3], &args[5]);
649 break;
650 CASE_OP_32_64(mulu2):
651 CASE_OP_32_64(muls2):
652 swap_commutative(args[0], &args[2], &args[3]);
653 break;
654 case INDEX_op_brcond2_i32:
655 if (swap_commutative2(&args[0], &args[2])) {
656 args[4] = tcg_swap_cond(args[4]);
658 break;
659 case INDEX_op_setcond2_i32:
660 if (swap_commutative2(&args[1], &args[3])) {
661 args[5] = tcg_swap_cond(args[5]);
663 break;
664 default:
665 break;
668 /* Simplify expressions for "shift/rot r, 0, a => movi r, 0",
669 and "sub r, 0, a => neg r, a" case. */
670 switch (opc) {
671 CASE_OP_32_64(shl):
672 CASE_OP_32_64(shr):
673 CASE_OP_32_64(sar):
674 CASE_OP_32_64(rotl):
675 CASE_OP_32_64(rotr):
676 if (temp_is_const(args[1]) && temps[args[1]].val == 0) {
677 tcg_opt_gen_movi(s, op, args, args[0], 0);
678 continue;
680 break;
681 CASE_OP_32_64(sub):
683 TCGOpcode neg_op;
684 bool have_neg;
686 if (temp_is_const(args[2])) {
687 /* Proceed with possible constant folding. */
688 break;
690 if (opc == INDEX_op_sub_i32) {
691 neg_op = INDEX_op_neg_i32;
692 have_neg = TCG_TARGET_HAS_neg_i32;
693 } else {
694 neg_op = INDEX_op_neg_i64;
695 have_neg = TCG_TARGET_HAS_neg_i64;
697 if (!have_neg) {
698 break;
700 if (temp_is_const(args[1]) && temps[args[1]].val == 0) {
701 op->opc = neg_op;
702 reset_temp(args[0]);
703 args[1] = args[2];
704 continue;
707 break;
708 CASE_OP_32_64(xor):
709 CASE_OP_32_64(nand):
710 if (!temp_is_const(args[1])
711 && temp_is_const(args[2]) && temps[args[2]].val == -1) {
712 i = 1;
713 goto try_not;
715 break;
716 CASE_OP_32_64(nor):
717 if (!temp_is_const(args[1])
718 && temp_is_const(args[2]) && temps[args[2]].val == 0) {
719 i = 1;
720 goto try_not;
722 break;
723 CASE_OP_32_64(andc):
724 if (!temp_is_const(args[2])
725 && temp_is_const(args[1]) && temps[args[1]].val == -1) {
726 i = 2;
727 goto try_not;
729 break;
730 CASE_OP_32_64(orc):
731 CASE_OP_32_64(eqv):
732 if (!temp_is_const(args[2])
733 && temp_is_const(args[1]) && temps[args[1]].val == 0) {
734 i = 2;
735 goto try_not;
737 break;
738 try_not:
740 TCGOpcode not_op;
741 bool have_not;
743 if (def->flags & TCG_OPF_64BIT) {
744 not_op = INDEX_op_not_i64;
745 have_not = TCG_TARGET_HAS_not_i64;
746 } else {
747 not_op = INDEX_op_not_i32;
748 have_not = TCG_TARGET_HAS_not_i32;
750 if (!have_not) {
751 break;
753 op->opc = not_op;
754 reset_temp(args[0]);
755 args[1] = args[i];
756 continue;
758 default:
759 break;
762 /* Simplify expression for "op r, a, const => mov r, a" cases */
763 switch (opc) {
764 CASE_OP_32_64(add):
765 CASE_OP_32_64(sub):
766 CASE_OP_32_64(shl):
767 CASE_OP_32_64(shr):
768 CASE_OP_32_64(sar):
769 CASE_OP_32_64(rotl):
770 CASE_OP_32_64(rotr):
771 CASE_OP_32_64(or):
772 CASE_OP_32_64(xor):
773 CASE_OP_32_64(andc):
774 if (!temp_is_const(args[1])
775 && temp_is_const(args[2]) && temps[args[2]].val == 0) {
776 tcg_opt_gen_mov(s, op, args, args[0], args[1]);
777 continue;
779 break;
780 CASE_OP_32_64(and):
781 CASE_OP_32_64(orc):
782 CASE_OP_32_64(eqv):
783 if (!temp_is_const(args[1])
784 && temp_is_const(args[2]) && temps[args[2]].val == -1) {
785 tcg_opt_gen_mov(s, op, args, args[0], args[1]);
786 continue;
788 break;
789 default:
790 break;
793 /* Simplify using known-zero bits. Currently only ops with a single
794 output argument is supported. */
795 mask = -1;
796 affected = -1;
797 switch (opc) {
798 CASE_OP_32_64(ext8s):
799 if ((temps[args[1]].mask & 0x80) != 0) {
800 break;
802 CASE_OP_32_64(ext8u):
803 mask = 0xff;
804 goto and_const;
805 CASE_OP_32_64(ext16s):
806 if ((temps[args[1]].mask & 0x8000) != 0) {
807 break;
809 CASE_OP_32_64(ext16u):
810 mask = 0xffff;
811 goto and_const;
812 case INDEX_op_ext32s_i64:
813 if ((temps[args[1]].mask & 0x80000000) != 0) {
814 break;
816 case INDEX_op_ext32u_i64:
817 mask = 0xffffffffU;
818 goto and_const;
820 CASE_OP_32_64(and):
821 mask = temps[args[2]].mask;
822 if (temp_is_const(args[2])) {
823 and_const:
824 affected = temps[args[1]].mask & ~mask;
826 mask = temps[args[1]].mask & mask;
827 break;
829 case INDEX_op_ext_i32_i64:
830 if ((temps[args[1]].mask & 0x80000000) != 0) {
831 break;
833 case INDEX_op_extu_i32_i64:
834 /* We do not compute affected as it is a size changing op. */
835 mask = (uint32_t)temps[args[1]].mask;
836 break;
838 CASE_OP_32_64(andc):
839 /* Known-zeros does not imply known-ones. Therefore unless
840 args[2] is constant, we can't infer anything from it. */
841 if (temp_is_const(args[2])) {
842 mask = ~temps[args[2]].mask;
843 goto and_const;
845 /* But we certainly know nothing outside args[1] may be set. */
846 mask = temps[args[1]].mask;
847 break;
849 case INDEX_op_sar_i32:
850 if (temp_is_const(args[2])) {
851 tmp = temps[args[2]].val & 31;
852 mask = (int32_t)temps[args[1]].mask >> tmp;
854 break;
855 case INDEX_op_sar_i64:
856 if (temp_is_const(args[2])) {
857 tmp = temps[args[2]].val & 63;
858 mask = (int64_t)temps[args[1]].mask >> tmp;
860 break;
862 case INDEX_op_shr_i32:
863 if (temp_is_const(args[2])) {
864 tmp = temps[args[2]].val & 31;
865 mask = (uint32_t)temps[args[1]].mask >> tmp;
867 break;
868 case INDEX_op_shr_i64:
869 if (temp_is_const(args[2])) {
870 tmp = temps[args[2]].val & 63;
871 mask = (uint64_t)temps[args[1]].mask >> tmp;
873 break;
875 case INDEX_op_extrl_i64_i32:
876 mask = (uint32_t)temps[args[1]].mask;
877 break;
878 case INDEX_op_extrh_i64_i32:
879 mask = (uint64_t)temps[args[1]].mask >> 32;
880 break;
882 CASE_OP_32_64(shl):
883 if (temp_is_const(args[2])) {
884 tmp = temps[args[2]].val & (TCG_TARGET_REG_BITS - 1);
885 mask = temps[args[1]].mask << tmp;
887 break;
889 CASE_OP_32_64(neg):
890 /* Set to 1 all bits to the left of the rightmost. */
891 mask = -(temps[args[1]].mask & -temps[args[1]].mask);
892 break;
894 CASE_OP_32_64(deposit):
895 mask = deposit64(temps[args[1]].mask, args[3], args[4],
896 temps[args[2]].mask);
897 break;
899 CASE_OP_32_64(extract):
900 mask = extract64(temps[args[1]].mask, args[2], args[3]);
901 if (args[2] == 0) {
902 affected = temps[args[1]].mask & ~mask;
904 break;
905 CASE_OP_32_64(sextract):
906 mask = sextract64(temps[args[1]].mask, args[2], args[3]);
907 if (args[2] == 0 && (tcg_target_long)mask >= 0) {
908 affected = temps[args[1]].mask & ~mask;
910 break;
912 CASE_OP_32_64(or):
913 CASE_OP_32_64(xor):
914 mask = temps[args[1]].mask | temps[args[2]].mask;
915 break;
917 case INDEX_op_clz_i32:
918 case INDEX_op_ctz_i32:
919 mask = temps[args[2]].mask | 31;
920 break;
922 case INDEX_op_clz_i64:
923 case INDEX_op_ctz_i64:
924 mask = temps[args[2]].mask | 63;
925 break;
927 case INDEX_op_ctpop_i32:
928 mask = 32 | 31;
929 break;
930 case INDEX_op_ctpop_i64:
931 mask = 64 | 63;
932 break;
934 CASE_OP_32_64(setcond):
935 case INDEX_op_setcond2_i32:
936 mask = 1;
937 break;
939 CASE_OP_32_64(movcond):
940 mask = temps[args[3]].mask | temps[args[4]].mask;
941 break;
943 CASE_OP_32_64(ld8u):
944 mask = 0xff;
945 break;
946 CASE_OP_32_64(ld16u):
947 mask = 0xffff;
948 break;
949 case INDEX_op_ld32u_i64:
950 mask = 0xffffffffu;
951 break;
953 CASE_OP_32_64(qemu_ld):
955 TCGMemOpIdx oi = args[nb_oargs + nb_iargs];
956 TCGMemOp mop = get_memop(oi);
957 if (!(mop & MO_SIGN)) {
958 mask = (2ULL << ((8 << (mop & MO_SIZE)) - 1)) - 1;
961 break;
963 default:
964 break;
967 /* 32-bit ops generate 32-bit results. For the result is zero test
968 below, we can ignore high bits, but for further optimizations we
969 need to record that the high bits contain garbage. */
970 partmask = mask;
971 if (!(def->flags & TCG_OPF_64BIT)) {
972 mask |= ~(tcg_target_ulong)0xffffffffu;
973 partmask &= 0xffffffffu;
974 affected &= 0xffffffffu;
977 if (partmask == 0) {
978 tcg_debug_assert(nb_oargs == 1);
979 tcg_opt_gen_movi(s, op, args, args[0], 0);
980 continue;
982 if (affected == 0) {
983 tcg_debug_assert(nb_oargs == 1);
984 tcg_opt_gen_mov(s, op, args, args[0], args[1]);
985 continue;
988 /* Simplify expression for "op r, a, 0 => movi r, 0" cases */
989 switch (opc) {
990 CASE_OP_32_64(and):
991 CASE_OP_32_64(mul):
992 CASE_OP_32_64(muluh):
993 CASE_OP_32_64(mulsh):
994 if ((temp_is_const(args[2]) && temps[args[2]].val == 0)) {
995 tcg_opt_gen_movi(s, op, args, args[0], 0);
996 continue;
998 break;
999 default:
1000 break;
1003 /* Simplify expression for "op r, a, a => mov r, a" cases */
1004 switch (opc) {
1005 CASE_OP_32_64(or):
1006 CASE_OP_32_64(and):
1007 if (temps_are_copies(args[1], args[2])) {
1008 tcg_opt_gen_mov(s, op, args, args[0], args[1]);
1009 continue;
1011 break;
1012 default:
1013 break;
1016 /* Simplify expression for "op r, a, a => movi r, 0" cases */
1017 switch (opc) {
1018 CASE_OP_32_64(andc):
1019 CASE_OP_32_64(sub):
1020 CASE_OP_32_64(xor):
1021 if (temps_are_copies(args[1], args[2])) {
1022 tcg_opt_gen_movi(s, op, args, args[0], 0);
1023 continue;
1025 break;
1026 default:
1027 break;
1030 /* Propagate constants through copy operations and do constant
1031 folding. Constants will be substituted to arguments by register
1032 allocator where needed and possible. Also detect copies. */
1033 switch (opc) {
1034 CASE_OP_32_64(mov):
1035 tcg_opt_gen_mov(s, op, args, args[0], args[1]);
1036 break;
1037 CASE_OP_32_64(movi):
1038 tcg_opt_gen_movi(s, op, args, args[0], args[1]);
1039 break;
1041 CASE_OP_32_64(not):
1042 CASE_OP_32_64(neg):
1043 CASE_OP_32_64(ext8s):
1044 CASE_OP_32_64(ext8u):
1045 CASE_OP_32_64(ext16s):
1046 CASE_OP_32_64(ext16u):
1047 CASE_OP_32_64(ctpop):
1048 case INDEX_op_ext32s_i64:
1049 case INDEX_op_ext32u_i64:
1050 case INDEX_op_ext_i32_i64:
1051 case INDEX_op_extu_i32_i64:
1052 case INDEX_op_extrl_i64_i32:
1053 case INDEX_op_extrh_i64_i32:
1054 if (temp_is_const(args[1])) {
1055 tmp = do_constant_folding(opc, temps[args[1]].val, 0);
1056 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1057 break;
1059 goto do_default;
1061 CASE_OP_32_64(add):
1062 CASE_OP_32_64(sub):
1063 CASE_OP_32_64(mul):
1064 CASE_OP_32_64(or):
1065 CASE_OP_32_64(and):
1066 CASE_OP_32_64(xor):
1067 CASE_OP_32_64(shl):
1068 CASE_OP_32_64(shr):
1069 CASE_OP_32_64(sar):
1070 CASE_OP_32_64(rotl):
1071 CASE_OP_32_64(rotr):
1072 CASE_OP_32_64(andc):
1073 CASE_OP_32_64(orc):
1074 CASE_OP_32_64(eqv):
1075 CASE_OP_32_64(nand):
1076 CASE_OP_32_64(nor):
1077 CASE_OP_32_64(muluh):
1078 CASE_OP_32_64(mulsh):
1079 CASE_OP_32_64(div):
1080 CASE_OP_32_64(divu):
1081 CASE_OP_32_64(rem):
1082 CASE_OP_32_64(remu):
1083 if (temp_is_const(args[1]) && temp_is_const(args[2])) {
1084 tmp = do_constant_folding(opc, temps[args[1]].val,
1085 temps[args[2]].val);
1086 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1087 break;
1089 goto do_default;
1091 CASE_OP_32_64(clz):
1092 CASE_OP_32_64(ctz):
1093 if (temp_is_const(args[1])) {
1094 TCGArg v = temps[args[1]].val;
1095 if (v != 0) {
1096 tmp = do_constant_folding(opc, v, 0);
1097 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1098 } else {
1099 tcg_opt_gen_mov(s, op, args, args[0], args[2]);
1101 break;
1103 goto do_default;
1105 CASE_OP_32_64(deposit):
1106 if (temp_is_const(args[1]) && temp_is_const(args[2])) {
1107 tmp = deposit64(temps[args[1]].val, args[3], args[4],
1108 temps[args[2]].val);
1109 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1110 break;
1112 goto do_default;
1114 CASE_OP_32_64(extract):
1115 if (temp_is_const(args[1])) {
1116 tmp = extract64(temps[args[1]].val, args[2], args[3]);
1117 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1118 break;
1120 goto do_default;
1122 CASE_OP_32_64(sextract):
1123 if (temp_is_const(args[1])) {
1124 tmp = sextract64(temps[args[1]].val, args[2], args[3]);
1125 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1126 break;
1128 goto do_default;
1130 CASE_OP_32_64(setcond):
1131 tmp = do_constant_folding_cond(opc, args[1], args[2], args[3]);
1132 if (tmp != 2) {
1133 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1134 break;
1136 goto do_default;
1138 CASE_OP_32_64(brcond):
1139 tmp = do_constant_folding_cond(opc, args[0], args[1], args[2]);
1140 if (tmp != 2) {
1141 if (tmp) {
1142 reset_all_temps(nb_temps);
1143 op->opc = INDEX_op_br;
1144 args[0] = args[3];
1145 } else {
1146 tcg_op_remove(s, op);
1148 break;
1150 goto do_default;
1152 CASE_OP_32_64(movcond):
1153 tmp = do_constant_folding_cond(opc, args[1], args[2], args[5]);
1154 if (tmp != 2) {
1155 tcg_opt_gen_mov(s, op, args, args[0], args[4-tmp]);
1156 break;
1158 if (temp_is_const(args[3]) && temp_is_const(args[4])) {
1159 tcg_target_ulong tv = temps[args[3]].val;
1160 tcg_target_ulong fv = temps[args[4]].val;
1161 TCGCond cond = args[5];
1162 if (fv == 1 && tv == 0) {
1163 cond = tcg_invert_cond(cond);
1164 } else if (!(tv == 1 && fv == 0)) {
1165 goto do_default;
1167 args[3] = cond;
1168 op->opc = opc = (opc == INDEX_op_movcond_i32
1169 ? INDEX_op_setcond_i32
1170 : INDEX_op_setcond_i64);
1171 nb_iargs = 2;
1173 goto do_default;
1175 case INDEX_op_add2_i32:
1176 case INDEX_op_sub2_i32:
1177 if (temp_is_const(args[2]) && temp_is_const(args[3])
1178 && temp_is_const(args[4]) && temp_is_const(args[5])) {
1179 uint32_t al = temps[args[2]].val;
1180 uint32_t ah = temps[args[3]].val;
1181 uint32_t bl = temps[args[4]].val;
1182 uint32_t bh = temps[args[5]].val;
1183 uint64_t a = ((uint64_t)ah << 32) | al;
1184 uint64_t b = ((uint64_t)bh << 32) | bl;
1185 TCGArg rl, rh;
1186 TCGOp *op2 = tcg_op_insert_before(s, op, INDEX_op_movi_i32, 2);
1187 TCGArg *args2 = &s->gen_opparam_buf[op2->args];
1189 if (opc == INDEX_op_add2_i32) {
1190 a += b;
1191 } else {
1192 a -= b;
1195 rl = args[0];
1196 rh = args[1];
1197 tcg_opt_gen_movi(s, op, args, rl, (int32_t)a);
1198 tcg_opt_gen_movi(s, op2, args2, rh, (int32_t)(a >> 32));
1200 /* We've done all we need to do with the movi. Skip it. */
1201 oi_next = op2->next;
1202 break;
1204 goto do_default;
1206 case INDEX_op_mulu2_i32:
1207 if (temp_is_const(args[2]) && temp_is_const(args[3])) {
1208 uint32_t a = temps[args[2]].val;
1209 uint32_t b = temps[args[3]].val;
1210 uint64_t r = (uint64_t)a * b;
1211 TCGArg rl, rh;
1212 TCGOp *op2 = tcg_op_insert_before(s, op, INDEX_op_movi_i32, 2);
1213 TCGArg *args2 = &s->gen_opparam_buf[op2->args];
1215 rl = args[0];
1216 rh = args[1];
1217 tcg_opt_gen_movi(s, op, args, rl, (int32_t)r);
1218 tcg_opt_gen_movi(s, op2, args2, rh, (int32_t)(r >> 32));
1220 /* We've done all we need to do with the movi. Skip it. */
1221 oi_next = op2->next;
1222 break;
1224 goto do_default;
1226 case INDEX_op_brcond2_i32:
1227 tmp = do_constant_folding_cond2(&args[0], &args[2], args[4]);
1228 if (tmp != 2) {
1229 if (tmp) {
1230 do_brcond_true:
1231 reset_all_temps(nb_temps);
1232 op->opc = INDEX_op_br;
1233 args[0] = args[5];
1234 } else {
1235 do_brcond_false:
1236 tcg_op_remove(s, op);
1238 } else if ((args[4] == TCG_COND_LT || args[4] == TCG_COND_GE)
1239 && temp_is_const(args[2]) && temps[args[2]].val == 0
1240 && temp_is_const(args[3]) && temps[args[3]].val == 0) {
1241 /* Simplify LT/GE comparisons vs zero to a single compare
1242 vs the high word of the input. */
1243 do_brcond_high:
1244 reset_all_temps(nb_temps);
1245 op->opc = INDEX_op_brcond_i32;
1246 args[0] = args[1];
1247 args[1] = args[3];
1248 args[2] = args[4];
1249 args[3] = args[5];
1250 } else if (args[4] == TCG_COND_EQ) {
1251 /* Simplify EQ comparisons where one of the pairs
1252 can be simplified. */
1253 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1254 args[0], args[2], TCG_COND_EQ);
1255 if (tmp == 0) {
1256 goto do_brcond_false;
1257 } else if (tmp == 1) {
1258 goto do_brcond_high;
1260 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1261 args[1], args[3], TCG_COND_EQ);
1262 if (tmp == 0) {
1263 goto do_brcond_false;
1264 } else if (tmp != 1) {
1265 goto do_default;
1267 do_brcond_low:
1268 reset_all_temps(nb_temps);
1269 op->opc = INDEX_op_brcond_i32;
1270 args[1] = args[2];
1271 args[2] = args[4];
1272 args[3] = args[5];
1273 } else if (args[4] == TCG_COND_NE) {
1274 /* Simplify NE comparisons where one of the pairs
1275 can be simplified. */
1276 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1277 args[0], args[2], TCG_COND_NE);
1278 if (tmp == 0) {
1279 goto do_brcond_high;
1280 } else if (tmp == 1) {
1281 goto do_brcond_true;
1283 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1284 args[1], args[3], TCG_COND_NE);
1285 if (tmp == 0) {
1286 goto do_brcond_low;
1287 } else if (tmp == 1) {
1288 goto do_brcond_true;
1290 goto do_default;
1291 } else {
1292 goto do_default;
1294 break;
1296 case INDEX_op_setcond2_i32:
1297 tmp = do_constant_folding_cond2(&args[1], &args[3], args[5]);
1298 if (tmp != 2) {
1299 do_setcond_const:
1300 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1301 } else if ((args[5] == TCG_COND_LT || args[5] == TCG_COND_GE)
1302 && temp_is_const(args[3]) && temps[args[3]].val == 0
1303 && temp_is_const(args[4]) && temps[args[4]].val == 0) {
1304 /* Simplify LT/GE comparisons vs zero to a single compare
1305 vs the high word of the input. */
1306 do_setcond_high:
1307 reset_temp(args[0]);
1308 temps[args[0]].mask = 1;
1309 op->opc = INDEX_op_setcond_i32;
1310 args[1] = args[2];
1311 args[2] = args[4];
1312 args[3] = args[5];
1313 } else if (args[5] == TCG_COND_EQ) {
1314 /* Simplify EQ comparisons where one of the pairs
1315 can be simplified. */
1316 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1317 args[1], args[3], TCG_COND_EQ);
1318 if (tmp == 0) {
1319 goto do_setcond_const;
1320 } else if (tmp == 1) {
1321 goto do_setcond_high;
1323 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1324 args[2], args[4], TCG_COND_EQ);
1325 if (tmp == 0) {
1326 goto do_setcond_high;
1327 } else if (tmp != 1) {
1328 goto do_default;
1330 do_setcond_low:
1331 reset_temp(args[0]);
1332 temps[args[0]].mask = 1;
1333 op->opc = INDEX_op_setcond_i32;
1334 args[2] = args[3];
1335 args[3] = args[5];
1336 } else if (args[5] == TCG_COND_NE) {
1337 /* Simplify NE comparisons where one of the pairs
1338 can be simplified. */
1339 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1340 args[1], args[3], TCG_COND_NE);
1341 if (tmp == 0) {
1342 goto do_setcond_high;
1343 } else if (tmp == 1) {
1344 goto do_setcond_const;
1346 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1347 args[2], args[4], TCG_COND_NE);
1348 if (tmp == 0) {
1349 goto do_setcond_low;
1350 } else if (tmp == 1) {
1351 goto do_setcond_const;
1353 goto do_default;
1354 } else {
1355 goto do_default;
1357 break;
1359 case INDEX_op_call:
1360 if (!(args[nb_oargs + nb_iargs + 1]
1361 & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) {
1362 for (i = 0; i < nb_globals; i++) {
1363 if (test_bit(i, temps_used.l)) {
1364 reset_temp(i);
1368 goto do_reset_output;
1370 default:
1371 do_default:
1372 /* Default case: we know nothing about operation (or were unable
1373 to compute the operation result) so no propagation is done.
1374 We trash everything if the operation is the end of a basic
1375 block, otherwise we only trash the output args. "mask" is
1376 the non-zero bits mask for the first output arg. */
1377 if (def->flags & TCG_OPF_BB_END) {
1378 reset_all_temps(nb_temps);
1379 } else {
1380 do_reset_output:
1381 for (i = 0; i < nb_oargs; i++) {
1382 reset_temp(args[i]);
1383 /* Save the corresponding known-zero bits mask for the
1384 first output argument (only one supported so far). */
1385 if (i == 0) {
1386 temps[args[i]].mask = mask;
1390 break;
1393 /* Eliminate duplicate and redundant fence instructions. */
1394 if (prev_mb_args) {
1395 switch (opc) {
1396 case INDEX_op_mb:
1397 /* Merge two barriers of the same type into one,
1398 * or a weaker barrier into a stronger one,
1399 * or two weaker barriers into a stronger one.
1400 * mb X; mb Y => mb X|Y
1401 * mb; strl => mb; st
1402 * ldaq; mb => ld; mb
1403 * ldaq; strl => ld; mb; st
1404 * Other combinations are also merged into a strong
1405 * barrier. This is stricter than specified but for
1406 * the purposes of TCG is better than not optimizing.
1408 prev_mb_args[0] |= args[0];
1409 tcg_op_remove(s, op);
1410 break;
1412 default:
1413 /* Opcodes that end the block stop the optimization. */
1414 if ((def->flags & TCG_OPF_BB_END) == 0) {
1415 break;
1417 /* fallthru */
1418 case INDEX_op_qemu_ld_i32:
1419 case INDEX_op_qemu_ld_i64:
1420 case INDEX_op_qemu_st_i32:
1421 case INDEX_op_qemu_st_i64:
1422 case INDEX_op_call:
1423 /* Opcodes that touch guest memory stop the optimization. */
1424 prev_mb_args = NULL;
1425 break;
1427 } else if (opc == INDEX_op_mb) {
1428 prev_mb_args = args;