2 * Optimizations for Tiny Code Generator for QEMU
4 * Copyright (c) 2010 Samsung Electronics.
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
26 #include "qemu/osdep.h"
27 #include "qemu-common.h"
28 #include "exec/cpu-common.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 {
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)) {
95 return INDEX_op_mov_i32;
97 return INDEX_op_mov_i64;
99 fprintf(stderr, "op_to_mov: unexpected return value of "
100 "function op_bits.\n");
105 static TCGOpcode op_to_movi(TCGOpcode op)
107 switch (op_bits(op)) {
109 return INDEX_op_movi_i32;
111 return INDEX_op_movi_i64;
113 fprintf(stderr, "op_to_movi: unexpected return value of "
114 "function op_bits.\n");
119 static TCGArg find_better_copy(TCGContext *s, TCGArg temp)
123 /* If this is already a global, we can't do better. */
124 if (temp < s->nb_globals) {
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) {
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) {
144 /* Failure to find a better representation, return the same temp. */
148 static bool temps_are_copies(TCGArg arg1, TCGArg arg2)
156 if (!temp_is_copy(arg1) || !temp_is_copy(arg2)) {
160 for (i = temps[arg1].next_copy ; i != arg1 ; i = temps[i].next_copy) {
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;
178 temps[dst].is_const = true;
179 temps[dst].val = 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;
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);
199 TCGOpcode new_op = op_to_mov(op->opc);
200 tcg_target_ulong mask;
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;
225 static TCGArg do_constant_folding_2(TCGOpcode op, TCGArg x, TCGArg 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);
299 CASE_OP_32_64(ext8s):
302 CASE_OP_32_64(ext16s):
305 CASE_OP_32_64(ext8u):
308 CASE_OP_32_64(ext16u):
311 case INDEX_op_ext_i32_i64:
312 case INDEX_op_ext32s_i64:
315 case INDEX_op_extu_i32_i64:
316 case INDEX_op_extrl_i64_i32:
317 case INDEX_op_ext32u_i64:
320 case INDEX_op_extrh_i64_i32:
321 return (uint64_t)x >> 32;
323 case INDEX_op_muluh_i32:
324 return ((uint64_t)(uint32_t)x * (uint32_t)y) >> 32;
325 case INDEX_op_mulsh_i32:
326 return ((int64_t)(int32_t)x * (int32_t)y) >> 32;
328 case INDEX_op_muluh_i64:
329 mulu64(&l64, &h64, x, y);
331 case INDEX_op_mulsh_i64:
332 muls64(&l64, &h64, x, y);
335 case INDEX_op_div_i32:
336 /* Avoid crashing on divide by zero, otherwise undefined. */
337 return (int32_t)x / ((int32_t)y ? : 1);
338 case INDEX_op_divu_i32:
339 return (uint32_t)x / ((uint32_t)y ? : 1);
340 case INDEX_op_div_i64:
341 return (int64_t)x / ((int64_t)y ? : 1);
342 case INDEX_op_divu_i64:
343 return (uint64_t)x / ((uint64_t)y ? : 1);
345 case INDEX_op_rem_i32:
346 return (int32_t)x % ((int32_t)y ? : 1);
347 case INDEX_op_remu_i32:
348 return (uint32_t)x % ((uint32_t)y ? : 1);
349 case INDEX_op_rem_i64:
350 return (int64_t)x % ((int64_t)y ? : 1);
351 case INDEX_op_remu_i64:
352 return (uint64_t)x % ((uint64_t)y ? : 1);
356 "Unrecognized operation %d in do_constant_folding.\n", op);
361 static TCGArg do_constant_folding(TCGOpcode op, TCGArg x, TCGArg y)
363 TCGArg res = do_constant_folding_2(op, x, y);
364 if (op_bits(op) == 32) {
370 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c)
378 return (int32_t)x < (int32_t)y;
380 return (int32_t)x >= (int32_t)y;
382 return (int32_t)x <= (int32_t)y;
384 return (int32_t)x > (int32_t)y;
398 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c)
406 return (int64_t)x < (int64_t)y;
408 return (int64_t)x >= (int64_t)y;
410 return (int64_t)x <= (int64_t)y;
412 return (int64_t)x > (int64_t)y;
426 static bool do_constant_folding_cond_eq(TCGCond c)
446 /* Return 2 if the condition can't be simplified, and the result
447 of the condition (0 or 1) if it can */
448 static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x,
451 if (temp_is_const(x) && temp_is_const(y)) {
452 switch (op_bits(op)) {
454 return do_constant_folding_cond_32(temps[x].val, temps[y].val, c);
456 return do_constant_folding_cond_64(temps[x].val, temps[y].val, c);
460 } else if (temps_are_copies(x, y)) {
461 return do_constant_folding_cond_eq(c);
462 } else if (temp_is_const(y) && temps[y].val == 0) {
475 /* Return 2 if the condition can't be simplified, and the result
476 of the condition (0 or 1) if it can */
477 static TCGArg do_constant_folding_cond2(TCGArg *p1, TCGArg *p2, TCGCond c)
479 TCGArg al = p1[0], ah = p1[1];
480 TCGArg bl = p2[0], bh = p2[1];
482 if (temp_is_const(bl) && temp_is_const(bh)) {
483 uint64_t b = ((uint64_t)temps[bh].val << 32) | (uint32_t)temps[bl].val;
485 if (temp_is_const(al) && temp_is_const(ah)) {
487 a = ((uint64_t)temps[ah].val << 32) | (uint32_t)temps[al].val;
488 return do_constant_folding_cond_64(a, b, c);
501 if (temps_are_copies(al, bl) && temps_are_copies(ah, bh)) {
502 return do_constant_folding_cond_eq(c);
507 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
509 TCGArg a1 = *p1, a2 = *p2;
511 sum += temp_is_const(a1);
512 sum -= temp_is_const(a2);
514 /* Prefer the constant in second argument, and then the form
515 op a, a, b, which is better handled on non-RISC hosts. */
516 if (sum > 0 || (sum == 0 && dest == a2)) {
524 static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
527 sum += temp_is_const(p1[0]);
528 sum += temp_is_const(p1[1]);
529 sum -= temp_is_const(p2[0]);
530 sum -= temp_is_const(p2[1]);
533 t = p1[0], p1[0] = p2[0], p2[0] = t;
534 t = p1[1], p1[1] = p2[1], p2[1] = t;
540 /* Propagate constants and copies, fold constant expressions. */
541 void tcg_optimize(TCGContext *s)
543 int oi, oi_next, nb_temps, nb_globals;
544 TCGArg *prev_mb_args = NULL;
546 /* Array VALS has an element for each temp.
547 If this temp holds a constant then its value is kept in VALS' element.
548 If this temp is a copy of other ones then the other copies are
549 available through the doubly linked circular list. */
551 nb_temps = s->nb_temps;
552 nb_globals = s->nb_globals;
553 reset_all_temps(nb_temps);
555 for (oi = s->gen_op_buf[0].next; oi != 0; oi = oi_next) {
556 tcg_target_ulong mask, partmask, affected;
557 int nb_oargs, nb_iargs, i;
560 TCGOp * const op = &s->gen_op_buf[oi];
561 TCGArg * const args = &s->gen_opparam_buf[op->args];
562 TCGOpcode opc = op->opc;
563 const TCGOpDef *def = &tcg_op_defs[opc];
567 /* Count the arguments, and initialize the temps that are
569 if (opc == INDEX_op_call) {
570 nb_oargs = op->callo;
571 nb_iargs = op->calli;
572 for (i = 0; i < nb_oargs + nb_iargs; i++) {
574 if (tmp != TCG_CALL_DUMMY_ARG) {
579 nb_oargs = def->nb_oargs;
580 nb_iargs = def->nb_iargs;
581 for (i = 0; i < nb_oargs + nb_iargs; i++) {
582 init_temp_info(args[i]);
586 /* Do copy propagation */
587 for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
588 if (temp_is_copy(args[i])) {
589 args[i] = find_better_copy(s, args[i]);
593 /* For commutative operations make constant second argument */
603 CASE_OP_32_64(muluh):
604 CASE_OP_32_64(mulsh):
605 swap_commutative(args[0], &args[1], &args[2]);
607 CASE_OP_32_64(brcond):
608 if (swap_commutative(-1, &args[0], &args[1])) {
609 args[2] = tcg_swap_cond(args[2]);
612 CASE_OP_32_64(setcond):
613 if (swap_commutative(args[0], &args[1], &args[2])) {
614 args[3] = tcg_swap_cond(args[3]);
617 CASE_OP_32_64(movcond):
618 if (swap_commutative(-1, &args[1], &args[2])) {
619 args[5] = tcg_swap_cond(args[5]);
621 /* For movcond, we canonicalize the "false" input reg to match
622 the destination reg so that the tcg backend can implement
623 a "move if true" operation. */
624 if (swap_commutative(args[0], &args[4], &args[3])) {
625 args[5] = tcg_invert_cond(args[5]);
629 swap_commutative(args[0], &args[2], &args[4]);
630 swap_commutative(args[1], &args[3], &args[5]);
632 CASE_OP_32_64(mulu2):
633 CASE_OP_32_64(muls2):
634 swap_commutative(args[0], &args[2], &args[3]);
636 case INDEX_op_brcond2_i32:
637 if (swap_commutative2(&args[0], &args[2])) {
638 args[4] = tcg_swap_cond(args[4]);
641 case INDEX_op_setcond2_i32:
642 if (swap_commutative2(&args[1], &args[3])) {
643 args[5] = tcg_swap_cond(args[5]);
650 /* Simplify expressions for "shift/rot r, 0, a => movi r, 0",
651 and "sub r, 0, a => neg r, a" case. */
658 if (temp_is_const(args[1]) && temps[args[1]].val == 0) {
659 tcg_opt_gen_movi(s, op, args, args[0], 0);
668 if (temp_is_const(args[2])) {
669 /* Proceed with possible constant folding. */
672 if (opc == INDEX_op_sub_i32) {
673 neg_op = INDEX_op_neg_i32;
674 have_neg = TCG_TARGET_HAS_neg_i32;
676 neg_op = INDEX_op_neg_i64;
677 have_neg = TCG_TARGET_HAS_neg_i64;
682 if (temp_is_const(args[1]) && temps[args[1]].val == 0) {
692 if (!temp_is_const(args[1])
693 && temp_is_const(args[2]) && temps[args[2]].val == -1) {
699 if (!temp_is_const(args[1])
700 && temp_is_const(args[2]) && temps[args[2]].val == 0) {
706 if (!temp_is_const(args[2])
707 && temp_is_const(args[1]) && temps[args[1]].val == -1) {
714 if (!temp_is_const(args[2])
715 && temp_is_const(args[1]) && temps[args[1]].val == 0) {
725 if (def->flags & TCG_OPF_64BIT) {
726 not_op = INDEX_op_not_i64;
727 have_not = TCG_TARGET_HAS_not_i64;
729 not_op = INDEX_op_not_i32;
730 have_not = TCG_TARGET_HAS_not_i32;
744 /* Simplify expression for "op r, a, const => mov r, a" cases */
756 if (!temp_is_const(args[1])
757 && temp_is_const(args[2]) && temps[args[2]].val == 0) {
758 tcg_opt_gen_mov(s, op, args, args[0], args[1]);
765 if (!temp_is_const(args[1])
766 && temp_is_const(args[2]) && temps[args[2]].val == -1) {
767 tcg_opt_gen_mov(s, op, args, args[0], args[1]);
775 /* Simplify using known-zero bits. Currently only ops with a single
776 output argument is supported. */
780 CASE_OP_32_64(ext8s):
781 if ((temps[args[1]].mask & 0x80) != 0) {
784 CASE_OP_32_64(ext8u):
787 CASE_OP_32_64(ext16s):
788 if ((temps[args[1]].mask & 0x8000) != 0) {
791 CASE_OP_32_64(ext16u):
794 case INDEX_op_ext32s_i64:
795 if ((temps[args[1]].mask & 0x80000000) != 0) {
798 case INDEX_op_ext32u_i64:
803 mask = temps[args[2]].mask;
804 if (temp_is_const(args[2])) {
806 affected = temps[args[1]].mask & ~mask;
808 mask = temps[args[1]].mask & mask;
811 case INDEX_op_ext_i32_i64:
812 if ((temps[args[1]].mask & 0x80000000) != 0) {
815 case INDEX_op_extu_i32_i64:
816 /* We do not compute affected as it is a size changing op. */
817 mask = (uint32_t)temps[args[1]].mask;
821 /* Known-zeros does not imply known-ones. Therefore unless
822 args[2] is constant, we can't infer anything from it. */
823 if (temp_is_const(args[2])) {
824 mask = ~temps[args[2]].mask;
827 /* But we certainly know nothing outside args[1] may be set. */
828 mask = temps[args[1]].mask;
831 case INDEX_op_sar_i32:
832 if (temp_is_const(args[2])) {
833 tmp = temps[args[2]].val & 31;
834 mask = (int32_t)temps[args[1]].mask >> tmp;
837 case INDEX_op_sar_i64:
838 if (temp_is_const(args[2])) {
839 tmp = temps[args[2]].val & 63;
840 mask = (int64_t)temps[args[1]].mask >> tmp;
844 case INDEX_op_shr_i32:
845 if (temp_is_const(args[2])) {
846 tmp = temps[args[2]].val & 31;
847 mask = (uint32_t)temps[args[1]].mask >> tmp;
850 case INDEX_op_shr_i64:
851 if (temp_is_const(args[2])) {
852 tmp = temps[args[2]].val & 63;
853 mask = (uint64_t)temps[args[1]].mask >> tmp;
857 case INDEX_op_extrl_i64_i32:
858 mask = (uint32_t)temps[args[1]].mask;
860 case INDEX_op_extrh_i64_i32:
861 mask = (uint64_t)temps[args[1]].mask >> 32;
865 if (temp_is_const(args[2])) {
866 tmp = temps[args[2]].val & (TCG_TARGET_REG_BITS - 1);
867 mask = temps[args[1]].mask << tmp;
872 /* Set to 1 all bits to the left of the rightmost. */
873 mask = -(temps[args[1]].mask & -temps[args[1]].mask);
876 CASE_OP_32_64(deposit):
877 mask = deposit64(temps[args[1]].mask, args[3], args[4],
878 temps[args[2]].mask);
883 mask = temps[args[1]].mask | temps[args[2]].mask;
886 CASE_OP_32_64(setcond):
887 case INDEX_op_setcond2_i32:
891 CASE_OP_32_64(movcond):
892 mask = temps[args[3]].mask | temps[args[4]].mask;
898 CASE_OP_32_64(ld16u):
901 case INDEX_op_ld32u_i64:
905 CASE_OP_32_64(qemu_ld):
907 TCGMemOpIdx oi = args[nb_oargs + nb_iargs];
908 TCGMemOp mop = get_memop(oi);
909 if (!(mop & MO_SIGN)) {
910 mask = (2ULL << ((8 << (mop & MO_SIZE)) - 1)) - 1;
919 /* 32-bit ops generate 32-bit results. For the result is zero test
920 below, we can ignore high bits, but for further optimizations we
921 need to record that the high bits contain garbage. */
923 if (!(def->flags & TCG_OPF_64BIT)) {
924 mask |= ~(tcg_target_ulong)0xffffffffu;
925 partmask &= 0xffffffffu;
926 affected &= 0xffffffffu;
930 tcg_debug_assert(nb_oargs == 1);
931 tcg_opt_gen_movi(s, op, args, args[0], 0);
935 tcg_debug_assert(nb_oargs == 1);
936 tcg_opt_gen_mov(s, op, args, args[0], args[1]);
940 /* Simplify expression for "op r, a, 0 => movi r, 0" cases */
944 CASE_OP_32_64(muluh):
945 CASE_OP_32_64(mulsh):
946 if ((temp_is_const(args[2]) && temps[args[2]].val == 0)) {
947 tcg_opt_gen_movi(s, op, args, args[0], 0);
955 /* Simplify expression for "op r, a, a => mov r, a" cases */
959 if (temps_are_copies(args[1], args[2])) {
960 tcg_opt_gen_mov(s, op, args, args[0], args[1]);
968 /* Simplify expression for "op r, a, a => movi r, 0" cases */
973 if (temps_are_copies(args[1], args[2])) {
974 tcg_opt_gen_movi(s, op, args, args[0], 0);
982 /* Propagate constants through copy operations and do constant
983 folding. Constants will be substituted to arguments by register
984 allocator where needed and possible. Also detect copies. */
987 tcg_opt_gen_mov(s, op, args, args[0], args[1]);
990 tcg_opt_gen_movi(s, op, args, args[0], args[1]);
995 CASE_OP_32_64(ext8s):
996 CASE_OP_32_64(ext8u):
997 CASE_OP_32_64(ext16s):
998 CASE_OP_32_64(ext16u):
999 case INDEX_op_ext32s_i64:
1000 case INDEX_op_ext32u_i64:
1001 case INDEX_op_ext_i32_i64:
1002 case INDEX_op_extu_i32_i64:
1003 case INDEX_op_extrl_i64_i32:
1004 case INDEX_op_extrh_i64_i32:
1005 if (temp_is_const(args[1])) {
1006 tmp = do_constant_folding(opc, temps[args[1]].val, 0);
1007 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1021 CASE_OP_32_64(rotl):
1022 CASE_OP_32_64(rotr):
1023 CASE_OP_32_64(andc):
1026 CASE_OP_32_64(nand):
1028 CASE_OP_32_64(muluh):
1029 CASE_OP_32_64(mulsh):
1031 CASE_OP_32_64(divu):
1033 CASE_OP_32_64(remu):
1034 if (temp_is_const(args[1]) && temp_is_const(args[2])) {
1035 tmp = do_constant_folding(opc, temps[args[1]].val,
1036 temps[args[2]].val);
1037 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1042 CASE_OP_32_64(deposit):
1043 if (temp_is_const(args[1]) && temp_is_const(args[2])) {
1044 tmp = deposit64(temps[args[1]].val, args[3], args[4],
1045 temps[args[2]].val);
1046 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1051 CASE_OP_32_64(setcond):
1052 tmp = do_constant_folding_cond(opc, args[1], args[2], args[3]);
1054 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1059 CASE_OP_32_64(brcond):
1060 tmp = do_constant_folding_cond(opc, args[0], args[1], args[2]);
1063 reset_all_temps(nb_temps);
1064 op->opc = INDEX_op_br;
1067 tcg_op_remove(s, op);
1073 CASE_OP_32_64(movcond):
1074 tmp = do_constant_folding_cond(opc, args[1], args[2], args[5]);
1076 tcg_opt_gen_mov(s, op, args, args[0], args[4-tmp]);
1081 case INDEX_op_add2_i32:
1082 case INDEX_op_sub2_i32:
1083 if (temp_is_const(args[2]) && temp_is_const(args[3])
1084 && temp_is_const(args[4]) && temp_is_const(args[5])) {
1085 uint32_t al = temps[args[2]].val;
1086 uint32_t ah = temps[args[3]].val;
1087 uint32_t bl = temps[args[4]].val;
1088 uint32_t bh = temps[args[5]].val;
1089 uint64_t a = ((uint64_t)ah << 32) | al;
1090 uint64_t b = ((uint64_t)bh << 32) | bl;
1092 TCGOp *op2 = tcg_op_insert_before(s, op, INDEX_op_movi_i32, 2);
1093 TCGArg *args2 = &s->gen_opparam_buf[op2->args];
1095 if (opc == INDEX_op_add2_i32) {
1103 tcg_opt_gen_movi(s, op, args, rl, (int32_t)a);
1104 tcg_opt_gen_movi(s, op2, args2, rh, (int32_t)(a >> 32));
1106 /* We've done all we need to do with the movi. Skip it. */
1107 oi_next = op2->next;
1112 case INDEX_op_mulu2_i32:
1113 if (temp_is_const(args[2]) && temp_is_const(args[3])) {
1114 uint32_t a = temps[args[2]].val;
1115 uint32_t b = temps[args[3]].val;
1116 uint64_t r = (uint64_t)a * b;
1118 TCGOp *op2 = tcg_op_insert_before(s, op, INDEX_op_movi_i32, 2);
1119 TCGArg *args2 = &s->gen_opparam_buf[op2->args];
1123 tcg_opt_gen_movi(s, op, args, rl, (int32_t)r);
1124 tcg_opt_gen_movi(s, op2, args2, rh, (int32_t)(r >> 32));
1126 /* We've done all we need to do with the movi. Skip it. */
1127 oi_next = op2->next;
1132 case INDEX_op_brcond2_i32:
1133 tmp = do_constant_folding_cond2(&args[0], &args[2], args[4]);
1137 reset_all_temps(nb_temps);
1138 op->opc = INDEX_op_br;
1142 tcg_op_remove(s, op);
1144 } else if ((args[4] == TCG_COND_LT || args[4] == TCG_COND_GE)
1145 && temp_is_const(args[2]) && temps[args[2]].val == 0
1146 && temp_is_const(args[3]) && temps[args[3]].val == 0) {
1147 /* Simplify LT/GE comparisons vs zero to a single compare
1148 vs the high word of the input. */
1150 reset_all_temps(nb_temps);
1151 op->opc = INDEX_op_brcond_i32;
1156 } else if (args[4] == TCG_COND_EQ) {
1157 /* Simplify EQ comparisons where one of the pairs
1158 can be simplified. */
1159 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1160 args[0], args[2], TCG_COND_EQ);
1162 goto do_brcond_false;
1163 } else if (tmp == 1) {
1164 goto do_brcond_high;
1166 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1167 args[1], args[3], TCG_COND_EQ);
1169 goto do_brcond_false;
1170 } else if (tmp != 1) {
1174 reset_all_temps(nb_temps);
1175 op->opc = INDEX_op_brcond_i32;
1179 } else if (args[4] == TCG_COND_NE) {
1180 /* Simplify NE comparisons where one of the pairs
1181 can be simplified. */
1182 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1183 args[0], args[2], TCG_COND_NE);
1185 goto do_brcond_high;
1186 } else if (tmp == 1) {
1187 goto do_brcond_true;
1189 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1190 args[1], args[3], TCG_COND_NE);
1193 } else if (tmp == 1) {
1194 goto do_brcond_true;
1202 case INDEX_op_setcond2_i32:
1203 tmp = do_constant_folding_cond2(&args[1], &args[3], args[5]);
1206 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1207 } else if ((args[5] == TCG_COND_LT || args[5] == TCG_COND_GE)
1208 && temp_is_const(args[3]) && temps[args[3]].val == 0
1209 && temp_is_const(args[4]) && temps[args[4]].val == 0) {
1210 /* Simplify LT/GE comparisons vs zero to a single compare
1211 vs the high word of the input. */
1213 reset_temp(args[0]);
1214 temps[args[0]].mask = 1;
1215 op->opc = INDEX_op_setcond_i32;
1219 } else if (args[5] == TCG_COND_EQ) {
1220 /* Simplify EQ comparisons where one of the pairs
1221 can be simplified. */
1222 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1223 args[1], args[3], TCG_COND_EQ);
1225 goto do_setcond_const;
1226 } else if (tmp == 1) {
1227 goto do_setcond_high;
1229 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1230 args[2], args[4], TCG_COND_EQ);
1232 goto do_setcond_high;
1233 } else if (tmp != 1) {
1237 reset_temp(args[0]);
1238 temps[args[0]].mask = 1;
1239 op->opc = INDEX_op_setcond_i32;
1242 } else if (args[5] == TCG_COND_NE) {
1243 /* Simplify NE comparisons where one of the pairs
1244 can be simplified. */
1245 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1246 args[1], args[3], TCG_COND_NE);
1248 goto do_setcond_high;
1249 } else if (tmp == 1) {
1250 goto do_setcond_const;
1252 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1253 args[2], args[4], TCG_COND_NE);
1255 goto do_setcond_low;
1256 } else if (tmp == 1) {
1257 goto do_setcond_const;
1266 if (!(args[nb_oargs + nb_iargs + 1]
1267 & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) {
1268 for (i = 0; i < nb_globals; i++) {
1269 if (test_bit(i, temps_used.l)) {
1274 goto do_reset_output;
1278 /* Default case: we know nothing about operation (or were unable
1279 to compute the operation result) so no propagation is done.
1280 We trash everything if the operation is the end of a basic
1281 block, otherwise we only trash the output args. "mask" is
1282 the non-zero bits mask for the first output arg. */
1283 if (def->flags & TCG_OPF_BB_END) {
1284 reset_all_temps(nb_temps);
1287 for (i = 0; i < nb_oargs; i++) {
1288 reset_temp(args[i]);
1289 /* Save the corresponding known-zero bits mask for the
1290 first output argument (only one supported so far). */
1292 temps[args[i]].mask = mask;
1299 /* Eliminate duplicate and redundant fence instructions. */
1303 /* Merge two barriers of the same type into one,
1304 * or a weaker barrier into a stronger one,
1305 * or two weaker barriers into a stronger one.
1306 * mb X; mb Y => mb X|Y
1307 * mb; strl => mb; st
1308 * ldaq; mb => ld; mb
1309 * ldaq; strl => ld; mb; st
1310 * Other combinations are also merged into a strong
1311 * barrier. This is stricter than specified but for
1312 * the purposes of TCG is better than not optimizing.
1314 prev_mb_args[0] |= args[0];
1315 tcg_op_remove(s, op);
1319 /* Opcodes that end the block stop the optimization. */
1320 if ((def->flags & TCG_OPF_BB_END) == 0) {
1324 case INDEX_op_qemu_ld_i32:
1325 case INDEX_op_qemu_ld_i64:
1326 case INDEX_op_qemu_st_i32:
1327 case INDEX_op_qemu_st_i64:
1329 /* Opcodes that touch guest memory stop the optimization. */
1330 prev_mb_args = NULL;
1333 } else if (opc == INDEX_op_mb) {
1334 prev_mb_args = args;