tcg/optimize.c

   1 /*
   2  * Optimizations for Tiny Code Generator for QEMU
   3  *
   4  * Copyright (c) 2010 Samsung Electronics.
   5  * Contributed by Kirill Batuzov <batuzovk@ispras.ru>
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a copy
   8  * of this software and associated documentation files (the "Software"), to deal
   9  * in the Software without restriction, including without limitation the rights
  10  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  11  * copies of the Software, and to permit persons to whom the Software is
  12  * furnished to do so, subject to the following conditions:
  13  *
  14  * The above copyright notice and this permission notice shall be included in
  15  * all copies or substantial portions of the Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  23  * THE SOFTWARE.
  24  */
  25
  26 #include "qemu/osdep.h"
  27 #include "qemu-common.h"
  28 #include "exec/cpu-common.h"
  29 #include "tcg-op.h"
  30
  31 #define CASE_OP_32_64(x)                        \
  32         glue(glue(case INDEX_op_, x), _i32):    \
  33         glue(glue(case INDEX_op_, x), _i64)
  34
  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;
  41 };
  42
  43 static struct tcg_temp_info temps[TCG_MAX_TEMPS];
  44 static TCGTempSet temps_used;
  45
  46 static inline bool temp_is_const(TCGArg arg)
  47 {
  48     return temps[arg].is_const;
  49 }
  50
  51 static inline bool temp_is_copy(TCGArg arg)
  52 {
  53     return temps[arg].next_copy != arg;
  54 }
  55
  56 /* Reset TEMP's state, possibly removing the temp for the list of copies.  */
  57 static void reset_temp(TCGArg temp)
  58 {
  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;
  65 }
  66
  67 /* Reset all temporaries, given that there are NB_TEMPS of them.  */
  68 static void reset_all_temps(int nb_temps)
  69 {
  70     bitmap_zero(temps_used.l, nb_temps);
  71 }
  72
  73 /* Initialize and activate a temporary.  */
  74 static void init_temp_info(TCGArg temp)
  75 {
  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);
  82     }
  83 }
  84
  85 static int op_bits(TCGOpcode op)
  86 {
  87     const TCGOpDef *def = &tcg_op_defs[op];
  88     return def->flags & TCG_OPF_64BIT ? 64 : 32;
  89 }
  90
  91 static TCGOpcode op_to_mov(TCGOpcode op)
  92 {
  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();
 102     }
 103 }
 104
 105 static TCGOpcode op_to_movi(TCGOpcode op)
 106 {
 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();
 116     }
 117 }
 118
 119 static TCGArg find_better_copy(TCGContext *s, TCGArg temp)
 120 {
 121     TCGArg i;
 122
 123     /* If this is already a global, we can't do better. */
 124     if (temp < s->nb_globals) {
 125         return temp;
 126     }
 127
 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;
 132         }
 133     }
 134
 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;
 140             }
 141         }
 142     }
 143
 144     /* Failure to find a better representation, return the same temp. */
 145     return temp;
 146 }
 147
 148 static bool temps_are_copies(TCGArg arg1, TCGArg arg2)
 149 {
 150     TCGArg i;
 151
 152     if (arg1 == arg2) {
 153         return true;
 154     }
 155
 156     if (!temp_is_copy(arg1) || !temp_is_copy(arg2)) {
 157         return false;
 158     }
 159
 160     for (i = temps[arg1].next_copy ; i != arg1 ; i = temps[i].next_copy) {
 161         if (i == arg2) {
 162             return true;
 163         }
 164     }
 165
 166     return false;
 167 }
 168
 169 static void tcg_opt_gen_movi(TCGContext *s, TCGOp *op, TCGArg *args,
 170                              TCGArg dst, TCGArg val)
 171 {
 172     TCGOpcode new_op = op_to_movi(op->opc);
 173     tcg_target_ulong mask;
 174
 175     op->opc = new_op;
 176
 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;
 184     }
 185     temps[dst].mask = mask;
 186
 187     args[0] = dst;
 188     args[1] = val;
 189 }
 190
 191 static void tcg_opt_gen_mov(TCGContext *s, TCGOp *op, TCGArg *args,
 192                             TCGArg dst, TCGArg src)
 193 {
 194     if (temps_are_copies(dst, src)) {
 195         tcg_op_remove(s, op);
 196         return;
 197     }
 198
 199     TCGOpcode new_op = op_to_mov(op->opc);
 200     tcg_target_ulong mask;
 201
 202     op->opc = new_op;
 203
 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;
 209     }
 210     temps[dst].mask = mask;
 211
 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;
 219     }
 220
 221     args[0] = dst;
 222     args[1] = src;
 223 }
 224
 225 static TCGArg do_constant_folding_2(TCGOpcode op, TCGArg x, TCGArg y)
 226 {
 227     uint64_t l64, h64;
 228
 229     switch (op) {
 230     CASE_OP_32_64(add):
 231         return x + y;
 232
 233     CASE_OP_32_64(sub):
 234         return x - y;
 235
 236     CASE_OP_32_64(mul):
 237         return x * y;
 238
 239     CASE_OP_32_64(and):
 240         return x & y;
 241
 242     CASE_OP_32_64(or):
 243         return x | y;
 244
 245     CASE_OP_32_64(xor):
 246         return x ^ y;
 247
 248     case INDEX_op_shl_i32:
 249         return (uint32_t)x << (y & 31);
 250
 251     case INDEX_op_shl_i64:
 252         return (uint64_t)x << (y & 63);
 253
 254     case INDEX_op_shr_i32:
 255         return (uint32_t)x >> (y & 31);
 256
 257     case INDEX_op_shr_i64:
 258         return (uint64_t)x >> (y & 63);
 259
 260     case INDEX_op_sar_i32:
 261         return (int32_t)x >> (y & 31);
 262
 263     case INDEX_op_sar_i64:
 264         return (int64_t)x >> (y & 63);
 265
 266     case INDEX_op_rotr_i32:
 267         return ror32(x, y & 31);
 268
 269     case INDEX_op_rotr_i64:
 270         return ror64(x, y & 63);
 271
 272     case INDEX_op_rotl_i32:
 273         return rol32(x, y & 31);
 274
 275     case INDEX_op_rotl_i64:
 276         return rol64(x, y & 63);
 277
 278     CASE_OP_32_64(not):
 279         return ~x;
 280
 281     CASE_OP_32_64(neg):
 282         return -x;
 283
 284     CASE_OP_32_64(andc):
 285         return x & ~y;
 286
 287     CASE_OP_32_64(orc):
 288         return x | ~y;
 289
 290     CASE_OP_32_64(eqv):
 291         return ~(x ^ y);
 292
 293     CASE_OP_32_64(nand):
 294         return ~(x & y);
 295
 296     CASE_OP_32_64(nor):
 297         return ~(x | y);
 298
 299     CASE_OP_32_64(ext8s):
 300         return (int8_t)x;
 301
 302     CASE_OP_32_64(ext16s):
 303         return (int16_t)x;
 304
 305     CASE_OP_32_64(ext8u):
 306         return (uint8_t)x;
 307
 308     CASE_OP_32_64(ext16u):
 309         return (uint16_t)x;
 310
 311     case INDEX_op_ext_i32_i64:
 312     case INDEX_op_ext32s_i64:
 313         return (int32_t)x;
 314
 315     case INDEX_op_extu_i32_i64:
 316     case INDEX_op_extrl_i64_i32:
 317     case INDEX_op_ext32u_i64:
 318         return (uint32_t)x;
 319
 320     case INDEX_op_extrh_i64_i32:
 321         return (uint64_t)x >> 32;
 322
 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;
 327
 328     case INDEX_op_muluh_i64:
 329         mulu64(&l64, &h64, x, y);
 330         return h64;
 331     case INDEX_op_mulsh_i64:
 332         muls64(&l64, &h64, x, y);
 333         return h64;
 334
 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);
 344
 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);
 353
 354     default:
 355         fprintf(stderr,
 356                 "Unrecognized operation %d in do_constant_folding.\n", op);
 357         tcg_abort();
 358     }
 359 }
 360
 361 static TCGArg do_constant_folding(TCGOpcode op, TCGArg x, TCGArg y)
 362 {
 363     TCGArg res = do_constant_folding_2(op, x, y);
 364     if (op_bits(op) == 32) {
 365         res = (int32_t)res;
 366     }
 367     return res;
 368 }
 369
 370 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c)
 371 {
 372     switch (c) {
 373     case TCG_COND_EQ:
 374         return x == y;
 375     case TCG_COND_NE:
 376         return x != y;
 377     case TCG_COND_LT:
 378         return (int32_t)x < (int32_t)y;
 379     case TCG_COND_GE:
 380         return (int32_t)x >= (int32_t)y;
 381     case TCG_COND_LE:
 382         return (int32_t)x <= (int32_t)y;
 383     case TCG_COND_GT:
 384         return (int32_t)x > (int32_t)y;
 385     case TCG_COND_LTU:
 386         return x < y;
 387     case TCG_COND_GEU:
 388         return x >= y;
 389     case TCG_COND_LEU:
 390         return x <= y;
 391     case TCG_COND_GTU:
 392         return x > y;
 393     default:
 394         tcg_abort();
 395     }
 396 }
 397
 398 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c)
 399 {
 400     switch (c) {
 401     case TCG_COND_EQ:
 402         return x == y;
 403     case TCG_COND_NE:
 404         return x != y;
 405     case TCG_COND_LT:
 406         return (int64_t)x < (int64_t)y;
 407     case TCG_COND_GE:
 408         return (int64_t)x >= (int64_t)y;
 409     case TCG_COND_LE:
 410         return (int64_t)x <= (int64_t)y;
 411     case TCG_COND_GT:
 412         return (int64_t)x > (int64_t)y;
 413     case TCG_COND_LTU:
 414         return x < y;
 415     case TCG_COND_GEU:
 416         return x >= y;
 417     case TCG_COND_LEU:
 418         return x <= y;
 419     case TCG_COND_GTU:
 420         return x > y;
 421     default:
 422         tcg_abort();
 423     }
 424 }
 425
 426 static bool do_constant_folding_cond_eq(TCGCond c)
 427 {
 428     switch (c) {
 429     case TCG_COND_GT:
 430     case TCG_COND_LTU:
 431     case TCG_COND_LT:
 432     case TCG_COND_GTU:
 433     case TCG_COND_NE:
 434         return 0;
 435     case TCG_COND_GE:
 436     case TCG_COND_GEU:
 437     case TCG_COND_LE:
 438     case TCG_COND_LEU:
 439     case TCG_COND_EQ:
 440         return 1;
 441     default:
 442         tcg_abort();
 443     }
 444 }
 445
 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,
 449                                        TCGArg y, TCGCond c)
 450 {
 451     if (temp_is_const(x) && temp_is_const(y)) {
 452         switch (op_bits(op)) {
 453         case 32:
 454             return do_constant_folding_cond_32(temps[x].val, temps[y].val, c);
 455         case 64:
 456             return do_constant_folding_cond_64(temps[x].val, temps[y].val, c);
 457         default:
 458             tcg_abort();
 459         }
 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) {
 463         switch (c) {
 464         case TCG_COND_LTU:
 465             return 0;
 466         case TCG_COND_GEU:
 467             return 1;
 468         default:
 469             return 2;
 470         }
 471     } else {
 472         return 2;
 473     }
 474 }
 475
 476 /* Return 2 if the condition can't be simplified, and the result
 477    of the condition (0 or 1) if it can */
 478 static TCGArg do_constant_folding_cond2(TCGArg *p1, TCGArg *p2, TCGCond c)
 479 {
 480     TCGArg al = p1[0], ah = p1[1];
 481     TCGArg bl = p2[0], bh = p2[1];
 482
 483     if (temp_is_const(bl) && temp_is_const(bh)) {
 484         uint64_t b = ((uint64_t)temps[bh].val << 32) | (uint32_t)temps[bl].val;
 485
 486         if (temp_is_const(al) && temp_is_const(ah)) {
 487             uint64_t a;
 488             a = ((uint64_t)temps[ah].val << 32) | (uint32_t)temps[al].val;
 489             return do_constant_folding_cond_64(a, b, c);
 490         }
 491         if (b == 0) {
 492             switch (c) {
 493             case TCG_COND_LTU:
 494                 return 0;
 495             case TCG_COND_GEU:
 496                 return 1;
 497             default:
 498                 break;
 499             }
 500         }
 501     }
 502     if (temps_are_copies(al, bl) && temps_are_copies(ah, bh)) {
 503         return do_constant_folding_cond_eq(c);
 504     }
 505     return 2;
 506 }
 507
 508 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2)
 509 {
 510     TCGArg a1 = *p1, a2 = *p2;
 511     int sum = 0;
 512     sum += temp_is_const(a1);
 513     sum -= temp_is_const(a2);
 514
 515     /* Prefer the constant in second argument, and then the form
 516        op a, a, b, which is better handled on non-RISC hosts. */
 517     if (sum > 0 || (sum == 0 && dest == a2)) {
 518         *p1 = a2;
 519         *p2 = a1;
 520         return true;
 521     }
 522     return false;
 523 }
 524
 525 static bool swap_commutative2(TCGArg *p1, TCGArg *p2)
 526 {
 527     int sum = 0;
 528     sum += temp_is_const(p1[0]);
 529     sum += temp_is_const(p1[1]);
 530     sum -= temp_is_const(p2[0]);
 531     sum -= temp_is_const(p2[1]);
 532     if (sum > 0) {
 533         TCGArg t;
 534         t = p1[0], p1[0] = p2[0], p2[0] = t;
 535         t = p1[1], p1[1] = p2[1], p2[1] = t;
 536         return true;
 537     }
 538     return false;
 539 }
 540
 541 /* Propagate constants and copies, fold constant expressions. */
 542 void tcg_optimize(TCGContext *s)
 543 {
 544     int oi, oi_next, nb_temps, nb_globals;
 545     TCGArg *prev_mb_args = NULL;
 546
 547     /* Array VALS has an element for each temp.
 548        If this temp holds a constant then its value is kept in VALS' element.
 549        If this temp is a copy of other ones then the other copies are
 550        available through the doubly linked circular list. */
 551
 552     nb_temps = s->nb_temps;
 553     nb_globals = s->nb_globals;
 554     reset_all_temps(nb_temps);
 555
 556     for (oi = s->gen_op_buf[0].next; oi != 0; oi = oi_next) {
 557         tcg_target_ulong mask, partmask, affected;
 558         int nb_oargs, nb_iargs, i;
 559         TCGArg tmp;
 560
 561         TCGOp * const op = &s->gen_op_buf[oi];
 562         TCGArg * const args = &s->gen_opparam_buf[op->args];
 563         TCGOpcode opc = op->opc;
 564         const TCGOpDef *def = &tcg_op_defs[opc];
 565
 566         oi_next = op->next;
 567
 568         /* Count the arguments, and initialize the temps that are
 569            going to be used */
 570         if (opc == INDEX_op_call) {
 571             nb_oargs = op->callo;
 572             nb_iargs = op->calli;
 573             for (i = 0; i < nb_oargs + nb_iargs; i++) {
 574                 tmp = args[i];
 575                 if (tmp != TCG_CALL_DUMMY_ARG) {
 576                     init_temp_info(tmp);
 577                 }
 578             }
 579         } else {
 580             nb_oargs = def->nb_oargs;
 581             nb_iargs = def->nb_iargs;
 582             for (i = 0; i < nb_oargs + nb_iargs; i++) {
 583                 init_temp_info(args[i]);
 584             }
 585         }
 586
 587         /* Do copy propagation */
 588         for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) {
 589             if (temp_is_copy(args[i])) {
 590                 args[i] = find_better_copy(s, args[i]);
 591             }
 592         }
 593
 594         /* For commutative operations make constant second argument */
 595         switch (opc) {
 596         CASE_OP_32_64(add):
 597         CASE_OP_32_64(mul):
 598         CASE_OP_32_64(and):
 599         CASE_OP_32_64(or):
 600         CASE_OP_32_64(xor):
 601         CASE_OP_32_64(eqv):
 602         CASE_OP_32_64(nand):
 603         CASE_OP_32_64(nor):
 604         CASE_OP_32_64(muluh):
 605         CASE_OP_32_64(mulsh):
 606             swap_commutative(args[0], &args[1], &args[2]);
 607             break;
 608         CASE_OP_32_64(brcond):
 609             if (swap_commutative(-1, &args[0], &args[1])) {
 610                 args[2] = tcg_swap_cond(args[2]);
 611             }
 612             break;
 613         CASE_OP_32_64(setcond):
 614             if (swap_commutative(args[0], &args[1], &args[2])) {
 615                 args[3] = tcg_swap_cond(args[3]);
 616             }
 617             break;
 618         CASE_OP_32_64(movcond):
 619             if (swap_commutative(-1, &args[1], &args[2])) {
 620                 args[5] = tcg_swap_cond(args[5]);
 621             }
 622             /* For movcond, we canonicalize the "false" input reg to match
 623                the destination reg so that the tcg backend can implement
 624                a "move if true" operation.  */
 625             if (swap_commutative(args[0], &args[4], &args[3])) {
 626                 args[5] = tcg_invert_cond(args[5]);
 627             }
 628             break;
 629         CASE_OP_32_64(add2):
 630             swap_commutative(args[0], &args[2], &args[4]);
 631             swap_commutative(args[1], &args[3], &args[5]);
 632             break;
 633         CASE_OP_32_64(mulu2):
 634         CASE_OP_32_64(muls2):
 635             swap_commutative(args[0], &args[2], &args[3]);
 636             break;
 637         case INDEX_op_brcond2_i32:
 638             if (swap_commutative2(&args[0], &args[2])) {
 639                 args[4] = tcg_swap_cond(args[4]);
 640             }
 641             break;
 642         case INDEX_op_setcond2_i32:
 643             if (swap_commutative2(&args[1], &args[3])) {
 644                 args[5] = tcg_swap_cond(args[5]);
 645             }
 646             break;
 647         default:
 648             break;
 649         }
 650
 651         /* Simplify expressions for "shift/rot r, 0, a => movi r, 0",
 652            and "sub r, 0, a => neg r, a" case.  */
 653         switch (opc) {
 654         CASE_OP_32_64(shl):
 655         CASE_OP_32_64(shr):
 656         CASE_OP_32_64(sar):
 657         CASE_OP_32_64(rotl):
 658         CASE_OP_32_64(rotr):
 659             if (temp_is_const(args[1]) && temps[args[1]].val == 0) {
 660                 tcg_opt_gen_movi(s, op, args, args[0], 0);
 661                 continue;
 662             }
 663             break;
 664         CASE_OP_32_64(sub):
 665             {
 666                 TCGOpcode neg_op;
 667                 bool have_neg;
 668
 669                 if (temp_is_const(args[2])) {
 670                     /* Proceed with possible constant folding. */
 671                     break;
 672                 }
 673                 if (opc == INDEX_op_sub_i32) {
 674                     neg_op = INDEX_op_neg_i32;
 675                     have_neg = TCG_TARGET_HAS_neg_i32;
 676                 } else {
 677                     neg_op = INDEX_op_neg_i64;
 678                     have_neg = TCG_TARGET_HAS_neg_i64;
 679                 }
 680                 if (!have_neg) {
 681                     break;
 682                 }
 683                 if (temp_is_const(args[1]) && temps[args[1]].val == 0) {
 684                     op->opc = neg_op;
 685                     reset_temp(args[0]);
 686                     args[1] = args[2];
 687                     continue;
 688                 }
 689             }
 690             break;
 691         CASE_OP_32_64(xor):
 692         CASE_OP_32_64(nand):
 693             if (!temp_is_const(args[1])
 694                 && temp_is_const(args[2]) && temps[args[2]].val == -1) {
 695                 i = 1;
 696                 goto try_not;
 697             }
 698             break;
 699         CASE_OP_32_64(nor):
 700             if (!temp_is_const(args[1])
 701                 && temp_is_const(args[2]) && temps[args[2]].val == 0) {
 702                 i = 1;
 703                 goto try_not;
 704             }
 705             break;
 706         CASE_OP_32_64(andc):
 707             if (!temp_is_const(args[2])
 708                 && temp_is_const(args[1]) && temps[args[1]].val == -1) {
 709                 i = 2;
 710                 goto try_not;
 711             }
 712             break;
 713         CASE_OP_32_64(orc):
 714         CASE_OP_32_64(eqv):
 715             if (!temp_is_const(args[2])
 716                 && temp_is_const(args[1]) && temps[args[1]].val == 0) {
 717                 i = 2;
 718                 goto try_not;
 719             }
 720             break;
 721         try_not:
 722             {
 723                 TCGOpcode not_op;
 724                 bool have_not;
 725
 726                 if (def->flags & TCG_OPF_64BIT) {
 727                     not_op = INDEX_op_not_i64;
 728                     have_not = TCG_TARGET_HAS_not_i64;
 729                 } else {
 730                     not_op = INDEX_op_not_i32;
 731                     have_not = TCG_TARGET_HAS_not_i32;
 732                 }
 733                 if (!have_not) {
 734                     break;
 735                 }
 736                 op->opc = not_op;
 737                 reset_temp(args[0]);
 738                 args[1] = args[i];
 739                 continue;
 740             }
 741         default:
 742             break;
 743         }
 744
 745         /* Simplify expression for "op r, a, const => mov r, a" cases */
 746         switch (opc) {
 747         CASE_OP_32_64(add):
 748         CASE_OP_32_64(sub):
 749         CASE_OP_32_64(shl):
 750         CASE_OP_32_64(shr):
 751         CASE_OP_32_64(sar):
 752         CASE_OP_32_64(rotl):
 753         CASE_OP_32_64(rotr):
 754         CASE_OP_32_64(or):
 755         CASE_OP_32_64(xor):
 756         CASE_OP_32_64(andc):
 757             if (!temp_is_const(args[1])
 758                 && temp_is_const(args[2]) && temps[args[2]].val == 0) {
 759                 tcg_opt_gen_mov(s, op, args, args[0], args[1]);
 760                 continue;
 761             }
 762             break;
 763         CASE_OP_32_64(and):
 764         CASE_OP_32_64(orc):
 765         CASE_OP_32_64(eqv):
 766             if (!temp_is_const(args[1])
 767                 && temp_is_const(args[2]) && temps[args[2]].val == -1) {
 768                 tcg_opt_gen_mov(s, op, args, args[0], args[1]);
 769                 continue;
 770             }
 771             break;
 772         default:
 773             break;
 774         }
 775
 776         /* Simplify using known-zero bits. Currently only ops with a single
 777            output argument is supported. */
 778         mask = -1;
 779         affected = -1;
 780         switch (opc) {
 781         CASE_OP_32_64(ext8s):
 782             if ((temps[args[1]].mask & 0x80) != 0) {
 783                 break;
 784             }
 785         CASE_OP_32_64(ext8u):
 786             mask = 0xff;
 787             goto and_const;
 788         CASE_OP_32_64(ext16s):
 789             if ((temps[args[1]].mask & 0x8000) != 0) {
 790                 break;
 791             }
 792         CASE_OP_32_64(ext16u):
 793             mask = 0xffff;
 794             goto and_const;
 795         case INDEX_op_ext32s_i64:
 796             if ((temps[args[1]].mask & 0x80000000) != 0) {
 797                 break;
 798             }
 799         case INDEX_op_ext32u_i64:
 800             mask = 0xffffffffU;
 801             goto and_const;
 802
 803         CASE_OP_32_64(and):
 804             mask = temps[args[2]].mask;
 805             if (temp_is_const(args[2])) {
 806         and_const:
 807                 affected = temps[args[1]].mask & ~mask;
 808             }
 809             mask = temps[args[1]].mask & mask;
 810             break;
 811
 812         case INDEX_op_ext_i32_i64:
 813             if ((temps[args[1]].mask & 0x80000000) != 0) {
 814                 break;
 815             }
 816         case INDEX_op_extu_i32_i64:
 817             /* We do not compute affected as it is a size changing op.  */
 818             mask = (uint32_t)temps[args[1]].mask;
 819             break;
 820
 821         CASE_OP_32_64(andc):
 822             /* Known-zeros does not imply known-ones.  Therefore unless
 823                args[2] is constant, we can't infer anything from it.  */
 824             if (temp_is_const(args[2])) {
 825                 mask = ~temps[args[2]].mask;
 826                 goto and_const;
 827             }
 828             /* But we certainly know nothing outside args[1] may be set. */
 829             mask = temps[args[1]].mask;
 830             break;
 831
 832         case INDEX_op_sar_i32:
 833             if (temp_is_const(args[2])) {
 834                 tmp = temps[args[2]].val & 31;
 835                 mask = (int32_t)temps[args[1]].mask >> tmp;
 836             }
 837             break;
 838         case INDEX_op_sar_i64:
 839             if (temp_is_const(args[2])) {
 840                 tmp = temps[args[2]].val & 63;
 841                 mask = (int64_t)temps[args[1]].mask >> tmp;
 842             }
 843             break;
 844
 845         case INDEX_op_shr_i32:
 846             if (temp_is_const(args[2])) {
 847                 tmp = temps[args[2]].val & 31;
 848                 mask = (uint32_t)temps[args[1]].mask >> tmp;
 849             }
 850             break;
 851         case INDEX_op_shr_i64:
 852             if (temp_is_const(args[2])) {
 853                 tmp = temps[args[2]].val & 63;
 854                 mask = (uint64_t)temps[args[1]].mask >> tmp;
 855             }
 856             break;
 857
 858         case INDEX_op_extrl_i64_i32:
 859             mask = (uint32_t)temps[args[1]].mask;
 860             break;
 861         case INDEX_op_extrh_i64_i32:
 862             mask = (uint64_t)temps[args[1]].mask >> 32;
 863             break;
 864
 865         CASE_OP_32_64(shl):
 866             if (temp_is_const(args[2])) {
 867                 tmp = temps[args[2]].val & (TCG_TARGET_REG_BITS - 1);
 868                 mask = temps[args[1]].mask << tmp;
 869             }
 870             break;
 871
 872         CASE_OP_32_64(neg):
 873             /* Set to 1 all bits to the left of the rightmost.  */
 874             mask = -(temps[args[1]].mask & -temps[args[1]].mask);
 875             break;
 876
 877         CASE_OP_32_64(deposit):
 878             mask = deposit64(temps[args[1]].mask, args[3], args[4],
 879                              temps[args[2]].mask);
 880             break;
 881
 882         CASE_OP_32_64(or):
 883         CASE_OP_32_64(xor):
 884             mask = temps[args[1]].mask | temps[args[2]].mask;
 885             break;
 886
 887         CASE_OP_32_64(setcond):
 888         case INDEX_op_setcond2_i32:
 889             mask = 1;
 890             break;
 891
 892         CASE_OP_32_64(movcond):
 893             mask = temps[args[3]].mask | temps[args[4]].mask;
 894             break;
 895
 896         CASE_OP_32_64(ld8u):
 897             mask = 0xff;
 898             break;
 899         CASE_OP_32_64(ld16u):
 900             mask = 0xffff;
 901             break;
 902         case INDEX_op_ld32u_i64:
 903             mask = 0xffffffffu;
 904             break;
 905
 906         CASE_OP_32_64(qemu_ld):
 907             {
 908                 TCGMemOpIdx oi = args[nb_oargs + nb_iargs];
 909                 TCGMemOp mop = get_memop(oi);
 910                 if (!(mop & MO_SIGN)) {
 911                     mask = (2ULL << ((8 << (mop & MO_SIZE)) - 1)) - 1;
 912                 }
 913             }
 914             break;
 915
 916         default:
 917             break;
 918         }
 919
 920         /* 32-bit ops generate 32-bit results.  For the result is zero test
 921            below, we can ignore high bits, but for further optimizations we
 922            need to record that the high bits contain garbage.  */
 923         partmask = mask;
 924         if (!(def->flags & TCG_OPF_64BIT)) {
 925             mask |= ~(tcg_target_ulong)0xffffffffu;
 926             partmask &= 0xffffffffu;
 927             affected &= 0xffffffffu;
 928         }
 929
 930         if (partmask == 0) {
 931             tcg_debug_assert(nb_oargs == 1);
 932             tcg_opt_gen_movi(s, op, args, args[0], 0);
 933             continue;
 934         }
 935         if (affected == 0) {
 936             tcg_debug_assert(nb_oargs == 1);
 937             tcg_opt_gen_mov(s, op, args, args[0], args[1]);
 938             continue;
 939         }
 940
 941         /* Simplify expression for "op r, a, 0 => movi r, 0" cases */
 942         switch (opc) {
 943         CASE_OP_32_64(and):
 944         CASE_OP_32_64(mul):
 945         CASE_OP_32_64(muluh):
 946         CASE_OP_32_64(mulsh):
 947             if ((temp_is_const(args[2]) && temps[args[2]].val == 0)) {
 948                 tcg_opt_gen_movi(s, op, args, args[0], 0);
 949                 continue;
 950             }
 951             break;
 952         default:
 953             break;
 954         }
 955
 956         /* Simplify expression for "op r, a, a => mov r, a" cases */
 957         switch (opc) {
 958         CASE_OP_32_64(or):
 959         CASE_OP_32_64(and):
 960             if (temps_are_copies(args[1], args[2])) {
 961                 tcg_opt_gen_mov(s, op, args, args[0], args[1]);
 962                 continue;
 963             }
 964             break;
 965         default:
 966             break;
 967         }
 968
 969         /* Simplify expression for "op r, a, a => movi r, 0" cases */
 970         switch (opc) {
 971         CASE_OP_32_64(andc):
 972         CASE_OP_32_64(sub):
 973         CASE_OP_32_64(xor):
 974             if (temps_are_copies(args[1], args[2])) {
 975                 tcg_opt_gen_movi(s, op, args, args[0], 0);
 976                 continue;
 977             }
 978             break;
 979         default:
 980             break;
 981         }
 982
 983         /* Propagate constants through copy operations and do constant
 984            folding.  Constants will be substituted to arguments by register
 985            allocator where needed and possible.  Also detect copies. */
 986         switch (opc) {
 987         CASE_OP_32_64(mov):
 988             tcg_opt_gen_mov(s, op, args, args[0], args[1]);
 989             break;
 990         CASE_OP_32_64(movi):
 991             tcg_opt_gen_movi(s, op, args, args[0], args[1]);
 992             break;
 993
 994         CASE_OP_32_64(not):
 995         CASE_OP_32_64(neg):
 996         CASE_OP_32_64(ext8s):
 997         CASE_OP_32_64(ext8u):
 998         CASE_OP_32_64(ext16s):
 999         CASE_OP_32_64(ext16u):
1000         case INDEX_op_ext32s_i64:
1001         case INDEX_op_ext32u_i64:
1002         case INDEX_op_ext_i32_i64:
1003         case INDEX_op_extu_i32_i64:
1004         case INDEX_op_extrl_i64_i32:
1005         case INDEX_op_extrh_i64_i32:
1006             if (temp_is_const(args[1])) {
1007                 tmp = do_constant_folding(opc, temps[args[1]].val, 0);
1008                 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1009                 break;
1010             }
1011             goto do_default;
1012
1013         CASE_OP_32_64(add):
1014         CASE_OP_32_64(sub):
1015         CASE_OP_32_64(mul):
1016         CASE_OP_32_64(or):
1017         CASE_OP_32_64(and):
1018         CASE_OP_32_64(xor):
1019         CASE_OP_32_64(shl):
1020         CASE_OP_32_64(shr):
1021         CASE_OP_32_64(sar):
1022         CASE_OP_32_64(rotl):
1023         CASE_OP_32_64(rotr):
1024         CASE_OP_32_64(andc):
1025         CASE_OP_32_64(orc):
1026         CASE_OP_32_64(eqv):
1027         CASE_OP_32_64(nand):
1028         CASE_OP_32_64(nor):
1029         CASE_OP_32_64(muluh):
1030         CASE_OP_32_64(mulsh):
1031         CASE_OP_32_64(div):
1032         CASE_OP_32_64(divu):
1033         CASE_OP_32_64(rem):
1034         CASE_OP_32_64(remu):
1035             if (temp_is_const(args[1]) && temp_is_const(args[2])) {
1036                 tmp = do_constant_folding(opc, temps[args[1]].val,
1037                                           temps[args[2]].val);
1038                 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1039                 break;
1040             }
1041             goto do_default;
1042
1043         CASE_OP_32_64(deposit):
1044             if (temp_is_const(args[1]) && temp_is_const(args[2])) {
1045                 tmp = deposit64(temps[args[1]].val, args[3], args[4],
1046                                 temps[args[2]].val);
1047                 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1048                 break;
1049             }
1050             goto do_default;
1051
1052         CASE_OP_32_64(setcond):
1053             tmp = do_constant_folding_cond(opc, args[1], args[2], args[3]);
1054             if (tmp != 2) {
1055                 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1056                 break;
1057             }
1058             goto do_default;
1059
1060         CASE_OP_32_64(brcond):
1061             tmp = do_constant_folding_cond(opc, args[0], args[1], args[2]);
1062             if (tmp != 2) {
1063                 if (tmp) {
1064                     reset_all_temps(nb_temps);
1065                     op->opc = INDEX_op_br;
1066                     args[0] = args[3];
1067                 } else {
1068                     tcg_op_remove(s, op);
1069                 }
1070                 break;
1071             }
1072             goto do_default;
1073
1074         CASE_OP_32_64(movcond):
1075             tmp = do_constant_folding_cond(opc, args[1], args[2], args[5]);
1076             if (tmp != 2) {
1077                 tcg_opt_gen_mov(s, op, args, args[0], args[4-tmp]);
1078                 break;
1079             }
1080             goto do_default;
1081
1082         case INDEX_op_add2_i32:
1083         case INDEX_op_sub2_i32:
1084             if (temp_is_const(args[2]) && temp_is_const(args[3])
1085                 && temp_is_const(args[4]) && temp_is_const(args[5])) {
1086                 uint32_t al = temps[args[2]].val;
1087                 uint32_t ah = temps[args[3]].val;
1088                 uint32_t bl = temps[args[4]].val;
1089                 uint32_t bh = temps[args[5]].val;
1090                 uint64_t a = ((uint64_t)ah << 32) | al;
1091                 uint64_t b = ((uint64_t)bh << 32) | bl;
1092                 TCGArg rl, rh;
1093                 TCGOp *op2 = tcg_op_insert_before(s, op, INDEX_op_movi_i32, 2);
1094                 TCGArg *args2 = &s->gen_opparam_buf[op2->args];
1095
1096                 if (opc == INDEX_op_add2_i32) {
1097                     a += b;
1098                 } else {
1099                     a -= b;
1100                 }
1101
1102                 rl = args[0];
1103                 rh = args[1];
1104                 tcg_opt_gen_movi(s, op, args, rl, (int32_t)a);
1105                 tcg_opt_gen_movi(s, op2, args2, rh, (int32_t)(a >> 32));
1106
1107                 /* We've done all we need to do with the movi.  Skip it.  */
1108                 oi_next = op2->next;
1109                 break;
1110             }
1111             goto do_default;
1112
1113         case INDEX_op_mulu2_i32:
1114             if (temp_is_const(args[2]) && temp_is_const(args[3])) {
1115                 uint32_t a = temps[args[2]].val;
1116                 uint32_t b = temps[args[3]].val;
1117                 uint64_t r = (uint64_t)a * b;
1118                 TCGArg rl, rh;
1119                 TCGOp *op2 = tcg_op_insert_before(s, op, INDEX_op_movi_i32, 2);
1120                 TCGArg *args2 = &s->gen_opparam_buf[op2->args];
1121
1122                 rl = args[0];
1123                 rh = args[1];
1124                 tcg_opt_gen_movi(s, op, args, rl, (int32_t)r);
1125                 tcg_opt_gen_movi(s, op2, args2, rh, (int32_t)(r >> 32));
1126
1127                 /* We've done all we need to do with the movi.  Skip it.  */
1128                 oi_next = op2->next;
1129                 break;
1130             }
1131             goto do_default;
1132
1133         case INDEX_op_brcond2_i32:
1134             tmp = do_constant_folding_cond2(&args[0], &args[2], args[4]);
1135             if (tmp != 2) {
1136                 if (tmp) {
1137             do_brcond_true:
1138                     reset_all_temps(nb_temps);
1139                     op->opc = INDEX_op_br;
1140                     args[0] = args[5];
1141                 } else {
1142             do_brcond_false:
1143                     tcg_op_remove(s, op);
1144                 }
1145             } else if ((args[4] == TCG_COND_LT || args[4] == TCG_COND_GE)
1146                        && temp_is_const(args[2]) && temps[args[2]].val == 0
1147                        && temp_is_const(args[3]) && temps[args[3]].val == 0) {
1148                 /* Simplify LT/GE comparisons vs zero to a single compare
1149                    vs the high word of the input.  */
1150             do_brcond_high:
1151                 reset_all_temps(nb_temps);
1152                 op->opc = INDEX_op_brcond_i32;
1153                 args[0] = args[1];
1154                 args[1] = args[3];
1155                 args[2] = args[4];
1156                 args[3] = args[5];
1157             } else if (args[4] == TCG_COND_EQ) {
1158                 /* Simplify EQ comparisons where one of the pairs
1159                    can be simplified.  */
1160                 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1161                                                args[0], args[2], TCG_COND_EQ);
1162                 if (tmp == 0) {
1163                     goto do_brcond_false;
1164                 } else if (tmp == 1) {
1165                     goto do_brcond_high;
1166                 }
1167                 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1168                                                args[1], args[3], TCG_COND_EQ);
1169                 if (tmp == 0) {
1170                     goto do_brcond_false;
1171                 } else if (tmp != 1) {
1172                     goto do_default;
1173                 }
1174             do_brcond_low:
1175                 reset_all_temps(nb_temps);
1176                 op->opc = INDEX_op_brcond_i32;
1177                 args[1] = args[2];
1178                 args[2] = args[4];
1179                 args[3] = args[5];
1180             } else if (args[4] == TCG_COND_NE) {
1181                 /* Simplify NE comparisons where one of the pairs
1182                    can be simplified.  */
1183                 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1184                                                args[0], args[2], TCG_COND_NE);
1185                 if (tmp == 0) {
1186                     goto do_brcond_high;
1187                 } else if (tmp == 1) {
1188                     goto do_brcond_true;
1189                 }
1190                 tmp = do_constant_folding_cond(INDEX_op_brcond_i32,
1191                                                args[1], args[3], TCG_COND_NE);
1192                 if (tmp == 0) {
1193                     goto do_brcond_low;
1194                 } else if (tmp == 1) {
1195                     goto do_brcond_true;
1196                 }
1197                 goto do_default;
1198             } else {
1199                 goto do_default;
1200             }
1201             break;
1202
1203         case INDEX_op_setcond2_i32:
1204             tmp = do_constant_folding_cond2(&args[1], &args[3], args[5]);
1205             if (tmp != 2) {
1206             do_setcond_const:
1207                 tcg_opt_gen_movi(s, op, args, args[0], tmp);
1208             } else if ((args[5] == TCG_COND_LT || args[5] == TCG_COND_GE)
1209                        && temp_is_const(args[3]) && temps[args[3]].val == 0
1210                        && temp_is_const(args[4]) && temps[args[4]].val == 0) {
1211                 /* Simplify LT/GE comparisons vs zero to a single compare
1212                    vs the high word of the input.  */
1213             do_setcond_high:
1214                 reset_temp(args[0]);
1215                 temps[args[0]].mask = 1;
1216                 op->opc = INDEX_op_setcond_i32;
1217                 args[1] = args[2];
1218                 args[2] = args[4];
1219                 args[3] = args[5];
1220             } else if (args[5] == TCG_COND_EQ) {
1221                 /* Simplify EQ comparisons where one of the pairs
1222                    can be simplified.  */
1223                 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1224                                                args[1], args[3], TCG_COND_EQ);
1225                 if (tmp == 0) {
1226                     goto do_setcond_const;
1227                 } else if (tmp == 1) {
1228                     goto do_setcond_high;
1229                 }
1230                 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1231                                                args[2], args[4], TCG_COND_EQ);
1232                 if (tmp == 0) {
1233                     goto do_setcond_high;
1234                 } else if (tmp != 1) {
1235                     goto do_default;
1236                 }
1237             do_setcond_low:
1238                 reset_temp(args[0]);
1239                 temps[args[0]].mask = 1;
1240                 op->opc = INDEX_op_setcond_i32;
1241                 args[2] = args[3];
1242                 args[3] = args[5];
1243             } else if (args[5] == TCG_COND_NE) {
1244                 /* Simplify NE comparisons where one of the pairs
1245                    can be simplified.  */
1246                 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1247                                                args[1], args[3], TCG_COND_NE);
1248                 if (tmp == 0) {
1249                     goto do_setcond_high;
1250                 } else if (tmp == 1) {
1251                     goto do_setcond_const;
1252                 }
1253                 tmp = do_constant_folding_cond(INDEX_op_setcond_i32,
1254                                                args[2], args[4], TCG_COND_NE);
1255                 if (tmp == 0) {
1256                     goto do_setcond_low;
1257                 } else if (tmp == 1) {
1258                     goto do_setcond_const;
1259                 }
1260                 goto do_default;
1261             } else {
1262                 goto do_default;
1263             }
1264             break;
1265
1266         case INDEX_op_call:
1267             if (!(args[nb_oargs + nb_iargs + 1]
1268                   & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) {
1269                 for (i = 0; i < nb_globals; i++) {
1270                     if (test_bit(i, temps_used.l)) {
1271                         reset_temp(i);
1272                     }
1273                 }
1274             }
1275             goto do_reset_output;
1276
1277         default:
1278         do_default:
1279             /* Default case: we know nothing about operation (or were unable
1280                to compute the operation result) so no propagation is done.
1281                We trash everything if the operation is the end of a basic
1282                block, otherwise we only trash the output args.  "mask" is
1283                the non-zero bits mask for the first output arg.  */
1284             if (def->flags & TCG_OPF_BB_END) {
1285                 reset_all_temps(nb_temps);
1286             } else {
1287         do_reset_output:
1288                 for (i = 0; i < nb_oargs; i++) {
1289                     reset_temp(args[i]);
1290                     /* Save the corresponding known-zero bits mask for the
1291                        first output argument (only one supported so far). */
1292                     if (i == 0) {
1293                         temps[args[i]].mask = mask;
1294                     }
1295                 }
1296             }
1297             break;
1298         }
1299
1300         /* Eliminate duplicate and redundant fence instructions.  */
1301         if (prev_mb_args) {
1302             switch (opc) {
1303             case INDEX_op_mb:
1304                 /* Merge two barriers of the same type into one,
1305                  * or a weaker barrier into a stronger one,
1306                  * or two weaker barriers into a stronger one.
1307                  *   mb X; mb Y => mb X|Y
1308                  *   mb; strl => mb; st
1309                  *   ldaq; mb => ld; mb
1310                  *   ldaq; strl => ld; mb; st
1311                  * Other combinations are also merged into a strong
1312                  * barrier.  This is stricter than specified but for
1313                  * the purposes of TCG is better than not optimizing.
1314                  */
1315                 prev_mb_args[0] |= args[0];
1316                 tcg_op_remove(s, op);
1317                 break;
1318
1319             default:
1320                 /* Opcodes that end the block stop the optimization.  */
1321                 if ((def->flags & TCG_OPF_BB_END) == 0) {
1322                     break;
1323                 }
1324                 /* fallthru */
1325             case INDEX_op_qemu_ld_i32:
1326             case INDEX_op_qemu_ld_i64:
1327             case INDEX_op_qemu_st_i32:
1328             case INDEX_op_qemu_st_i64:
1329             case INDEX_op_call:
1330                 /* Opcodes that touch guest memory stop the optimization.  */
1331                 prev_mb_args = NULL;
1332                 break;
1333             }
1334         } else if (opc == INDEX_op_mb) {
1335             prev_mb_args = args;
1336         }
1337     }
1338 }