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/int128.h" 28 #include "qemu/interval-tree.h" 29 #include "tcg/tcg-op-common.h" 30 #include "tcg-internal.h" 31 #include "tcg-has.h" 32 33 #define CASE_OP_32_64(x) \ 34 glue(glue(case INDEX_op_, x), _i32): \ 35 glue(glue(case INDEX_op_, x), _i64) 36 37 #define CASE_OP_32_64_VEC(x) \ 38 glue(glue(case INDEX_op_, x), _i32): \ 39 glue(glue(case INDEX_op_, x), _i64): \ 40 glue(glue(case INDEX_op_, x), _vec) 41 42 typedef struct MemCopyInfo { 43 IntervalTreeNode itree; 44 QSIMPLEQ_ENTRY (MemCopyInfo) next; 45 TCGTemp *ts; 46 TCGType type; 47 } MemCopyInfo; 48 49 typedef struct TempOptInfo { 50 bool is_const; 51 TCGTemp *prev_copy; 52 TCGTemp *next_copy; 53 QSIMPLEQ_HEAD(, MemCopyInfo) mem_copy; 54 uint64_t val; 55 uint64_t z_mask; /* mask bit is 0 if and only if value bit is 0 */ 56 uint64_t s_mask; /* mask bit is 1 if value bit matches msb */ 57 } TempOptInfo; 58 59 typedef struct OptContext { 60 TCGContext *tcg; 61 TCGOp *prev_mb; 62 TCGTempSet temps_used; 63 64 IntervalTreeRoot mem_copy; 65 QSIMPLEQ_HEAD(, MemCopyInfo) mem_free; 66 67 /* In flight values from optimization. */ 68 TCGType type; 69 } OptContext; 70 71 static inline TempOptInfo *ts_info(TCGTemp *ts) 72 { 73 return ts->state_ptr; 74 } 75 76 static inline TempOptInfo *arg_info(TCGArg arg) 77 { 78 return ts_info(arg_temp(arg)); 79 } 80 81 static inline bool ti_is_const(TempOptInfo *ti) 82 { 83 return ti->is_const; 84 } 85 86 static inline uint64_t ti_const_val(TempOptInfo *ti) 87 { 88 return ti->val; 89 } 90 91 static inline bool ti_is_const_val(TempOptInfo *ti, uint64_t val) 92 { 93 return ti_is_const(ti) && ti_const_val(ti) == val; 94 } 95 96 static inline bool ts_is_const(TCGTemp *ts) 97 { 98 return ti_is_const(ts_info(ts)); 99 } 100 101 static inline bool ts_is_const_val(TCGTemp *ts, uint64_t val) 102 { 103 return ti_is_const_val(ts_info(ts), val); 104 } 105 106 static inline bool arg_is_const(TCGArg arg) 107 { 108 return ts_is_const(arg_temp(arg)); 109 } 110 111 static inline bool arg_is_const_val(TCGArg arg, uint64_t val) 112 { 113 return ts_is_const_val(arg_temp(arg), val); 114 } 115 116 static inline bool ts_is_copy(TCGTemp *ts) 117 { 118 return ts_info(ts)->next_copy != ts; 119 } 120 121 static TCGTemp *cmp_better_copy(TCGTemp *a, TCGTemp *b) 122 { 123 return a->kind < b->kind ? b : a; 124 } 125 126 /* Initialize and activate a temporary. */ 127 static void init_ts_info(OptContext *ctx, TCGTemp *ts) 128 { 129 size_t idx = temp_idx(ts); 130 TempOptInfo *ti; 131 132 if (test_bit(idx, ctx->temps_used.l)) { 133 return; 134 } 135 set_bit(idx, ctx->temps_used.l); 136 137 ti = ts->state_ptr; 138 if (ti == NULL) { 139 ti = tcg_malloc(sizeof(TempOptInfo)); 140 ts->state_ptr = ti; 141 } 142 143 ti->next_copy = ts; 144 ti->prev_copy = ts; 145 QSIMPLEQ_INIT(&ti->mem_copy); 146 if (ts->kind == TEMP_CONST) { 147 ti->is_const = true; 148 ti->val = ts->val; 149 ti->z_mask = ts->val; 150 ti->s_mask = INT64_MIN >> clrsb64(ts->val); 151 } else { 152 ti->is_const = false; 153 ti->z_mask = -1; 154 ti->s_mask = 0; 155 } 156 } 157 158 static MemCopyInfo *mem_copy_first(OptContext *ctx, intptr_t s, intptr_t l) 159 { 160 IntervalTreeNode *r = interval_tree_iter_first(&ctx->mem_copy, s, l); 161 return r ? container_of(r, MemCopyInfo, itree) : NULL; 162 } 163 164 static MemCopyInfo *mem_copy_next(MemCopyInfo *mem, intptr_t s, intptr_t l) 165 { 166 IntervalTreeNode *r = interval_tree_iter_next(&mem->itree, s, l); 167 return r ? container_of(r, MemCopyInfo, itree) : NULL; 168 } 169 170 static void remove_mem_copy(OptContext *ctx, MemCopyInfo *mc) 171 { 172 TCGTemp *ts = mc->ts; 173 TempOptInfo *ti = ts_info(ts); 174 175 interval_tree_remove(&mc->itree, &ctx->mem_copy); 176 QSIMPLEQ_REMOVE(&ti->mem_copy, mc, MemCopyInfo, next); 177 QSIMPLEQ_INSERT_TAIL(&ctx->mem_free, mc, next); 178 } 179 180 static void remove_mem_copy_in(OptContext *ctx, intptr_t s, intptr_t l) 181 { 182 while (true) { 183 MemCopyInfo *mc = mem_copy_first(ctx, s, l); 184 if (!mc) { 185 break; 186 } 187 remove_mem_copy(ctx, mc); 188 } 189 } 190 191 static void remove_mem_copy_all(OptContext *ctx) 192 { 193 remove_mem_copy_in(ctx, 0, -1); 194 tcg_debug_assert(interval_tree_is_empty(&ctx->mem_copy)); 195 } 196 197 static TCGTemp *find_better_copy(TCGTemp *ts) 198 { 199 TCGTemp *i, *ret; 200 201 /* If this is already readonly, we can't do better. */ 202 if (temp_readonly(ts)) { 203 return ts; 204 } 205 206 ret = ts; 207 for (i = ts_info(ts)->next_copy; i != ts; i = ts_info(i)->next_copy) { 208 ret = cmp_better_copy(ret, i); 209 } 210 return ret; 211 } 212 213 static void move_mem_copies(TCGTemp *dst_ts, TCGTemp *src_ts) 214 { 215 TempOptInfo *si = ts_info(src_ts); 216 TempOptInfo *di = ts_info(dst_ts); 217 MemCopyInfo *mc; 218 219 QSIMPLEQ_FOREACH(mc, &si->mem_copy, next) { 220 tcg_debug_assert(mc->ts == src_ts); 221 mc->ts = dst_ts; 222 } 223 QSIMPLEQ_CONCAT(&di->mem_copy, &si->mem_copy); 224 } 225 226 /* Reset TEMP's state, possibly removing the temp for the list of copies. */ 227 static void reset_ts(OptContext *ctx, TCGTemp *ts) 228 { 229 TempOptInfo *ti = ts_info(ts); 230 TCGTemp *pts = ti->prev_copy; 231 TCGTemp *nts = ti->next_copy; 232 TempOptInfo *pi = ts_info(pts); 233 TempOptInfo *ni = ts_info(nts); 234 235 ni->prev_copy = ti->prev_copy; 236 pi->next_copy = ti->next_copy; 237 ti->next_copy = ts; 238 ti->prev_copy = ts; 239 ti->is_const = false; 240 ti->z_mask = -1; 241 ti->s_mask = 0; 242 243 if (!QSIMPLEQ_EMPTY(&ti->mem_copy)) { 244 if (ts == nts) { 245 /* Last temp copy being removed, the mem copies die. */ 246 MemCopyInfo *mc; 247 QSIMPLEQ_FOREACH(mc, &ti->mem_copy, next) { 248 interval_tree_remove(&mc->itree, &ctx->mem_copy); 249 } 250 QSIMPLEQ_CONCAT(&ctx->mem_free, &ti->mem_copy); 251 } else { 252 move_mem_copies(find_better_copy(nts), ts); 253 } 254 } 255 } 256 257 static void reset_temp(OptContext *ctx, TCGArg arg) 258 { 259 reset_ts(ctx, arg_temp(arg)); 260 } 261 262 static void record_mem_copy(OptContext *ctx, TCGType type, 263 TCGTemp *ts, intptr_t start, intptr_t last) 264 { 265 MemCopyInfo *mc; 266 TempOptInfo *ti; 267 268 mc = QSIMPLEQ_FIRST(&ctx->mem_free); 269 if (mc) { 270 QSIMPLEQ_REMOVE_HEAD(&ctx->mem_free, next); 271 } else { 272 mc = tcg_malloc(sizeof(*mc)); 273 } 274 275 memset(mc, 0, sizeof(*mc)); 276 mc->itree.start = start; 277 mc->itree.last = last; 278 mc->type = type; 279 interval_tree_insert(&mc->itree, &ctx->mem_copy); 280 281 ts = find_better_copy(ts); 282 ti = ts_info(ts); 283 mc->ts = ts; 284 QSIMPLEQ_INSERT_TAIL(&ti->mem_copy, mc, next); 285 } 286 287 static bool ts_are_copies(TCGTemp *ts1, TCGTemp *ts2) 288 { 289 TCGTemp *i; 290 291 if (ts1 == ts2) { 292 return true; 293 } 294 295 if (!ts_is_copy(ts1) || !ts_is_copy(ts2)) { 296 return false; 297 } 298 299 for (i = ts_info(ts1)->next_copy; i != ts1; i = ts_info(i)->next_copy) { 300 if (i == ts2) { 301 return true; 302 } 303 } 304 305 return false; 306 } 307 308 static bool args_are_copies(TCGArg arg1, TCGArg arg2) 309 { 310 return ts_are_copies(arg_temp(arg1), arg_temp(arg2)); 311 } 312 313 static TCGTemp *find_mem_copy_for(OptContext *ctx, TCGType type, intptr_t s) 314 { 315 MemCopyInfo *mc; 316 317 for (mc = mem_copy_first(ctx, s, s); mc; mc = mem_copy_next(mc, s, s)) { 318 if (mc->itree.start == s && mc->type == type) { 319 return find_better_copy(mc->ts); 320 } 321 } 322 return NULL; 323 } 324 325 static TCGArg arg_new_constant(OptContext *ctx, uint64_t val) 326 { 327 TCGType type = ctx->type; 328 TCGTemp *ts; 329 330 if (type == TCG_TYPE_I32) { 331 val = (int32_t)val; 332 } 333 334 ts = tcg_constant_internal(type, val); 335 init_ts_info(ctx, ts); 336 337 return temp_arg(ts); 338 } 339 340 static TCGArg arg_new_temp(OptContext *ctx) 341 { 342 TCGTemp *ts = tcg_temp_new_internal(ctx->type, TEMP_EBB); 343 init_ts_info(ctx, ts); 344 return temp_arg(ts); 345 } 346 347 static TCGOp *opt_insert_after(OptContext *ctx, TCGOp *op, 348 TCGOpcode opc, unsigned narg) 349 { 350 return tcg_op_insert_after(ctx->tcg, op, opc, ctx->type, narg); 351 } 352 353 static TCGOp *opt_insert_before(OptContext *ctx, TCGOp *op, 354 TCGOpcode opc, unsigned narg) 355 { 356 return tcg_op_insert_before(ctx->tcg, op, opc, ctx->type, narg); 357 } 358 359 static bool tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src) 360 { 361 TCGTemp *dst_ts = arg_temp(dst); 362 TCGTemp *src_ts = arg_temp(src); 363 TempOptInfo *di; 364 TempOptInfo *si; 365 TCGOpcode new_op; 366 367 if (ts_are_copies(dst_ts, src_ts)) { 368 tcg_op_remove(ctx->tcg, op); 369 return true; 370 } 371 372 reset_ts(ctx, dst_ts); 373 di = ts_info(dst_ts); 374 si = ts_info(src_ts); 375 376 switch (ctx->type) { 377 case TCG_TYPE_I32: 378 case TCG_TYPE_I64: 379 new_op = INDEX_op_mov; 380 break; 381 case TCG_TYPE_V64: 382 case TCG_TYPE_V128: 383 case TCG_TYPE_V256: 384 /* TCGOP_TYPE and TCGOP_VECE remain unchanged. */ 385 new_op = INDEX_op_mov_vec; 386 break; 387 default: 388 g_assert_not_reached(); 389 } 390 op->opc = new_op; 391 op->args[0] = dst; 392 op->args[1] = src; 393 394 di->z_mask = si->z_mask; 395 di->s_mask = si->s_mask; 396 397 if (src_ts->type == dst_ts->type) { 398 TempOptInfo *ni = ts_info(si->next_copy); 399 400 di->next_copy = si->next_copy; 401 di->prev_copy = src_ts; 402 ni->prev_copy = dst_ts; 403 si->next_copy = dst_ts; 404 di->is_const = si->is_const; 405 di->val = si->val; 406 407 if (!QSIMPLEQ_EMPTY(&si->mem_copy) 408 && cmp_better_copy(src_ts, dst_ts) == dst_ts) { 409 move_mem_copies(dst_ts, src_ts); 410 } 411 } 412 return true; 413 } 414 415 static bool tcg_opt_gen_movi(OptContext *ctx, TCGOp *op, 416 TCGArg dst, uint64_t val) 417 { 418 /* Convert movi to mov with constant temp. */ 419 return tcg_opt_gen_mov(ctx, op, dst, arg_new_constant(ctx, val)); 420 } 421 422 static uint64_t do_constant_folding_2(TCGOpcode op, uint64_t x, uint64_t y) 423 { 424 uint64_t l64, h64; 425 426 switch (op) { 427 case INDEX_op_add: 428 return x + y; 429 430 CASE_OP_32_64(sub): 431 return x - y; 432 433 CASE_OP_32_64(mul): 434 return x * y; 435 436 case INDEX_op_and: 437 case INDEX_op_and_vec: 438 return x & y; 439 440 case INDEX_op_or: 441 case INDEX_op_or_vec: 442 return x | y; 443 444 case INDEX_op_xor: 445 case INDEX_op_xor_vec: 446 return x ^ y; 447 448 case INDEX_op_shl_i32: 449 return (uint32_t)x << (y & 31); 450 451 case INDEX_op_shl_i64: 452 return (uint64_t)x << (y & 63); 453 454 case INDEX_op_shr_i32: 455 return (uint32_t)x >> (y & 31); 456 457 case INDEX_op_shr_i64: 458 return (uint64_t)x >> (y & 63); 459 460 case INDEX_op_sar_i32: 461 return (int32_t)x >> (y & 31); 462 463 case INDEX_op_sar_i64: 464 return (int64_t)x >> (y & 63); 465 466 case INDEX_op_rotr_i32: 467 return ror32(x, y & 31); 468 469 case INDEX_op_rotr_i64: 470 return ror64(x, y & 63); 471 472 case INDEX_op_rotl_i32: 473 return rol32(x, y & 31); 474 475 case INDEX_op_rotl_i64: 476 return rol64(x, y & 63); 477 478 CASE_OP_32_64_VEC(not): 479 return ~x; 480 481 CASE_OP_32_64(neg): 482 return -x; 483 484 case INDEX_op_andc: 485 case INDEX_op_andc_vec: 486 return x & ~y; 487 488 case INDEX_op_orc: 489 case INDEX_op_orc_vec: 490 return x | ~y; 491 492 case INDEX_op_eqv: 493 case INDEX_op_eqv_vec: 494 return ~(x ^ y); 495 496 CASE_OP_32_64_VEC(nand): 497 return ~(x & y); 498 499 CASE_OP_32_64_VEC(nor): 500 return ~(x | y); 501 502 case INDEX_op_clz_i32: 503 return (uint32_t)x ? clz32(x) : y; 504 505 case INDEX_op_clz_i64: 506 return x ? clz64(x) : y; 507 508 case INDEX_op_ctz_i32: 509 return (uint32_t)x ? ctz32(x) : y; 510 511 case INDEX_op_ctz_i64: 512 return x ? ctz64(x) : y; 513 514 case INDEX_op_ctpop_i32: 515 return ctpop32(x); 516 517 case INDEX_op_ctpop_i64: 518 return ctpop64(x); 519 520 CASE_OP_32_64(bswap16): 521 x = bswap16(x); 522 return y & TCG_BSWAP_OS ? (int16_t)x : x; 523 524 CASE_OP_32_64(bswap32): 525 x = bswap32(x); 526 return y & TCG_BSWAP_OS ? (int32_t)x : x; 527 528 case INDEX_op_bswap64_i64: 529 return bswap64(x); 530 531 case INDEX_op_ext_i32_i64: 532 return (int32_t)x; 533 534 case INDEX_op_extu_i32_i64: 535 case INDEX_op_extrl_i64_i32: 536 return (uint32_t)x; 537 538 case INDEX_op_extrh_i64_i32: 539 return (uint64_t)x >> 32; 540 541 case INDEX_op_muluh_i32: 542 return ((uint64_t)(uint32_t)x * (uint32_t)y) >> 32; 543 case INDEX_op_mulsh_i32: 544 return ((int64_t)(int32_t)x * (int32_t)y) >> 32; 545 546 case INDEX_op_muluh_i64: 547 mulu64(&l64, &h64, x, y); 548 return h64; 549 case INDEX_op_mulsh_i64: 550 muls64(&l64, &h64, x, y); 551 return h64; 552 553 case INDEX_op_div_i32: 554 /* Avoid crashing on divide by zero, otherwise undefined. */ 555 return (int32_t)x / ((int32_t)y ? : 1); 556 case INDEX_op_divu_i32: 557 return (uint32_t)x / ((uint32_t)y ? : 1); 558 case INDEX_op_div_i64: 559 return (int64_t)x / ((int64_t)y ? : 1); 560 case INDEX_op_divu_i64: 561 return (uint64_t)x / ((uint64_t)y ? : 1); 562 563 case INDEX_op_rem_i32: 564 return (int32_t)x % ((int32_t)y ? : 1); 565 case INDEX_op_remu_i32: 566 return (uint32_t)x % ((uint32_t)y ? : 1); 567 case INDEX_op_rem_i64: 568 return (int64_t)x % ((int64_t)y ? : 1); 569 case INDEX_op_remu_i64: 570 return (uint64_t)x % ((uint64_t)y ? : 1); 571 572 default: 573 g_assert_not_reached(); 574 } 575 } 576 577 static uint64_t do_constant_folding(TCGOpcode op, TCGType type, 578 uint64_t x, uint64_t y) 579 { 580 uint64_t res = do_constant_folding_2(op, x, y); 581 if (type == TCG_TYPE_I32) { 582 res = (int32_t)res; 583 } 584 return res; 585 } 586 587 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c) 588 { 589 switch (c) { 590 case TCG_COND_EQ: 591 return x == y; 592 case TCG_COND_NE: 593 return x != y; 594 case TCG_COND_LT: 595 return (int32_t)x < (int32_t)y; 596 case TCG_COND_GE: 597 return (int32_t)x >= (int32_t)y; 598 case TCG_COND_LE: 599 return (int32_t)x <= (int32_t)y; 600 case TCG_COND_GT: 601 return (int32_t)x > (int32_t)y; 602 case TCG_COND_LTU: 603 return x < y; 604 case TCG_COND_GEU: 605 return x >= y; 606 case TCG_COND_LEU: 607 return x <= y; 608 case TCG_COND_GTU: 609 return x > y; 610 case TCG_COND_TSTEQ: 611 return (x & y) == 0; 612 case TCG_COND_TSTNE: 613 return (x & y) != 0; 614 case TCG_COND_ALWAYS: 615 case TCG_COND_NEVER: 616 break; 617 } 618 g_assert_not_reached(); 619 } 620 621 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c) 622 { 623 switch (c) { 624 case TCG_COND_EQ: 625 return x == y; 626 case TCG_COND_NE: 627 return x != y; 628 case TCG_COND_LT: 629 return (int64_t)x < (int64_t)y; 630 case TCG_COND_GE: 631 return (int64_t)x >= (int64_t)y; 632 case TCG_COND_LE: 633 return (int64_t)x <= (int64_t)y; 634 case TCG_COND_GT: 635 return (int64_t)x > (int64_t)y; 636 case TCG_COND_LTU: 637 return x < y; 638 case TCG_COND_GEU: 639 return x >= y; 640 case TCG_COND_LEU: 641 return x <= y; 642 case TCG_COND_GTU: 643 return x > y; 644 case TCG_COND_TSTEQ: 645 return (x & y) == 0; 646 case TCG_COND_TSTNE: 647 return (x & y) != 0; 648 case TCG_COND_ALWAYS: 649 case TCG_COND_NEVER: 650 break; 651 } 652 g_assert_not_reached(); 653 } 654 655 static int do_constant_folding_cond_eq(TCGCond c) 656 { 657 switch (c) { 658 case TCG_COND_GT: 659 case TCG_COND_LTU: 660 case TCG_COND_LT: 661 case TCG_COND_GTU: 662 case TCG_COND_NE: 663 return 0; 664 case TCG_COND_GE: 665 case TCG_COND_GEU: 666 case TCG_COND_LE: 667 case TCG_COND_LEU: 668 case TCG_COND_EQ: 669 return 1; 670 case TCG_COND_TSTEQ: 671 case TCG_COND_TSTNE: 672 return -1; 673 case TCG_COND_ALWAYS: 674 case TCG_COND_NEVER: 675 break; 676 } 677 g_assert_not_reached(); 678 } 679 680 /* 681 * Return -1 if the condition can't be simplified, 682 * and the result of the condition (0 or 1) if it can. 683 */ 684 static int do_constant_folding_cond(TCGType type, TCGArg x, 685 TCGArg y, TCGCond c) 686 { 687 if (arg_is_const(x) && arg_is_const(y)) { 688 uint64_t xv = arg_info(x)->val; 689 uint64_t yv = arg_info(y)->val; 690 691 switch (type) { 692 case TCG_TYPE_I32: 693 return do_constant_folding_cond_32(xv, yv, c); 694 case TCG_TYPE_I64: 695 return do_constant_folding_cond_64(xv, yv, c); 696 default: 697 /* Only scalar comparisons are optimizable */ 698 return -1; 699 } 700 } else if (args_are_copies(x, y)) { 701 return do_constant_folding_cond_eq(c); 702 } else if (arg_is_const_val(y, 0)) { 703 switch (c) { 704 case TCG_COND_LTU: 705 case TCG_COND_TSTNE: 706 return 0; 707 case TCG_COND_GEU: 708 case TCG_COND_TSTEQ: 709 return 1; 710 default: 711 return -1; 712 } 713 } 714 return -1; 715 } 716 717 /** 718 * swap_commutative: 719 * @dest: TCGArg of the destination argument, or NO_DEST. 720 * @p1: first paired argument 721 * @p2: second paired argument 722 * 723 * If *@p1 is a constant and *@p2 is not, swap. 724 * If *@p2 matches @dest, swap. 725 * Return true if a swap was performed. 726 */ 727 728 #define NO_DEST temp_arg(NULL) 729 730 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2) 731 { 732 TCGArg a1 = *p1, a2 = *p2; 733 int sum = 0; 734 sum += arg_is_const(a1); 735 sum -= arg_is_const(a2); 736 737 /* Prefer the constant in second argument, and then the form 738 op a, a, b, which is better handled on non-RISC hosts. */ 739 if (sum > 0 || (sum == 0 && dest == a2)) { 740 *p1 = a2; 741 *p2 = a1; 742 return true; 743 } 744 return false; 745 } 746 747 static bool swap_commutative2(TCGArg *p1, TCGArg *p2) 748 { 749 int sum = 0; 750 sum += arg_is_const(p1[0]); 751 sum += arg_is_const(p1[1]); 752 sum -= arg_is_const(p2[0]); 753 sum -= arg_is_const(p2[1]); 754 if (sum > 0) { 755 TCGArg t; 756 t = p1[0], p1[0] = p2[0], p2[0] = t; 757 t = p1[1], p1[1] = p2[1], p2[1] = t; 758 return true; 759 } 760 return false; 761 } 762 763 /* 764 * Return -1 if the condition can't be simplified, 765 * and the result of the condition (0 or 1) if it can. 766 */ 767 static int do_constant_folding_cond1(OptContext *ctx, TCGOp *op, TCGArg dest, 768 TCGArg *p1, TCGArg *p2, TCGArg *pcond) 769 { 770 TCGCond cond; 771 TempOptInfo *i1; 772 bool swap; 773 int r; 774 775 swap = swap_commutative(dest, p1, p2); 776 cond = *pcond; 777 if (swap) { 778 *pcond = cond = tcg_swap_cond(cond); 779 } 780 781 r = do_constant_folding_cond(ctx->type, *p1, *p2, cond); 782 if (r >= 0) { 783 return r; 784 } 785 if (!is_tst_cond(cond)) { 786 return -1; 787 } 788 789 i1 = arg_info(*p1); 790 791 /* 792 * TSTNE x,x -> NE x,0 793 * TSTNE x,i -> NE x,0 if i includes all nonzero bits of x 794 */ 795 if (args_are_copies(*p1, *p2) || 796 (arg_is_const(*p2) && (i1->z_mask & ~arg_info(*p2)->val) == 0)) { 797 *p2 = arg_new_constant(ctx, 0); 798 *pcond = tcg_tst_eqne_cond(cond); 799 return -1; 800 } 801 802 /* TSTNE x,i -> LT x,0 if i only includes sign bit copies */ 803 if (arg_is_const(*p2) && (arg_info(*p2)->val & ~i1->s_mask) == 0) { 804 *p2 = arg_new_constant(ctx, 0); 805 *pcond = tcg_tst_ltge_cond(cond); 806 return -1; 807 } 808 809 /* Expand to AND with a temporary if no backend support. */ 810 if (!TCG_TARGET_HAS_tst) { 811 TCGOp *op2 = opt_insert_before(ctx, op, INDEX_op_and, 3); 812 TCGArg tmp = arg_new_temp(ctx); 813 814 op2->args[0] = tmp; 815 op2->args[1] = *p1; 816 op2->args[2] = *p2; 817 818 *p1 = tmp; 819 *p2 = arg_new_constant(ctx, 0); 820 *pcond = tcg_tst_eqne_cond(cond); 821 } 822 return -1; 823 } 824 825 static int do_constant_folding_cond2(OptContext *ctx, TCGOp *op, TCGArg *args) 826 { 827 TCGArg al, ah, bl, bh; 828 TCGCond c; 829 bool swap; 830 int r; 831 832 swap = swap_commutative2(args, args + 2); 833 c = args[4]; 834 if (swap) { 835 args[4] = c = tcg_swap_cond(c); 836 } 837 838 al = args[0]; 839 ah = args[1]; 840 bl = args[2]; 841 bh = args[3]; 842 843 if (arg_is_const(bl) && arg_is_const(bh)) { 844 tcg_target_ulong blv = arg_info(bl)->val; 845 tcg_target_ulong bhv = arg_info(bh)->val; 846 uint64_t b = deposit64(blv, 32, 32, bhv); 847 848 if (arg_is_const(al) && arg_is_const(ah)) { 849 tcg_target_ulong alv = arg_info(al)->val; 850 tcg_target_ulong ahv = arg_info(ah)->val; 851 uint64_t a = deposit64(alv, 32, 32, ahv); 852 853 r = do_constant_folding_cond_64(a, b, c); 854 if (r >= 0) { 855 return r; 856 } 857 } 858 859 if (b == 0) { 860 switch (c) { 861 case TCG_COND_LTU: 862 case TCG_COND_TSTNE: 863 return 0; 864 case TCG_COND_GEU: 865 case TCG_COND_TSTEQ: 866 return 1; 867 default: 868 break; 869 } 870 } 871 872 /* TSTNE x,-1 -> NE x,0 */ 873 if (b == -1 && is_tst_cond(c)) { 874 args[3] = args[2] = arg_new_constant(ctx, 0); 875 args[4] = tcg_tst_eqne_cond(c); 876 return -1; 877 } 878 879 /* TSTNE x,sign -> LT x,0 */ 880 if (b == INT64_MIN && is_tst_cond(c)) { 881 /* bl must be 0, so copy that to bh */ 882 args[3] = bl; 883 args[4] = tcg_tst_ltge_cond(c); 884 return -1; 885 } 886 } 887 888 if (args_are_copies(al, bl) && args_are_copies(ah, bh)) { 889 r = do_constant_folding_cond_eq(c); 890 if (r >= 0) { 891 return r; 892 } 893 894 /* TSTNE x,x -> NE x,0 */ 895 if (is_tst_cond(c)) { 896 args[3] = args[2] = arg_new_constant(ctx, 0); 897 args[4] = tcg_tst_eqne_cond(c); 898 return -1; 899 } 900 } 901 902 /* Expand to AND with a temporary if no backend support. */ 903 if (!TCG_TARGET_HAS_tst && is_tst_cond(c)) { 904 TCGOp *op1 = opt_insert_before(ctx, op, INDEX_op_and, 3); 905 TCGOp *op2 = opt_insert_before(ctx, op, INDEX_op_and, 3); 906 TCGArg t1 = arg_new_temp(ctx); 907 TCGArg t2 = arg_new_temp(ctx); 908 909 op1->args[0] = t1; 910 op1->args[1] = al; 911 op1->args[2] = bl; 912 op2->args[0] = t2; 913 op2->args[1] = ah; 914 op2->args[2] = bh; 915 916 args[0] = t1; 917 args[1] = t2; 918 args[3] = args[2] = arg_new_constant(ctx, 0); 919 args[4] = tcg_tst_eqne_cond(c); 920 } 921 return -1; 922 } 923 924 static void init_arguments(OptContext *ctx, TCGOp *op, int nb_args) 925 { 926 for (int i = 0; i < nb_args; i++) { 927 TCGTemp *ts = arg_temp(op->args[i]); 928 init_ts_info(ctx, ts); 929 } 930 } 931 932 static void copy_propagate(OptContext *ctx, TCGOp *op, 933 int nb_oargs, int nb_iargs) 934 { 935 for (int i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 936 TCGTemp *ts = arg_temp(op->args[i]); 937 if (ts_is_copy(ts)) { 938 op->args[i] = temp_arg(find_better_copy(ts)); 939 } 940 } 941 } 942 943 static void finish_bb(OptContext *ctx) 944 { 945 /* We only optimize memory barriers across basic blocks. */ 946 ctx->prev_mb = NULL; 947 } 948 949 static void finish_ebb(OptContext *ctx) 950 { 951 finish_bb(ctx); 952 /* We only optimize across extended basic blocks. */ 953 memset(&ctx->temps_used, 0, sizeof(ctx->temps_used)); 954 remove_mem_copy_all(ctx); 955 } 956 957 static bool finish_folding(OptContext *ctx, TCGOp *op) 958 { 959 const TCGOpDef *def = &tcg_op_defs[op->opc]; 960 int i, nb_oargs; 961 962 nb_oargs = def->nb_oargs; 963 for (i = 0; i < nb_oargs; i++) { 964 TCGTemp *ts = arg_temp(op->args[i]); 965 reset_ts(ctx, ts); 966 } 967 return true; 968 } 969 970 /* 971 * The fold_* functions return true when processing is complete, 972 * usually by folding the operation to a constant or to a copy, 973 * and calling tcg_opt_gen_{mov,movi}. They may do other things, 974 * like collect information about the value produced, for use in 975 * optimizing a subsequent operation. 976 * 977 * These first fold_* functions are all helpers, used by other 978 * folders for more specific operations. 979 */ 980 981 static bool fold_const1(OptContext *ctx, TCGOp *op) 982 { 983 if (arg_is_const(op->args[1])) { 984 uint64_t t; 985 986 t = arg_info(op->args[1])->val; 987 t = do_constant_folding(op->opc, ctx->type, t, 0); 988 return tcg_opt_gen_movi(ctx, op, op->args[0], t); 989 } 990 return false; 991 } 992 993 static bool fold_const2(OptContext *ctx, TCGOp *op) 994 { 995 if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { 996 uint64_t t1 = arg_info(op->args[1])->val; 997 uint64_t t2 = arg_info(op->args[2])->val; 998 999 t1 = do_constant_folding(op->opc, ctx->type, t1, t2); 1000 return tcg_opt_gen_movi(ctx, op, op->args[0], t1); 1001 } 1002 return false; 1003 } 1004 1005 static bool fold_commutative(OptContext *ctx, TCGOp *op) 1006 { 1007 swap_commutative(op->args[0], &op->args[1], &op->args[2]); 1008 return false; 1009 } 1010 1011 static bool fold_const2_commutative(OptContext *ctx, TCGOp *op) 1012 { 1013 swap_commutative(op->args[0], &op->args[1], &op->args[2]); 1014 return fold_const2(ctx, op); 1015 } 1016 1017 /* 1018 * Record "zero" and "sign" masks for the single output of @op. 1019 * See TempOptInfo definition of z_mask and s_mask. 1020 * If z_mask allows, fold the output to constant zero. 1021 * The passed s_mask may be augmented by z_mask. 1022 */ 1023 static bool fold_masks_zs(OptContext *ctx, TCGOp *op, 1024 uint64_t z_mask, int64_t s_mask) 1025 { 1026 const TCGOpDef *def = &tcg_op_defs[op->opc]; 1027 TCGTemp *ts; 1028 TempOptInfo *ti; 1029 int rep; 1030 1031 /* Only single-output opcodes are supported here. */ 1032 tcg_debug_assert(def->nb_oargs == 1); 1033 1034 /* 1035 * 32-bit ops generate 32-bit results, which for the purpose of 1036 * simplifying tcg are sign-extended. Certainly that's how we 1037 * represent our constants elsewhere. Note that the bits will 1038 * be reset properly for a 64-bit value when encountering the 1039 * type changing opcodes. 1040 */ 1041 if (ctx->type == TCG_TYPE_I32) { 1042 z_mask = (int32_t)z_mask; 1043 s_mask |= INT32_MIN; 1044 } 1045 1046 if (z_mask == 0) { 1047 return tcg_opt_gen_movi(ctx, op, op->args[0], 0); 1048 } 1049 1050 ts = arg_temp(op->args[0]); 1051 reset_ts(ctx, ts); 1052 1053 ti = ts_info(ts); 1054 ti->z_mask = z_mask; 1055 1056 /* Canonicalize s_mask and incorporate data from z_mask. */ 1057 rep = clz64(~s_mask); 1058 rep = MAX(rep, clz64(z_mask)); 1059 rep = MAX(rep - 1, 0); 1060 ti->s_mask = INT64_MIN >> rep; 1061 1062 return true; 1063 } 1064 1065 static bool fold_masks_z(OptContext *ctx, TCGOp *op, uint64_t z_mask) 1066 { 1067 return fold_masks_zs(ctx, op, z_mask, 0); 1068 } 1069 1070 static bool fold_masks_s(OptContext *ctx, TCGOp *op, uint64_t s_mask) 1071 { 1072 return fold_masks_zs(ctx, op, -1, s_mask); 1073 } 1074 1075 /* 1076 * An "affected" mask bit is 0 if and only if the result is identical 1077 * to the first input. Thus if the entire mask is 0, the operation 1078 * is equivalent to a copy. 1079 */ 1080 static bool fold_affected_mask(OptContext *ctx, TCGOp *op, uint64_t a_mask) 1081 { 1082 if (ctx->type == TCG_TYPE_I32) { 1083 a_mask = (uint32_t)a_mask; 1084 } 1085 if (a_mask == 0) { 1086 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); 1087 } 1088 return false; 1089 } 1090 1091 /* 1092 * Convert @op to NOT, if NOT is supported by the host. 1093 * Return true f the conversion is successful, which will still 1094 * indicate that the processing is complete. 1095 */ 1096 static bool fold_not(OptContext *ctx, TCGOp *op); 1097 static bool fold_to_not(OptContext *ctx, TCGOp *op, int idx) 1098 { 1099 TCGOpcode not_op; 1100 bool have_not; 1101 1102 switch (ctx->type) { 1103 case TCG_TYPE_I32: 1104 not_op = INDEX_op_not_i32; 1105 have_not = TCG_TARGET_HAS_not_i32; 1106 break; 1107 case TCG_TYPE_I64: 1108 not_op = INDEX_op_not_i64; 1109 have_not = TCG_TARGET_HAS_not_i64; 1110 break; 1111 case TCG_TYPE_V64: 1112 case TCG_TYPE_V128: 1113 case TCG_TYPE_V256: 1114 not_op = INDEX_op_not_vec; 1115 have_not = TCG_TARGET_HAS_not_vec; 1116 break; 1117 default: 1118 g_assert_not_reached(); 1119 } 1120 if (have_not) { 1121 op->opc = not_op; 1122 op->args[1] = op->args[idx]; 1123 return fold_not(ctx, op); 1124 } 1125 return false; 1126 } 1127 1128 /* If the binary operation has first argument @i, fold to @i. */ 1129 static bool fold_ix_to_i(OptContext *ctx, TCGOp *op, uint64_t i) 1130 { 1131 if (arg_is_const_val(op->args[1], i)) { 1132 return tcg_opt_gen_movi(ctx, op, op->args[0], i); 1133 } 1134 return false; 1135 } 1136 1137 /* If the binary operation has first argument @i, fold to NOT. */ 1138 static bool fold_ix_to_not(OptContext *ctx, TCGOp *op, uint64_t i) 1139 { 1140 if (arg_is_const_val(op->args[1], i)) { 1141 return fold_to_not(ctx, op, 2); 1142 } 1143 return false; 1144 } 1145 1146 /* If the binary operation has second argument @i, fold to @i. */ 1147 static bool fold_xi_to_i(OptContext *ctx, TCGOp *op, uint64_t i) 1148 { 1149 if (arg_is_const_val(op->args[2], i)) { 1150 return tcg_opt_gen_movi(ctx, op, op->args[0], i); 1151 } 1152 return false; 1153 } 1154 1155 /* If the binary operation has second argument @i, fold to identity. */ 1156 static bool fold_xi_to_x(OptContext *ctx, TCGOp *op, uint64_t i) 1157 { 1158 if (arg_is_const_val(op->args[2], i)) { 1159 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); 1160 } 1161 return false; 1162 } 1163 1164 /* If the binary operation has second argument @i, fold to NOT. */ 1165 static bool fold_xi_to_not(OptContext *ctx, TCGOp *op, uint64_t i) 1166 { 1167 if (arg_is_const_val(op->args[2], i)) { 1168 return fold_to_not(ctx, op, 1); 1169 } 1170 return false; 1171 } 1172 1173 /* If the binary operation has both arguments equal, fold to @i. */ 1174 static bool fold_xx_to_i(OptContext *ctx, TCGOp *op, uint64_t i) 1175 { 1176 if (args_are_copies(op->args[1], op->args[2])) { 1177 return tcg_opt_gen_movi(ctx, op, op->args[0], i); 1178 } 1179 return false; 1180 } 1181 1182 /* If the binary operation has both arguments equal, fold to identity. */ 1183 static bool fold_xx_to_x(OptContext *ctx, TCGOp *op) 1184 { 1185 if (args_are_copies(op->args[1], op->args[2])) { 1186 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); 1187 } 1188 return false; 1189 } 1190 1191 /* 1192 * These outermost fold_<op> functions are sorted alphabetically. 1193 * 1194 * The ordering of the transformations should be: 1195 * 1) those that produce a constant 1196 * 2) those that produce a copy 1197 * 3) those that produce information about the result value. 1198 */ 1199 1200 static bool fold_or(OptContext *ctx, TCGOp *op); 1201 static bool fold_orc(OptContext *ctx, TCGOp *op); 1202 static bool fold_xor(OptContext *ctx, TCGOp *op); 1203 1204 static bool fold_add(OptContext *ctx, TCGOp *op) 1205 { 1206 if (fold_const2_commutative(ctx, op) || 1207 fold_xi_to_x(ctx, op, 0)) { 1208 return true; 1209 } 1210 return finish_folding(ctx, op); 1211 } 1212 1213 /* We cannot as yet do_constant_folding with vectors. */ 1214 static bool fold_add_vec(OptContext *ctx, TCGOp *op) 1215 { 1216 if (fold_commutative(ctx, op) || 1217 fold_xi_to_x(ctx, op, 0)) { 1218 return true; 1219 } 1220 return finish_folding(ctx, op); 1221 } 1222 1223 static bool fold_addsub2(OptContext *ctx, TCGOp *op, bool add) 1224 { 1225 bool a_const = arg_is_const(op->args[2]) && arg_is_const(op->args[3]); 1226 bool b_const = arg_is_const(op->args[4]) && arg_is_const(op->args[5]); 1227 1228 if (a_const && b_const) { 1229 uint64_t al = arg_info(op->args[2])->val; 1230 uint64_t ah = arg_info(op->args[3])->val; 1231 uint64_t bl = arg_info(op->args[4])->val; 1232 uint64_t bh = arg_info(op->args[5])->val; 1233 TCGArg rl, rh; 1234 TCGOp *op2; 1235 1236 if (ctx->type == TCG_TYPE_I32) { 1237 uint64_t a = deposit64(al, 32, 32, ah); 1238 uint64_t b = deposit64(bl, 32, 32, bh); 1239 1240 if (add) { 1241 a += b; 1242 } else { 1243 a -= b; 1244 } 1245 1246 al = sextract64(a, 0, 32); 1247 ah = sextract64(a, 32, 32); 1248 } else { 1249 Int128 a = int128_make128(al, ah); 1250 Int128 b = int128_make128(bl, bh); 1251 1252 if (add) { 1253 a = int128_add(a, b); 1254 } else { 1255 a = int128_sub(a, b); 1256 } 1257 1258 al = int128_getlo(a); 1259 ah = int128_gethi(a); 1260 } 1261 1262 rl = op->args[0]; 1263 rh = op->args[1]; 1264 1265 /* The proper opcode is supplied by tcg_opt_gen_mov. */ 1266 op2 = opt_insert_before(ctx, op, 0, 2); 1267 1268 tcg_opt_gen_movi(ctx, op, rl, al); 1269 tcg_opt_gen_movi(ctx, op2, rh, ah); 1270 return true; 1271 } 1272 1273 /* Fold sub2 r,x,i to add2 r,x,-i */ 1274 if (!add && b_const) { 1275 uint64_t bl = arg_info(op->args[4])->val; 1276 uint64_t bh = arg_info(op->args[5])->val; 1277 1278 /* Negate the two parts without assembling and disassembling. */ 1279 bl = -bl; 1280 bh = ~bh + !bl; 1281 1282 op->opc = (ctx->type == TCG_TYPE_I32 1283 ? INDEX_op_add2_i32 : INDEX_op_add2_i64); 1284 op->args[4] = arg_new_constant(ctx, bl); 1285 op->args[5] = arg_new_constant(ctx, bh); 1286 } 1287 return finish_folding(ctx, op); 1288 } 1289 1290 static bool fold_add2(OptContext *ctx, TCGOp *op) 1291 { 1292 /* Note that the high and low parts may be independently swapped. */ 1293 swap_commutative(op->args[0], &op->args[2], &op->args[4]); 1294 swap_commutative(op->args[1], &op->args[3], &op->args[5]); 1295 1296 return fold_addsub2(ctx, op, true); 1297 } 1298 1299 static bool fold_and(OptContext *ctx, TCGOp *op) 1300 { 1301 uint64_t z1, z2, z_mask, s_mask; 1302 TempOptInfo *t1, *t2; 1303 1304 if (fold_const2_commutative(ctx, op) || 1305 fold_xi_to_i(ctx, op, 0) || 1306 fold_xi_to_x(ctx, op, -1) || 1307 fold_xx_to_x(ctx, op)) { 1308 return true; 1309 } 1310 1311 t1 = arg_info(op->args[1]); 1312 t2 = arg_info(op->args[2]); 1313 z1 = t1->z_mask; 1314 z2 = t2->z_mask; 1315 1316 /* 1317 * Known-zeros does not imply known-ones. Therefore unless 1318 * arg2 is constant, we can't infer affected bits from it. 1319 */ 1320 if (ti_is_const(t2) && fold_affected_mask(ctx, op, z1 & ~z2)) { 1321 return true; 1322 } 1323 1324 z_mask = z1 & z2; 1325 1326 /* 1327 * Sign repetitions are perforce all identical, whether they are 1 or 0. 1328 * Bitwise operations preserve the relative quantity of the repetitions. 1329 */ 1330 s_mask = t1->s_mask & t2->s_mask; 1331 1332 return fold_masks_zs(ctx, op, z_mask, s_mask); 1333 } 1334 1335 static bool fold_andc(OptContext *ctx, TCGOp *op) 1336 { 1337 uint64_t z_mask, s_mask; 1338 TempOptInfo *t1, *t2; 1339 1340 if (fold_const2(ctx, op) || 1341 fold_xx_to_i(ctx, op, 0) || 1342 fold_xi_to_x(ctx, op, 0) || 1343 fold_ix_to_not(ctx, op, -1)) { 1344 return true; 1345 } 1346 1347 t1 = arg_info(op->args[1]); 1348 t2 = arg_info(op->args[2]); 1349 z_mask = t1->z_mask; 1350 1351 if (ti_is_const(t2)) { 1352 /* Fold andc r,x,i to and r,x,~i. */ 1353 switch (ctx->type) { 1354 case TCG_TYPE_I32: 1355 case TCG_TYPE_I64: 1356 op->opc = INDEX_op_and; 1357 break; 1358 case TCG_TYPE_V64: 1359 case TCG_TYPE_V128: 1360 case TCG_TYPE_V256: 1361 op->opc = INDEX_op_and_vec; 1362 break; 1363 default: 1364 g_assert_not_reached(); 1365 } 1366 op->args[2] = arg_new_constant(ctx, ~ti_const_val(t2)); 1367 return fold_and(ctx, op); 1368 } 1369 1370 /* 1371 * Known-zeros does not imply known-ones. Therefore unless 1372 * arg2 is constant, we can't infer anything from it. 1373 */ 1374 if (ti_is_const(t2)) { 1375 uint64_t v2 = ti_const_val(t2); 1376 if (fold_affected_mask(ctx, op, z_mask & v2)) { 1377 return true; 1378 } 1379 z_mask &= ~v2; 1380 } 1381 1382 s_mask = t1->s_mask & t2->s_mask; 1383 return fold_masks_zs(ctx, op, z_mask, s_mask); 1384 } 1385 1386 static bool fold_bitsel_vec(OptContext *ctx, TCGOp *op) 1387 { 1388 /* If true and false values are the same, eliminate the cmp. */ 1389 if (args_are_copies(op->args[2], op->args[3])) { 1390 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]); 1391 } 1392 1393 if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) { 1394 uint64_t tv = arg_info(op->args[2])->val; 1395 uint64_t fv = arg_info(op->args[3])->val; 1396 1397 if (tv == -1 && fv == 0) { 1398 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); 1399 } 1400 if (tv == 0 && fv == -1) { 1401 if (TCG_TARGET_HAS_not_vec) { 1402 op->opc = INDEX_op_not_vec; 1403 return fold_not(ctx, op); 1404 } else { 1405 op->opc = INDEX_op_xor_vec; 1406 op->args[2] = arg_new_constant(ctx, -1); 1407 return fold_xor(ctx, op); 1408 } 1409 } 1410 } 1411 if (arg_is_const(op->args[2])) { 1412 uint64_t tv = arg_info(op->args[2])->val; 1413 if (tv == -1) { 1414 op->opc = INDEX_op_or_vec; 1415 op->args[2] = op->args[3]; 1416 return fold_or(ctx, op); 1417 } 1418 if (tv == 0 && TCG_TARGET_HAS_andc_vec) { 1419 op->opc = INDEX_op_andc_vec; 1420 op->args[2] = op->args[1]; 1421 op->args[1] = op->args[3]; 1422 return fold_andc(ctx, op); 1423 } 1424 } 1425 if (arg_is_const(op->args[3])) { 1426 uint64_t fv = arg_info(op->args[3])->val; 1427 if (fv == 0) { 1428 op->opc = INDEX_op_and_vec; 1429 return fold_and(ctx, op); 1430 } 1431 if (fv == -1 && TCG_TARGET_HAS_orc_vec) { 1432 op->opc = INDEX_op_orc_vec; 1433 op->args[2] = op->args[1]; 1434 op->args[1] = op->args[3]; 1435 return fold_orc(ctx, op); 1436 } 1437 } 1438 return finish_folding(ctx, op); 1439 } 1440 1441 static bool fold_brcond(OptContext *ctx, TCGOp *op) 1442 { 1443 int i = do_constant_folding_cond1(ctx, op, NO_DEST, &op->args[0], 1444 &op->args[1], &op->args[2]); 1445 if (i == 0) { 1446 tcg_op_remove(ctx->tcg, op); 1447 return true; 1448 } 1449 if (i > 0) { 1450 op->opc = INDEX_op_br; 1451 op->args[0] = op->args[3]; 1452 finish_ebb(ctx); 1453 } else { 1454 finish_bb(ctx); 1455 } 1456 return true; 1457 } 1458 1459 static bool fold_brcond2(OptContext *ctx, TCGOp *op) 1460 { 1461 TCGCond cond; 1462 TCGArg label; 1463 int i, inv = 0; 1464 1465 i = do_constant_folding_cond2(ctx, op, &op->args[0]); 1466 cond = op->args[4]; 1467 label = op->args[5]; 1468 if (i >= 0) { 1469 goto do_brcond_const; 1470 } 1471 1472 switch (cond) { 1473 case TCG_COND_LT: 1474 case TCG_COND_GE: 1475 /* 1476 * Simplify LT/GE comparisons vs zero to a single compare 1477 * vs the high word of the input. 1478 */ 1479 if (arg_is_const_val(op->args[2], 0) && 1480 arg_is_const_val(op->args[3], 0)) { 1481 goto do_brcond_high; 1482 } 1483 break; 1484 1485 case TCG_COND_NE: 1486 inv = 1; 1487 QEMU_FALLTHROUGH; 1488 case TCG_COND_EQ: 1489 /* 1490 * Simplify EQ/NE comparisons where one of the pairs 1491 * can be simplified. 1492 */ 1493 i = do_constant_folding_cond(TCG_TYPE_I32, op->args[0], 1494 op->args[2], cond); 1495 switch (i ^ inv) { 1496 case 0: 1497 goto do_brcond_const; 1498 case 1: 1499 goto do_brcond_high; 1500 } 1501 1502 i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1], 1503 op->args[3], cond); 1504 switch (i ^ inv) { 1505 case 0: 1506 goto do_brcond_const; 1507 case 1: 1508 goto do_brcond_low; 1509 } 1510 break; 1511 1512 case TCG_COND_TSTEQ: 1513 case TCG_COND_TSTNE: 1514 if (arg_is_const_val(op->args[2], 0)) { 1515 goto do_brcond_high; 1516 } 1517 if (arg_is_const_val(op->args[3], 0)) { 1518 goto do_brcond_low; 1519 } 1520 break; 1521 1522 default: 1523 break; 1524 1525 do_brcond_low: 1526 op->opc = INDEX_op_brcond_i32; 1527 op->args[1] = op->args[2]; 1528 op->args[2] = cond; 1529 op->args[3] = label; 1530 return fold_brcond(ctx, op); 1531 1532 do_brcond_high: 1533 op->opc = INDEX_op_brcond_i32; 1534 op->args[0] = op->args[1]; 1535 op->args[1] = op->args[3]; 1536 op->args[2] = cond; 1537 op->args[3] = label; 1538 return fold_brcond(ctx, op); 1539 1540 do_brcond_const: 1541 if (i == 0) { 1542 tcg_op_remove(ctx->tcg, op); 1543 return true; 1544 } 1545 op->opc = INDEX_op_br; 1546 op->args[0] = label; 1547 finish_ebb(ctx); 1548 return true; 1549 } 1550 1551 finish_bb(ctx); 1552 return true; 1553 } 1554 1555 static bool fold_bswap(OptContext *ctx, TCGOp *op) 1556 { 1557 uint64_t z_mask, s_mask, sign; 1558 TempOptInfo *t1 = arg_info(op->args[1]); 1559 1560 if (ti_is_const(t1)) { 1561 return tcg_opt_gen_movi(ctx, op, op->args[0], 1562 do_constant_folding(op->opc, ctx->type, 1563 ti_const_val(t1), 1564 op->args[2])); 1565 } 1566 1567 z_mask = t1->z_mask; 1568 switch (op->opc) { 1569 case INDEX_op_bswap16_i32: 1570 case INDEX_op_bswap16_i64: 1571 z_mask = bswap16(z_mask); 1572 sign = INT16_MIN; 1573 break; 1574 case INDEX_op_bswap32_i32: 1575 case INDEX_op_bswap32_i64: 1576 z_mask = bswap32(z_mask); 1577 sign = INT32_MIN; 1578 break; 1579 case INDEX_op_bswap64_i64: 1580 z_mask = bswap64(z_mask); 1581 sign = INT64_MIN; 1582 break; 1583 default: 1584 g_assert_not_reached(); 1585 } 1586 1587 s_mask = 0; 1588 switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) { 1589 case TCG_BSWAP_OZ: 1590 break; 1591 case TCG_BSWAP_OS: 1592 /* If the sign bit may be 1, force all the bits above to 1. */ 1593 if (z_mask & sign) { 1594 z_mask |= sign; 1595 } 1596 /* The value and therefore s_mask is explicitly sign-extended. */ 1597 s_mask = sign; 1598 break; 1599 default: 1600 /* The high bits are undefined: force all bits above the sign to 1. */ 1601 z_mask |= sign << 1; 1602 break; 1603 } 1604 1605 return fold_masks_zs(ctx, op, z_mask, s_mask); 1606 } 1607 1608 static bool fold_call(OptContext *ctx, TCGOp *op) 1609 { 1610 TCGContext *s = ctx->tcg; 1611 int nb_oargs = TCGOP_CALLO(op); 1612 int nb_iargs = TCGOP_CALLI(op); 1613 int flags, i; 1614 1615 init_arguments(ctx, op, nb_oargs + nb_iargs); 1616 copy_propagate(ctx, op, nb_oargs, nb_iargs); 1617 1618 /* If the function reads or writes globals, reset temp data. */ 1619 flags = tcg_call_flags(op); 1620 if (!(flags & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) { 1621 int nb_globals = s->nb_globals; 1622 1623 for (i = 0; i < nb_globals; i++) { 1624 if (test_bit(i, ctx->temps_used.l)) { 1625 reset_ts(ctx, &ctx->tcg->temps[i]); 1626 } 1627 } 1628 } 1629 1630 /* If the function has side effects, reset mem data. */ 1631 if (!(flags & TCG_CALL_NO_SIDE_EFFECTS)) { 1632 remove_mem_copy_all(ctx); 1633 } 1634 1635 /* Reset temp data for outputs. */ 1636 for (i = 0; i < nb_oargs; i++) { 1637 reset_temp(ctx, op->args[i]); 1638 } 1639 1640 /* Stop optimizing MB across calls. */ 1641 ctx->prev_mb = NULL; 1642 return true; 1643 } 1644 1645 static bool fold_cmp_vec(OptContext *ctx, TCGOp *op) 1646 { 1647 /* Canonicalize the comparison to put immediate second. */ 1648 if (swap_commutative(NO_DEST, &op->args[1], &op->args[2])) { 1649 op->args[3] = tcg_swap_cond(op->args[3]); 1650 } 1651 return finish_folding(ctx, op); 1652 } 1653 1654 static bool fold_cmpsel_vec(OptContext *ctx, TCGOp *op) 1655 { 1656 /* If true and false values are the same, eliminate the cmp. */ 1657 if (args_are_copies(op->args[3], op->args[4])) { 1658 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[3]); 1659 } 1660 1661 /* Canonicalize the comparison to put immediate second. */ 1662 if (swap_commutative(NO_DEST, &op->args[1], &op->args[2])) { 1663 op->args[5] = tcg_swap_cond(op->args[5]); 1664 } 1665 /* 1666 * Canonicalize the "false" input reg to match the destination, 1667 * so that the tcg backend can implement "move if true". 1668 */ 1669 if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) { 1670 op->args[5] = tcg_invert_cond(op->args[5]); 1671 } 1672 return finish_folding(ctx, op); 1673 } 1674 1675 static bool fold_count_zeros(OptContext *ctx, TCGOp *op) 1676 { 1677 uint64_t z_mask, s_mask; 1678 TempOptInfo *t1 = arg_info(op->args[1]); 1679 TempOptInfo *t2 = arg_info(op->args[2]); 1680 1681 if (ti_is_const(t1)) { 1682 uint64_t t = ti_const_val(t1); 1683 1684 if (t != 0) { 1685 t = do_constant_folding(op->opc, ctx->type, t, 0); 1686 return tcg_opt_gen_movi(ctx, op, op->args[0], t); 1687 } 1688 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]); 1689 } 1690 1691 switch (ctx->type) { 1692 case TCG_TYPE_I32: 1693 z_mask = 31; 1694 break; 1695 case TCG_TYPE_I64: 1696 z_mask = 63; 1697 break; 1698 default: 1699 g_assert_not_reached(); 1700 } 1701 s_mask = ~z_mask; 1702 z_mask |= t2->z_mask; 1703 s_mask &= t2->s_mask; 1704 1705 return fold_masks_zs(ctx, op, z_mask, s_mask); 1706 } 1707 1708 static bool fold_ctpop(OptContext *ctx, TCGOp *op) 1709 { 1710 uint64_t z_mask; 1711 1712 if (fold_const1(ctx, op)) { 1713 return true; 1714 } 1715 1716 switch (ctx->type) { 1717 case TCG_TYPE_I32: 1718 z_mask = 32 | 31; 1719 break; 1720 case TCG_TYPE_I64: 1721 z_mask = 64 | 63; 1722 break; 1723 default: 1724 g_assert_not_reached(); 1725 } 1726 return fold_masks_z(ctx, op, z_mask); 1727 } 1728 1729 static bool fold_deposit(OptContext *ctx, TCGOp *op) 1730 { 1731 TempOptInfo *t1 = arg_info(op->args[1]); 1732 TempOptInfo *t2 = arg_info(op->args[2]); 1733 int ofs = op->args[3]; 1734 int len = op->args[4]; 1735 int width = 8 * tcg_type_size(ctx->type); 1736 uint64_t z_mask, s_mask; 1737 1738 if (ti_is_const(t1) && ti_is_const(t2)) { 1739 return tcg_opt_gen_movi(ctx, op, op->args[0], 1740 deposit64(ti_const_val(t1), ofs, len, 1741 ti_const_val(t2))); 1742 } 1743 1744 /* Inserting a value into zero at offset 0. */ 1745 if (ti_is_const_val(t1, 0) && ofs == 0) { 1746 uint64_t mask = MAKE_64BIT_MASK(0, len); 1747 1748 op->opc = INDEX_op_and; 1749 op->args[1] = op->args[2]; 1750 op->args[2] = arg_new_constant(ctx, mask); 1751 return fold_and(ctx, op); 1752 } 1753 1754 /* Inserting zero into a value. */ 1755 if (ti_is_const_val(t2, 0)) { 1756 uint64_t mask = deposit64(-1, ofs, len, 0); 1757 1758 op->opc = INDEX_op_and; 1759 op->args[2] = arg_new_constant(ctx, mask); 1760 return fold_and(ctx, op); 1761 } 1762 1763 /* The s_mask from the top portion of the deposit is still valid. */ 1764 if (ofs + len == width) { 1765 s_mask = t2->s_mask << ofs; 1766 } else { 1767 s_mask = t1->s_mask & ~MAKE_64BIT_MASK(0, ofs + len); 1768 } 1769 1770 z_mask = deposit64(t1->z_mask, ofs, len, t2->z_mask); 1771 return fold_masks_zs(ctx, op, z_mask, s_mask); 1772 } 1773 1774 static bool fold_divide(OptContext *ctx, TCGOp *op) 1775 { 1776 if (fold_const2(ctx, op) || 1777 fold_xi_to_x(ctx, op, 1)) { 1778 return true; 1779 } 1780 return finish_folding(ctx, op); 1781 } 1782 1783 static bool fold_dup(OptContext *ctx, TCGOp *op) 1784 { 1785 if (arg_is_const(op->args[1])) { 1786 uint64_t t = arg_info(op->args[1])->val; 1787 t = dup_const(TCGOP_VECE(op), t); 1788 return tcg_opt_gen_movi(ctx, op, op->args[0], t); 1789 } 1790 return finish_folding(ctx, op); 1791 } 1792 1793 static bool fold_dup2(OptContext *ctx, TCGOp *op) 1794 { 1795 if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { 1796 uint64_t t = deposit64(arg_info(op->args[1])->val, 32, 32, 1797 arg_info(op->args[2])->val); 1798 return tcg_opt_gen_movi(ctx, op, op->args[0], t); 1799 } 1800 1801 if (args_are_copies(op->args[1], op->args[2])) { 1802 op->opc = INDEX_op_dup_vec; 1803 TCGOP_VECE(op) = MO_32; 1804 } 1805 return finish_folding(ctx, op); 1806 } 1807 1808 static bool fold_eqv(OptContext *ctx, TCGOp *op) 1809 { 1810 uint64_t s_mask; 1811 TempOptInfo *t1, *t2; 1812 1813 if (fold_const2_commutative(ctx, op) || 1814 fold_xi_to_x(ctx, op, -1) || 1815 fold_xi_to_not(ctx, op, 0)) { 1816 return true; 1817 } 1818 1819 t2 = arg_info(op->args[2]); 1820 if (ti_is_const(t2)) { 1821 /* Fold eqv r,x,i to xor r,x,~i. */ 1822 switch (ctx->type) { 1823 case TCG_TYPE_I32: 1824 case TCG_TYPE_I64: 1825 op->opc = INDEX_op_xor; 1826 break; 1827 case TCG_TYPE_V64: 1828 case TCG_TYPE_V128: 1829 case TCG_TYPE_V256: 1830 op->opc = INDEX_op_xor_vec; 1831 break; 1832 default: 1833 g_assert_not_reached(); 1834 } 1835 op->args[2] = arg_new_constant(ctx, ~ti_const_val(t2)); 1836 return fold_xor(ctx, op); 1837 } 1838 1839 t1 = arg_info(op->args[1]); 1840 s_mask = t1->s_mask & t2->s_mask; 1841 return fold_masks_s(ctx, op, s_mask); 1842 } 1843 1844 static bool fold_extract(OptContext *ctx, TCGOp *op) 1845 { 1846 uint64_t z_mask_old, z_mask; 1847 TempOptInfo *t1 = arg_info(op->args[1]); 1848 int pos = op->args[2]; 1849 int len = op->args[3]; 1850 1851 if (ti_is_const(t1)) { 1852 return tcg_opt_gen_movi(ctx, op, op->args[0], 1853 extract64(ti_const_val(t1), pos, len)); 1854 } 1855 1856 z_mask_old = t1->z_mask; 1857 z_mask = extract64(z_mask_old, pos, len); 1858 if (pos == 0 && fold_affected_mask(ctx, op, z_mask_old ^ z_mask)) { 1859 return true; 1860 } 1861 1862 return fold_masks_z(ctx, op, z_mask); 1863 } 1864 1865 static bool fold_extract2(OptContext *ctx, TCGOp *op) 1866 { 1867 if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { 1868 uint64_t v1 = arg_info(op->args[1])->val; 1869 uint64_t v2 = arg_info(op->args[2])->val; 1870 int shr = op->args[3]; 1871 1872 if (op->opc == INDEX_op_extract2_i64) { 1873 v1 >>= shr; 1874 v2 <<= 64 - shr; 1875 } else { 1876 v1 = (uint32_t)v1 >> shr; 1877 v2 = (uint64_t)((int32_t)v2 << (32 - shr)); 1878 } 1879 return tcg_opt_gen_movi(ctx, op, op->args[0], v1 | v2); 1880 } 1881 return finish_folding(ctx, op); 1882 } 1883 1884 static bool fold_exts(OptContext *ctx, TCGOp *op) 1885 { 1886 uint64_t s_mask, z_mask; 1887 TempOptInfo *t1; 1888 1889 if (fold_const1(ctx, op)) { 1890 return true; 1891 } 1892 1893 t1 = arg_info(op->args[1]); 1894 z_mask = t1->z_mask; 1895 s_mask = t1->s_mask; 1896 1897 switch (op->opc) { 1898 case INDEX_op_ext_i32_i64: 1899 s_mask |= INT32_MIN; 1900 z_mask = (int32_t)z_mask; 1901 break; 1902 default: 1903 g_assert_not_reached(); 1904 } 1905 return fold_masks_zs(ctx, op, z_mask, s_mask); 1906 } 1907 1908 static bool fold_extu(OptContext *ctx, TCGOp *op) 1909 { 1910 uint64_t z_mask; 1911 1912 if (fold_const1(ctx, op)) { 1913 return true; 1914 } 1915 1916 z_mask = arg_info(op->args[1])->z_mask; 1917 switch (op->opc) { 1918 case INDEX_op_extrl_i64_i32: 1919 case INDEX_op_extu_i32_i64: 1920 z_mask = (uint32_t)z_mask; 1921 break; 1922 case INDEX_op_extrh_i64_i32: 1923 z_mask >>= 32; 1924 break; 1925 default: 1926 g_assert_not_reached(); 1927 } 1928 return fold_masks_z(ctx, op, z_mask); 1929 } 1930 1931 static bool fold_mb(OptContext *ctx, TCGOp *op) 1932 { 1933 /* Eliminate duplicate and redundant fence instructions. */ 1934 if (ctx->prev_mb) { 1935 /* 1936 * Merge two barriers of the same type into one, 1937 * or a weaker barrier into a stronger one, 1938 * or two weaker barriers into a stronger one. 1939 * mb X; mb Y => mb X|Y 1940 * mb; strl => mb; st 1941 * ldaq; mb => ld; mb 1942 * ldaq; strl => ld; mb; st 1943 * Other combinations are also merged into a strong 1944 * barrier. This is stricter than specified but for 1945 * the purposes of TCG is better than not optimizing. 1946 */ 1947 ctx->prev_mb->args[0] |= op->args[0]; 1948 tcg_op_remove(ctx->tcg, op); 1949 } else { 1950 ctx->prev_mb = op; 1951 } 1952 return true; 1953 } 1954 1955 static bool fold_mov(OptContext *ctx, TCGOp *op) 1956 { 1957 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); 1958 } 1959 1960 static bool fold_movcond(OptContext *ctx, TCGOp *op) 1961 { 1962 uint64_t z_mask, s_mask; 1963 TempOptInfo *tt, *ft; 1964 int i; 1965 1966 /* If true and false values are the same, eliminate the cmp. */ 1967 if (args_are_copies(op->args[3], op->args[4])) { 1968 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[3]); 1969 } 1970 1971 /* 1972 * Canonicalize the "false" input reg to match the destination reg so 1973 * that the tcg backend can implement a "move if true" operation. 1974 */ 1975 if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) { 1976 op->args[5] = tcg_invert_cond(op->args[5]); 1977 } 1978 1979 i = do_constant_folding_cond1(ctx, op, NO_DEST, &op->args[1], 1980 &op->args[2], &op->args[5]); 1981 if (i >= 0) { 1982 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[4 - i]); 1983 } 1984 1985 tt = arg_info(op->args[3]); 1986 ft = arg_info(op->args[4]); 1987 z_mask = tt->z_mask | ft->z_mask; 1988 s_mask = tt->s_mask & ft->s_mask; 1989 1990 if (ti_is_const(tt) && ti_is_const(ft)) { 1991 uint64_t tv = ti_const_val(tt); 1992 uint64_t fv = ti_const_val(ft); 1993 TCGOpcode opc, negopc = 0; 1994 TCGCond cond = op->args[5]; 1995 1996 switch (ctx->type) { 1997 case TCG_TYPE_I32: 1998 opc = INDEX_op_setcond_i32; 1999 if (TCG_TARGET_HAS_negsetcond_i32) { 2000 negopc = INDEX_op_negsetcond_i32; 2001 } 2002 tv = (int32_t)tv; 2003 fv = (int32_t)fv; 2004 break; 2005 case TCG_TYPE_I64: 2006 opc = INDEX_op_setcond_i64; 2007 if (TCG_TARGET_HAS_negsetcond_i64) { 2008 negopc = INDEX_op_negsetcond_i64; 2009 } 2010 break; 2011 default: 2012 g_assert_not_reached(); 2013 } 2014 2015 if (tv == 1 && fv == 0) { 2016 op->opc = opc; 2017 op->args[3] = cond; 2018 } else if (fv == 1 && tv == 0) { 2019 op->opc = opc; 2020 op->args[3] = tcg_invert_cond(cond); 2021 } else if (negopc) { 2022 if (tv == -1 && fv == 0) { 2023 op->opc = negopc; 2024 op->args[3] = cond; 2025 } else if (fv == -1 && tv == 0) { 2026 op->opc = negopc; 2027 op->args[3] = tcg_invert_cond(cond); 2028 } 2029 } 2030 } 2031 2032 return fold_masks_zs(ctx, op, z_mask, s_mask); 2033 } 2034 2035 static bool fold_mul(OptContext *ctx, TCGOp *op) 2036 { 2037 if (fold_const2(ctx, op) || 2038 fold_xi_to_i(ctx, op, 0) || 2039 fold_xi_to_x(ctx, op, 1)) { 2040 return true; 2041 } 2042 return finish_folding(ctx, op); 2043 } 2044 2045 static bool fold_mul_highpart(OptContext *ctx, TCGOp *op) 2046 { 2047 if (fold_const2_commutative(ctx, op) || 2048 fold_xi_to_i(ctx, op, 0)) { 2049 return true; 2050 } 2051 return finish_folding(ctx, op); 2052 } 2053 2054 static bool fold_multiply2(OptContext *ctx, TCGOp *op) 2055 { 2056 swap_commutative(op->args[0], &op->args[2], &op->args[3]); 2057 2058 if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) { 2059 uint64_t a = arg_info(op->args[2])->val; 2060 uint64_t b = arg_info(op->args[3])->val; 2061 uint64_t h, l; 2062 TCGArg rl, rh; 2063 TCGOp *op2; 2064 2065 switch (op->opc) { 2066 case INDEX_op_mulu2_i32: 2067 l = (uint64_t)(uint32_t)a * (uint32_t)b; 2068 h = (int32_t)(l >> 32); 2069 l = (int32_t)l; 2070 break; 2071 case INDEX_op_muls2_i32: 2072 l = (int64_t)(int32_t)a * (int32_t)b; 2073 h = l >> 32; 2074 l = (int32_t)l; 2075 break; 2076 case INDEX_op_mulu2_i64: 2077 mulu64(&l, &h, a, b); 2078 break; 2079 case INDEX_op_muls2_i64: 2080 muls64(&l, &h, a, b); 2081 break; 2082 default: 2083 g_assert_not_reached(); 2084 } 2085 2086 rl = op->args[0]; 2087 rh = op->args[1]; 2088 2089 /* The proper opcode is supplied by tcg_opt_gen_mov. */ 2090 op2 = opt_insert_before(ctx, op, 0, 2); 2091 2092 tcg_opt_gen_movi(ctx, op, rl, l); 2093 tcg_opt_gen_movi(ctx, op2, rh, h); 2094 return true; 2095 } 2096 return finish_folding(ctx, op); 2097 } 2098 2099 static bool fold_nand(OptContext *ctx, TCGOp *op) 2100 { 2101 uint64_t s_mask; 2102 2103 if (fold_const2_commutative(ctx, op) || 2104 fold_xi_to_not(ctx, op, -1)) { 2105 return true; 2106 } 2107 2108 s_mask = arg_info(op->args[1])->s_mask 2109 & arg_info(op->args[2])->s_mask; 2110 return fold_masks_s(ctx, op, s_mask); 2111 } 2112 2113 static bool fold_neg_no_const(OptContext *ctx, TCGOp *op) 2114 { 2115 /* Set to 1 all bits to the left of the rightmost. */ 2116 uint64_t z_mask = arg_info(op->args[1])->z_mask; 2117 z_mask = -(z_mask & -z_mask); 2118 2119 return fold_masks_z(ctx, op, z_mask); 2120 } 2121 2122 static bool fold_neg(OptContext *ctx, TCGOp *op) 2123 { 2124 return fold_const1(ctx, op) || fold_neg_no_const(ctx, op); 2125 } 2126 2127 static bool fold_nor(OptContext *ctx, TCGOp *op) 2128 { 2129 uint64_t s_mask; 2130 2131 if (fold_const2_commutative(ctx, op) || 2132 fold_xi_to_not(ctx, op, 0)) { 2133 return true; 2134 } 2135 2136 s_mask = arg_info(op->args[1])->s_mask 2137 & arg_info(op->args[2])->s_mask; 2138 return fold_masks_s(ctx, op, s_mask); 2139 } 2140 2141 static bool fold_not(OptContext *ctx, TCGOp *op) 2142 { 2143 if (fold_const1(ctx, op)) { 2144 return true; 2145 } 2146 return fold_masks_s(ctx, op, arg_info(op->args[1])->s_mask); 2147 } 2148 2149 static bool fold_or(OptContext *ctx, TCGOp *op) 2150 { 2151 uint64_t z_mask, s_mask; 2152 TempOptInfo *t1, *t2; 2153 2154 if (fold_const2_commutative(ctx, op) || 2155 fold_xi_to_x(ctx, op, 0) || 2156 fold_xx_to_x(ctx, op)) { 2157 return true; 2158 } 2159 2160 t1 = arg_info(op->args[1]); 2161 t2 = arg_info(op->args[2]); 2162 z_mask = t1->z_mask | t2->z_mask; 2163 s_mask = t1->s_mask & t2->s_mask; 2164 return fold_masks_zs(ctx, op, z_mask, s_mask); 2165 } 2166 2167 static bool fold_orc(OptContext *ctx, TCGOp *op) 2168 { 2169 uint64_t s_mask; 2170 TempOptInfo *t1, *t2; 2171 2172 if (fold_const2(ctx, op) || 2173 fold_xx_to_i(ctx, op, -1) || 2174 fold_xi_to_x(ctx, op, -1) || 2175 fold_ix_to_not(ctx, op, 0)) { 2176 return true; 2177 } 2178 2179 t2 = arg_info(op->args[2]); 2180 if (ti_is_const(t2)) { 2181 /* Fold orc r,x,i to or r,x,~i. */ 2182 switch (ctx->type) { 2183 case TCG_TYPE_I32: 2184 case TCG_TYPE_I64: 2185 op->opc = INDEX_op_or; 2186 break; 2187 case TCG_TYPE_V64: 2188 case TCG_TYPE_V128: 2189 case TCG_TYPE_V256: 2190 op->opc = INDEX_op_or_vec; 2191 break; 2192 default: 2193 g_assert_not_reached(); 2194 } 2195 op->args[2] = arg_new_constant(ctx, ~ti_const_val(t2)); 2196 return fold_or(ctx, op); 2197 } 2198 2199 t1 = arg_info(op->args[1]); 2200 s_mask = t1->s_mask & t2->s_mask; 2201 return fold_masks_s(ctx, op, s_mask); 2202 } 2203 2204 static bool fold_qemu_ld_1reg(OptContext *ctx, TCGOp *op) 2205 { 2206 const TCGOpDef *def = &tcg_op_defs[op->opc]; 2207 MemOpIdx oi = op->args[def->nb_oargs + def->nb_iargs]; 2208 MemOp mop = get_memop(oi); 2209 int width = 8 * memop_size(mop); 2210 uint64_t z_mask = -1, s_mask = 0; 2211 2212 if (width < 64) { 2213 if (mop & MO_SIGN) { 2214 s_mask = MAKE_64BIT_MASK(width - 1, 64 - (width - 1)); 2215 } else { 2216 z_mask = MAKE_64BIT_MASK(0, width); 2217 } 2218 } 2219 2220 /* Opcodes that touch guest memory stop the mb optimization. */ 2221 ctx->prev_mb = NULL; 2222 2223 return fold_masks_zs(ctx, op, z_mask, s_mask); 2224 } 2225 2226 static bool fold_qemu_ld_2reg(OptContext *ctx, TCGOp *op) 2227 { 2228 /* Opcodes that touch guest memory stop the mb optimization. */ 2229 ctx->prev_mb = NULL; 2230 return finish_folding(ctx, op); 2231 } 2232 2233 static bool fold_qemu_st(OptContext *ctx, TCGOp *op) 2234 { 2235 /* Opcodes that touch guest memory stop the mb optimization. */ 2236 ctx->prev_mb = NULL; 2237 return true; 2238 } 2239 2240 static bool fold_remainder(OptContext *ctx, TCGOp *op) 2241 { 2242 if (fold_const2(ctx, op) || 2243 fold_xx_to_i(ctx, op, 0)) { 2244 return true; 2245 } 2246 return finish_folding(ctx, op); 2247 } 2248 2249 /* Return 1 if finished, -1 if simplified, 0 if unchanged. */ 2250 static int fold_setcond_zmask(OptContext *ctx, TCGOp *op, bool neg) 2251 { 2252 uint64_t a_zmask, b_val; 2253 TCGCond cond; 2254 2255 if (!arg_is_const(op->args[2])) { 2256 return false; 2257 } 2258 2259 a_zmask = arg_info(op->args[1])->z_mask; 2260 b_val = arg_info(op->args[2])->val; 2261 cond = op->args[3]; 2262 2263 if (ctx->type == TCG_TYPE_I32) { 2264 a_zmask = (uint32_t)a_zmask; 2265 b_val = (uint32_t)b_val; 2266 } 2267 2268 /* 2269 * A with only low bits set vs B with high bits set means that A < B. 2270 */ 2271 if (a_zmask < b_val) { 2272 bool inv = false; 2273 2274 switch (cond) { 2275 case TCG_COND_NE: 2276 case TCG_COND_LEU: 2277 case TCG_COND_LTU: 2278 inv = true; 2279 /* fall through */ 2280 case TCG_COND_GTU: 2281 case TCG_COND_GEU: 2282 case TCG_COND_EQ: 2283 return tcg_opt_gen_movi(ctx, op, op->args[0], neg ? -inv : inv); 2284 default: 2285 break; 2286 } 2287 } 2288 2289 /* 2290 * A with only lsb set is already boolean. 2291 */ 2292 if (a_zmask <= 1) { 2293 bool convert = false; 2294 bool inv = false; 2295 2296 switch (cond) { 2297 case TCG_COND_EQ: 2298 inv = true; 2299 /* fall through */ 2300 case TCG_COND_NE: 2301 convert = (b_val == 0); 2302 break; 2303 case TCG_COND_LTU: 2304 case TCG_COND_TSTEQ: 2305 inv = true; 2306 /* fall through */ 2307 case TCG_COND_GEU: 2308 case TCG_COND_TSTNE: 2309 convert = (b_val == 1); 2310 break; 2311 default: 2312 break; 2313 } 2314 if (convert) { 2315 TCGOpcode neg_opc; 2316 2317 if (!inv && !neg) { 2318 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); 2319 } 2320 2321 switch (ctx->type) { 2322 case TCG_TYPE_I32: 2323 neg_opc = INDEX_op_neg_i32; 2324 break; 2325 case TCG_TYPE_I64: 2326 neg_opc = INDEX_op_neg_i64; 2327 break; 2328 default: 2329 g_assert_not_reached(); 2330 } 2331 2332 if (!inv) { 2333 op->opc = neg_opc; 2334 } else if (neg) { 2335 op->opc = INDEX_op_add; 2336 op->args[2] = arg_new_constant(ctx, -1); 2337 } else { 2338 op->opc = INDEX_op_xor; 2339 op->args[2] = arg_new_constant(ctx, 1); 2340 } 2341 return -1; 2342 } 2343 } 2344 return 0; 2345 } 2346 2347 static void fold_setcond_tst_pow2(OptContext *ctx, TCGOp *op, bool neg) 2348 { 2349 TCGOpcode neg_opc, shr_opc; 2350 TCGOpcode uext_opc = 0, sext_opc = 0; 2351 TCGCond cond = op->args[3]; 2352 TCGArg ret, src1, src2; 2353 TCGOp *op2; 2354 uint64_t val; 2355 int sh; 2356 bool inv; 2357 2358 if (!is_tst_cond(cond) || !arg_is_const(op->args[2])) { 2359 return; 2360 } 2361 2362 src2 = op->args[2]; 2363 val = arg_info(src2)->val; 2364 if (!is_power_of_2(val)) { 2365 return; 2366 } 2367 sh = ctz64(val); 2368 2369 switch (ctx->type) { 2370 case TCG_TYPE_I32: 2371 shr_opc = INDEX_op_shr_i32; 2372 neg_opc = INDEX_op_neg_i32; 2373 if (TCG_TARGET_extract_valid(TCG_TYPE_I32, sh, 1)) { 2374 uext_opc = INDEX_op_extract_i32; 2375 } 2376 if (TCG_TARGET_sextract_valid(TCG_TYPE_I32, sh, 1)) { 2377 sext_opc = INDEX_op_sextract_i32; 2378 } 2379 break; 2380 case TCG_TYPE_I64: 2381 shr_opc = INDEX_op_shr_i64; 2382 neg_opc = INDEX_op_neg_i64; 2383 if (TCG_TARGET_extract_valid(TCG_TYPE_I64, sh, 1)) { 2384 uext_opc = INDEX_op_extract_i64; 2385 } 2386 if (TCG_TARGET_sextract_valid(TCG_TYPE_I64, sh, 1)) { 2387 sext_opc = INDEX_op_sextract_i64; 2388 } 2389 break; 2390 default: 2391 g_assert_not_reached(); 2392 } 2393 2394 ret = op->args[0]; 2395 src1 = op->args[1]; 2396 inv = cond == TCG_COND_TSTEQ; 2397 2398 if (sh && sext_opc && neg && !inv) { 2399 op->opc = sext_opc; 2400 op->args[1] = src1; 2401 op->args[2] = sh; 2402 op->args[3] = 1; 2403 return; 2404 } else if (sh && uext_opc) { 2405 op->opc = uext_opc; 2406 op->args[1] = src1; 2407 op->args[2] = sh; 2408 op->args[3] = 1; 2409 } else { 2410 if (sh) { 2411 op2 = opt_insert_before(ctx, op, shr_opc, 3); 2412 op2->args[0] = ret; 2413 op2->args[1] = src1; 2414 op2->args[2] = arg_new_constant(ctx, sh); 2415 src1 = ret; 2416 } 2417 op->opc = INDEX_op_and; 2418 op->args[1] = src1; 2419 op->args[2] = arg_new_constant(ctx, 1); 2420 } 2421 2422 if (neg && inv) { 2423 op2 = opt_insert_after(ctx, op, INDEX_op_add, 3); 2424 op2->args[0] = ret; 2425 op2->args[1] = ret; 2426 op2->args[2] = arg_new_constant(ctx, -1); 2427 } else if (inv) { 2428 op2 = opt_insert_after(ctx, op, INDEX_op_xor, 3); 2429 op2->args[0] = ret; 2430 op2->args[1] = ret; 2431 op2->args[2] = arg_new_constant(ctx, 1); 2432 } else if (neg) { 2433 op2 = opt_insert_after(ctx, op, neg_opc, 2); 2434 op2->args[0] = ret; 2435 op2->args[1] = ret; 2436 } 2437 } 2438 2439 static bool fold_setcond(OptContext *ctx, TCGOp *op) 2440 { 2441 int i = do_constant_folding_cond1(ctx, op, op->args[0], &op->args[1], 2442 &op->args[2], &op->args[3]); 2443 if (i >= 0) { 2444 return tcg_opt_gen_movi(ctx, op, op->args[0], i); 2445 } 2446 2447 i = fold_setcond_zmask(ctx, op, false); 2448 if (i > 0) { 2449 return true; 2450 } 2451 if (i == 0) { 2452 fold_setcond_tst_pow2(ctx, op, false); 2453 } 2454 2455 return fold_masks_z(ctx, op, 1); 2456 } 2457 2458 static bool fold_negsetcond(OptContext *ctx, TCGOp *op) 2459 { 2460 int i = do_constant_folding_cond1(ctx, op, op->args[0], &op->args[1], 2461 &op->args[2], &op->args[3]); 2462 if (i >= 0) { 2463 return tcg_opt_gen_movi(ctx, op, op->args[0], -i); 2464 } 2465 2466 i = fold_setcond_zmask(ctx, op, true); 2467 if (i > 0) { 2468 return true; 2469 } 2470 if (i == 0) { 2471 fold_setcond_tst_pow2(ctx, op, true); 2472 } 2473 2474 /* Value is {0,-1} so all bits are repetitions of the sign. */ 2475 return fold_masks_s(ctx, op, -1); 2476 } 2477 2478 static bool fold_setcond2(OptContext *ctx, TCGOp *op) 2479 { 2480 TCGCond cond; 2481 int i, inv = 0; 2482 2483 i = do_constant_folding_cond2(ctx, op, &op->args[1]); 2484 cond = op->args[5]; 2485 if (i >= 0) { 2486 goto do_setcond_const; 2487 } 2488 2489 switch (cond) { 2490 case TCG_COND_LT: 2491 case TCG_COND_GE: 2492 /* 2493 * Simplify LT/GE comparisons vs zero to a single compare 2494 * vs the high word of the input. 2495 */ 2496 if (arg_is_const_val(op->args[3], 0) && 2497 arg_is_const_val(op->args[4], 0)) { 2498 goto do_setcond_high; 2499 } 2500 break; 2501 2502 case TCG_COND_NE: 2503 inv = 1; 2504 QEMU_FALLTHROUGH; 2505 case TCG_COND_EQ: 2506 /* 2507 * Simplify EQ/NE comparisons where one of the pairs 2508 * can be simplified. 2509 */ 2510 i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1], 2511 op->args[3], cond); 2512 switch (i ^ inv) { 2513 case 0: 2514 goto do_setcond_const; 2515 case 1: 2516 goto do_setcond_high; 2517 } 2518 2519 i = do_constant_folding_cond(TCG_TYPE_I32, op->args[2], 2520 op->args[4], cond); 2521 switch (i ^ inv) { 2522 case 0: 2523 goto do_setcond_const; 2524 case 1: 2525 goto do_setcond_low; 2526 } 2527 break; 2528 2529 case TCG_COND_TSTEQ: 2530 case TCG_COND_TSTNE: 2531 if (arg_is_const_val(op->args[3], 0)) { 2532 goto do_setcond_high; 2533 } 2534 if (arg_is_const_val(op->args[4], 0)) { 2535 goto do_setcond_low; 2536 } 2537 break; 2538 2539 default: 2540 break; 2541 2542 do_setcond_low: 2543 op->args[2] = op->args[3]; 2544 op->args[3] = cond; 2545 op->opc = INDEX_op_setcond_i32; 2546 return fold_setcond(ctx, op); 2547 2548 do_setcond_high: 2549 op->args[1] = op->args[2]; 2550 op->args[2] = op->args[4]; 2551 op->args[3] = cond; 2552 op->opc = INDEX_op_setcond_i32; 2553 return fold_setcond(ctx, op); 2554 } 2555 2556 return fold_masks_z(ctx, op, 1); 2557 2558 do_setcond_const: 2559 return tcg_opt_gen_movi(ctx, op, op->args[0], i); 2560 } 2561 2562 static bool fold_sextract(OptContext *ctx, TCGOp *op) 2563 { 2564 uint64_t z_mask, s_mask, s_mask_old; 2565 TempOptInfo *t1 = arg_info(op->args[1]); 2566 int pos = op->args[2]; 2567 int len = op->args[3]; 2568 2569 if (ti_is_const(t1)) { 2570 return tcg_opt_gen_movi(ctx, op, op->args[0], 2571 sextract64(ti_const_val(t1), pos, len)); 2572 } 2573 2574 s_mask_old = t1->s_mask; 2575 s_mask = s_mask_old >> pos; 2576 s_mask |= -1ull << (len - 1); 2577 2578 if (pos == 0 && fold_affected_mask(ctx, op, s_mask & ~s_mask_old)) { 2579 return true; 2580 } 2581 2582 z_mask = sextract64(t1->z_mask, pos, len); 2583 return fold_masks_zs(ctx, op, z_mask, s_mask); 2584 } 2585 2586 static bool fold_shift(OptContext *ctx, TCGOp *op) 2587 { 2588 uint64_t s_mask, z_mask; 2589 TempOptInfo *t1, *t2; 2590 2591 if (fold_const2(ctx, op) || 2592 fold_ix_to_i(ctx, op, 0) || 2593 fold_xi_to_x(ctx, op, 0)) { 2594 return true; 2595 } 2596 2597 t1 = arg_info(op->args[1]); 2598 t2 = arg_info(op->args[2]); 2599 s_mask = t1->s_mask; 2600 z_mask = t1->z_mask; 2601 2602 if (ti_is_const(t2)) { 2603 int sh = ti_const_val(t2); 2604 2605 z_mask = do_constant_folding(op->opc, ctx->type, z_mask, sh); 2606 s_mask = do_constant_folding(op->opc, ctx->type, s_mask, sh); 2607 2608 return fold_masks_zs(ctx, op, z_mask, s_mask); 2609 } 2610 2611 switch (op->opc) { 2612 CASE_OP_32_64(sar): 2613 /* 2614 * Arithmetic right shift will not reduce the number of 2615 * input sign repetitions. 2616 */ 2617 return fold_masks_s(ctx, op, s_mask); 2618 CASE_OP_32_64(shr): 2619 /* 2620 * If the sign bit is known zero, then logical right shift 2621 * will not reduce the number of input sign repetitions. 2622 */ 2623 if (~z_mask & -s_mask) { 2624 return fold_masks_s(ctx, op, s_mask); 2625 } 2626 break; 2627 default: 2628 break; 2629 } 2630 2631 return finish_folding(ctx, op); 2632 } 2633 2634 static bool fold_sub_to_neg(OptContext *ctx, TCGOp *op) 2635 { 2636 TCGOpcode neg_op; 2637 bool have_neg; 2638 2639 if (!arg_is_const(op->args[1]) || arg_info(op->args[1])->val != 0) { 2640 return false; 2641 } 2642 2643 switch (ctx->type) { 2644 case TCG_TYPE_I32: 2645 neg_op = INDEX_op_neg_i32; 2646 have_neg = true; 2647 break; 2648 case TCG_TYPE_I64: 2649 neg_op = INDEX_op_neg_i64; 2650 have_neg = true; 2651 break; 2652 case TCG_TYPE_V64: 2653 case TCG_TYPE_V128: 2654 case TCG_TYPE_V256: 2655 neg_op = INDEX_op_neg_vec; 2656 have_neg = (TCG_TARGET_HAS_neg_vec && 2657 tcg_can_emit_vec_op(neg_op, ctx->type, TCGOP_VECE(op)) > 0); 2658 break; 2659 default: 2660 g_assert_not_reached(); 2661 } 2662 if (have_neg) { 2663 op->opc = neg_op; 2664 op->args[1] = op->args[2]; 2665 return fold_neg_no_const(ctx, op); 2666 } 2667 return false; 2668 } 2669 2670 /* We cannot as yet do_constant_folding with vectors. */ 2671 static bool fold_sub_vec(OptContext *ctx, TCGOp *op) 2672 { 2673 if (fold_xx_to_i(ctx, op, 0) || 2674 fold_xi_to_x(ctx, op, 0) || 2675 fold_sub_to_neg(ctx, op)) { 2676 return true; 2677 } 2678 return finish_folding(ctx, op); 2679 } 2680 2681 static bool fold_sub(OptContext *ctx, TCGOp *op) 2682 { 2683 if (fold_const2(ctx, op) || 2684 fold_xx_to_i(ctx, op, 0) || 2685 fold_xi_to_x(ctx, op, 0) || 2686 fold_sub_to_neg(ctx, op)) { 2687 return true; 2688 } 2689 2690 /* Fold sub r,x,i to add r,x,-i */ 2691 if (arg_is_const(op->args[2])) { 2692 uint64_t val = arg_info(op->args[2])->val; 2693 2694 op->opc = INDEX_op_add; 2695 op->args[2] = arg_new_constant(ctx, -val); 2696 } 2697 return finish_folding(ctx, op); 2698 } 2699 2700 static bool fold_sub2(OptContext *ctx, TCGOp *op) 2701 { 2702 return fold_addsub2(ctx, op, false); 2703 } 2704 2705 static bool fold_tcg_ld(OptContext *ctx, TCGOp *op) 2706 { 2707 uint64_t z_mask = -1, s_mask = 0; 2708 2709 /* We can't do any folding with a load, but we can record bits. */ 2710 switch (op->opc) { 2711 CASE_OP_32_64(ld8s): 2712 s_mask = INT8_MIN; 2713 break; 2714 CASE_OP_32_64(ld8u): 2715 z_mask = MAKE_64BIT_MASK(0, 8); 2716 break; 2717 CASE_OP_32_64(ld16s): 2718 s_mask = INT16_MIN; 2719 break; 2720 CASE_OP_32_64(ld16u): 2721 z_mask = MAKE_64BIT_MASK(0, 16); 2722 break; 2723 case INDEX_op_ld32s_i64: 2724 s_mask = INT32_MIN; 2725 break; 2726 case INDEX_op_ld32u_i64: 2727 z_mask = MAKE_64BIT_MASK(0, 32); 2728 break; 2729 default: 2730 g_assert_not_reached(); 2731 } 2732 return fold_masks_zs(ctx, op, z_mask, s_mask); 2733 } 2734 2735 static bool fold_tcg_ld_memcopy(OptContext *ctx, TCGOp *op) 2736 { 2737 TCGTemp *dst, *src; 2738 intptr_t ofs; 2739 TCGType type; 2740 2741 if (op->args[1] != tcgv_ptr_arg(tcg_env)) { 2742 return finish_folding(ctx, op); 2743 } 2744 2745 type = ctx->type; 2746 ofs = op->args[2]; 2747 dst = arg_temp(op->args[0]); 2748 src = find_mem_copy_for(ctx, type, ofs); 2749 if (src && src->base_type == type) { 2750 return tcg_opt_gen_mov(ctx, op, temp_arg(dst), temp_arg(src)); 2751 } 2752 2753 reset_ts(ctx, dst); 2754 record_mem_copy(ctx, type, dst, ofs, ofs + tcg_type_size(type) - 1); 2755 return true; 2756 } 2757 2758 static bool fold_tcg_st(OptContext *ctx, TCGOp *op) 2759 { 2760 intptr_t ofs = op->args[2]; 2761 intptr_t lm1; 2762 2763 if (op->args[1] != tcgv_ptr_arg(tcg_env)) { 2764 remove_mem_copy_all(ctx); 2765 return true; 2766 } 2767 2768 switch (op->opc) { 2769 CASE_OP_32_64(st8): 2770 lm1 = 0; 2771 break; 2772 CASE_OP_32_64(st16): 2773 lm1 = 1; 2774 break; 2775 case INDEX_op_st32_i64: 2776 case INDEX_op_st_i32: 2777 lm1 = 3; 2778 break; 2779 case INDEX_op_st_i64: 2780 lm1 = 7; 2781 break; 2782 case INDEX_op_st_vec: 2783 lm1 = tcg_type_size(ctx->type) - 1; 2784 break; 2785 default: 2786 g_assert_not_reached(); 2787 } 2788 remove_mem_copy_in(ctx, ofs, ofs + lm1); 2789 return true; 2790 } 2791 2792 static bool fold_tcg_st_memcopy(OptContext *ctx, TCGOp *op) 2793 { 2794 TCGTemp *src; 2795 intptr_t ofs, last; 2796 TCGType type; 2797 2798 if (op->args[1] != tcgv_ptr_arg(tcg_env)) { 2799 return fold_tcg_st(ctx, op); 2800 } 2801 2802 src = arg_temp(op->args[0]); 2803 ofs = op->args[2]; 2804 type = ctx->type; 2805 2806 /* 2807 * Eliminate duplicate stores of a constant. 2808 * This happens frequently when the target ISA zero-extends. 2809 */ 2810 if (ts_is_const(src)) { 2811 TCGTemp *prev = find_mem_copy_for(ctx, type, ofs); 2812 if (src == prev) { 2813 tcg_op_remove(ctx->tcg, op); 2814 return true; 2815 } 2816 } 2817 2818 last = ofs + tcg_type_size(type) - 1; 2819 remove_mem_copy_in(ctx, ofs, last); 2820 record_mem_copy(ctx, type, src, ofs, last); 2821 return true; 2822 } 2823 2824 static bool fold_xor(OptContext *ctx, TCGOp *op) 2825 { 2826 uint64_t z_mask, s_mask; 2827 TempOptInfo *t1, *t2; 2828 2829 if (fold_const2_commutative(ctx, op) || 2830 fold_xx_to_i(ctx, op, 0) || 2831 fold_xi_to_x(ctx, op, 0) || 2832 fold_xi_to_not(ctx, op, -1)) { 2833 return true; 2834 } 2835 2836 t1 = arg_info(op->args[1]); 2837 t2 = arg_info(op->args[2]); 2838 z_mask = t1->z_mask | t2->z_mask; 2839 s_mask = t1->s_mask & t2->s_mask; 2840 return fold_masks_zs(ctx, op, z_mask, s_mask); 2841 } 2842 2843 /* Propagate constants and copies, fold constant expressions. */ 2844 void tcg_optimize(TCGContext *s) 2845 { 2846 int nb_temps, i; 2847 TCGOp *op, *op_next; 2848 OptContext ctx = { .tcg = s }; 2849 2850 QSIMPLEQ_INIT(&ctx.mem_free); 2851 2852 /* Array VALS has an element for each temp. 2853 If this temp holds a constant then its value is kept in VALS' element. 2854 If this temp is a copy of other ones then the other copies are 2855 available through the doubly linked circular list. */ 2856 2857 nb_temps = s->nb_temps; 2858 for (i = 0; i < nb_temps; ++i) { 2859 s->temps[i].state_ptr = NULL; 2860 } 2861 2862 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) { 2863 TCGOpcode opc = op->opc; 2864 const TCGOpDef *def; 2865 bool done = false; 2866 2867 /* Calls are special. */ 2868 if (opc == INDEX_op_call) { 2869 fold_call(&ctx, op); 2870 continue; 2871 } 2872 2873 def = &tcg_op_defs[opc]; 2874 init_arguments(&ctx, op, def->nb_oargs + def->nb_iargs); 2875 copy_propagate(&ctx, op, def->nb_oargs, def->nb_iargs); 2876 2877 /* Pre-compute the type of the operation. */ 2878 ctx.type = TCGOP_TYPE(op); 2879 2880 /* 2881 * Process each opcode. 2882 * Sorted alphabetically by opcode as much as possible. 2883 */ 2884 switch (opc) { 2885 case INDEX_op_add: 2886 done = fold_add(&ctx, op); 2887 break; 2888 case INDEX_op_add_vec: 2889 done = fold_add_vec(&ctx, op); 2890 break; 2891 CASE_OP_32_64(add2): 2892 done = fold_add2(&ctx, op); 2893 break; 2894 case INDEX_op_and: 2895 case INDEX_op_and_vec: 2896 done = fold_and(&ctx, op); 2897 break; 2898 case INDEX_op_andc: 2899 case INDEX_op_andc_vec: 2900 done = fold_andc(&ctx, op); 2901 break; 2902 CASE_OP_32_64(brcond): 2903 done = fold_brcond(&ctx, op); 2904 break; 2905 case INDEX_op_brcond2_i32: 2906 done = fold_brcond2(&ctx, op); 2907 break; 2908 CASE_OP_32_64(bswap16): 2909 CASE_OP_32_64(bswap32): 2910 case INDEX_op_bswap64_i64: 2911 done = fold_bswap(&ctx, op); 2912 break; 2913 CASE_OP_32_64(clz): 2914 CASE_OP_32_64(ctz): 2915 done = fold_count_zeros(&ctx, op); 2916 break; 2917 CASE_OP_32_64(ctpop): 2918 done = fold_ctpop(&ctx, op); 2919 break; 2920 CASE_OP_32_64(deposit): 2921 done = fold_deposit(&ctx, op); 2922 break; 2923 CASE_OP_32_64(div): 2924 CASE_OP_32_64(divu): 2925 done = fold_divide(&ctx, op); 2926 break; 2927 case INDEX_op_dup_vec: 2928 done = fold_dup(&ctx, op); 2929 break; 2930 case INDEX_op_dup2_vec: 2931 done = fold_dup2(&ctx, op); 2932 break; 2933 case INDEX_op_eqv: 2934 case INDEX_op_eqv_vec: 2935 done = fold_eqv(&ctx, op); 2936 break; 2937 CASE_OP_32_64(extract): 2938 done = fold_extract(&ctx, op); 2939 break; 2940 CASE_OP_32_64(extract2): 2941 done = fold_extract2(&ctx, op); 2942 break; 2943 case INDEX_op_ext_i32_i64: 2944 done = fold_exts(&ctx, op); 2945 break; 2946 case INDEX_op_extu_i32_i64: 2947 case INDEX_op_extrl_i64_i32: 2948 case INDEX_op_extrh_i64_i32: 2949 done = fold_extu(&ctx, op); 2950 break; 2951 CASE_OP_32_64(ld8s): 2952 CASE_OP_32_64(ld8u): 2953 CASE_OP_32_64(ld16s): 2954 CASE_OP_32_64(ld16u): 2955 case INDEX_op_ld32s_i64: 2956 case INDEX_op_ld32u_i64: 2957 done = fold_tcg_ld(&ctx, op); 2958 break; 2959 case INDEX_op_ld_i32: 2960 case INDEX_op_ld_i64: 2961 case INDEX_op_ld_vec: 2962 done = fold_tcg_ld_memcopy(&ctx, op); 2963 break; 2964 CASE_OP_32_64(st8): 2965 CASE_OP_32_64(st16): 2966 case INDEX_op_st32_i64: 2967 done = fold_tcg_st(&ctx, op); 2968 break; 2969 case INDEX_op_st_i32: 2970 case INDEX_op_st_i64: 2971 case INDEX_op_st_vec: 2972 done = fold_tcg_st_memcopy(&ctx, op); 2973 break; 2974 case INDEX_op_mb: 2975 done = fold_mb(&ctx, op); 2976 break; 2977 case INDEX_op_mov: 2978 case INDEX_op_mov_vec: 2979 done = fold_mov(&ctx, op); 2980 break; 2981 CASE_OP_32_64(movcond): 2982 done = fold_movcond(&ctx, op); 2983 break; 2984 CASE_OP_32_64(mul): 2985 done = fold_mul(&ctx, op); 2986 break; 2987 CASE_OP_32_64(mulsh): 2988 CASE_OP_32_64(muluh): 2989 done = fold_mul_highpart(&ctx, op); 2990 break; 2991 CASE_OP_32_64(muls2): 2992 CASE_OP_32_64(mulu2): 2993 done = fold_multiply2(&ctx, op); 2994 break; 2995 CASE_OP_32_64_VEC(nand): 2996 done = fold_nand(&ctx, op); 2997 break; 2998 CASE_OP_32_64(neg): 2999 done = fold_neg(&ctx, op); 3000 break; 3001 CASE_OP_32_64_VEC(nor): 3002 done = fold_nor(&ctx, op); 3003 break; 3004 CASE_OP_32_64_VEC(not): 3005 done = fold_not(&ctx, op); 3006 break; 3007 case INDEX_op_or: 3008 case INDEX_op_or_vec: 3009 done = fold_or(&ctx, op); 3010 break; 3011 case INDEX_op_orc: 3012 case INDEX_op_orc_vec: 3013 done = fold_orc(&ctx, op); 3014 break; 3015 case INDEX_op_qemu_ld_i32: 3016 done = fold_qemu_ld_1reg(&ctx, op); 3017 break; 3018 case INDEX_op_qemu_ld_i64: 3019 if (TCG_TARGET_REG_BITS == 64) { 3020 done = fold_qemu_ld_1reg(&ctx, op); 3021 break; 3022 } 3023 QEMU_FALLTHROUGH; 3024 case INDEX_op_qemu_ld_i128: 3025 done = fold_qemu_ld_2reg(&ctx, op); 3026 break; 3027 case INDEX_op_qemu_st8_i32: 3028 case INDEX_op_qemu_st_i32: 3029 case INDEX_op_qemu_st_i64: 3030 case INDEX_op_qemu_st_i128: 3031 done = fold_qemu_st(&ctx, op); 3032 break; 3033 CASE_OP_32_64(rem): 3034 CASE_OP_32_64(remu): 3035 done = fold_remainder(&ctx, op); 3036 break; 3037 CASE_OP_32_64(rotl): 3038 CASE_OP_32_64(rotr): 3039 CASE_OP_32_64(sar): 3040 CASE_OP_32_64(shl): 3041 CASE_OP_32_64(shr): 3042 done = fold_shift(&ctx, op); 3043 break; 3044 CASE_OP_32_64(setcond): 3045 done = fold_setcond(&ctx, op); 3046 break; 3047 CASE_OP_32_64(negsetcond): 3048 done = fold_negsetcond(&ctx, op); 3049 break; 3050 case INDEX_op_setcond2_i32: 3051 done = fold_setcond2(&ctx, op); 3052 break; 3053 case INDEX_op_cmp_vec: 3054 done = fold_cmp_vec(&ctx, op); 3055 break; 3056 case INDEX_op_cmpsel_vec: 3057 done = fold_cmpsel_vec(&ctx, op); 3058 break; 3059 case INDEX_op_bitsel_vec: 3060 done = fold_bitsel_vec(&ctx, op); 3061 break; 3062 CASE_OP_32_64(sextract): 3063 done = fold_sextract(&ctx, op); 3064 break; 3065 CASE_OP_32_64(sub): 3066 done = fold_sub(&ctx, op); 3067 break; 3068 case INDEX_op_sub_vec: 3069 done = fold_sub_vec(&ctx, op); 3070 break; 3071 CASE_OP_32_64(sub2): 3072 done = fold_sub2(&ctx, op); 3073 break; 3074 case INDEX_op_xor: 3075 case INDEX_op_xor_vec: 3076 done = fold_xor(&ctx, op); 3077 break; 3078 case INDEX_op_set_label: 3079 case INDEX_op_br: 3080 case INDEX_op_exit_tb: 3081 case INDEX_op_goto_tb: 3082 case INDEX_op_goto_ptr: 3083 finish_ebb(&ctx); 3084 done = true; 3085 break; 3086 default: 3087 done = finish_folding(&ctx, op); 3088 break; 3089 } 3090 tcg_debug_assert(done); 3091 } 3092 } 3093