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