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 INDEX_op_nand: 497 case INDEX_op_nand_vec: 498 return ~(x & y); 499 500 case INDEX_op_nor: 501 case INDEX_op_nor_vec: 502 return ~(x | y); 503 504 case INDEX_op_clz_i32: 505 return (uint32_t)x ? clz32(x) : y; 506 507 case INDEX_op_clz_i64: 508 return x ? clz64(x) : y; 509 510 case INDEX_op_ctz_i32: 511 return (uint32_t)x ? ctz32(x) : y; 512 513 case INDEX_op_ctz_i64: 514 return x ? ctz64(x) : y; 515 516 case INDEX_op_ctpop_i32: 517 return ctpop32(x); 518 519 case INDEX_op_ctpop_i64: 520 return ctpop64(x); 521 522 CASE_OP_32_64(bswap16): 523 x = bswap16(x); 524 return y & TCG_BSWAP_OS ? (int16_t)x : x; 525 526 CASE_OP_32_64(bswap32): 527 x = bswap32(x); 528 return y & TCG_BSWAP_OS ? (int32_t)x : x; 529 530 case INDEX_op_bswap64_i64: 531 return bswap64(x); 532 533 case INDEX_op_ext_i32_i64: 534 return (int32_t)x; 535 536 case INDEX_op_extu_i32_i64: 537 case INDEX_op_extrl_i64_i32: 538 return (uint32_t)x; 539 540 case INDEX_op_extrh_i64_i32: 541 return (uint64_t)x >> 32; 542 543 case INDEX_op_muluh_i32: 544 return ((uint64_t)(uint32_t)x * (uint32_t)y) >> 32; 545 case INDEX_op_mulsh_i32: 546 return ((int64_t)(int32_t)x * (int32_t)y) >> 32; 547 548 case INDEX_op_muluh_i64: 549 mulu64(&l64, &h64, x, y); 550 return h64; 551 case INDEX_op_mulsh_i64: 552 muls64(&l64, &h64, x, y); 553 return h64; 554 555 case INDEX_op_div_i32: 556 /* Avoid crashing on divide by zero, otherwise undefined. */ 557 return (int32_t)x / ((int32_t)y ? : 1); 558 case INDEX_op_divu_i32: 559 return (uint32_t)x / ((uint32_t)y ? : 1); 560 case INDEX_op_div_i64: 561 return (int64_t)x / ((int64_t)y ? : 1); 562 case INDEX_op_divu_i64: 563 return (uint64_t)x / ((uint64_t)y ? : 1); 564 565 case INDEX_op_rem_i32: 566 return (int32_t)x % ((int32_t)y ? : 1); 567 case INDEX_op_remu_i32: 568 return (uint32_t)x % ((uint32_t)y ? : 1); 569 case INDEX_op_rem_i64: 570 return (int64_t)x % ((int64_t)y ? : 1); 571 case INDEX_op_remu_i64: 572 return (uint64_t)x % ((uint64_t)y ? : 1); 573 574 default: 575 g_assert_not_reached(); 576 } 577 } 578 579 static uint64_t do_constant_folding(TCGOpcode op, TCGType type, 580 uint64_t x, uint64_t y) 581 { 582 uint64_t res = do_constant_folding_2(op, x, y); 583 if (type == TCG_TYPE_I32) { 584 res = (int32_t)res; 585 } 586 return res; 587 } 588 589 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c) 590 { 591 switch (c) { 592 case TCG_COND_EQ: 593 return x == y; 594 case TCG_COND_NE: 595 return x != y; 596 case TCG_COND_LT: 597 return (int32_t)x < (int32_t)y; 598 case TCG_COND_GE: 599 return (int32_t)x >= (int32_t)y; 600 case TCG_COND_LE: 601 return (int32_t)x <= (int32_t)y; 602 case TCG_COND_GT: 603 return (int32_t)x > (int32_t)y; 604 case TCG_COND_LTU: 605 return x < y; 606 case TCG_COND_GEU: 607 return x >= y; 608 case TCG_COND_LEU: 609 return x <= y; 610 case TCG_COND_GTU: 611 return x > y; 612 case TCG_COND_TSTEQ: 613 return (x & y) == 0; 614 case TCG_COND_TSTNE: 615 return (x & y) != 0; 616 case TCG_COND_ALWAYS: 617 case TCG_COND_NEVER: 618 break; 619 } 620 g_assert_not_reached(); 621 } 622 623 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c) 624 { 625 switch (c) { 626 case TCG_COND_EQ: 627 return x == y; 628 case TCG_COND_NE: 629 return x != y; 630 case TCG_COND_LT: 631 return (int64_t)x < (int64_t)y; 632 case TCG_COND_GE: 633 return (int64_t)x >= (int64_t)y; 634 case TCG_COND_LE: 635 return (int64_t)x <= (int64_t)y; 636 case TCG_COND_GT: 637 return (int64_t)x > (int64_t)y; 638 case TCG_COND_LTU: 639 return x < y; 640 case TCG_COND_GEU: 641 return x >= y; 642 case TCG_COND_LEU: 643 return x <= y; 644 case TCG_COND_GTU: 645 return x > y; 646 case TCG_COND_TSTEQ: 647 return (x & y) == 0; 648 case TCG_COND_TSTNE: 649 return (x & y) != 0; 650 case TCG_COND_ALWAYS: 651 case TCG_COND_NEVER: 652 break; 653 } 654 g_assert_not_reached(); 655 } 656 657 static int do_constant_folding_cond_eq(TCGCond c) 658 { 659 switch (c) { 660 case TCG_COND_GT: 661 case TCG_COND_LTU: 662 case TCG_COND_LT: 663 case TCG_COND_GTU: 664 case TCG_COND_NE: 665 return 0; 666 case TCG_COND_GE: 667 case TCG_COND_GEU: 668 case TCG_COND_LE: 669 case TCG_COND_LEU: 670 case TCG_COND_EQ: 671 return 1; 672 case TCG_COND_TSTEQ: 673 case TCG_COND_TSTNE: 674 return -1; 675 case TCG_COND_ALWAYS: 676 case TCG_COND_NEVER: 677 break; 678 } 679 g_assert_not_reached(); 680 } 681 682 /* 683 * Return -1 if the condition can't be simplified, 684 * and the result of the condition (0 or 1) if it can. 685 */ 686 static int do_constant_folding_cond(TCGType type, TCGArg x, 687 TCGArg y, TCGCond c) 688 { 689 if (arg_is_const(x) && arg_is_const(y)) { 690 uint64_t xv = arg_info(x)->val; 691 uint64_t yv = arg_info(y)->val; 692 693 switch (type) { 694 case TCG_TYPE_I32: 695 return do_constant_folding_cond_32(xv, yv, c); 696 case TCG_TYPE_I64: 697 return do_constant_folding_cond_64(xv, yv, c); 698 default: 699 /* Only scalar comparisons are optimizable */ 700 return -1; 701 } 702 } else if (args_are_copies(x, y)) { 703 return do_constant_folding_cond_eq(c); 704 } else if (arg_is_const_val(y, 0)) { 705 switch (c) { 706 case TCG_COND_LTU: 707 case TCG_COND_TSTNE: 708 return 0; 709 case TCG_COND_GEU: 710 case TCG_COND_TSTEQ: 711 return 1; 712 default: 713 return -1; 714 } 715 } 716 return -1; 717 } 718 719 /** 720 * swap_commutative: 721 * @dest: TCGArg of the destination argument, or NO_DEST. 722 * @p1: first paired argument 723 * @p2: second paired argument 724 * 725 * If *@p1 is a constant and *@p2 is not, swap. 726 * If *@p2 matches @dest, swap. 727 * Return true if a swap was performed. 728 */ 729 730 #define NO_DEST temp_arg(NULL) 731 732 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2) 733 { 734 TCGArg a1 = *p1, a2 = *p2; 735 int sum = 0; 736 sum += arg_is_const(a1); 737 sum -= arg_is_const(a2); 738 739 /* Prefer the constant in second argument, and then the form 740 op a, a, b, which is better handled on non-RISC hosts. */ 741 if (sum > 0 || (sum == 0 && dest == a2)) { 742 *p1 = a2; 743 *p2 = a1; 744 return true; 745 } 746 return false; 747 } 748 749 static bool swap_commutative2(TCGArg *p1, TCGArg *p2) 750 { 751 int sum = 0; 752 sum += arg_is_const(p1[0]); 753 sum += arg_is_const(p1[1]); 754 sum -= arg_is_const(p2[0]); 755 sum -= arg_is_const(p2[1]); 756 if (sum > 0) { 757 TCGArg t; 758 t = p1[0], p1[0] = p2[0], p2[0] = t; 759 t = p1[1], p1[1] = p2[1], p2[1] = t; 760 return true; 761 } 762 return false; 763 } 764 765 /* 766 * Return -1 if the condition can't be simplified, 767 * and the result of the condition (0 or 1) if it can. 768 */ 769 static int do_constant_folding_cond1(OptContext *ctx, TCGOp *op, TCGArg dest, 770 TCGArg *p1, TCGArg *p2, TCGArg *pcond) 771 { 772 TCGCond cond; 773 TempOptInfo *i1; 774 bool swap; 775 int r; 776 777 swap = swap_commutative(dest, p1, p2); 778 cond = *pcond; 779 if (swap) { 780 *pcond = cond = tcg_swap_cond(cond); 781 } 782 783 r = do_constant_folding_cond(ctx->type, *p1, *p2, cond); 784 if (r >= 0) { 785 return r; 786 } 787 if (!is_tst_cond(cond)) { 788 return -1; 789 } 790 791 i1 = arg_info(*p1); 792 793 /* 794 * TSTNE x,x -> NE x,0 795 * TSTNE x,i -> NE x,0 if i includes all nonzero bits of x 796 */ 797 if (args_are_copies(*p1, *p2) || 798 (arg_is_const(*p2) && (i1->z_mask & ~arg_info(*p2)->val) == 0)) { 799 *p2 = arg_new_constant(ctx, 0); 800 *pcond = tcg_tst_eqne_cond(cond); 801 return -1; 802 } 803 804 /* TSTNE x,i -> LT x,0 if i only includes sign bit copies */ 805 if (arg_is_const(*p2) && (arg_info(*p2)->val & ~i1->s_mask) == 0) { 806 *p2 = arg_new_constant(ctx, 0); 807 *pcond = tcg_tst_ltge_cond(cond); 808 return -1; 809 } 810 811 /* Expand to AND with a temporary if no backend support. */ 812 if (!TCG_TARGET_HAS_tst) { 813 TCGOp *op2 = opt_insert_before(ctx, op, INDEX_op_and, 3); 814 TCGArg tmp = arg_new_temp(ctx); 815 816 op2->args[0] = tmp; 817 op2->args[1] = *p1; 818 op2->args[2] = *p2; 819 820 *p1 = tmp; 821 *p2 = arg_new_constant(ctx, 0); 822 *pcond = tcg_tst_eqne_cond(cond); 823 } 824 return -1; 825 } 826 827 static int do_constant_folding_cond2(OptContext *ctx, TCGOp *op, TCGArg *args) 828 { 829 TCGArg al, ah, bl, bh; 830 TCGCond c; 831 bool swap; 832 int r; 833 834 swap = swap_commutative2(args, args + 2); 835 c = args[4]; 836 if (swap) { 837 args[4] = c = tcg_swap_cond(c); 838 } 839 840 al = args[0]; 841 ah = args[1]; 842 bl = args[2]; 843 bh = args[3]; 844 845 if (arg_is_const(bl) && arg_is_const(bh)) { 846 tcg_target_ulong blv = arg_info(bl)->val; 847 tcg_target_ulong bhv = arg_info(bh)->val; 848 uint64_t b = deposit64(blv, 32, 32, bhv); 849 850 if (arg_is_const(al) && arg_is_const(ah)) { 851 tcg_target_ulong alv = arg_info(al)->val; 852 tcg_target_ulong ahv = arg_info(ah)->val; 853 uint64_t a = deposit64(alv, 32, 32, ahv); 854 855 r = do_constant_folding_cond_64(a, b, c); 856 if (r >= 0) { 857 return r; 858 } 859 } 860 861 if (b == 0) { 862 switch (c) { 863 case TCG_COND_LTU: 864 case TCG_COND_TSTNE: 865 return 0; 866 case TCG_COND_GEU: 867 case TCG_COND_TSTEQ: 868 return 1; 869 default: 870 break; 871 } 872 } 873 874 /* TSTNE x,-1 -> NE x,0 */ 875 if (b == -1 && is_tst_cond(c)) { 876 args[3] = args[2] = arg_new_constant(ctx, 0); 877 args[4] = tcg_tst_eqne_cond(c); 878 return -1; 879 } 880 881 /* TSTNE x,sign -> LT x,0 */ 882 if (b == INT64_MIN && is_tst_cond(c)) { 883 /* bl must be 0, so copy that to bh */ 884 args[3] = bl; 885 args[4] = tcg_tst_ltge_cond(c); 886 return -1; 887 } 888 } 889 890 if (args_are_copies(al, bl) && args_are_copies(ah, bh)) { 891 r = do_constant_folding_cond_eq(c); 892 if (r >= 0) { 893 return r; 894 } 895 896 /* TSTNE x,x -> NE x,0 */ 897 if (is_tst_cond(c)) { 898 args[3] = args[2] = arg_new_constant(ctx, 0); 899 args[4] = tcg_tst_eqne_cond(c); 900 return -1; 901 } 902 } 903 904 /* Expand to AND with a temporary if no backend support. */ 905 if (!TCG_TARGET_HAS_tst && is_tst_cond(c)) { 906 TCGOp *op1 = opt_insert_before(ctx, op, INDEX_op_and, 3); 907 TCGOp *op2 = opt_insert_before(ctx, op, INDEX_op_and, 3); 908 TCGArg t1 = arg_new_temp(ctx); 909 TCGArg t2 = arg_new_temp(ctx); 910 911 op1->args[0] = t1; 912 op1->args[1] = al; 913 op1->args[2] = bl; 914 op2->args[0] = t2; 915 op2->args[1] = ah; 916 op2->args[2] = bh; 917 918 args[0] = t1; 919 args[1] = t2; 920 args[3] = args[2] = arg_new_constant(ctx, 0); 921 args[4] = tcg_tst_eqne_cond(c); 922 } 923 return -1; 924 } 925 926 static void init_arguments(OptContext *ctx, TCGOp *op, int nb_args) 927 { 928 for (int i = 0; i < nb_args; i++) { 929 TCGTemp *ts = arg_temp(op->args[i]); 930 init_ts_info(ctx, ts); 931 } 932 } 933 934 static void copy_propagate(OptContext *ctx, TCGOp *op, 935 int nb_oargs, int nb_iargs) 936 { 937 for (int i = nb_oargs; i < nb_oargs + nb_iargs; i++) { 938 TCGTemp *ts = arg_temp(op->args[i]); 939 if (ts_is_copy(ts)) { 940 op->args[i] = temp_arg(find_better_copy(ts)); 941 } 942 } 943 } 944 945 static void finish_bb(OptContext *ctx) 946 { 947 /* We only optimize memory barriers across basic blocks. */ 948 ctx->prev_mb = NULL; 949 } 950 951 static void finish_ebb(OptContext *ctx) 952 { 953 finish_bb(ctx); 954 /* We only optimize across extended basic blocks. */ 955 memset(&ctx->temps_used, 0, sizeof(ctx->temps_used)); 956 remove_mem_copy_all(ctx); 957 } 958 959 static bool finish_folding(OptContext *ctx, TCGOp *op) 960 { 961 const TCGOpDef *def = &tcg_op_defs[op->opc]; 962 int i, nb_oargs; 963 964 nb_oargs = def->nb_oargs; 965 for (i = 0; i < nb_oargs; i++) { 966 TCGTemp *ts = arg_temp(op->args[i]); 967 reset_ts(ctx, ts); 968 } 969 return true; 970 } 971 972 /* 973 * The fold_* functions return true when processing is complete, 974 * usually by folding the operation to a constant or to a copy, 975 * and calling tcg_opt_gen_{mov,movi}. They may do other things, 976 * like collect information about the value produced, for use in 977 * optimizing a subsequent operation. 978 * 979 * These first fold_* functions are all helpers, used by other 980 * folders for more specific operations. 981 */ 982 983 static bool fold_const1(OptContext *ctx, TCGOp *op) 984 { 985 if (arg_is_const(op->args[1])) { 986 uint64_t t; 987 988 t = arg_info(op->args[1])->val; 989 t = do_constant_folding(op->opc, ctx->type, t, 0); 990 return tcg_opt_gen_movi(ctx, op, op->args[0], t); 991 } 992 return false; 993 } 994 995 static bool fold_const2(OptContext *ctx, TCGOp *op) 996 { 997 if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { 998 uint64_t t1 = arg_info(op->args[1])->val; 999 uint64_t t2 = arg_info(op->args[2])->val; 1000 1001 t1 = do_constant_folding(op->opc, ctx->type, t1, t2); 1002 return tcg_opt_gen_movi(ctx, op, op->args[0], t1); 1003 } 1004 return false; 1005 } 1006 1007 static bool fold_commutative(OptContext *ctx, TCGOp *op) 1008 { 1009 swap_commutative(op->args[0], &op->args[1], &op->args[2]); 1010 return false; 1011 } 1012 1013 static bool fold_const2_commutative(OptContext *ctx, TCGOp *op) 1014 { 1015 swap_commutative(op->args[0], &op->args[1], &op->args[2]); 1016 return fold_const2(ctx, op); 1017 } 1018 1019 /* 1020 * Record "zero" and "sign" masks for the single output of @op. 1021 * See TempOptInfo definition of z_mask and s_mask. 1022 * If z_mask allows, fold the output to constant zero. 1023 * The passed s_mask may be augmented by z_mask. 1024 */ 1025 static bool fold_masks_zs(OptContext *ctx, TCGOp *op, 1026 uint64_t z_mask, int64_t s_mask) 1027 { 1028 const TCGOpDef *def = &tcg_op_defs[op->opc]; 1029 TCGTemp *ts; 1030 TempOptInfo *ti; 1031 int rep; 1032 1033 /* Only single-output opcodes are supported here. */ 1034 tcg_debug_assert(def->nb_oargs == 1); 1035 1036 /* 1037 * 32-bit ops generate 32-bit results, which for the purpose of 1038 * simplifying tcg are sign-extended. Certainly that's how we 1039 * represent our constants elsewhere. Note that the bits will 1040 * be reset properly for a 64-bit value when encountering the 1041 * type changing opcodes. 1042 */ 1043 if (ctx->type == TCG_TYPE_I32) { 1044 z_mask = (int32_t)z_mask; 1045 s_mask |= INT32_MIN; 1046 } 1047 1048 if (z_mask == 0) { 1049 return tcg_opt_gen_movi(ctx, op, op->args[0], 0); 1050 } 1051 1052 ts = arg_temp(op->args[0]); 1053 reset_ts(ctx, ts); 1054 1055 ti = ts_info(ts); 1056 ti->z_mask = z_mask; 1057 1058 /* Canonicalize s_mask and incorporate data from z_mask. */ 1059 rep = clz64(~s_mask); 1060 rep = MAX(rep, clz64(z_mask)); 1061 rep = MAX(rep - 1, 0); 1062 ti->s_mask = INT64_MIN >> rep; 1063 1064 return true; 1065 } 1066 1067 static bool fold_masks_z(OptContext *ctx, TCGOp *op, uint64_t z_mask) 1068 { 1069 return fold_masks_zs(ctx, op, z_mask, 0); 1070 } 1071 1072 static bool fold_masks_s(OptContext *ctx, TCGOp *op, uint64_t s_mask) 1073 { 1074 return fold_masks_zs(ctx, op, -1, s_mask); 1075 } 1076 1077 /* 1078 * An "affected" mask bit is 0 if and only if the result is identical 1079 * to the first input. Thus if the entire mask is 0, the operation 1080 * is equivalent to a copy. 1081 */ 1082 static bool fold_affected_mask(OptContext *ctx, TCGOp *op, uint64_t a_mask) 1083 { 1084 if (ctx->type == TCG_TYPE_I32) { 1085 a_mask = (uint32_t)a_mask; 1086 } 1087 if (a_mask == 0) { 1088 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); 1089 } 1090 return false; 1091 } 1092 1093 /* 1094 * Convert @op to NOT, if NOT is supported by the host. 1095 * Return true f the conversion is successful, which will still 1096 * indicate that the processing is complete. 1097 */ 1098 static bool fold_not(OptContext *ctx, TCGOp *op); 1099 static bool fold_to_not(OptContext *ctx, TCGOp *op, int idx) 1100 { 1101 TCGOpcode not_op; 1102 bool have_not; 1103 1104 switch (ctx->type) { 1105 case TCG_TYPE_I32: 1106 not_op = INDEX_op_not_i32; 1107 have_not = TCG_TARGET_HAS_not_i32; 1108 break; 1109 case TCG_TYPE_I64: 1110 not_op = INDEX_op_not_i64; 1111 have_not = TCG_TARGET_HAS_not_i64; 1112 break; 1113 case TCG_TYPE_V64: 1114 case TCG_TYPE_V128: 1115 case TCG_TYPE_V256: 1116 not_op = INDEX_op_not_vec; 1117 have_not = TCG_TARGET_HAS_not_vec; 1118 break; 1119 default: 1120 g_assert_not_reached(); 1121 } 1122 if (have_not) { 1123 op->opc = not_op; 1124 op->args[1] = op->args[idx]; 1125 return fold_not(ctx, op); 1126 } 1127 return false; 1128 } 1129 1130 /* If the binary operation has first argument @i, fold to @i. */ 1131 static bool fold_ix_to_i(OptContext *ctx, TCGOp *op, uint64_t i) 1132 { 1133 if (arg_is_const_val(op->args[1], i)) { 1134 return tcg_opt_gen_movi(ctx, op, op->args[0], i); 1135 } 1136 return false; 1137 } 1138 1139 /* If the binary operation has first argument @i, fold to NOT. */ 1140 static bool fold_ix_to_not(OptContext *ctx, TCGOp *op, uint64_t i) 1141 { 1142 if (arg_is_const_val(op->args[1], i)) { 1143 return fold_to_not(ctx, op, 2); 1144 } 1145 return false; 1146 } 1147 1148 /* If the binary operation has second argument @i, fold to @i. */ 1149 static bool fold_xi_to_i(OptContext *ctx, TCGOp *op, uint64_t i) 1150 { 1151 if (arg_is_const_val(op->args[2], i)) { 1152 return tcg_opt_gen_movi(ctx, op, op->args[0], i); 1153 } 1154 return false; 1155 } 1156 1157 /* If the binary operation has second argument @i, fold to identity. */ 1158 static bool fold_xi_to_x(OptContext *ctx, TCGOp *op, uint64_t i) 1159 { 1160 if (arg_is_const_val(op->args[2], i)) { 1161 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); 1162 } 1163 return false; 1164 } 1165 1166 /* If the binary operation has second argument @i, fold to NOT. */ 1167 static bool fold_xi_to_not(OptContext *ctx, TCGOp *op, uint64_t i) 1168 { 1169 if (arg_is_const_val(op->args[2], i)) { 1170 return fold_to_not(ctx, op, 1); 1171 } 1172 return false; 1173 } 1174 1175 /* If the binary operation has both arguments equal, fold to @i. */ 1176 static bool fold_xx_to_i(OptContext *ctx, TCGOp *op, uint64_t i) 1177 { 1178 if (args_are_copies(op->args[1], op->args[2])) { 1179 return tcg_opt_gen_movi(ctx, op, op->args[0], i); 1180 } 1181 return false; 1182 } 1183 1184 /* If the binary operation has both arguments equal, fold to identity. */ 1185 static bool fold_xx_to_x(OptContext *ctx, TCGOp *op) 1186 { 1187 if (args_are_copies(op->args[1], op->args[2])) { 1188 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); 1189 } 1190 return false; 1191 } 1192 1193 /* 1194 * These outermost fold_<op> functions are sorted alphabetically. 1195 * 1196 * The ordering of the transformations should be: 1197 * 1) those that produce a constant 1198 * 2) those that produce a copy 1199 * 3) those that produce information about the result value. 1200 */ 1201 1202 static bool fold_or(OptContext *ctx, TCGOp *op); 1203 static bool fold_orc(OptContext *ctx, TCGOp *op); 1204 static bool fold_xor(OptContext *ctx, TCGOp *op); 1205 1206 static bool fold_add(OptContext *ctx, TCGOp *op) 1207 { 1208 if (fold_const2_commutative(ctx, op) || 1209 fold_xi_to_x(ctx, op, 0)) { 1210 return true; 1211 } 1212 return finish_folding(ctx, op); 1213 } 1214 1215 /* We cannot as yet do_constant_folding with vectors. */ 1216 static bool fold_add_vec(OptContext *ctx, TCGOp *op) 1217 { 1218 if (fold_commutative(ctx, op) || 1219 fold_xi_to_x(ctx, op, 0)) { 1220 return true; 1221 } 1222 return finish_folding(ctx, op); 1223 } 1224 1225 static bool fold_addsub2(OptContext *ctx, TCGOp *op, bool add) 1226 { 1227 bool a_const = arg_is_const(op->args[2]) && arg_is_const(op->args[3]); 1228 bool b_const = arg_is_const(op->args[4]) && arg_is_const(op->args[5]); 1229 1230 if (a_const && b_const) { 1231 uint64_t al = arg_info(op->args[2])->val; 1232 uint64_t ah = arg_info(op->args[3])->val; 1233 uint64_t bl = arg_info(op->args[4])->val; 1234 uint64_t bh = arg_info(op->args[5])->val; 1235 TCGArg rl, rh; 1236 TCGOp *op2; 1237 1238 if (ctx->type == TCG_TYPE_I32) { 1239 uint64_t a = deposit64(al, 32, 32, ah); 1240 uint64_t b = deposit64(bl, 32, 32, bh); 1241 1242 if (add) { 1243 a += b; 1244 } else { 1245 a -= b; 1246 } 1247 1248 al = sextract64(a, 0, 32); 1249 ah = sextract64(a, 32, 32); 1250 } else { 1251 Int128 a = int128_make128(al, ah); 1252 Int128 b = int128_make128(bl, bh); 1253 1254 if (add) { 1255 a = int128_add(a, b); 1256 } else { 1257 a = int128_sub(a, b); 1258 } 1259 1260 al = int128_getlo(a); 1261 ah = int128_gethi(a); 1262 } 1263 1264 rl = op->args[0]; 1265 rh = op->args[1]; 1266 1267 /* The proper opcode is supplied by tcg_opt_gen_mov. */ 1268 op2 = opt_insert_before(ctx, op, 0, 2); 1269 1270 tcg_opt_gen_movi(ctx, op, rl, al); 1271 tcg_opt_gen_movi(ctx, op2, rh, ah); 1272 return true; 1273 } 1274 1275 /* Fold sub2 r,x,i to add2 r,x,-i */ 1276 if (!add && b_const) { 1277 uint64_t bl = arg_info(op->args[4])->val; 1278 uint64_t bh = arg_info(op->args[5])->val; 1279 1280 /* Negate the two parts without assembling and disassembling. */ 1281 bl = -bl; 1282 bh = ~bh + !bl; 1283 1284 op->opc = (ctx->type == TCG_TYPE_I32 1285 ? INDEX_op_add2_i32 : INDEX_op_add2_i64); 1286 op->args[4] = arg_new_constant(ctx, bl); 1287 op->args[5] = arg_new_constant(ctx, bh); 1288 } 1289 return finish_folding(ctx, op); 1290 } 1291 1292 static bool fold_add2(OptContext *ctx, TCGOp *op) 1293 { 1294 /* Note that the high and low parts may be independently swapped. */ 1295 swap_commutative(op->args[0], &op->args[2], &op->args[4]); 1296 swap_commutative(op->args[1], &op->args[3], &op->args[5]); 1297 1298 return fold_addsub2(ctx, op, true); 1299 } 1300 1301 static bool fold_and(OptContext *ctx, TCGOp *op) 1302 { 1303 uint64_t z1, z2, z_mask, s_mask; 1304 TempOptInfo *t1, *t2; 1305 1306 if (fold_const2_commutative(ctx, op) || 1307 fold_xi_to_i(ctx, op, 0) || 1308 fold_xi_to_x(ctx, op, -1) || 1309 fold_xx_to_x(ctx, op)) { 1310 return true; 1311 } 1312 1313 t1 = arg_info(op->args[1]); 1314 t2 = arg_info(op->args[2]); 1315 z1 = t1->z_mask; 1316 z2 = t2->z_mask; 1317 1318 /* 1319 * Known-zeros does not imply known-ones. Therefore unless 1320 * arg2 is constant, we can't infer affected bits from it. 1321 */ 1322 if (ti_is_const(t2) && fold_affected_mask(ctx, op, z1 & ~z2)) { 1323 return true; 1324 } 1325 1326 z_mask = z1 & z2; 1327 1328 /* 1329 * Sign repetitions are perforce all identical, whether they are 1 or 0. 1330 * Bitwise operations preserve the relative quantity of the repetitions. 1331 */ 1332 s_mask = t1->s_mask & t2->s_mask; 1333 1334 return fold_masks_zs(ctx, op, z_mask, s_mask); 1335 } 1336 1337 static bool fold_andc(OptContext *ctx, TCGOp *op) 1338 { 1339 uint64_t z_mask, s_mask; 1340 TempOptInfo *t1, *t2; 1341 1342 if (fold_const2(ctx, op) || 1343 fold_xx_to_i(ctx, op, 0) || 1344 fold_xi_to_x(ctx, op, 0) || 1345 fold_ix_to_not(ctx, op, -1)) { 1346 return true; 1347 } 1348 1349 t1 = arg_info(op->args[1]); 1350 t2 = arg_info(op->args[2]); 1351 z_mask = t1->z_mask; 1352 1353 if (ti_is_const(t2)) { 1354 /* Fold andc r,x,i to and r,x,~i. */ 1355 switch (ctx->type) { 1356 case TCG_TYPE_I32: 1357 case TCG_TYPE_I64: 1358 op->opc = INDEX_op_and; 1359 break; 1360 case TCG_TYPE_V64: 1361 case TCG_TYPE_V128: 1362 case TCG_TYPE_V256: 1363 op->opc = INDEX_op_and_vec; 1364 break; 1365 default: 1366 g_assert_not_reached(); 1367 } 1368 op->args[2] = arg_new_constant(ctx, ~ti_const_val(t2)); 1369 return fold_and(ctx, op); 1370 } 1371 1372 /* 1373 * Known-zeros does not imply known-ones. Therefore unless 1374 * arg2 is constant, we can't infer anything from it. 1375 */ 1376 if (ti_is_const(t2)) { 1377 uint64_t v2 = ti_const_val(t2); 1378 if (fold_affected_mask(ctx, op, z_mask & v2)) { 1379 return true; 1380 } 1381 z_mask &= ~v2; 1382 } 1383 1384 s_mask = t1->s_mask & t2->s_mask; 1385 return fold_masks_zs(ctx, op, z_mask, s_mask); 1386 } 1387 1388 static bool fold_bitsel_vec(OptContext *ctx, TCGOp *op) 1389 { 1390 /* If true and false values are the same, eliminate the cmp. */ 1391 if (args_are_copies(op->args[2], op->args[3])) { 1392 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]); 1393 } 1394 1395 if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) { 1396 uint64_t tv = arg_info(op->args[2])->val; 1397 uint64_t fv = arg_info(op->args[3])->val; 1398 1399 if (tv == -1 && fv == 0) { 1400 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); 1401 } 1402 if (tv == 0 && fv == -1) { 1403 if (TCG_TARGET_HAS_not_vec) { 1404 op->opc = INDEX_op_not_vec; 1405 return fold_not(ctx, op); 1406 } else { 1407 op->opc = INDEX_op_xor_vec; 1408 op->args[2] = arg_new_constant(ctx, -1); 1409 return fold_xor(ctx, op); 1410 } 1411 } 1412 } 1413 if (arg_is_const(op->args[2])) { 1414 uint64_t tv = arg_info(op->args[2])->val; 1415 if (tv == -1) { 1416 op->opc = INDEX_op_or_vec; 1417 op->args[2] = op->args[3]; 1418 return fold_or(ctx, op); 1419 } 1420 if (tv == 0 && TCG_TARGET_HAS_andc_vec) { 1421 op->opc = INDEX_op_andc_vec; 1422 op->args[2] = op->args[1]; 1423 op->args[1] = op->args[3]; 1424 return fold_andc(ctx, op); 1425 } 1426 } 1427 if (arg_is_const(op->args[3])) { 1428 uint64_t fv = arg_info(op->args[3])->val; 1429 if (fv == 0) { 1430 op->opc = INDEX_op_and_vec; 1431 return fold_and(ctx, op); 1432 } 1433 if (fv == -1 && TCG_TARGET_HAS_orc_vec) { 1434 op->opc = INDEX_op_orc_vec; 1435 op->args[2] = op->args[1]; 1436 op->args[1] = op->args[3]; 1437 return fold_orc(ctx, op); 1438 } 1439 } 1440 return finish_folding(ctx, op); 1441 } 1442 1443 static bool fold_brcond(OptContext *ctx, TCGOp *op) 1444 { 1445 int i = do_constant_folding_cond1(ctx, op, NO_DEST, &op->args[0], 1446 &op->args[1], &op->args[2]); 1447 if (i == 0) { 1448 tcg_op_remove(ctx->tcg, op); 1449 return true; 1450 } 1451 if (i > 0) { 1452 op->opc = INDEX_op_br; 1453 op->args[0] = op->args[3]; 1454 finish_ebb(ctx); 1455 } else { 1456 finish_bb(ctx); 1457 } 1458 return true; 1459 } 1460 1461 static bool fold_brcond2(OptContext *ctx, TCGOp *op) 1462 { 1463 TCGCond cond; 1464 TCGArg label; 1465 int i, inv = 0; 1466 1467 i = do_constant_folding_cond2(ctx, op, &op->args[0]); 1468 cond = op->args[4]; 1469 label = op->args[5]; 1470 if (i >= 0) { 1471 goto do_brcond_const; 1472 } 1473 1474 switch (cond) { 1475 case TCG_COND_LT: 1476 case TCG_COND_GE: 1477 /* 1478 * Simplify LT/GE comparisons vs zero to a single compare 1479 * vs the high word of the input. 1480 */ 1481 if (arg_is_const_val(op->args[2], 0) && 1482 arg_is_const_val(op->args[3], 0)) { 1483 goto do_brcond_high; 1484 } 1485 break; 1486 1487 case TCG_COND_NE: 1488 inv = 1; 1489 QEMU_FALLTHROUGH; 1490 case TCG_COND_EQ: 1491 /* 1492 * Simplify EQ/NE comparisons where one of the pairs 1493 * can be simplified. 1494 */ 1495 i = do_constant_folding_cond(TCG_TYPE_I32, op->args[0], 1496 op->args[2], cond); 1497 switch (i ^ inv) { 1498 case 0: 1499 goto do_brcond_const; 1500 case 1: 1501 goto do_brcond_high; 1502 } 1503 1504 i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1], 1505 op->args[3], cond); 1506 switch (i ^ inv) { 1507 case 0: 1508 goto do_brcond_const; 1509 case 1: 1510 goto do_brcond_low; 1511 } 1512 break; 1513 1514 case TCG_COND_TSTEQ: 1515 case TCG_COND_TSTNE: 1516 if (arg_is_const_val(op->args[2], 0)) { 1517 goto do_brcond_high; 1518 } 1519 if (arg_is_const_val(op->args[3], 0)) { 1520 goto do_brcond_low; 1521 } 1522 break; 1523 1524 default: 1525 break; 1526 1527 do_brcond_low: 1528 op->opc = INDEX_op_brcond_i32; 1529 op->args[1] = op->args[2]; 1530 op->args[2] = cond; 1531 op->args[3] = label; 1532 return fold_brcond(ctx, op); 1533 1534 do_brcond_high: 1535 op->opc = INDEX_op_brcond_i32; 1536 op->args[0] = op->args[1]; 1537 op->args[1] = op->args[3]; 1538 op->args[2] = cond; 1539 op->args[3] = label; 1540 return fold_brcond(ctx, op); 1541 1542 do_brcond_const: 1543 if (i == 0) { 1544 tcg_op_remove(ctx->tcg, op); 1545 return true; 1546 } 1547 op->opc = INDEX_op_br; 1548 op->args[0] = label; 1549 finish_ebb(ctx); 1550 return true; 1551 } 1552 1553 finish_bb(ctx); 1554 return true; 1555 } 1556 1557 static bool fold_bswap(OptContext *ctx, TCGOp *op) 1558 { 1559 uint64_t z_mask, s_mask, sign; 1560 TempOptInfo *t1 = arg_info(op->args[1]); 1561 1562 if (ti_is_const(t1)) { 1563 return tcg_opt_gen_movi(ctx, op, op->args[0], 1564 do_constant_folding(op->opc, ctx->type, 1565 ti_const_val(t1), 1566 op->args[2])); 1567 } 1568 1569 z_mask = t1->z_mask; 1570 switch (op->opc) { 1571 case INDEX_op_bswap16_i32: 1572 case INDEX_op_bswap16_i64: 1573 z_mask = bswap16(z_mask); 1574 sign = INT16_MIN; 1575 break; 1576 case INDEX_op_bswap32_i32: 1577 case INDEX_op_bswap32_i64: 1578 z_mask = bswap32(z_mask); 1579 sign = INT32_MIN; 1580 break; 1581 case INDEX_op_bswap64_i64: 1582 z_mask = bswap64(z_mask); 1583 sign = INT64_MIN; 1584 break; 1585 default: 1586 g_assert_not_reached(); 1587 } 1588 1589 s_mask = 0; 1590 switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) { 1591 case TCG_BSWAP_OZ: 1592 break; 1593 case TCG_BSWAP_OS: 1594 /* If the sign bit may be 1, force all the bits above to 1. */ 1595 if (z_mask & sign) { 1596 z_mask |= sign; 1597 } 1598 /* The value and therefore s_mask is explicitly sign-extended. */ 1599 s_mask = sign; 1600 break; 1601 default: 1602 /* The high bits are undefined: force all bits above the sign to 1. */ 1603 z_mask |= sign << 1; 1604 break; 1605 } 1606 1607 return fold_masks_zs(ctx, op, z_mask, s_mask); 1608 } 1609 1610 static bool fold_call(OptContext *ctx, TCGOp *op) 1611 { 1612 TCGContext *s = ctx->tcg; 1613 int nb_oargs = TCGOP_CALLO(op); 1614 int nb_iargs = TCGOP_CALLI(op); 1615 int flags, i; 1616 1617 init_arguments(ctx, op, nb_oargs + nb_iargs); 1618 copy_propagate(ctx, op, nb_oargs, nb_iargs); 1619 1620 /* If the function reads or writes globals, reset temp data. */ 1621 flags = tcg_call_flags(op); 1622 if (!(flags & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) { 1623 int nb_globals = s->nb_globals; 1624 1625 for (i = 0; i < nb_globals; i++) { 1626 if (test_bit(i, ctx->temps_used.l)) { 1627 reset_ts(ctx, &ctx->tcg->temps[i]); 1628 } 1629 } 1630 } 1631 1632 /* If the function has side effects, reset mem data. */ 1633 if (!(flags & TCG_CALL_NO_SIDE_EFFECTS)) { 1634 remove_mem_copy_all(ctx); 1635 } 1636 1637 /* Reset temp data for outputs. */ 1638 for (i = 0; i < nb_oargs; i++) { 1639 reset_temp(ctx, op->args[i]); 1640 } 1641 1642 /* Stop optimizing MB across calls. */ 1643 ctx->prev_mb = NULL; 1644 return true; 1645 } 1646 1647 static bool fold_cmp_vec(OptContext *ctx, TCGOp *op) 1648 { 1649 /* Canonicalize the comparison to put immediate second. */ 1650 if (swap_commutative(NO_DEST, &op->args[1], &op->args[2])) { 1651 op->args[3] = tcg_swap_cond(op->args[3]); 1652 } 1653 return finish_folding(ctx, op); 1654 } 1655 1656 static bool fold_cmpsel_vec(OptContext *ctx, TCGOp *op) 1657 { 1658 /* If true and false values are the same, eliminate the cmp. */ 1659 if (args_are_copies(op->args[3], op->args[4])) { 1660 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[3]); 1661 } 1662 1663 /* Canonicalize the comparison to put immediate second. */ 1664 if (swap_commutative(NO_DEST, &op->args[1], &op->args[2])) { 1665 op->args[5] = tcg_swap_cond(op->args[5]); 1666 } 1667 /* 1668 * Canonicalize the "false" input reg to match the destination, 1669 * so that the tcg backend can implement "move if true". 1670 */ 1671 if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) { 1672 op->args[5] = tcg_invert_cond(op->args[5]); 1673 } 1674 return finish_folding(ctx, op); 1675 } 1676 1677 static bool fold_count_zeros(OptContext *ctx, TCGOp *op) 1678 { 1679 uint64_t z_mask, s_mask; 1680 TempOptInfo *t1 = arg_info(op->args[1]); 1681 TempOptInfo *t2 = arg_info(op->args[2]); 1682 1683 if (ti_is_const(t1)) { 1684 uint64_t t = ti_const_val(t1); 1685 1686 if (t != 0) { 1687 t = do_constant_folding(op->opc, ctx->type, t, 0); 1688 return tcg_opt_gen_movi(ctx, op, op->args[0], t); 1689 } 1690 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]); 1691 } 1692 1693 switch (ctx->type) { 1694 case TCG_TYPE_I32: 1695 z_mask = 31; 1696 break; 1697 case TCG_TYPE_I64: 1698 z_mask = 63; 1699 break; 1700 default: 1701 g_assert_not_reached(); 1702 } 1703 s_mask = ~z_mask; 1704 z_mask |= t2->z_mask; 1705 s_mask &= t2->s_mask; 1706 1707 return fold_masks_zs(ctx, op, z_mask, s_mask); 1708 } 1709 1710 static bool fold_ctpop(OptContext *ctx, TCGOp *op) 1711 { 1712 uint64_t z_mask; 1713 1714 if (fold_const1(ctx, op)) { 1715 return true; 1716 } 1717 1718 switch (ctx->type) { 1719 case TCG_TYPE_I32: 1720 z_mask = 32 | 31; 1721 break; 1722 case TCG_TYPE_I64: 1723 z_mask = 64 | 63; 1724 break; 1725 default: 1726 g_assert_not_reached(); 1727 } 1728 return fold_masks_z(ctx, op, z_mask); 1729 } 1730 1731 static bool fold_deposit(OptContext *ctx, TCGOp *op) 1732 { 1733 TempOptInfo *t1 = arg_info(op->args[1]); 1734 TempOptInfo *t2 = arg_info(op->args[2]); 1735 int ofs = op->args[3]; 1736 int len = op->args[4]; 1737 int width = 8 * tcg_type_size(ctx->type); 1738 uint64_t z_mask, s_mask; 1739 1740 if (ti_is_const(t1) && ti_is_const(t2)) { 1741 return tcg_opt_gen_movi(ctx, op, op->args[0], 1742 deposit64(ti_const_val(t1), ofs, len, 1743 ti_const_val(t2))); 1744 } 1745 1746 /* Inserting a value into zero at offset 0. */ 1747 if (ti_is_const_val(t1, 0) && ofs == 0) { 1748 uint64_t mask = MAKE_64BIT_MASK(0, len); 1749 1750 op->opc = INDEX_op_and; 1751 op->args[1] = op->args[2]; 1752 op->args[2] = arg_new_constant(ctx, mask); 1753 return fold_and(ctx, op); 1754 } 1755 1756 /* Inserting zero into a value. */ 1757 if (ti_is_const_val(t2, 0)) { 1758 uint64_t mask = deposit64(-1, ofs, len, 0); 1759 1760 op->opc = INDEX_op_and; 1761 op->args[2] = arg_new_constant(ctx, mask); 1762 return fold_and(ctx, op); 1763 } 1764 1765 /* The s_mask from the top portion of the deposit is still valid. */ 1766 if (ofs + len == width) { 1767 s_mask = t2->s_mask << ofs; 1768 } else { 1769 s_mask = t1->s_mask & ~MAKE_64BIT_MASK(0, ofs + len); 1770 } 1771 1772 z_mask = deposit64(t1->z_mask, ofs, len, t2->z_mask); 1773 return fold_masks_zs(ctx, op, z_mask, s_mask); 1774 } 1775 1776 static bool fold_divide(OptContext *ctx, TCGOp *op) 1777 { 1778 if (fold_const2(ctx, op) || 1779 fold_xi_to_x(ctx, op, 1)) { 1780 return true; 1781 } 1782 return finish_folding(ctx, op); 1783 } 1784 1785 static bool fold_dup(OptContext *ctx, TCGOp *op) 1786 { 1787 if (arg_is_const(op->args[1])) { 1788 uint64_t t = arg_info(op->args[1])->val; 1789 t = dup_const(TCGOP_VECE(op), t); 1790 return tcg_opt_gen_movi(ctx, op, op->args[0], t); 1791 } 1792 return finish_folding(ctx, op); 1793 } 1794 1795 static bool fold_dup2(OptContext *ctx, TCGOp *op) 1796 { 1797 if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { 1798 uint64_t t = deposit64(arg_info(op->args[1])->val, 32, 32, 1799 arg_info(op->args[2])->val); 1800 return tcg_opt_gen_movi(ctx, op, op->args[0], t); 1801 } 1802 1803 if (args_are_copies(op->args[1], op->args[2])) { 1804 op->opc = INDEX_op_dup_vec; 1805 TCGOP_VECE(op) = MO_32; 1806 } 1807 return finish_folding(ctx, op); 1808 } 1809 1810 static bool fold_eqv(OptContext *ctx, TCGOp *op) 1811 { 1812 uint64_t s_mask; 1813 TempOptInfo *t1, *t2; 1814 1815 if (fold_const2_commutative(ctx, op) || 1816 fold_xi_to_x(ctx, op, -1) || 1817 fold_xi_to_not(ctx, op, 0)) { 1818 return true; 1819 } 1820 1821 t2 = arg_info(op->args[2]); 1822 if (ti_is_const(t2)) { 1823 /* Fold eqv r,x,i to xor r,x,~i. */ 1824 switch (ctx->type) { 1825 case TCG_TYPE_I32: 1826 case TCG_TYPE_I64: 1827 op->opc = INDEX_op_xor; 1828 break; 1829 case TCG_TYPE_V64: 1830 case TCG_TYPE_V128: 1831 case TCG_TYPE_V256: 1832 op->opc = INDEX_op_xor_vec; 1833 break; 1834 default: 1835 g_assert_not_reached(); 1836 } 1837 op->args[2] = arg_new_constant(ctx, ~ti_const_val(t2)); 1838 return fold_xor(ctx, op); 1839 } 1840 1841 t1 = arg_info(op->args[1]); 1842 s_mask = t1->s_mask & t2->s_mask; 1843 return fold_masks_s(ctx, op, s_mask); 1844 } 1845 1846 static bool fold_extract(OptContext *ctx, TCGOp *op) 1847 { 1848 uint64_t z_mask_old, z_mask; 1849 TempOptInfo *t1 = arg_info(op->args[1]); 1850 int pos = op->args[2]; 1851 int len = op->args[3]; 1852 1853 if (ti_is_const(t1)) { 1854 return tcg_opt_gen_movi(ctx, op, op->args[0], 1855 extract64(ti_const_val(t1), pos, len)); 1856 } 1857 1858 z_mask_old = t1->z_mask; 1859 z_mask = extract64(z_mask_old, pos, len); 1860 if (pos == 0 && fold_affected_mask(ctx, op, z_mask_old ^ z_mask)) { 1861 return true; 1862 } 1863 1864 return fold_masks_z(ctx, op, z_mask); 1865 } 1866 1867 static bool fold_extract2(OptContext *ctx, TCGOp *op) 1868 { 1869 if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { 1870 uint64_t v1 = arg_info(op->args[1])->val; 1871 uint64_t v2 = arg_info(op->args[2])->val; 1872 int shr = op->args[3]; 1873 1874 if (op->opc == INDEX_op_extract2_i64) { 1875 v1 >>= shr; 1876 v2 <<= 64 - shr; 1877 } else { 1878 v1 = (uint32_t)v1 >> shr; 1879 v2 = (uint64_t)((int32_t)v2 << (32 - shr)); 1880 } 1881 return tcg_opt_gen_movi(ctx, op, op->args[0], v1 | v2); 1882 } 1883 return finish_folding(ctx, op); 1884 } 1885 1886 static bool fold_exts(OptContext *ctx, TCGOp *op) 1887 { 1888 uint64_t s_mask, z_mask; 1889 TempOptInfo *t1; 1890 1891 if (fold_const1(ctx, op)) { 1892 return true; 1893 } 1894 1895 t1 = arg_info(op->args[1]); 1896 z_mask = t1->z_mask; 1897 s_mask = t1->s_mask; 1898 1899 switch (op->opc) { 1900 case INDEX_op_ext_i32_i64: 1901 s_mask |= INT32_MIN; 1902 z_mask = (int32_t)z_mask; 1903 break; 1904 default: 1905 g_assert_not_reached(); 1906 } 1907 return fold_masks_zs(ctx, op, z_mask, s_mask); 1908 } 1909 1910 static bool fold_extu(OptContext *ctx, TCGOp *op) 1911 { 1912 uint64_t z_mask; 1913 1914 if (fold_const1(ctx, op)) { 1915 return true; 1916 } 1917 1918 z_mask = arg_info(op->args[1])->z_mask; 1919 switch (op->opc) { 1920 case INDEX_op_extrl_i64_i32: 1921 case INDEX_op_extu_i32_i64: 1922 z_mask = (uint32_t)z_mask; 1923 break; 1924 case INDEX_op_extrh_i64_i32: 1925 z_mask >>= 32; 1926 break; 1927 default: 1928 g_assert_not_reached(); 1929 } 1930 return fold_masks_z(ctx, op, z_mask); 1931 } 1932 1933 static bool fold_mb(OptContext *ctx, TCGOp *op) 1934 { 1935 /* Eliminate duplicate and redundant fence instructions. */ 1936 if (ctx->prev_mb) { 1937 /* 1938 * Merge two barriers of the same type into one, 1939 * or a weaker barrier into a stronger one, 1940 * or two weaker barriers into a stronger one. 1941 * mb X; mb Y => mb X|Y 1942 * mb; strl => mb; st 1943 * ldaq; mb => ld; mb 1944 * ldaq; strl => ld; mb; st 1945 * Other combinations are also merged into a strong 1946 * barrier. This is stricter than specified but for 1947 * the purposes of TCG is better than not optimizing. 1948 */ 1949 ctx->prev_mb->args[0] |= op->args[0]; 1950 tcg_op_remove(ctx->tcg, op); 1951 } else { 1952 ctx->prev_mb = op; 1953 } 1954 return true; 1955 } 1956 1957 static bool fold_mov(OptContext *ctx, TCGOp *op) 1958 { 1959 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); 1960 } 1961 1962 static bool fold_movcond(OptContext *ctx, TCGOp *op) 1963 { 1964 uint64_t z_mask, s_mask; 1965 TempOptInfo *tt, *ft; 1966 int i; 1967 1968 /* If true and false values are the same, eliminate the cmp. */ 1969 if (args_are_copies(op->args[3], op->args[4])) { 1970 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[3]); 1971 } 1972 1973 /* 1974 * Canonicalize the "false" input reg to match the destination reg so 1975 * that the tcg backend can implement a "move if true" operation. 1976 */ 1977 if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) { 1978 op->args[5] = tcg_invert_cond(op->args[5]); 1979 } 1980 1981 i = do_constant_folding_cond1(ctx, op, NO_DEST, &op->args[1], 1982 &op->args[2], &op->args[5]); 1983 if (i >= 0) { 1984 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[4 - i]); 1985 } 1986 1987 tt = arg_info(op->args[3]); 1988 ft = arg_info(op->args[4]); 1989 z_mask = tt->z_mask | ft->z_mask; 1990 s_mask = tt->s_mask & ft->s_mask; 1991 1992 if (ti_is_const(tt) && ti_is_const(ft)) { 1993 uint64_t tv = ti_const_val(tt); 1994 uint64_t fv = ti_const_val(ft); 1995 TCGOpcode opc, negopc = 0; 1996 TCGCond cond = op->args[5]; 1997 1998 switch (ctx->type) { 1999 case TCG_TYPE_I32: 2000 opc = INDEX_op_setcond_i32; 2001 if (TCG_TARGET_HAS_negsetcond_i32) { 2002 negopc = INDEX_op_negsetcond_i32; 2003 } 2004 tv = (int32_t)tv; 2005 fv = (int32_t)fv; 2006 break; 2007 case TCG_TYPE_I64: 2008 opc = INDEX_op_setcond_i64; 2009 if (TCG_TARGET_HAS_negsetcond_i64) { 2010 negopc = INDEX_op_negsetcond_i64; 2011 } 2012 break; 2013 default: 2014 g_assert_not_reached(); 2015 } 2016 2017 if (tv == 1 && fv == 0) { 2018 op->opc = opc; 2019 op->args[3] = cond; 2020 } else if (fv == 1 && tv == 0) { 2021 op->opc = opc; 2022 op->args[3] = tcg_invert_cond(cond); 2023 } else if (negopc) { 2024 if (tv == -1 && fv == 0) { 2025 op->opc = negopc; 2026 op->args[3] = cond; 2027 } else if (fv == -1 && tv == 0) { 2028 op->opc = negopc; 2029 op->args[3] = tcg_invert_cond(cond); 2030 } 2031 } 2032 } 2033 2034 return fold_masks_zs(ctx, op, z_mask, s_mask); 2035 } 2036 2037 static bool fold_mul(OptContext *ctx, TCGOp *op) 2038 { 2039 if (fold_const2(ctx, op) || 2040 fold_xi_to_i(ctx, op, 0) || 2041 fold_xi_to_x(ctx, op, 1)) { 2042 return true; 2043 } 2044 return finish_folding(ctx, op); 2045 } 2046 2047 static bool fold_mul_highpart(OptContext *ctx, TCGOp *op) 2048 { 2049 if (fold_const2_commutative(ctx, op) || 2050 fold_xi_to_i(ctx, op, 0)) { 2051 return true; 2052 } 2053 return finish_folding(ctx, op); 2054 } 2055 2056 static bool fold_multiply2(OptContext *ctx, TCGOp *op) 2057 { 2058 swap_commutative(op->args[0], &op->args[2], &op->args[3]); 2059 2060 if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) { 2061 uint64_t a = arg_info(op->args[2])->val; 2062 uint64_t b = arg_info(op->args[3])->val; 2063 uint64_t h, l; 2064 TCGArg rl, rh; 2065 TCGOp *op2; 2066 2067 switch (op->opc) { 2068 case INDEX_op_mulu2_i32: 2069 l = (uint64_t)(uint32_t)a * (uint32_t)b; 2070 h = (int32_t)(l >> 32); 2071 l = (int32_t)l; 2072 break; 2073 case INDEX_op_muls2_i32: 2074 l = (int64_t)(int32_t)a * (int32_t)b; 2075 h = l >> 32; 2076 l = (int32_t)l; 2077 break; 2078 case INDEX_op_mulu2_i64: 2079 mulu64(&l, &h, a, b); 2080 break; 2081 case INDEX_op_muls2_i64: 2082 muls64(&l, &h, a, b); 2083 break; 2084 default: 2085 g_assert_not_reached(); 2086 } 2087 2088 rl = op->args[0]; 2089 rh = op->args[1]; 2090 2091 /* The proper opcode is supplied by tcg_opt_gen_mov. */ 2092 op2 = opt_insert_before(ctx, op, 0, 2); 2093 2094 tcg_opt_gen_movi(ctx, op, rl, l); 2095 tcg_opt_gen_movi(ctx, op2, rh, h); 2096 return true; 2097 } 2098 return finish_folding(ctx, op); 2099 } 2100 2101 static bool fold_nand(OptContext *ctx, TCGOp *op) 2102 { 2103 uint64_t s_mask; 2104 2105 if (fold_const2_commutative(ctx, op) || 2106 fold_xi_to_not(ctx, op, -1)) { 2107 return true; 2108 } 2109 2110 s_mask = arg_info(op->args[1])->s_mask 2111 & arg_info(op->args[2])->s_mask; 2112 return fold_masks_s(ctx, op, s_mask); 2113 } 2114 2115 static bool fold_neg_no_const(OptContext *ctx, TCGOp *op) 2116 { 2117 /* Set to 1 all bits to the left of the rightmost. */ 2118 uint64_t z_mask = arg_info(op->args[1])->z_mask; 2119 z_mask = -(z_mask & -z_mask); 2120 2121 return fold_masks_z(ctx, op, z_mask); 2122 } 2123 2124 static bool fold_neg(OptContext *ctx, TCGOp *op) 2125 { 2126 return fold_const1(ctx, op) || fold_neg_no_const(ctx, op); 2127 } 2128 2129 static bool fold_nor(OptContext *ctx, TCGOp *op) 2130 { 2131 uint64_t s_mask; 2132 2133 if (fold_const2_commutative(ctx, op) || 2134 fold_xi_to_not(ctx, op, 0)) { 2135 return true; 2136 } 2137 2138 s_mask = arg_info(op->args[1])->s_mask 2139 & arg_info(op->args[2])->s_mask; 2140 return fold_masks_s(ctx, op, s_mask); 2141 } 2142 2143 static bool fold_not(OptContext *ctx, TCGOp *op) 2144 { 2145 if (fold_const1(ctx, op)) { 2146 return true; 2147 } 2148 return fold_masks_s(ctx, op, arg_info(op->args[1])->s_mask); 2149 } 2150 2151 static bool fold_or(OptContext *ctx, TCGOp *op) 2152 { 2153 uint64_t z_mask, s_mask; 2154 TempOptInfo *t1, *t2; 2155 2156 if (fold_const2_commutative(ctx, op) || 2157 fold_xi_to_x(ctx, op, 0) || 2158 fold_xx_to_x(ctx, op)) { 2159 return true; 2160 } 2161 2162 t1 = arg_info(op->args[1]); 2163 t2 = arg_info(op->args[2]); 2164 z_mask = t1->z_mask | t2->z_mask; 2165 s_mask = t1->s_mask & t2->s_mask; 2166 return fold_masks_zs(ctx, op, z_mask, s_mask); 2167 } 2168 2169 static bool fold_orc(OptContext *ctx, TCGOp *op) 2170 { 2171 uint64_t s_mask; 2172 TempOptInfo *t1, *t2; 2173 2174 if (fold_const2(ctx, op) || 2175 fold_xx_to_i(ctx, op, -1) || 2176 fold_xi_to_x(ctx, op, -1) || 2177 fold_ix_to_not(ctx, op, 0)) { 2178 return true; 2179 } 2180 2181 t2 = arg_info(op->args[2]); 2182 if (ti_is_const(t2)) { 2183 /* Fold orc r,x,i to or r,x,~i. */ 2184 switch (ctx->type) { 2185 case TCG_TYPE_I32: 2186 case TCG_TYPE_I64: 2187 op->opc = INDEX_op_or; 2188 break; 2189 case TCG_TYPE_V64: 2190 case TCG_TYPE_V128: 2191 case TCG_TYPE_V256: 2192 op->opc = INDEX_op_or_vec; 2193 break; 2194 default: 2195 g_assert_not_reached(); 2196 } 2197 op->args[2] = arg_new_constant(ctx, ~ti_const_val(t2)); 2198 return fold_or(ctx, op); 2199 } 2200 2201 t1 = arg_info(op->args[1]); 2202 s_mask = t1->s_mask & t2->s_mask; 2203 return fold_masks_s(ctx, op, s_mask); 2204 } 2205 2206 static bool fold_qemu_ld_1reg(OptContext *ctx, TCGOp *op) 2207 { 2208 const TCGOpDef *def = &tcg_op_defs[op->opc]; 2209 MemOpIdx oi = op->args[def->nb_oargs + def->nb_iargs]; 2210 MemOp mop = get_memop(oi); 2211 int width = 8 * memop_size(mop); 2212 uint64_t z_mask = -1, s_mask = 0; 2213 2214 if (width < 64) { 2215 if (mop & MO_SIGN) { 2216 s_mask = MAKE_64BIT_MASK(width - 1, 64 - (width - 1)); 2217 } else { 2218 z_mask = MAKE_64BIT_MASK(0, width); 2219 } 2220 } 2221 2222 /* Opcodes that touch guest memory stop the mb optimization. */ 2223 ctx->prev_mb = NULL; 2224 2225 return fold_masks_zs(ctx, op, z_mask, s_mask); 2226 } 2227 2228 static bool fold_qemu_ld_2reg(OptContext *ctx, TCGOp *op) 2229 { 2230 /* Opcodes that touch guest memory stop the mb optimization. */ 2231 ctx->prev_mb = NULL; 2232 return finish_folding(ctx, op); 2233 } 2234 2235 static bool fold_qemu_st(OptContext *ctx, TCGOp *op) 2236 { 2237 /* Opcodes that touch guest memory stop the mb optimization. */ 2238 ctx->prev_mb = NULL; 2239 return true; 2240 } 2241 2242 static bool fold_remainder(OptContext *ctx, TCGOp *op) 2243 { 2244 if (fold_const2(ctx, op) || 2245 fold_xx_to_i(ctx, op, 0)) { 2246 return true; 2247 } 2248 return finish_folding(ctx, op); 2249 } 2250 2251 /* Return 1 if finished, -1 if simplified, 0 if unchanged. */ 2252 static int fold_setcond_zmask(OptContext *ctx, TCGOp *op, bool neg) 2253 { 2254 uint64_t a_zmask, b_val; 2255 TCGCond cond; 2256 2257 if (!arg_is_const(op->args[2])) { 2258 return false; 2259 } 2260 2261 a_zmask = arg_info(op->args[1])->z_mask; 2262 b_val = arg_info(op->args[2])->val; 2263 cond = op->args[3]; 2264 2265 if (ctx->type == TCG_TYPE_I32) { 2266 a_zmask = (uint32_t)a_zmask; 2267 b_val = (uint32_t)b_val; 2268 } 2269 2270 /* 2271 * A with only low bits set vs B with high bits set means that A < B. 2272 */ 2273 if (a_zmask < b_val) { 2274 bool inv = false; 2275 2276 switch (cond) { 2277 case TCG_COND_NE: 2278 case TCG_COND_LEU: 2279 case TCG_COND_LTU: 2280 inv = true; 2281 /* fall through */ 2282 case TCG_COND_GTU: 2283 case TCG_COND_GEU: 2284 case TCG_COND_EQ: 2285 return tcg_opt_gen_movi(ctx, op, op->args[0], neg ? -inv : inv); 2286 default: 2287 break; 2288 } 2289 } 2290 2291 /* 2292 * A with only lsb set is already boolean. 2293 */ 2294 if (a_zmask <= 1) { 2295 bool convert = false; 2296 bool inv = false; 2297 2298 switch (cond) { 2299 case TCG_COND_EQ: 2300 inv = true; 2301 /* fall through */ 2302 case TCG_COND_NE: 2303 convert = (b_val == 0); 2304 break; 2305 case TCG_COND_LTU: 2306 case TCG_COND_TSTEQ: 2307 inv = true; 2308 /* fall through */ 2309 case TCG_COND_GEU: 2310 case TCG_COND_TSTNE: 2311 convert = (b_val == 1); 2312 break; 2313 default: 2314 break; 2315 } 2316 if (convert) { 2317 TCGOpcode neg_opc; 2318 2319 if (!inv && !neg) { 2320 return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); 2321 } 2322 2323 switch (ctx->type) { 2324 case TCG_TYPE_I32: 2325 neg_opc = INDEX_op_neg_i32; 2326 break; 2327 case TCG_TYPE_I64: 2328 neg_opc = INDEX_op_neg_i64; 2329 break; 2330 default: 2331 g_assert_not_reached(); 2332 } 2333 2334 if (!inv) { 2335 op->opc = neg_opc; 2336 } else if (neg) { 2337 op->opc = INDEX_op_add; 2338 op->args[2] = arg_new_constant(ctx, -1); 2339 } else { 2340 op->opc = INDEX_op_xor; 2341 op->args[2] = arg_new_constant(ctx, 1); 2342 } 2343 return -1; 2344 } 2345 } 2346 return 0; 2347 } 2348 2349 static void fold_setcond_tst_pow2(OptContext *ctx, TCGOp *op, bool neg) 2350 { 2351 TCGOpcode neg_opc, shr_opc; 2352 TCGOpcode uext_opc = 0, sext_opc = 0; 2353 TCGCond cond = op->args[3]; 2354 TCGArg ret, src1, src2; 2355 TCGOp *op2; 2356 uint64_t val; 2357 int sh; 2358 bool inv; 2359 2360 if (!is_tst_cond(cond) || !arg_is_const(op->args[2])) { 2361 return; 2362 } 2363 2364 src2 = op->args[2]; 2365 val = arg_info(src2)->val; 2366 if (!is_power_of_2(val)) { 2367 return; 2368 } 2369 sh = ctz64(val); 2370 2371 switch (ctx->type) { 2372 case TCG_TYPE_I32: 2373 shr_opc = INDEX_op_shr_i32; 2374 neg_opc = INDEX_op_neg_i32; 2375 if (TCG_TARGET_extract_valid(TCG_TYPE_I32, sh, 1)) { 2376 uext_opc = INDEX_op_extract_i32; 2377 } 2378 if (TCG_TARGET_sextract_valid(TCG_TYPE_I32, sh, 1)) { 2379 sext_opc = INDEX_op_sextract_i32; 2380 } 2381 break; 2382 case TCG_TYPE_I64: 2383 shr_opc = INDEX_op_shr_i64; 2384 neg_opc = INDEX_op_neg_i64; 2385 if (TCG_TARGET_extract_valid(TCG_TYPE_I64, sh, 1)) { 2386 uext_opc = INDEX_op_extract_i64; 2387 } 2388 if (TCG_TARGET_sextract_valid(TCG_TYPE_I64, sh, 1)) { 2389 sext_opc = INDEX_op_sextract_i64; 2390 } 2391 break; 2392 default: 2393 g_assert_not_reached(); 2394 } 2395 2396 ret = op->args[0]; 2397 src1 = op->args[1]; 2398 inv = cond == TCG_COND_TSTEQ; 2399 2400 if (sh && sext_opc && neg && !inv) { 2401 op->opc = sext_opc; 2402 op->args[1] = src1; 2403 op->args[2] = sh; 2404 op->args[3] = 1; 2405 return; 2406 } else if (sh && uext_opc) { 2407 op->opc = uext_opc; 2408 op->args[1] = src1; 2409 op->args[2] = sh; 2410 op->args[3] = 1; 2411 } else { 2412 if (sh) { 2413 op2 = opt_insert_before(ctx, op, shr_opc, 3); 2414 op2->args[0] = ret; 2415 op2->args[1] = src1; 2416 op2->args[2] = arg_new_constant(ctx, sh); 2417 src1 = ret; 2418 } 2419 op->opc = INDEX_op_and; 2420 op->args[1] = src1; 2421 op->args[2] = arg_new_constant(ctx, 1); 2422 } 2423 2424 if (neg && inv) { 2425 op2 = opt_insert_after(ctx, op, INDEX_op_add, 3); 2426 op2->args[0] = ret; 2427 op2->args[1] = ret; 2428 op2->args[2] = arg_new_constant(ctx, -1); 2429 } else if (inv) { 2430 op2 = opt_insert_after(ctx, op, INDEX_op_xor, 3); 2431 op2->args[0] = ret; 2432 op2->args[1] = ret; 2433 op2->args[2] = arg_new_constant(ctx, 1); 2434 } else if (neg) { 2435 op2 = opt_insert_after(ctx, op, neg_opc, 2); 2436 op2->args[0] = ret; 2437 op2->args[1] = ret; 2438 } 2439 } 2440 2441 static bool fold_setcond(OptContext *ctx, TCGOp *op) 2442 { 2443 int i = do_constant_folding_cond1(ctx, op, op->args[0], &op->args[1], 2444 &op->args[2], &op->args[3]); 2445 if (i >= 0) { 2446 return tcg_opt_gen_movi(ctx, op, op->args[0], i); 2447 } 2448 2449 i = fold_setcond_zmask(ctx, op, false); 2450 if (i > 0) { 2451 return true; 2452 } 2453 if (i == 0) { 2454 fold_setcond_tst_pow2(ctx, op, false); 2455 } 2456 2457 return fold_masks_z(ctx, op, 1); 2458 } 2459 2460 static bool fold_negsetcond(OptContext *ctx, TCGOp *op) 2461 { 2462 int i = do_constant_folding_cond1(ctx, op, op->args[0], &op->args[1], 2463 &op->args[2], &op->args[3]); 2464 if (i >= 0) { 2465 return tcg_opt_gen_movi(ctx, op, op->args[0], -i); 2466 } 2467 2468 i = fold_setcond_zmask(ctx, op, true); 2469 if (i > 0) { 2470 return true; 2471 } 2472 if (i == 0) { 2473 fold_setcond_tst_pow2(ctx, op, true); 2474 } 2475 2476 /* Value is {0,-1} so all bits are repetitions of the sign. */ 2477 return fold_masks_s(ctx, op, -1); 2478 } 2479 2480 static bool fold_setcond2(OptContext *ctx, TCGOp *op) 2481 { 2482 TCGCond cond; 2483 int i, inv = 0; 2484 2485 i = do_constant_folding_cond2(ctx, op, &op->args[1]); 2486 cond = op->args[5]; 2487 if (i >= 0) { 2488 goto do_setcond_const; 2489 } 2490 2491 switch (cond) { 2492 case TCG_COND_LT: 2493 case TCG_COND_GE: 2494 /* 2495 * Simplify LT/GE comparisons vs zero to a single compare 2496 * vs the high word of the input. 2497 */ 2498 if (arg_is_const_val(op->args[3], 0) && 2499 arg_is_const_val(op->args[4], 0)) { 2500 goto do_setcond_high; 2501 } 2502 break; 2503 2504 case TCG_COND_NE: 2505 inv = 1; 2506 QEMU_FALLTHROUGH; 2507 case TCG_COND_EQ: 2508 /* 2509 * Simplify EQ/NE comparisons where one of the pairs 2510 * can be simplified. 2511 */ 2512 i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1], 2513 op->args[3], cond); 2514 switch (i ^ inv) { 2515 case 0: 2516 goto do_setcond_const; 2517 case 1: 2518 goto do_setcond_high; 2519 } 2520 2521 i = do_constant_folding_cond(TCG_TYPE_I32, op->args[2], 2522 op->args[4], cond); 2523 switch (i ^ inv) { 2524 case 0: 2525 goto do_setcond_const; 2526 case 1: 2527 goto do_setcond_low; 2528 } 2529 break; 2530 2531 case TCG_COND_TSTEQ: 2532 case TCG_COND_TSTNE: 2533 if (arg_is_const_val(op->args[3], 0)) { 2534 goto do_setcond_high; 2535 } 2536 if (arg_is_const_val(op->args[4], 0)) { 2537 goto do_setcond_low; 2538 } 2539 break; 2540 2541 default: 2542 break; 2543 2544 do_setcond_low: 2545 op->args[2] = op->args[3]; 2546 op->args[3] = cond; 2547 op->opc = INDEX_op_setcond_i32; 2548 return fold_setcond(ctx, op); 2549 2550 do_setcond_high: 2551 op->args[1] = op->args[2]; 2552 op->args[2] = op->args[4]; 2553 op->args[3] = cond; 2554 op->opc = INDEX_op_setcond_i32; 2555 return fold_setcond(ctx, op); 2556 } 2557 2558 return fold_masks_z(ctx, op, 1); 2559 2560 do_setcond_const: 2561 return tcg_opt_gen_movi(ctx, op, op->args[0], i); 2562 } 2563 2564 static bool fold_sextract(OptContext *ctx, TCGOp *op) 2565 { 2566 uint64_t z_mask, s_mask, s_mask_old; 2567 TempOptInfo *t1 = arg_info(op->args[1]); 2568 int pos = op->args[2]; 2569 int len = op->args[3]; 2570 2571 if (ti_is_const(t1)) { 2572 return tcg_opt_gen_movi(ctx, op, op->args[0], 2573 sextract64(ti_const_val(t1), pos, len)); 2574 } 2575 2576 s_mask_old = t1->s_mask; 2577 s_mask = s_mask_old >> pos; 2578 s_mask |= -1ull << (len - 1); 2579 2580 if (pos == 0 && fold_affected_mask(ctx, op, s_mask & ~s_mask_old)) { 2581 return true; 2582 } 2583 2584 z_mask = sextract64(t1->z_mask, pos, len); 2585 return fold_masks_zs(ctx, op, z_mask, s_mask); 2586 } 2587 2588 static bool fold_shift(OptContext *ctx, TCGOp *op) 2589 { 2590 uint64_t s_mask, z_mask; 2591 TempOptInfo *t1, *t2; 2592 2593 if (fold_const2(ctx, op) || 2594 fold_ix_to_i(ctx, op, 0) || 2595 fold_xi_to_x(ctx, op, 0)) { 2596 return true; 2597 } 2598 2599 t1 = arg_info(op->args[1]); 2600 t2 = arg_info(op->args[2]); 2601 s_mask = t1->s_mask; 2602 z_mask = t1->z_mask; 2603 2604 if (ti_is_const(t2)) { 2605 int sh = ti_const_val(t2); 2606 2607 z_mask = do_constant_folding(op->opc, ctx->type, z_mask, sh); 2608 s_mask = do_constant_folding(op->opc, ctx->type, s_mask, sh); 2609 2610 return fold_masks_zs(ctx, op, z_mask, s_mask); 2611 } 2612 2613 switch (op->opc) { 2614 CASE_OP_32_64(sar): 2615 /* 2616 * Arithmetic right shift will not reduce the number of 2617 * input sign repetitions. 2618 */ 2619 return fold_masks_s(ctx, op, s_mask); 2620 CASE_OP_32_64(shr): 2621 /* 2622 * If the sign bit is known zero, then logical right shift 2623 * will not reduce the number of input sign repetitions. 2624 */ 2625 if (~z_mask & -s_mask) { 2626 return fold_masks_s(ctx, op, s_mask); 2627 } 2628 break; 2629 default: 2630 break; 2631 } 2632 2633 return finish_folding(ctx, op); 2634 } 2635 2636 static bool fold_sub_to_neg(OptContext *ctx, TCGOp *op) 2637 { 2638 TCGOpcode neg_op; 2639 bool have_neg; 2640 2641 if (!arg_is_const(op->args[1]) || arg_info(op->args[1])->val != 0) { 2642 return false; 2643 } 2644 2645 switch (ctx->type) { 2646 case TCG_TYPE_I32: 2647 neg_op = INDEX_op_neg_i32; 2648 have_neg = true; 2649 break; 2650 case TCG_TYPE_I64: 2651 neg_op = INDEX_op_neg_i64; 2652 have_neg = true; 2653 break; 2654 case TCG_TYPE_V64: 2655 case TCG_TYPE_V128: 2656 case TCG_TYPE_V256: 2657 neg_op = INDEX_op_neg_vec; 2658 have_neg = (TCG_TARGET_HAS_neg_vec && 2659 tcg_can_emit_vec_op(neg_op, ctx->type, TCGOP_VECE(op)) > 0); 2660 break; 2661 default: 2662 g_assert_not_reached(); 2663 } 2664 if (have_neg) { 2665 op->opc = neg_op; 2666 op->args[1] = op->args[2]; 2667 return fold_neg_no_const(ctx, op); 2668 } 2669 return false; 2670 } 2671 2672 /* We cannot as yet do_constant_folding with vectors. */ 2673 static bool fold_sub_vec(OptContext *ctx, TCGOp *op) 2674 { 2675 if (fold_xx_to_i(ctx, op, 0) || 2676 fold_xi_to_x(ctx, op, 0) || 2677 fold_sub_to_neg(ctx, op)) { 2678 return true; 2679 } 2680 return finish_folding(ctx, op); 2681 } 2682 2683 static bool fold_sub(OptContext *ctx, TCGOp *op) 2684 { 2685 if (fold_const2(ctx, op) || 2686 fold_xx_to_i(ctx, op, 0) || 2687 fold_xi_to_x(ctx, op, 0) || 2688 fold_sub_to_neg(ctx, op)) { 2689 return true; 2690 } 2691 2692 /* Fold sub r,x,i to add r,x,-i */ 2693 if (arg_is_const(op->args[2])) { 2694 uint64_t val = arg_info(op->args[2])->val; 2695 2696 op->opc = INDEX_op_add; 2697 op->args[2] = arg_new_constant(ctx, -val); 2698 } 2699 return finish_folding(ctx, op); 2700 } 2701 2702 static bool fold_sub2(OptContext *ctx, TCGOp *op) 2703 { 2704 return fold_addsub2(ctx, op, false); 2705 } 2706 2707 static bool fold_tcg_ld(OptContext *ctx, TCGOp *op) 2708 { 2709 uint64_t z_mask = -1, s_mask = 0; 2710 2711 /* We can't do any folding with a load, but we can record bits. */ 2712 switch (op->opc) { 2713 CASE_OP_32_64(ld8s): 2714 s_mask = INT8_MIN; 2715 break; 2716 CASE_OP_32_64(ld8u): 2717 z_mask = MAKE_64BIT_MASK(0, 8); 2718 break; 2719 CASE_OP_32_64(ld16s): 2720 s_mask = INT16_MIN; 2721 break; 2722 CASE_OP_32_64(ld16u): 2723 z_mask = MAKE_64BIT_MASK(0, 16); 2724 break; 2725 case INDEX_op_ld32s_i64: 2726 s_mask = INT32_MIN; 2727 break; 2728 case INDEX_op_ld32u_i64: 2729 z_mask = MAKE_64BIT_MASK(0, 32); 2730 break; 2731 default: 2732 g_assert_not_reached(); 2733 } 2734 return fold_masks_zs(ctx, op, z_mask, s_mask); 2735 } 2736 2737 static bool fold_tcg_ld_memcopy(OptContext *ctx, TCGOp *op) 2738 { 2739 TCGTemp *dst, *src; 2740 intptr_t ofs; 2741 TCGType type; 2742 2743 if (op->args[1] != tcgv_ptr_arg(tcg_env)) { 2744 return finish_folding(ctx, op); 2745 } 2746 2747 type = ctx->type; 2748 ofs = op->args[2]; 2749 dst = arg_temp(op->args[0]); 2750 src = find_mem_copy_for(ctx, type, ofs); 2751 if (src && src->base_type == type) { 2752 return tcg_opt_gen_mov(ctx, op, temp_arg(dst), temp_arg(src)); 2753 } 2754 2755 reset_ts(ctx, dst); 2756 record_mem_copy(ctx, type, dst, ofs, ofs + tcg_type_size(type) - 1); 2757 return true; 2758 } 2759 2760 static bool fold_tcg_st(OptContext *ctx, TCGOp *op) 2761 { 2762 intptr_t ofs = op->args[2]; 2763 intptr_t lm1; 2764 2765 if (op->args[1] != tcgv_ptr_arg(tcg_env)) { 2766 remove_mem_copy_all(ctx); 2767 return true; 2768 } 2769 2770 switch (op->opc) { 2771 CASE_OP_32_64(st8): 2772 lm1 = 0; 2773 break; 2774 CASE_OP_32_64(st16): 2775 lm1 = 1; 2776 break; 2777 case INDEX_op_st32_i64: 2778 case INDEX_op_st_i32: 2779 lm1 = 3; 2780 break; 2781 case INDEX_op_st_i64: 2782 lm1 = 7; 2783 break; 2784 case INDEX_op_st_vec: 2785 lm1 = tcg_type_size(ctx->type) - 1; 2786 break; 2787 default: 2788 g_assert_not_reached(); 2789 } 2790 remove_mem_copy_in(ctx, ofs, ofs + lm1); 2791 return true; 2792 } 2793 2794 static bool fold_tcg_st_memcopy(OptContext *ctx, TCGOp *op) 2795 { 2796 TCGTemp *src; 2797 intptr_t ofs, last; 2798 TCGType type; 2799 2800 if (op->args[1] != tcgv_ptr_arg(tcg_env)) { 2801 return fold_tcg_st(ctx, op); 2802 } 2803 2804 src = arg_temp(op->args[0]); 2805 ofs = op->args[2]; 2806 type = ctx->type; 2807 2808 /* 2809 * Eliminate duplicate stores of a constant. 2810 * This happens frequently when the target ISA zero-extends. 2811 */ 2812 if (ts_is_const(src)) { 2813 TCGTemp *prev = find_mem_copy_for(ctx, type, ofs); 2814 if (src == prev) { 2815 tcg_op_remove(ctx->tcg, op); 2816 return true; 2817 } 2818 } 2819 2820 last = ofs + tcg_type_size(type) - 1; 2821 remove_mem_copy_in(ctx, ofs, last); 2822 record_mem_copy(ctx, type, src, ofs, last); 2823 return true; 2824 } 2825 2826 static bool fold_xor(OptContext *ctx, TCGOp *op) 2827 { 2828 uint64_t z_mask, s_mask; 2829 TempOptInfo *t1, *t2; 2830 2831 if (fold_const2_commutative(ctx, op) || 2832 fold_xx_to_i(ctx, op, 0) || 2833 fold_xi_to_x(ctx, op, 0) || 2834 fold_xi_to_not(ctx, op, -1)) { 2835 return true; 2836 } 2837 2838 t1 = arg_info(op->args[1]); 2839 t2 = arg_info(op->args[2]); 2840 z_mask = t1->z_mask | t2->z_mask; 2841 s_mask = t1->s_mask & t2->s_mask; 2842 return fold_masks_zs(ctx, op, z_mask, s_mask); 2843 } 2844 2845 /* Propagate constants and copies, fold constant expressions. */ 2846 void tcg_optimize(TCGContext *s) 2847 { 2848 int nb_temps, i; 2849 TCGOp *op, *op_next; 2850 OptContext ctx = { .tcg = s }; 2851 2852 QSIMPLEQ_INIT(&ctx.mem_free); 2853 2854 /* Array VALS has an element for each temp. 2855 If this temp holds a constant then its value is kept in VALS' element. 2856 If this temp is a copy of other ones then the other copies are 2857 available through the doubly linked circular list. */ 2858 2859 nb_temps = s->nb_temps; 2860 for (i = 0; i < nb_temps; ++i) { 2861 s->temps[i].state_ptr = NULL; 2862 } 2863 2864 QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) { 2865 TCGOpcode opc = op->opc; 2866 const TCGOpDef *def; 2867 bool done = false; 2868 2869 /* Calls are special. */ 2870 if (opc == INDEX_op_call) { 2871 fold_call(&ctx, op); 2872 continue; 2873 } 2874 2875 def = &tcg_op_defs[opc]; 2876 init_arguments(&ctx, op, def->nb_oargs + def->nb_iargs); 2877 copy_propagate(&ctx, op, def->nb_oargs, def->nb_iargs); 2878 2879 /* Pre-compute the type of the operation. */ 2880 ctx.type = TCGOP_TYPE(op); 2881 2882 /* 2883 * Process each opcode. 2884 * Sorted alphabetically by opcode as much as possible. 2885 */ 2886 switch (opc) { 2887 case INDEX_op_add: 2888 done = fold_add(&ctx, op); 2889 break; 2890 case INDEX_op_add_vec: 2891 done = fold_add_vec(&ctx, op); 2892 break; 2893 CASE_OP_32_64(add2): 2894 done = fold_add2(&ctx, op); 2895 break; 2896 case INDEX_op_and: 2897 case INDEX_op_and_vec: 2898 done = fold_and(&ctx, op); 2899 break; 2900 case INDEX_op_andc: 2901 case INDEX_op_andc_vec: 2902 done = fold_andc(&ctx, op); 2903 break; 2904 CASE_OP_32_64(brcond): 2905 done = fold_brcond(&ctx, op); 2906 break; 2907 case INDEX_op_brcond2_i32: 2908 done = fold_brcond2(&ctx, op); 2909 break; 2910 CASE_OP_32_64(bswap16): 2911 CASE_OP_32_64(bswap32): 2912 case INDEX_op_bswap64_i64: 2913 done = fold_bswap(&ctx, op); 2914 break; 2915 CASE_OP_32_64(clz): 2916 CASE_OP_32_64(ctz): 2917 done = fold_count_zeros(&ctx, op); 2918 break; 2919 CASE_OP_32_64(ctpop): 2920 done = fold_ctpop(&ctx, op); 2921 break; 2922 CASE_OP_32_64(deposit): 2923 done = fold_deposit(&ctx, op); 2924 break; 2925 CASE_OP_32_64(div): 2926 CASE_OP_32_64(divu): 2927 done = fold_divide(&ctx, op); 2928 break; 2929 case INDEX_op_dup_vec: 2930 done = fold_dup(&ctx, op); 2931 break; 2932 case INDEX_op_dup2_vec: 2933 done = fold_dup2(&ctx, op); 2934 break; 2935 case INDEX_op_eqv: 2936 case INDEX_op_eqv_vec: 2937 done = fold_eqv(&ctx, op); 2938 break; 2939 CASE_OP_32_64(extract): 2940 done = fold_extract(&ctx, op); 2941 break; 2942 CASE_OP_32_64(extract2): 2943 done = fold_extract2(&ctx, op); 2944 break; 2945 case INDEX_op_ext_i32_i64: 2946 done = fold_exts(&ctx, op); 2947 break; 2948 case INDEX_op_extu_i32_i64: 2949 case INDEX_op_extrl_i64_i32: 2950 case INDEX_op_extrh_i64_i32: 2951 done = fold_extu(&ctx, op); 2952 break; 2953 CASE_OP_32_64(ld8s): 2954 CASE_OP_32_64(ld8u): 2955 CASE_OP_32_64(ld16s): 2956 CASE_OP_32_64(ld16u): 2957 case INDEX_op_ld32s_i64: 2958 case INDEX_op_ld32u_i64: 2959 done = fold_tcg_ld(&ctx, op); 2960 break; 2961 case INDEX_op_ld_i32: 2962 case INDEX_op_ld_i64: 2963 case INDEX_op_ld_vec: 2964 done = fold_tcg_ld_memcopy(&ctx, op); 2965 break; 2966 CASE_OP_32_64(st8): 2967 CASE_OP_32_64(st16): 2968 case INDEX_op_st32_i64: 2969 done = fold_tcg_st(&ctx, op); 2970 break; 2971 case INDEX_op_st_i32: 2972 case INDEX_op_st_i64: 2973 case INDEX_op_st_vec: 2974 done = fold_tcg_st_memcopy(&ctx, op); 2975 break; 2976 case INDEX_op_mb: 2977 done = fold_mb(&ctx, op); 2978 break; 2979 case INDEX_op_mov: 2980 case INDEX_op_mov_vec: 2981 done = fold_mov(&ctx, op); 2982 break; 2983 CASE_OP_32_64(movcond): 2984 done = fold_movcond(&ctx, op); 2985 break; 2986 CASE_OP_32_64(mul): 2987 done = fold_mul(&ctx, op); 2988 break; 2989 CASE_OP_32_64(mulsh): 2990 CASE_OP_32_64(muluh): 2991 done = fold_mul_highpart(&ctx, op); 2992 break; 2993 CASE_OP_32_64(muls2): 2994 CASE_OP_32_64(mulu2): 2995 done = fold_multiply2(&ctx, op); 2996 break; 2997 case INDEX_op_nand: 2998 case INDEX_op_nand_vec: 2999 done = fold_nand(&ctx, op); 3000 break; 3001 CASE_OP_32_64(neg): 3002 done = fold_neg(&ctx, op); 3003 break; 3004 case INDEX_op_nor: 3005 case INDEX_op_nor_vec: 3006 done = fold_nor(&ctx, op); 3007 break; 3008 CASE_OP_32_64_VEC(not): 3009 done = fold_not(&ctx, op); 3010 break; 3011 case INDEX_op_or: 3012 case INDEX_op_or_vec: 3013 done = fold_or(&ctx, op); 3014 break; 3015 case INDEX_op_orc: 3016 case INDEX_op_orc_vec: 3017 done = fold_orc(&ctx, op); 3018 break; 3019 case INDEX_op_qemu_ld_i32: 3020 done = fold_qemu_ld_1reg(&ctx, op); 3021 break; 3022 case INDEX_op_qemu_ld_i64: 3023 if (TCG_TARGET_REG_BITS == 64) { 3024 done = fold_qemu_ld_1reg(&ctx, op); 3025 break; 3026 } 3027 QEMU_FALLTHROUGH; 3028 case INDEX_op_qemu_ld_i128: 3029 done = fold_qemu_ld_2reg(&ctx, op); 3030 break; 3031 case INDEX_op_qemu_st8_i32: 3032 case INDEX_op_qemu_st_i32: 3033 case INDEX_op_qemu_st_i64: 3034 case INDEX_op_qemu_st_i128: 3035 done = fold_qemu_st(&ctx, op); 3036 break; 3037 CASE_OP_32_64(rem): 3038 CASE_OP_32_64(remu): 3039 done = fold_remainder(&ctx, op); 3040 break; 3041 CASE_OP_32_64(rotl): 3042 CASE_OP_32_64(rotr): 3043 CASE_OP_32_64(sar): 3044 CASE_OP_32_64(shl): 3045 CASE_OP_32_64(shr): 3046 done = fold_shift(&ctx, op); 3047 break; 3048 CASE_OP_32_64(setcond): 3049 done = fold_setcond(&ctx, op); 3050 break; 3051 CASE_OP_32_64(negsetcond): 3052 done = fold_negsetcond(&ctx, op); 3053 break; 3054 case INDEX_op_setcond2_i32: 3055 done = fold_setcond2(&ctx, op); 3056 break; 3057 case INDEX_op_cmp_vec: 3058 done = fold_cmp_vec(&ctx, op); 3059 break; 3060 case INDEX_op_cmpsel_vec: 3061 done = fold_cmpsel_vec(&ctx, op); 3062 break; 3063 case INDEX_op_bitsel_vec: 3064 done = fold_bitsel_vec(&ctx, op); 3065 break; 3066 CASE_OP_32_64(sextract): 3067 done = fold_sextract(&ctx, op); 3068 break; 3069 CASE_OP_32_64(sub): 3070 done = fold_sub(&ctx, op); 3071 break; 3072 case INDEX_op_sub_vec: 3073 done = fold_sub_vec(&ctx, op); 3074 break; 3075 CASE_OP_32_64(sub2): 3076 done = fold_sub2(&ctx, op); 3077 break; 3078 case INDEX_op_xor: 3079 case INDEX_op_xor_vec: 3080 done = fold_xor(&ctx, op); 3081 break; 3082 case INDEX_op_set_label: 3083 case INDEX_op_br: 3084 case INDEX_op_exit_tb: 3085 case INDEX_op_goto_tb: 3086 case INDEX_op_goto_ptr: 3087 finish_ebb(&ctx); 3088 done = true; 3089 break; 3090 default: 3091 done = finish_folding(&ctx, op); 3092 break; 3093 } 3094 tcg_debug_assert(done); 3095 } 3096 } 3097