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 "config.h" 27 28 #include <stdlib.h> 29 #include <stdio.h> 30 31 #include "qemu-common.h" 32 #include "tcg-op.h" 33 34 #define CASE_OP_32_64(x) \ 35 glue(glue(case INDEX_op_, x), _i32): \ 36 glue(glue(case INDEX_op_, x), _i64) 37 38 typedef enum { 39 TCG_TEMP_UNDEF = 0, 40 TCG_TEMP_CONST, 41 TCG_TEMP_COPY, 42 } tcg_temp_state; 43 44 struct tcg_temp_info { 45 tcg_temp_state state; 46 uint16_t prev_copy; 47 uint16_t next_copy; 48 tcg_target_ulong val; 49 tcg_target_ulong mask; 50 }; 51 52 static struct tcg_temp_info temps[TCG_MAX_TEMPS]; 53 54 /* Reset TEMP's state to TCG_TEMP_UNDEF. If TEMP only had one copy, remove 55 the copy flag from the left temp. */ 56 static void reset_temp(TCGArg temp) 57 { 58 if (temps[temp].state == TCG_TEMP_COPY) { 59 if (temps[temp].prev_copy == temps[temp].next_copy) { 60 temps[temps[temp].next_copy].state = TCG_TEMP_UNDEF; 61 } else { 62 temps[temps[temp].next_copy].prev_copy = temps[temp].prev_copy; 63 temps[temps[temp].prev_copy].next_copy = temps[temp].next_copy; 64 } 65 } 66 temps[temp].state = TCG_TEMP_UNDEF; 67 temps[temp].mask = -1; 68 } 69 70 /* Reset all temporaries, given that there are NB_TEMPS of them. */ 71 static void reset_all_temps(int nb_temps) 72 { 73 int i; 74 for (i = 0; i < nb_temps; i++) { 75 temps[i].state = TCG_TEMP_UNDEF; 76 temps[i].mask = -1; 77 } 78 } 79 80 static int op_bits(TCGOpcode op) 81 { 82 const TCGOpDef *def = &tcg_op_defs[op]; 83 return def->flags & TCG_OPF_64BIT ? 64 : 32; 84 } 85 86 static TCGOpcode op_to_movi(TCGOpcode op) 87 { 88 switch (op_bits(op)) { 89 case 32: 90 return INDEX_op_movi_i32; 91 case 64: 92 return INDEX_op_movi_i64; 93 default: 94 fprintf(stderr, "op_to_movi: unexpected return value of " 95 "function op_bits.\n"); 96 tcg_abort(); 97 } 98 } 99 100 static TCGArg find_better_copy(TCGContext *s, TCGArg temp) 101 { 102 TCGArg i; 103 104 /* If this is already a global, we can't do better. */ 105 if (temp < s->nb_globals) { 106 return temp; 107 } 108 109 /* Search for a global first. */ 110 for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) { 111 if (i < s->nb_globals) { 112 return i; 113 } 114 } 115 116 /* If it is a temp, search for a temp local. */ 117 if (!s->temps[temp].temp_local) { 118 for (i = temps[temp].next_copy ; i != temp ; i = temps[i].next_copy) { 119 if (s->temps[i].temp_local) { 120 return i; 121 } 122 } 123 } 124 125 /* Failure to find a better representation, return the same temp. */ 126 return temp; 127 } 128 129 static bool temps_are_copies(TCGArg arg1, TCGArg arg2) 130 { 131 TCGArg i; 132 133 if (arg1 == arg2) { 134 return true; 135 } 136 137 if (temps[arg1].state != TCG_TEMP_COPY 138 || temps[arg2].state != TCG_TEMP_COPY) { 139 return false; 140 } 141 142 for (i = temps[arg1].next_copy ; i != arg1 ; i = temps[i].next_copy) { 143 if (i == arg2) { 144 return true; 145 } 146 } 147 148 return false; 149 } 150 151 static void tcg_opt_gen_mov(TCGContext *s, TCGArg *gen_args, 152 TCGArg dst, TCGArg src) 153 { 154 reset_temp(dst); 155 temps[dst].mask = temps[src].mask; 156 assert(temps[src].state != TCG_TEMP_CONST); 157 158 if (s->temps[src].type == s->temps[dst].type) { 159 if (temps[src].state != TCG_TEMP_COPY) { 160 temps[src].state = TCG_TEMP_COPY; 161 temps[src].next_copy = src; 162 temps[src].prev_copy = src; 163 } 164 temps[dst].state = TCG_TEMP_COPY; 165 temps[dst].next_copy = temps[src].next_copy; 166 temps[dst].prev_copy = src; 167 temps[temps[dst].next_copy].prev_copy = dst; 168 temps[src].next_copy = dst; 169 } 170 171 gen_args[0] = dst; 172 gen_args[1] = src; 173 } 174 175 static void tcg_opt_gen_movi(TCGArg *gen_args, TCGArg dst, TCGArg val) 176 { 177 reset_temp(dst); 178 temps[dst].state = TCG_TEMP_CONST; 179 temps[dst].val = val; 180 temps[dst].mask = val; 181 gen_args[0] = dst; 182 gen_args[1] = val; 183 } 184 185 static TCGOpcode op_to_mov(TCGOpcode op) 186 { 187 switch (op_bits(op)) { 188 case 32: 189 return INDEX_op_mov_i32; 190 case 64: 191 return INDEX_op_mov_i64; 192 default: 193 fprintf(stderr, "op_to_mov: unexpected return value of " 194 "function op_bits.\n"); 195 tcg_abort(); 196 } 197 } 198 199 static TCGArg do_constant_folding_2(TCGOpcode op, TCGArg x, TCGArg y) 200 { 201 uint64_t l64, h64; 202 203 switch (op) { 204 CASE_OP_32_64(add): 205 return x + y; 206 207 CASE_OP_32_64(sub): 208 return x - y; 209 210 CASE_OP_32_64(mul): 211 return x * y; 212 213 CASE_OP_32_64(and): 214 return x & y; 215 216 CASE_OP_32_64(or): 217 return x | y; 218 219 CASE_OP_32_64(xor): 220 return x ^ y; 221 222 case INDEX_op_shl_i32: 223 return (uint32_t)x << (uint32_t)y; 224 225 case INDEX_op_shl_i64: 226 return (uint64_t)x << (uint64_t)y; 227 228 case INDEX_op_shr_i32: 229 return (uint32_t)x >> (uint32_t)y; 230 231 case INDEX_op_shr_i64: 232 return (uint64_t)x >> (uint64_t)y; 233 234 case INDEX_op_sar_i32: 235 return (int32_t)x >> (int32_t)y; 236 237 case INDEX_op_sar_i64: 238 return (int64_t)x >> (int64_t)y; 239 240 case INDEX_op_rotr_i32: 241 return ror32(x, y); 242 243 case INDEX_op_rotr_i64: 244 return ror64(x, y); 245 246 case INDEX_op_rotl_i32: 247 return rol32(x, y); 248 249 case INDEX_op_rotl_i64: 250 return rol64(x, y); 251 252 CASE_OP_32_64(not): 253 return ~x; 254 255 CASE_OP_32_64(neg): 256 return -x; 257 258 CASE_OP_32_64(andc): 259 return x & ~y; 260 261 CASE_OP_32_64(orc): 262 return x | ~y; 263 264 CASE_OP_32_64(eqv): 265 return ~(x ^ y); 266 267 CASE_OP_32_64(nand): 268 return ~(x & y); 269 270 CASE_OP_32_64(nor): 271 return ~(x | y); 272 273 CASE_OP_32_64(ext8s): 274 return (int8_t)x; 275 276 CASE_OP_32_64(ext16s): 277 return (int16_t)x; 278 279 CASE_OP_32_64(ext8u): 280 return (uint8_t)x; 281 282 CASE_OP_32_64(ext16u): 283 return (uint16_t)x; 284 285 case INDEX_op_ext32s_i64: 286 return (int32_t)x; 287 288 case INDEX_op_ext32u_i64: 289 return (uint32_t)x; 290 291 case INDEX_op_muluh_i32: 292 return ((uint64_t)(uint32_t)x * (uint32_t)y) >> 32; 293 case INDEX_op_mulsh_i32: 294 return ((int64_t)(int32_t)x * (int32_t)y) >> 32; 295 296 case INDEX_op_muluh_i64: 297 mulu64(&l64, &h64, x, y); 298 return h64; 299 case INDEX_op_mulsh_i64: 300 muls64(&l64, &h64, x, y); 301 return h64; 302 303 case INDEX_op_div_i32: 304 /* Avoid crashing on divide by zero, otherwise undefined. */ 305 return (int32_t)x / ((int32_t)y ? : 1); 306 case INDEX_op_divu_i32: 307 return (uint32_t)x / ((uint32_t)y ? : 1); 308 case INDEX_op_div_i64: 309 return (int64_t)x / ((int64_t)y ? : 1); 310 case INDEX_op_divu_i64: 311 return (uint64_t)x / ((uint64_t)y ? : 1); 312 313 case INDEX_op_rem_i32: 314 return (int32_t)x % ((int32_t)y ? : 1); 315 case INDEX_op_remu_i32: 316 return (uint32_t)x % ((uint32_t)y ? : 1); 317 case INDEX_op_rem_i64: 318 return (int64_t)x % ((int64_t)y ? : 1); 319 case INDEX_op_remu_i64: 320 return (uint64_t)x % ((uint64_t)y ? : 1); 321 322 default: 323 fprintf(stderr, 324 "Unrecognized operation %d in do_constant_folding.\n", op); 325 tcg_abort(); 326 } 327 } 328 329 static TCGArg do_constant_folding(TCGOpcode op, TCGArg x, TCGArg y) 330 { 331 TCGArg res = do_constant_folding_2(op, x, y); 332 if (op_bits(op) == 32) { 333 res &= 0xffffffff; 334 } 335 return res; 336 } 337 338 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c) 339 { 340 switch (c) { 341 case TCG_COND_EQ: 342 return x == y; 343 case TCG_COND_NE: 344 return x != y; 345 case TCG_COND_LT: 346 return (int32_t)x < (int32_t)y; 347 case TCG_COND_GE: 348 return (int32_t)x >= (int32_t)y; 349 case TCG_COND_LE: 350 return (int32_t)x <= (int32_t)y; 351 case TCG_COND_GT: 352 return (int32_t)x > (int32_t)y; 353 case TCG_COND_LTU: 354 return x < y; 355 case TCG_COND_GEU: 356 return x >= y; 357 case TCG_COND_LEU: 358 return x <= y; 359 case TCG_COND_GTU: 360 return x > y; 361 default: 362 tcg_abort(); 363 } 364 } 365 366 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c) 367 { 368 switch (c) { 369 case TCG_COND_EQ: 370 return x == y; 371 case TCG_COND_NE: 372 return x != y; 373 case TCG_COND_LT: 374 return (int64_t)x < (int64_t)y; 375 case TCG_COND_GE: 376 return (int64_t)x >= (int64_t)y; 377 case TCG_COND_LE: 378 return (int64_t)x <= (int64_t)y; 379 case TCG_COND_GT: 380 return (int64_t)x > (int64_t)y; 381 case TCG_COND_LTU: 382 return x < y; 383 case TCG_COND_GEU: 384 return x >= y; 385 case TCG_COND_LEU: 386 return x <= y; 387 case TCG_COND_GTU: 388 return x > y; 389 default: 390 tcg_abort(); 391 } 392 } 393 394 static bool do_constant_folding_cond_eq(TCGCond c) 395 { 396 switch (c) { 397 case TCG_COND_GT: 398 case TCG_COND_LTU: 399 case TCG_COND_LT: 400 case TCG_COND_GTU: 401 case TCG_COND_NE: 402 return 0; 403 case TCG_COND_GE: 404 case TCG_COND_GEU: 405 case TCG_COND_LE: 406 case TCG_COND_LEU: 407 case TCG_COND_EQ: 408 return 1; 409 default: 410 tcg_abort(); 411 } 412 } 413 414 /* Return 2 if the condition can't be simplified, and the result 415 of the condition (0 or 1) if it can */ 416 static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x, 417 TCGArg y, TCGCond c) 418 { 419 if (temps[x].state == TCG_TEMP_CONST && temps[y].state == TCG_TEMP_CONST) { 420 switch (op_bits(op)) { 421 case 32: 422 return do_constant_folding_cond_32(temps[x].val, temps[y].val, c); 423 case 64: 424 return do_constant_folding_cond_64(temps[x].val, temps[y].val, c); 425 default: 426 tcg_abort(); 427 } 428 } else if (temps_are_copies(x, y)) { 429 return do_constant_folding_cond_eq(c); 430 } else if (temps[y].state == TCG_TEMP_CONST && temps[y].val == 0) { 431 switch (c) { 432 case TCG_COND_LTU: 433 return 0; 434 case TCG_COND_GEU: 435 return 1; 436 default: 437 return 2; 438 } 439 } else { 440 return 2; 441 } 442 } 443 444 /* Return 2 if the condition can't be simplified, and the result 445 of the condition (0 or 1) if it can */ 446 static TCGArg do_constant_folding_cond2(TCGArg *p1, TCGArg *p2, TCGCond c) 447 { 448 TCGArg al = p1[0], ah = p1[1]; 449 TCGArg bl = p2[0], bh = p2[1]; 450 451 if (temps[bl].state == TCG_TEMP_CONST 452 && temps[bh].state == TCG_TEMP_CONST) { 453 uint64_t b = ((uint64_t)temps[bh].val << 32) | (uint32_t)temps[bl].val; 454 455 if (temps[al].state == TCG_TEMP_CONST 456 && temps[ah].state == TCG_TEMP_CONST) { 457 uint64_t a; 458 a = ((uint64_t)temps[ah].val << 32) | (uint32_t)temps[al].val; 459 return do_constant_folding_cond_64(a, b, c); 460 } 461 if (b == 0) { 462 switch (c) { 463 case TCG_COND_LTU: 464 return 0; 465 case TCG_COND_GEU: 466 return 1; 467 default: 468 break; 469 } 470 } 471 } 472 if (temps_are_copies(al, bl) && temps_are_copies(ah, bh)) { 473 return do_constant_folding_cond_eq(c); 474 } 475 return 2; 476 } 477 478 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2) 479 { 480 TCGArg a1 = *p1, a2 = *p2; 481 int sum = 0; 482 sum += temps[a1].state == TCG_TEMP_CONST; 483 sum -= temps[a2].state == TCG_TEMP_CONST; 484 485 /* Prefer the constant in second argument, and then the form 486 op a, a, b, which is better handled on non-RISC hosts. */ 487 if (sum > 0 || (sum == 0 && dest == a2)) { 488 *p1 = a2; 489 *p2 = a1; 490 return true; 491 } 492 return false; 493 } 494 495 static bool swap_commutative2(TCGArg *p1, TCGArg *p2) 496 { 497 int sum = 0; 498 sum += temps[p1[0]].state == TCG_TEMP_CONST; 499 sum += temps[p1[1]].state == TCG_TEMP_CONST; 500 sum -= temps[p2[0]].state == TCG_TEMP_CONST; 501 sum -= temps[p2[1]].state == TCG_TEMP_CONST; 502 if (sum > 0) { 503 TCGArg t; 504 t = p1[0], p1[0] = p2[0], p2[0] = t; 505 t = p1[1], p1[1] = p2[1], p2[1] = t; 506 return true; 507 } 508 return false; 509 } 510 511 /* Propagate constants and copies, fold constant expressions. */ 512 static TCGArg *tcg_constant_folding(TCGContext *s, uint16_t *tcg_opc_ptr, 513 TCGArg *args, TCGOpDef *tcg_op_defs) 514 { 515 int i, nb_ops, op_index, nb_temps, nb_globals, nb_call_args; 516 tcg_target_ulong mask, affected; 517 TCGOpcode op; 518 const TCGOpDef *def; 519 TCGArg *gen_args; 520 TCGArg tmp; 521 522 /* Array VALS has an element for each temp. 523 If this temp holds a constant then its value is kept in VALS' element. 524 If this temp is a copy of other ones then the other copies are 525 available through the doubly linked circular list. */ 526 527 nb_temps = s->nb_temps; 528 nb_globals = s->nb_globals; 529 reset_all_temps(nb_temps); 530 531 nb_ops = tcg_opc_ptr - s->gen_opc_buf; 532 gen_args = args; 533 for (op_index = 0; op_index < nb_ops; op_index++) { 534 op = s->gen_opc_buf[op_index]; 535 def = &tcg_op_defs[op]; 536 /* Do copy propagation */ 537 if (op == INDEX_op_call) { 538 int nb_oargs = args[0] >> 16; 539 int nb_iargs = args[0] & 0xffff; 540 for (i = nb_oargs + 1; i < nb_oargs + nb_iargs + 1; i++) { 541 if (temps[args[i]].state == TCG_TEMP_COPY) { 542 args[i] = find_better_copy(s, args[i]); 543 } 544 } 545 } else { 546 for (i = def->nb_oargs; i < def->nb_oargs + def->nb_iargs; i++) { 547 if (temps[args[i]].state == TCG_TEMP_COPY) { 548 args[i] = find_better_copy(s, args[i]); 549 } 550 } 551 } 552 553 /* For commutative operations make constant second argument */ 554 switch (op) { 555 CASE_OP_32_64(add): 556 CASE_OP_32_64(mul): 557 CASE_OP_32_64(and): 558 CASE_OP_32_64(or): 559 CASE_OP_32_64(xor): 560 CASE_OP_32_64(eqv): 561 CASE_OP_32_64(nand): 562 CASE_OP_32_64(nor): 563 CASE_OP_32_64(muluh): 564 CASE_OP_32_64(mulsh): 565 swap_commutative(args[0], &args[1], &args[2]); 566 break; 567 CASE_OP_32_64(brcond): 568 if (swap_commutative(-1, &args[0], &args[1])) { 569 args[2] = tcg_swap_cond(args[2]); 570 } 571 break; 572 CASE_OP_32_64(setcond): 573 if (swap_commutative(args[0], &args[1], &args[2])) { 574 args[3] = tcg_swap_cond(args[3]); 575 } 576 break; 577 CASE_OP_32_64(movcond): 578 if (swap_commutative(-1, &args[1], &args[2])) { 579 args[5] = tcg_swap_cond(args[5]); 580 } 581 /* For movcond, we canonicalize the "false" input reg to match 582 the destination reg so that the tcg backend can implement 583 a "move if true" operation. */ 584 if (swap_commutative(args[0], &args[4], &args[3])) { 585 args[5] = tcg_invert_cond(args[5]); 586 } 587 break; 588 CASE_OP_32_64(add2): 589 swap_commutative(args[0], &args[2], &args[4]); 590 swap_commutative(args[1], &args[3], &args[5]); 591 break; 592 CASE_OP_32_64(mulu2): 593 CASE_OP_32_64(muls2): 594 swap_commutative(args[0], &args[2], &args[3]); 595 break; 596 case INDEX_op_brcond2_i32: 597 if (swap_commutative2(&args[0], &args[2])) { 598 args[4] = tcg_swap_cond(args[4]); 599 } 600 break; 601 case INDEX_op_setcond2_i32: 602 if (swap_commutative2(&args[1], &args[3])) { 603 args[5] = tcg_swap_cond(args[5]); 604 } 605 break; 606 default: 607 break; 608 } 609 610 /* Simplify expressions for "shift/rot r, 0, a => movi r, 0", 611 and "sub r, 0, a => neg r, a" case. */ 612 switch (op) { 613 CASE_OP_32_64(shl): 614 CASE_OP_32_64(shr): 615 CASE_OP_32_64(sar): 616 CASE_OP_32_64(rotl): 617 CASE_OP_32_64(rotr): 618 if (temps[args[1]].state == TCG_TEMP_CONST 619 && temps[args[1]].val == 0) { 620 s->gen_opc_buf[op_index] = op_to_movi(op); 621 tcg_opt_gen_movi(gen_args, args[0], 0); 622 args += 3; 623 gen_args += 2; 624 continue; 625 } 626 break; 627 CASE_OP_32_64(sub): 628 { 629 TCGOpcode neg_op; 630 bool have_neg; 631 632 if (temps[args[2]].state == TCG_TEMP_CONST) { 633 /* Proceed with possible constant folding. */ 634 break; 635 } 636 if (op == INDEX_op_sub_i32) { 637 neg_op = INDEX_op_neg_i32; 638 have_neg = TCG_TARGET_HAS_neg_i32; 639 } else { 640 neg_op = INDEX_op_neg_i64; 641 have_neg = TCG_TARGET_HAS_neg_i64; 642 } 643 if (!have_neg) { 644 break; 645 } 646 if (temps[args[1]].state == TCG_TEMP_CONST 647 && temps[args[1]].val == 0) { 648 s->gen_opc_buf[op_index] = neg_op; 649 reset_temp(args[0]); 650 gen_args[0] = args[0]; 651 gen_args[1] = args[2]; 652 args += 3; 653 gen_args += 2; 654 continue; 655 } 656 } 657 break; 658 CASE_OP_32_64(xor): 659 CASE_OP_32_64(nand): 660 if (temps[args[1]].state != TCG_TEMP_CONST 661 && temps[args[2]].state == TCG_TEMP_CONST 662 && temps[args[2]].val == -1) { 663 i = 1; 664 goto try_not; 665 } 666 break; 667 CASE_OP_32_64(nor): 668 if (temps[args[1]].state != TCG_TEMP_CONST 669 && temps[args[2]].state == TCG_TEMP_CONST 670 && temps[args[2]].val == 0) { 671 i = 1; 672 goto try_not; 673 } 674 break; 675 CASE_OP_32_64(andc): 676 if (temps[args[2]].state != TCG_TEMP_CONST 677 && temps[args[1]].state == TCG_TEMP_CONST 678 && temps[args[1]].val == -1) { 679 i = 2; 680 goto try_not; 681 } 682 break; 683 CASE_OP_32_64(orc): 684 CASE_OP_32_64(eqv): 685 if (temps[args[2]].state != TCG_TEMP_CONST 686 && temps[args[1]].state == TCG_TEMP_CONST 687 && temps[args[1]].val == 0) { 688 i = 2; 689 goto try_not; 690 } 691 break; 692 try_not: 693 { 694 TCGOpcode not_op; 695 bool have_not; 696 697 if (def->flags & TCG_OPF_64BIT) { 698 not_op = INDEX_op_not_i64; 699 have_not = TCG_TARGET_HAS_not_i64; 700 } else { 701 not_op = INDEX_op_not_i32; 702 have_not = TCG_TARGET_HAS_not_i32; 703 } 704 if (!have_not) { 705 break; 706 } 707 s->gen_opc_buf[op_index] = not_op; 708 reset_temp(args[0]); 709 gen_args[0] = args[0]; 710 gen_args[1] = args[i]; 711 args += 3; 712 gen_args += 2; 713 continue; 714 } 715 default: 716 break; 717 } 718 719 /* Simplify expression for "op r, a, const => mov r, a" cases */ 720 switch (op) { 721 CASE_OP_32_64(add): 722 CASE_OP_32_64(sub): 723 CASE_OP_32_64(shl): 724 CASE_OP_32_64(shr): 725 CASE_OP_32_64(sar): 726 CASE_OP_32_64(rotl): 727 CASE_OP_32_64(rotr): 728 CASE_OP_32_64(or): 729 CASE_OP_32_64(xor): 730 CASE_OP_32_64(andc): 731 if (temps[args[1]].state != TCG_TEMP_CONST 732 && temps[args[2]].state == TCG_TEMP_CONST 733 && temps[args[2]].val == 0) { 734 goto do_mov3; 735 } 736 break; 737 CASE_OP_32_64(and): 738 CASE_OP_32_64(orc): 739 CASE_OP_32_64(eqv): 740 if (temps[args[1]].state != TCG_TEMP_CONST 741 && temps[args[2]].state == TCG_TEMP_CONST 742 && temps[args[2]].val == -1) { 743 goto do_mov3; 744 } 745 break; 746 do_mov3: 747 if (temps_are_copies(args[0], args[1])) { 748 s->gen_opc_buf[op_index] = INDEX_op_nop; 749 } else { 750 s->gen_opc_buf[op_index] = op_to_mov(op); 751 tcg_opt_gen_mov(s, gen_args, args[0], args[1]); 752 gen_args += 2; 753 } 754 args += 3; 755 continue; 756 default: 757 break; 758 } 759 760 /* Simplify using known-zero bits. Currently only ops with a single 761 output argument is supported. */ 762 mask = -1; 763 affected = -1; 764 switch (op) { 765 CASE_OP_32_64(ext8s): 766 if ((temps[args[1]].mask & 0x80) != 0) { 767 break; 768 } 769 CASE_OP_32_64(ext8u): 770 mask = 0xff; 771 goto and_const; 772 CASE_OP_32_64(ext16s): 773 if ((temps[args[1]].mask & 0x8000) != 0) { 774 break; 775 } 776 CASE_OP_32_64(ext16u): 777 mask = 0xffff; 778 goto and_const; 779 case INDEX_op_ext32s_i64: 780 if ((temps[args[1]].mask & 0x80000000) != 0) { 781 break; 782 } 783 case INDEX_op_ext32u_i64: 784 mask = 0xffffffffU; 785 goto and_const; 786 787 CASE_OP_32_64(and): 788 mask = temps[args[2]].mask; 789 if (temps[args[2]].state == TCG_TEMP_CONST) { 790 and_const: 791 affected = temps[args[1]].mask & ~mask; 792 } 793 mask = temps[args[1]].mask & mask; 794 break; 795 796 CASE_OP_32_64(andc): 797 /* Known-zeros does not imply known-ones. Therefore unless 798 args[2] is constant, we can't infer anything from it. */ 799 if (temps[args[2]].state == TCG_TEMP_CONST) { 800 mask = ~temps[args[2]].mask; 801 goto and_const; 802 } 803 /* But we certainly know nothing outside args[1] may be set. */ 804 mask = temps[args[1]].mask; 805 break; 806 807 case INDEX_op_sar_i32: 808 if (temps[args[2]].state == TCG_TEMP_CONST) { 809 mask = (int32_t)temps[args[1]].mask >> temps[args[2]].val; 810 } 811 break; 812 case INDEX_op_sar_i64: 813 if (temps[args[2]].state == TCG_TEMP_CONST) { 814 mask = (int64_t)temps[args[1]].mask >> temps[args[2]].val; 815 } 816 break; 817 818 case INDEX_op_shr_i32: 819 if (temps[args[2]].state == TCG_TEMP_CONST) { 820 mask = (uint32_t)temps[args[1]].mask >> temps[args[2]].val; 821 } 822 break; 823 case INDEX_op_shr_i64: 824 if (temps[args[2]].state == TCG_TEMP_CONST) { 825 mask = (uint64_t)temps[args[1]].mask >> temps[args[2]].val; 826 } 827 break; 828 829 CASE_OP_32_64(shl): 830 if (temps[args[2]].state == TCG_TEMP_CONST) { 831 mask = temps[args[1]].mask << temps[args[2]].val; 832 } 833 break; 834 835 CASE_OP_32_64(neg): 836 /* Set to 1 all bits to the left of the rightmost. */ 837 mask = -(temps[args[1]].mask & -temps[args[1]].mask); 838 break; 839 840 CASE_OP_32_64(deposit): 841 tmp = ((1ull << args[4]) - 1); 842 mask = ((temps[args[1]].mask & ~(tmp << args[3])) 843 | ((temps[args[2]].mask & tmp) << args[3])); 844 break; 845 846 CASE_OP_32_64(or): 847 CASE_OP_32_64(xor): 848 mask = temps[args[1]].mask | temps[args[2]].mask; 849 break; 850 851 CASE_OP_32_64(setcond): 852 mask = 1; 853 break; 854 855 CASE_OP_32_64(movcond): 856 mask = temps[args[3]].mask | temps[args[4]].mask; 857 break; 858 859 CASE_OP_32_64(ld8u): 860 case INDEX_op_qemu_ld8u: 861 mask = 0xff; 862 break; 863 CASE_OP_32_64(ld16u): 864 case INDEX_op_qemu_ld16u: 865 mask = 0xffff; 866 break; 867 case INDEX_op_ld32u_i64: 868 #if TCG_TARGET_REG_BITS == 64 869 case INDEX_op_qemu_ld32u: 870 #endif 871 mask = 0xffffffffu; 872 break; 873 874 CASE_OP_32_64(qemu_ld): 875 { 876 TCGMemOp mop = args[def->nb_oargs + def->nb_iargs]; 877 if (!(mop & MO_SIGN)) { 878 mask = (2ULL << ((8 << (mop & MO_SIZE)) - 1)) - 1; 879 } 880 } 881 break; 882 883 default: 884 break; 885 } 886 887 /* 32-bit ops (non 64-bit ops and non load/store ops) generate 32-bit 888 results */ 889 if (!(def->flags & (TCG_OPF_CALL_CLOBBER | TCG_OPF_64BIT))) { 890 mask &= 0xffffffffu; 891 } 892 893 if (mask == 0) { 894 assert(def->nb_oargs == 1); 895 s->gen_opc_buf[op_index] = op_to_movi(op); 896 tcg_opt_gen_movi(gen_args, args[0], 0); 897 args += def->nb_oargs + def->nb_iargs + def->nb_cargs; 898 gen_args += 2; 899 continue; 900 } 901 if (affected == 0) { 902 assert(def->nb_oargs == 1); 903 if (temps_are_copies(args[0], args[1])) { 904 s->gen_opc_buf[op_index] = INDEX_op_nop; 905 } else if (temps[args[1]].state != TCG_TEMP_CONST) { 906 s->gen_opc_buf[op_index] = op_to_mov(op); 907 tcg_opt_gen_mov(s, gen_args, args[0], args[1]); 908 gen_args += 2; 909 } else { 910 s->gen_opc_buf[op_index] = op_to_movi(op); 911 tcg_opt_gen_movi(gen_args, args[0], temps[args[1]].val); 912 gen_args += 2; 913 } 914 args += def->nb_iargs + 1; 915 continue; 916 } 917 918 /* Simplify expression for "op r, a, 0 => movi r, 0" cases */ 919 switch (op) { 920 CASE_OP_32_64(and): 921 CASE_OP_32_64(mul): 922 CASE_OP_32_64(muluh): 923 CASE_OP_32_64(mulsh): 924 if ((temps[args[2]].state == TCG_TEMP_CONST 925 && temps[args[2]].val == 0)) { 926 s->gen_opc_buf[op_index] = op_to_movi(op); 927 tcg_opt_gen_movi(gen_args, args[0], 0); 928 args += 3; 929 gen_args += 2; 930 continue; 931 } 932 break; 933 default: 934 break; 935 } 936 937 /* Simplify expression for "op r, a, a => mov r, a" cases */ 938 switch (op) { 939 CASE_OP_32_64(or): 940 CASE_OP_32_64(and): 941 if (temps_are_copies(args[1], args[2])) { 942 if (temps_are_copies(args[0], args[1])) { 943 s->gen_opc_buf[op_index] = INDEX_op_nop; 944 } else { 945 s->gen_opc_buf[op_index] = op_to_mov(op); 946 tcg_opt_gen_mov(s, gen_args, args[0], args[1]); 947 gen_args += 2; 948 } 949 args += 3; 950 continue; 951 } 952 break; 953 default: 954 break; 955 } 956 957 /* Simplify expression for "op r, a, a => movi r, 0" cases */ 958 switch (op) { 959 CASE_OP_32_64(andc): 960 CASE_OP_32_64(sub): 961 CASE_OP_32_64(xor): 962 if (temps_are_copies(args[1], args[2])) { 963 s->gen_opc_buf[op_index] = op_to_movi(op); 964 tcg_opt_gen_movi(gen_args, args[0], 0); 965 gen_args += 2; 966 args += 3; 967 continue; 968 } 969 break; 970 default: 971 break; 972 } 973 974 /* Propagate constants through copy operations and do constant 975 folding. Constants will be substituted to arguments by register 976 allocator where needed and possible. Also detect copies. */ 977 switch (op) { 978 CASE_OP_32_64(mov): 979 if (temps_are_copies(args[0], args[1])) { 980 args += 2; 981 s->gen_opc_buf[op_index] = INDEX_op_nop; 982 break; 983 } 984 if (temps[args[1]].state != TCG_TEMP_CONST) { 985 tcg_opt_gen_mov(s, gen_args, args[0], args[1]); 986 gen_args += 2; 987 args += 2; 988 break; 989 } 990 /* Source argument is constant. Rewrite the operation and 991 let movi case handle it. */ 992 op = op_to_movi(op); 993 s->gen_opc_buf[op_index] = op; 994 args[1] = temps[args[1]].val; 995 /* fallthrough */ 996 CASE_OP_32_64(movi): 997 tcg_opt_gen_movi(gen_args, args[0], args[1]); 998 gen_args += 2; 999 args += 2; 1000 break; 1001 1002 CASE_OP_32_64(not): 1003 CASE_OP_32_64(neg): 1004 CASE_OP_32_64(ext8s): 1005 CASE_OP_32_64(ext8u): 1006 CASE_OP_32_64(ext16s): 1007 CASE_OP_32_64(ext16u): 1008 case INDEX_op_ext32s_i64: 1009 case INDEX_op_ext32u_i64: 1010 if (temps[args[1]].state == TCG_TEMP_CONST) { 1011 s->gen_opc_buf[op_index] = op_to_movi(op); 1012 tmp = do_constant_folding(op, temps[args[1]].val, 0); 1013 tcg_opt_gen_movi(gen_args, args[0], tmp); 1014 gen_args += 2; 1015 args += 2; 1016 break; 1017 } 1018 goto do_default; 1019 1020 CASE_OP_32_64(add): 1021 CASE_OP_32_64(sub): 1022 CASE_OP_32_64(mul): 1023 CASE_OP_32_64(or): 1024 CASE_OP_32_64(and): 1025 CASE_OP_32_64(xor): 1026 CASE_OP_32_64(shl): 1027 CASE_OP_32_64(shr): 1028 CASE_OP_32_64(sar): 1029 CASE_OP_32_64(rotl): 1030 CASE_OP_32_64(rotr): 1031 CASE_OP_32_64(andc): 1032 CASE_OP_32_64(orc): 1033 CASE_OP_32_64(eqv): 1034 CASE_OP_32_64(nand): 1035 CASE_OP_32_64(nor): 1036 CASE_OP_32_64(muluh): 1037 CASE_OP_32_64(mulsh): 1038 CASE_OP_32_64(div): 1039 CASE_OP_32_64(divu): 1040 CASE_OP_32_64(rem): 1041 CASE_OP_32_64(remu): 1042 if (temps[args[1]].state == TCG_TEMP_CONST 1043 && temps[args[2]].state == TCG_TEMP_CONST) { 1044 s->gen_opc_buf[op_index] = op_to_movi(op); 1045 tmp = do_constant_folding(op, temps[args[1]].val, 1046 temps[args[2]].val); 1047 tcg_opt_gen_movi(gen_args, args[0], tmp); 1048 gen_args += 2; 1049 args += 3; 1050 break; 1051 } 1052 goto do_default; 1053 1054 CASE_OP_32_64(deposit): 1055 if (temps[args[1]].state == TCG_TEMP_CONST 1056 && temps[args[2]].state == TCG_TEMP_CONST) { 1057 s->gen_opc_buf[op_index] = op_to_movi(op); 1058 tmp = ((1ull << args[4]) - 1); 1059 tmp = (temps[args[1]].val & ~(tmp << args[3])) 1060 | ((temps[args[2]].val & tmp) << args[3]); 1061 tcg_opt_gen_movi(gen_args, args[0], tmp); 1062 gen_args += 2; 1063 args += 5; 1064 break; 1065 } 1066 goto do_default; 1067 1068 CASE_OP_32_64(setcond): 1069 tmp = do_constant_folding_cond(op, args[1], args[2], args[3]); 1070 if (tmp != 2) { 1071 s->gen_opc_buf[op_index] = op_to_movi(op); 1072 tcg_opt_gen_movi(gen_args, args[0], tmp); 1073 gen_args += 2; 1074 args += 4; 1075 break; 1076 } 1077 goto do_default; 1078 1079 CASE_OP_32_64(brcond): 1080 tmp = do_constant_folding_cond(op, args[0], args[1], args[2]); 1081 if (tmp != 2) { 1082 if (tmp) { 1083 reset_all_temps(nb_temps); 1084 s->gen_opc_buf[op_index] = INDEX_op_br; 1085 gen_args[0] = args[3]; 1086 gen_args += 1; 1087 } else { 1088 s->gen_opc_buf[op_index] = INDEX_op_nop; 1089 } 1090 args += 4; 1091 break; 1092 } 1093 goto do_default; 1094 1095 CASE_OP_32_64(movcond): 1096 tmp = do_constant_folding_cond(op, args[1], args[2], args[5]); 1097 if (tmp != 2) { 1098 if (temps_are_copies(args[0], args[4-tmp])) { 1099 s->gen_opc_buf[op_index] = INDEX_op_nop; 1100 } else if (temps[args[4-tmp]].state == TCG_TEMP_CONST) { 1101 s->gen_opc_buf[op_index] = op_to_movi(op); 1102 tcg_opt_gen_movi(gen_args, args[0], temps[args[4-tmp]].val); 1103 gen_args += 2; 1104 } else { 1105 s->gen_opc_buf[op_index] = op_to_mov(op); 1106 tcg_opt_gen_mov(s, gen_args, args[0], args[4-tmp]); 1107 gen_args += 2; 1108 } 1109 args += 6; 1110 break; 1111 } 1112 goto do_default; 1113 1114 case INDEX_op_add2_i32: 1115 case INDEX_op_sub2_i32: 1116 if (temps[args[2]].state == TCG_TEMP_CONST 1117 && temps[args[3]].state == TCG_TEMP_CONST 1118 && temps[args[4]].state == TCG_TEMP_CONST 1119 && temps[args[5]].state == TCG_TEMP_CONST) { 1120 uint32_t al = temps[args[2]].val; 1121 uint32_t ah = temps[args[3]].val; 1122 uint32_t bl = temps[args[4]].val; 1123 uint32_t bh = temps[args[5]].val; 1124 uint64_t a = ((uint64_t)ah << 32) | al; 1125 uint64_t b = ((uint64_t)bh << 32) | bl; 1126 TCGArg rl, rh; 1127 1128 if (op == INDEX_op_add2_i32) { 1129 a += b; 1130 } else { 1131 a -= b; 1132 } 1133 1134 /* We emit the extra nop when we emit the add2/sub2. */ 1135 assert(s->gen_opc_buf[op_index + 1] == INDEX_op_nop); 1136 1137 rl = args[0]; 1138 rh = args[1]; 1139 s->gen_opc_buf[op_index] = INDEX_op_movi_i32; 1140 s->gen_opc_buf[++op_index] = INDEX_op_movi_i32; 1141 tcg_opt_gen_movi(&gen_args[0], rl, (uint32_t)a); 1142 tcg_opt_gen_movi(&gen_args[2], rh, (uint32_t)(a >> 32)); 1143 gen_args += 4; 1144 args += 6; 1145 break; 1146 } 1147 goto do_default; 1148 1149 case INDEX_op_mulu2_i32: 1150 if (temps[args[2]].state == TCG_TEMP_CONST 1151 && temps[args[3]].state == TCG_TEMP_CONST) { 1152 uint32_t a = temps[args[2]].val; 1153 uint32_t b = temps[args[3]].val; 1154 uint64_t r = (uint64_t)a * b; 1155 TCGArg rl, rh; 1156 1157 /* We emit the extra nop when we emit the mulu2. */ 1158 assert(s->gen_opc_buf[op_index + 1] == INDEX_op_nop); 1159 1160 rl = args[0]; 1161 rh = args[1]; 1162 s->gen_opc_buf[op_index] = INDEX_op_movi_i32; 1163 s->gen_opc_buf[++op_index] = INDEX_op_movi_i32; 1164 tcg_opt_gen_movi(&gen_args[0], rl, (uint32_t)r); 1165 tcg_opt_gen_movi(&gen_args[2], rh, (uint32_t)(r >> 32)); 1166 gen_args += 4; 1167 args += 4; 1168 break; 1169 } 1170 goto do_default; 1171 1172 case INDEX_op_brcond2_i32: 1173 tmp = do_constant_folding_cond2(&args[0], &args[2], args[4]); 1174 if (tmp != 2) { 1175 if (tmp) { 1176 reset_all_temps(nb_temps); 1177 s->gen_opc_buf[op_index] = INDEX_op_br; 1178 gen_args[0] = args[5]; 1179 gen_args += 1; 1180 } else { 1181 s->gen_opc_buf[op_index] = INDEX_op_nop; 1182 } 1183 } else if ((args[4] == TCG_COND_LT || args[4] == TCG_COND_GE) 1184 && temps[args[2]].state == TCG_TEMP_CONST 1185 && temps[args[3]].state == TCG_TEMP_CONST 1186 && temps[args[2]].val == 0 1187 && temps[args[3]].val == 0) { 1188 /* Simplify LT/GE comparisons vs zero to a single compare 1189 vs the high word of the input. */ 1190 reset_all_temps(nb_temps); 1191 s->gen_opc_buf[op_index] = INDEX_op_brcond_i32; 1192 gen_args[0] = args[1]; 1193 gen_args[1] = args[3]; 1194 gen_args[2] = args[4]; 1195 gen_args[3] = args[5]; 1196 gen_args += 4; 1197 } else { 1198 goto do_default; 1199 } 1200 args += 6; 1201 break; 1202 1203 case INDEX_op_setcond2_i32: 1204 tmp = do_constant_folding_cond2(&args[1], &args[3], args[5]); 1205 if (tmp != 2) { 1206 s->gen_opc_buf[op_index] = INDEX_op_movi_i32; 1207 tcg_opt_gen_movi(gen_args, args[0], tmp); 1208 gen_args += 2; 1209 } else if ((args[5] == TCG_COND_LT || args[5] == TCG_COND_GE) 1210 && temps[args[3]].state == TCG_TEMP_CONST 1211 && temps[args[4]].state == TCG_TEMP_CONST 1212 && temps[args[3]].val == 0 1213 && temps[args[4]].val == 0) { 1214 /* Simplify LT/GE comparisons vs zero to a single compare 1215 vs the high word of the input. */ 1216 s->gen_opc_buf[op_index] = INDEX_op_setcond_i32; 1217 reset_temp(args[0]); 1218 gen_args[0] = args[0]; 1219 gen_args[1] = args[2]; 1220 gen_args[2] = args[4]; 1221 gen_args[3] = args[5]; 1222 gen_args += 4; 1223 } else { 1224 goto do_default; 1225 } 1226 args += 6; 1227 break; 1228 1229 case INDEX_op_call: 1230 nb_call_args = (args[0] >> 16) + (args[0] & 0xffff); 1231 if (!(args[nb_call_args + 1] & (TCG_CALL_NO_READ_GLOBALS | 1232 TCG_CALL_NO_WRITE_GLOBALS))) { 1233 for (i = 0; i < nb_globals; i++) { 1234 reset_temp(i); 1235 } 1236 } 1237 for (i = 0; i < (args[0] >> 16); i++) { 1238 reset_temp(args[i + 1]); 1239 } 1240 i = nb_call_args + 3; 1241 while (i) { 1242 *gen_args = *args; 1243 args++; 1244 gen_args++; 1245 i--; 1246 } 1247 break; 1248 1249 default: 1250 do_default: 1251 /* Default case: we know nothing about operation (or were unable 1252 to compute the operation result) so no propagation is done. 1253 We trash everything if the operation is the end of a basic 1254 block, otherwise we only trash the output args. "mask" is 1255 the non-zero bits mask for the first output arg. */ 1256 if (def->flags & TCG_OPF_BB_END) { 1257 reset_all_temps(nb_temps); 1258 } else { 1259 for (i = 0; i < def->nb_oargs; i++) { 1260 reset_temp(args[i]); 1261 /* Save the corresponding known-zero bits mask for the 1262 first output argument (only one supported so far). */ 1263 if (i == 0) { 1264 temps[args[i]].mask = mask; 1265 } 1266 } 1267 } 1268 for (i = 0; i < def->nb_args; i++) { 1269 gen_args[i] = args[i]; 1270 } 1271 args += def->nb_args; 1272 gen_args += def->nb_args; 1273 break; 1274 } 1275 } 1276 1277 return gen_args; 1278 } 1279 1280 TCGArg *tcg_optimize(TCGContext *s, uint16_t *tcg_opc_ptr, 1281 TCGArg *args, TCGOpDef *tcg_op_defs) 1282 { 1283 TCGArg *res; 1284 res = tcg_constant_folding(s, tcg_opc_ptr, args, tcg_op_defs); 1285 return res; 1286 } 1287