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 switch (op) { 202 CASE_OP_32_64(add): 203 return x + y; 204 205 CASE_OP_32_64(sub): 206 return x - y; 207 208 CASE_OP_32_64(mul): 209 return x * y; 210 211 CASE_OP_32_64(and): 212 return x & y; 213 214 CASE_OP_32_64(or): 215 return x | y; 216 217 CASE_OP_32_64(xor): 218 return x ^ y; 219 220 case INDEX_op_shl_i32: 221 return (uint32_t)x << (uint32_t)y; 222 223 case INDEX_op_shl_i64: 224 return (uint64_t)x << (uint64_t)y; 225 226 case INDEX_op_shr_i32: 227 return (uint32_t)x >> (uint32_t)y; 228 229 case INDEX_op_shr_i64: 230 return (uint64_t)x >> (uint64_t)y; 231 232 case INDEX_op_sar_i32: 233 return (int32_t)x >> (int32_t)y; 234 235 case INDEX_op_sar_i64: 236 return (int64_t)x >> (int64_t)y; 237 238 case INDEX_op_rotr_i32: 239 x = ((uint32_t)x << (32 - y)) | ((uint32_t)x >> y); 240 return x; 241 242 case INDEX_op_rotr_i64: 243 x = ((uint64_t)x << (64 - y)) | ((uint64_t)x >> y); 244 return x; 245 246 case INDEX_op_rotl_i32: 247 x = ((uint32_t)x << y) | ((uint32_t)x >> (32 - y)); 248 return x; 249 250 case INDEX_op_rotl_i64: 251 x = ((uint64_t)x << y) | ((uint64_t)x >> (64 - y)); 252 return x; 253 254 CASE_OP_32_64(not): 255 return ~x; 256 257 CASE_OP_32_64(neg): 258 return -x; 259 260 CASE_OP_32_64(andc): 261 return x & ~y; 262 263 CASE_OP_32_64(orc): 264 return x | ~y; 265 266 CASE_OP_32_64(eqv): 267 return ~(x ^ y); 268 269 CASE_OP_32_64(nand): 270 return ~(x & y); 271 272 CASE_OP_32_64(nor): 273 return ~(x | y); 274 275 CASE_OP_32_64(ext8s): 276 return (int8_t)x; 277 278 CASE_OP_32_64(ext16s): 279 return (int16_t)x; 280 281 CASE_OP_32_64(ext8u): 282 return (uint8_t)x; 283 284 CASE_OP_32_64(ext16u): 285 return (uint16_t)x; 286 287 case INDEX_op_ext32s_i64: 288 return (int32_t)x; 289 290 case INDEX_op_ext32u_i64: 291 return (uint32_t)x; 292 293 default: 294 fprintf(stderr, 295 "Unrecognized operation %d in do_constant_folding.\n", op); 296 tcg_abort(); 297 } 298 } 299 300 static TCGArg do_constant_folding(TCGOpcode op, TCGArg x, TCGArg y) 301 { 302 TCGArg res = do_constant_folding_2(op, x, y); 303 if (op_bits(op) == 32) { 304 res &= 0xffffffff; 305 } 306 return res; 307 } 308 309 static bool do_constant_folding_cond_32(uint32_t x, uint32_t y, TCGCond c) 310 { 311 switch (c) { 312 case TCG_COND_EQ: 313 return x == y; 314 case TCG_COND_NE: 315 return x != y; 316 case TCG_COND_LT: 317 return (int32_t)x < (int32_t)y; 318 case TCG_COND_GE: 319 return (int32_t)x >= (int32_t)y; 320 case TCG_COND_LE: 321 return (int32_t)x <= (int32_t)y; 322 case TCG_COND_GT: 323 return (int32_t)x > (int32_t)y; 324 case TCG_COND_LTU: 325 return x < y; 326 case TCG_COND_GEU: 327 return x >= y; 328 case TCG_COND_LEU: 329 return x <= y; 330 case TCG_COND_GTU: 331 return x > y; 332 default: 333 tcg_abort(); 334 } 335 } 336 337 static bool do_constant_folding_cond_64(uint64_t x, uint64_t y, TCGCond c) 338 { 339 switch (c) { 340 case TCG_COND_EQ: 341 return x == y; 342 case TCG_COND_NE: 343 return x != y; 344 case TCG_COND_LT: 345 return (int64_t)x < (int64_t)y; 346 case TCG_COND_GE: 347 return (int64_t)x >= (int64_t)y; 348 case TCG_COND_LE: 349 return (int64_t)x <= (int64_t)y; 350 case TCG_COND_GT: 351 return (int64_t)x > (int64_t)y; 352 case TCG_COND_LTU: 353 return x < y; 354 case TCG_COND_GEU: 355 return x >= y; 356 case TCG_COND_LEU: 357 return x <= y; 358 case TCG_COND_GTU: 359 return x > y; 360 default: 361 tcg_abort(); 362 } 363 } 364 365 static bool do_constant_folding_cond_eq(TCGCond c) 366 { 367 switch (c) { 368 case TCG_COND_GT: 369 case TCG_COND_LTU: 370 case TCG_COND_LT: 371 case TCG_COND_GTU: 372 case TCG_COND_NE: 373 return 0; 374 case TCG_COND_GE: 375 case TCG_COND_GEU: 376 case TCG_COND_LE: 377 case TCG_COND_LEU: 378 case TCG_COND_EQ: 379 return 1; 380 default: 381 tcg_abort(); 382 } 383 } 384 385 /* Return 2 if the condition can't be simplified, and the result 386 of the condition (0 or 1) if it can */ 387 static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x, 388 TCGArg y, TCGCond c) 389 { 390 if (temps[x].state == TCG_TEMP_CONST && temps[y].state == TCG_TEMP_CONST) { 391 switch (op_bits(op)) { 392 case 32: 393 return do_constant_folding_cond_32(temps[x].val, temps[y].val, c); 394 case 64: 395 return do_constant_folding_cond_64(temps[x].val, temps[y].val, c); 396 default: 397 tcg_abort(); 398 } 399 } else if (temps_are_copies(x, y)) { 400 return do_constant_folding_cond_eq(c); 401 } else if (temps[y].state == TCG_TEMP_CONST && temps[y].val == 0) { 402 switch (c) { 403 case TCG_COND_LTU: 404 return 0; 405 case TCG_COND_GEU: 406 return 1; 407 default: 408 return 2; 409 } 410 } else { 411 return 2; 412 } 413 } 414 415 /* Return 2 if the condition can't be simplified, and the result 416 of the condition (0 or 1) if it can */ 417 static TCGArg do_constant_folding_cond2(TCGArg *p1, TCGArg *p2, TCGCond c) 418 { 419 TCGArg al = p1[0], ah = p1[1]; 420 TCGArg bl = p2[0], bh = p2[1]; 421 422 if (temps[bl].state == TCG_TEMP_CONST 423 && temps[bh].state == TCG_TEMP_CONST) { 424 uint64_t b = ((uint64_t)temps[bh].val << 32) | (uint32_t)temps[bl].val; 425 426 if (temps[al].state == TCG_TEMP_CONST 427 && temps[ah].state == TCG_TEMP_CONST) { 428 uint64_t a; 429 a = ((uint64_t)temps[ah].val << 32) | (uint32_t)temps[al].val; 430 return do_constant_folding_cond_64(a, b, c); 431 } 432 if (b == 0) { 433 switch (c) { 434 case TCG_COND_LTU: 435 return 0; 436 case TCG_COND_GEU: 437 return 1; 438 default: 439 break; 440 } 441 } 442 } 443 if (temps_are_copies(al, bl) && temps_are_copies(ah, bh)) { 444 return do_constant_folding_cond_eq(c); 445 } 446 return 2; 447 } 448 449 static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2) 450 { 451 TCGArg a1 = *p1, a2 = *p2; 452 int sum = 0; 453 sum += temps[a1].state == TCG_TEMP_CONST; 454 sum -= temps[a2].state == TCG_TEMP_CONST; 455 456 /* Prefer the constant in second argument, and then the form 457 op a, a, b, which is better handled on non-RISC hosts. */ 458 if (sum > 0 || (sum == 0 && dest == a2)) { 459 *p1 = a2; 460 *p2 = a1; 461 return true; 462 } 463 return false; 464 } 465 466 static bool swap_commutative2(TCGArg *p1, TCGArg *p2) 467 { 468 int sum = 0; 469 sum += temps[p1[0]].state == TCG_TEMP_CONST; 470 sum += temps[p1[1]].state == TCG_TEMP_CONST; 471 sum -= temps[p2[0]].state == TCG_TEMP_CONST; 472 sum -= temps[p2[1]].state == TCG_TEMP_CONST; 473 if (sum > 0) { 474 TCGArg t; 475 t = p1[0], p1[0] = p2[0], p2[0] = t; 476 t = p1[1], p1[1] = p2[1], p2[1] = t; 477 return true; 478 } 479 return false; 480 } 481 482 /* Propagate constants and copies, fold constant expressions. */ 483 static TCGArg *tcg_constant_folding(TCGContext *s, uint16_t *tcg_opc_ptr, 484 TCGArg *args, TCGOpDef *tcg_op_defs) 485 { 486 int i, nb_ops, op_index, nb_temps, nb_globals, nb_call_args; 487 tcg_target_ulong mask, affected; 488 TCGOpcode op; 489 const TCGOpDef *def; 490 TCGArg *gen_args; 491 TCGArg tmp; 492 493 /* Array VALS has an element for each temp. 494 If this temp holds a constant then its value is kept in VALS' element. 495 If this temp is a copy of other ones then the other copies are 496 available through the doubly linked circular list. */ 497 498 nb_temps = s->nb_temps; 499 nb_globals = s->nb_globals; 500 reset_all_temps(nb_temps); 501 502 nb_ops = tcg_opc_ptr - s->gen_opc_buf; 503 gen_args = args; 504 for (op_index = 0; op_index < nb_ops; op_index++) { 505 op = s->gen_opc_buf[op_index]; 506 def = &tcg_op_defs[op]; 507 /* Do copy propagation */ 508 if (op == INDEX_op_call) { 509 int nb_oargs = args[0] >> 16; 510 int nb_iargs = args[0] & 0xffff; 511 for (i = nb_oargs + 1; i < nb_oargs + nb_iargs + 1; i++) { 512 if (temps[args[i]].state == TCG_TEMP_COPY) { 513 args[i] = find_better_copy(s, args[i]); 514 } 515 } 516 } else { 517 for (i = def->nb_oargs; i < def->nb_oargs + def->nb_iargs; i++) { 518 if (temps[args[i]].state == TCG_TEMP_COPY) { 519 args[i] = find_better_copy(s, args[i]); 520 } 521 } 522 } 523 524 /* For commutative operations make constant second argument */ 525 switch (op) { 526 CASE_OP_32_64(add): 527 CASE_OP_32_64(mul): 528 CASE_OP_32_64(and): 529 CASE_OP_32_64(or): 530 CASE_OP_32_64(xor): 531 CASE_OP_32_64(eqv): 532 CASE_OP_32_64(nand): 533 CASE_OP_32_64(nor): 534 swap_commutative(args[0], &args[1], &args[2]); 535 break; 536 CASE_OP_32_64(brcond): 537 if (swap_commutative(-1, &args[0], &args[1])) { 538 args[2] = tcg_swap_cond(args[2]); 539 } 540 break; 541 CASE_OP_32_64(setcond): 542 if (swap_commutative(args[0], &args[1], &args[2])) { 543 args[3] = tcg_swap_cond(args[3]); 544 } 545 break; 546 CASE_OP_32_64(movcond): 547 if (swap_commutative(-1, &args[1], &args[2])) { 548 args[5] = tcg_swap_cond(args[5]); 549 } 550 /* For movcond, we canonicalize the "false" input reg to match 551 the destination reg so that the tcg backend can implement 552 a "move if true" operation. */ 553 if (swap_commutative(args[0], &args[4], &args[3])) { 554 args[5] = tcg_invert_cond(args[5]); 555 } 556 break; 557 case INDEX_op_add2_i32: 558 swap_commutative(args[0], &args[2], &args[4]); 559 swap_commutative(args[1], &args[3], &args[5]); 560 break; 561 case INDEX_op_mulu2_i32: 562 swap_commutative(args[0], &args[2], &args[3]); 563 break; 564 case INDEX_op_brcond2_i32: 565 if (swap_commutative2(&args[0], &args[2])) { 566 args[4] = tcg_swap_cond(args[4]); 567 } 568 break; 569 case INDEX_op_setcond2_i32: 570 if (swap_commutative2(&args[1], &args[3])) { 571 args[5] = tcg_swap_cond(args[5]); 572 } 573 break; 574 default: 575 break; 576 } 577 578 /* Simplify expressions for "shift/rot r, 0, a => movi r, 0" */ 579 switch (op) { 580 CASE_OP_32_64(shl): 581 CASE_OP_32_64(shr): 582 CASE_OP_32_64(sar): 583 CASE_OP_32_64(rotl): 584 CASE_OP_32_64(rotr): 585 if (temps[args[1]].state == TCG_TEMP_CONST 586 && temps[args[1]].val == 0) { 587 s->gen_opc_buf[op_index] = op_to_movi(op); 588 tcg_opt_gen_movi(gen_args, args[0], 0); 589 args += 3; 590 gen_args += 2; 591 continue; 592 } 593 break; 594 default: 595 break; 596 } 597 598 /* Simplify expression for "op r, a, 0 => mov r, a" cases */ 599 switch (op) { 600 CASE_OP_32_64(add): 601 CASE_OP_32_64(sub): 602 CASE_OP_32_64(shl): 603 CASE_OP_32_64(shr): 604 CASE_OP_32_64(sar): 605 CASE_OP_32_64(rotl): 606 CASE_OP_32_64(rotr): 607 CASE_OP_32_64(or): 608 CASE_OP_32_64(xor): 609 if (temps[args[1]].state == TCG_TEMP_CONST) { 610 /* Proceed with possible constant folding. */ 611 break; 612 } 613 if (temps[args[2]].state == TCG_TEMP_CONST 614 && temps[args[2]].val == 0) { 615 if (temps_are_copies(args[0], args[1])) { 616 s->gen_opc_buf[op_index] = INDEX_op_nop; 617 } else { 618 s->gen_opc_buf[op_index] = op_to_mov(op); 619 tcg_opt_gen_mov(s, gen_args, args[0], args[1]); 620 gen_args += 2; 621 } 622 args += 3; 623 continue; 624 } 625 break; 626 default: 627 break; 628 } 629 630 /* Simplify using known-zero bits */ 631 mask = -1; 632 affected = -1; 633 switch (op) { 634 CASE_OP_32_64(ext8s): 635 if ((temps[args[1]].mask & 0x80) != 0) { 636 break; 637 } 638 CASE_OP_32_64(ext8u): 639 mask = 0xff; 640 goto and_const; 641 CASE_OP_32_64(ext16s): 642 if ((temps[args[1]].mask & 0x8000) != 0) { 643 break; 644 } 645 CASE_OP_32_64(ext16u): 646 mask = 0xffff; 647 goto and_const; 648 case INDEX_op_ext32s_i64: 649 if ((temps[args[1]].mask & 0x80000000) != 0) { 650 break; 651 } 652 case INDEX_op_ext32u_i64: 653 mask = 0xffffffffU; 654 goto and_const; 655 656 CASE_OP_32_64(and): 657 mask = temps[args[2]].mask; 658 if (temps[args[2]].state == TCG_TEMP_CONST) { 659 and_const: 660 affected = temps[args[1]].mask & ~mask; 661 } 662 mask = temps[args[1]].mask & mask; 663 break; 664 665 CASE_OP_32_64(sar): 666 if (temps[args[2]].state == TCG_TEMP_CONST) { 667 mask = ((tcg_target_long)temps[args[1]].mask 668 >> temps[args[2]].val); 669 } 670 break; 671 672 CASE_OP_32_64(shr): 673 if (temps[args[2]].state == TCG_TEMP_CONST) { 674 mask = temps[args[1]].mask >> temps[args[2]].val; 675 } 676 break; 677 678 CASE_OP_32_64(shl): 679 if (temps[args[2]].state == TCG_TEMP_CONST) { 680 mask = temps[args[1]].mask << temps[args[2]].val; 681 } 682 break; 683 684 CASE_OP_32_64(neg): 685 /* Set to 1 all bits to the left of the rightmost. */ 686 mask = -(temps[args[1]].mask & -temps[args[1]].mask); 687 break; 688 689 CASE_OP_32_64(deposit): 690 tmp = ((1ull << args[4]) - 1); 691 mask = ((temps[args[1]].mask & ~(tmp << args[3])) 692 | ((temps[args[2]].mask & tmp) << args[3])); 693 break; 694 695 CASE_OP_32_64(or): 696 CASE_OP_32_64(xor): 697 mask = temps[args[1]].mask | temps[args[2]].mask; 698 break; 699 700 CASE_OP_32_64(setcond): 701 mask = 1; 702 break; 703 704 CASE_OP_32_64(movcond): 705 mask = temps[args[3]].mask | temps[args[4]].mask; 706 break; 707 708 default: 709 break; 710 } 711 712 if (mask == 0) { 713 assert(def->nb_oargs == 1); 714 s->gen_opc_buf[op_index] = op_to_movi(op); 715 tcg_opt_gen_movi(gen_args, args[0], 0); 716 args += def->nb_oargs + def->nb_iargs + def->nb_cargs; 717 gen_args += 2; 718 continue; 719 } 720 if (affected == 0) { 721 assert(def->nb_oargs == 1); 722 if (temps_are_copies(args[0], args[1])) { 723 s->gen_opc_buf[op_index] = INDEX_op_nop; 724 } else if (temps[args[1]].state != TCG_TEMP_CONST) { 725 s->gen_opc_buf[op_index] = op_to_mov(op); 726 tcg_opt_gen_mov(s, gen_args, args[0], args[1]); 727 gen_args += 2; 728 } else { 729 s->gen_opc_buf[op_index] = op_to_movi(op); 730 tcg_opt_gen_movi(gen_args, args[0], temps[args[1]].val); 731 gen_args += 2; 732 } 733 args += def->nb_iargs + 1; 734 continue; 735 } 736 737 /* Simplify expression for "op r, a, 0 => movi r, 0" cases */ 738 switch (op) { 739 CASE_OP_32_64(and): 740 CASE_OP_32_64(mul): 741 if ((temps[args[2]].state == TCG_TEMP_CONST 742 && temps[args[2]].val == 0)) { 743 s->gen_opc_buf[op_index] = op_to_movi(op); 744 tcg_opt_gen_movi(gen_args, args[0], 0); 745 args += 3; 746 gen_args += 2; 747 continue; 748 } 749 break; 750 default: 751 break; 752 } 753 754 /* Simplify expression for "op r, a, a => mov r, a" cases */ 755 switch (op) { 756 CASE_OP_32_64(or): 757 CASE_OP_32_64(and): 758 if (temps_are_copies(args[1], args[2])) { 759 if (temps_are_copies(args[0], args[1])) { 760 s->gen_opc_buf[op_index] = INDEX_op_nop; 761 } else { 762 s->gen_opc_buf[op_index] = op_to_mov(op); 763 tcg_opt_gen_mov(s, gen_args, args[0], args[1]); 764 gen_args += 2; 765 } 766 args += 3; 767 continue; 768 } 769 break; 770 default: 771 break; 772 } 773 774 /* Simplify expression for "op r, a, a => movi r, 0" cases */ 775 switch (op) { 776 CASE_OP_32_64(sub): 777 CASE_OP_32_64(xor): 778 if (temps_are_copies(args[1], args[2])) { 779 s->gen_opc_buf[op_index] = op_to_movi(op); 780 tcg_opt_gen_movi(gen_args, args[0], 0); 781 gen_args += 2; 782 args += 3; 783 continue; 784 } 785 break; 786 default: 787 break; 788 } 789 790 /* Propagate constants through copy operations and do constant 791 folding. Constants will be substituted to arguments by register 792 allocator where needed and possible. Also detect copies. */ 793 switch (op) { 794 CASE_OP_32_64(mov): 795 if (temps_are_copies(args[0], args[1])) { 796 args += 2; 797 s->gen_opc_buf[op_index] = INDEX_op_nop; 798 break; 799 } 800 if (temps[args[1]].state != TCG_TEMP_CONST) { 801 tcg_opt_gen_mov(s, gen_args, args[0], args[1]); 802 gen_args += 2; 803 args += 2; 804 break; 805 } 806 /* Source argument is constant. Rewrite the operation and 807 let movi case handle it. */ 808 op = op_to_movi(op); 809 s->gen_opc_buf[op_index] = op; 810 args[1] = temps[args[1]].val; 811 /* fallthrough */ 812 CASE_OP_32_64(movi): 813 tcg_opt_gen_movi(gen_args, args[0], args[1]); 814 gen_args += 2; 815 args += 2; 816 break; 817 818 CASE_OP_32_64(not): 819 CASE_OP_32_64(neg): 820 CASE_OP_32_64(ext8s): 821 CASE_OP_32_64(ext8u): 822 CASE_OP_32_64(ext16s): 823 CASE_OP_32_64(ext16u): 824 case INDEX_op_ext32s_i64: 825 case INDEX_op_ext32u_i64: 826 if (temps[args[1]].state == TCG_TEMP_CONST) { 827 s->gen_opc_buf[op_index] = op_to_movi(op); 828 tmp = do_constant_folding(op, temps[args[1]].val, 0); 829 tcg_opt_gen_movi(gen_args, args[0], tmp); 830 gen_args += 2; 831 args += 2; 832 break; 833 } 834 goto do_default; 835 836 CASE_OP_32_64(add): 837 CASE_OP_32_64(sub): 838 CASE_OP_32_64(mul): 839 CASE_OP_32_64(or): 840 CASE_OP_32_64(and): 841 CASE_OP_32_64(xor): 842 CASE_OP_32_64(shl): 843 CASE_OP_32_64(shr): 844 CASE_OP_32_64(sar): 845 CASE_OP_32_64(rotl): 846 CASE_OP_32_64(rotr): 847 CASE_OP_32_64(andc): 848 CASE_OP_32_64(orc): 849 CASE_OP_32_64(eqv): 850 CASE_OP_32_64(nand): 851 CASE_OP_32_64(nor): 852 if (temps[args[1]].state == TCG_TEMP_CONST 853 && temps[args[2]].state == TCG_TEMP_CONST) { 854 s->gen_opc_buf[op_index] = op_to_movi(op); 855 tmp = do_constant_folding(op, temps[args[1]].val, 856 temps[args[2]].val); 857 tcg_opt_gen_movi(gen_args, args[0], tmp); 858 gen_args += 2; 859 args += 3; 860 break; 861 } 862 goto do_default; 863 864 CASE_OP_32_64(deposit): 865 if (temps[args[1]].state == TCG_TEMP_CONST 866 && temps[args[2]].state == TCG_TEMP_CONST) { 867 s->gen_opc_buf[op_index] = op_to_movi(op); 868 tmp = ((1ull << args[4]) - 1); 869 tmp = (temps[args[1]].val & ~(tmp << args[3])) 870 | ((temps[args[2]].val & tmp) << args[3]); 871 tcg_opt_gen_movi(gen_args, args[0], tmp); 872 gen_args += 2; 873 args += 5; 874 break; 875 } 876 goto do_default; 877 878 CASE_OP_32_64(setcond): 879 tmp = do_constant_folding_cond(op, args[1], args[2], args[3]); 880 if (tmp != 2) { 881 s->gen_opc_buf[op_index] = op_to_movi(op); 882 tcg_opt_gen_movi(gen_args, args[0], tmp); 883 gen_args += 2; 884 args += 4; 885 break; 886 } 887 goto do_default; 888 889 CASE_OP_32_64(brcond): 890 tmp = do_constant_folding_cond(op, args[0], args[1], args[2]); 891 if (tmp != 2) { 892 if (tmp) { 893 reset_all_temps(nb_temps); 894 s->gen_opc_buf[op_index] = INDEX_op_br; 895 gen_args[0] = args[3]; 896 gen_args += 1; 897 } else { 898 s->gen_opc_buf[op_index] = INDEX_op_nop; 899 } 900 args += 4; 901 break; 902 } 903 goto do_default; 904 905 CASE_OP_32_64(movcond): 906 tmp = do_constant_folding_cond(op, args[1], args[2], args[5]); 907 if (tmp != 2) { 908 if (temps_are_copies(args[0], args[4-tmp])) { 909 s->gen_opc_buf[op_index] = INDEX_op_nop; 910 } else if (temps[args[4-tmp]].state == TCG_TEMP_CONST) { 911 s->gen_opc_buf[op_index] = op_to_movi(op); 912 tcg_opt_gen_movi(gen_args, args[0], temps[args[4-tmp]].val); 913 gen_args += 2; 914 } else { 915 s->gen_opc_buf[op_index] = op_to_mov(op); 916 tcg_opt_gen_mov(s, gen_args, args[0], args[4-tmp]); 917 gen_args += 2; 918 } 919 args += 6; 920 break; 921 } 922 goto do_default; 923 924 case INDEX_op_add2_i32: 925 case INDEX_op_sub2_i32: 926 if (temps[args[2]].state == TCG_TEMP_CONST 927 && temps[args[3]].state == TCG_TEMP_CONST 928 && temps[args[4]].state == TCG_TEMP_CONST 929 && temps[args[5]].state == TCG_TEMP_CONST) { 930 uint32_t al = temps[args[2]].val; 931 uint32_t ah = temps[args[3]].val; 932 uint32_t bl = temps[args[4]].val; 933 uint32_t bh = temps[args[5]].val; 934 uint64_t a = ((uint64_t)ah << 32) | al; 935 uint64_t b = ((uint64_t)bh << 32) | bl; 936 TCGArg rl, rh; 937 938 if (op == INDEX_op_add2_i32) { 939 a += b; 940 } else { 941 a -= b; 942 } 943 944 /* We emit the extra nop when we emit the add2/sub2. */ 945 assert(s->gen_opc_buf[op_index + 1] == INDEX_op_nop); 946 947 rl = args[0]; 948 rh = args[1]; 949 s->gen_opc_buf[op_index] = INDEX_op_movi_i32; 950 s->gen_opc_buf[++op_index] = INDEX_op_movi_i32; 951 tcg_opt_gen_movi(&gen_args[0], rl, (uint32_t)a); 952 tcg_opt_gen_movi(&gen_args[2], rh, (uint32_t)(a >> 32)); 953 gen_args += 4; 954 args += 6; 955 break; 956 } 957 goto do_default; 958 959 case INDEX_op_mulu2_i32: 960 if (temps[args[2]].state == TCG_TEMP_CONST 961 && temps[args[3]].state == TCG_TEMP_CONST) { 962 uint32_t a = temps[args[2]].val; 963 uint32_t b = temps[args[3]].val; 964 uint64_t r = (uint64_t)a * b; 965 TCGArg rl, rh; 966 967 /* We emit the extra nop when we emit the mulu2. */ 968 assert(s->gen_opc_buf[op_index + 1] == INDEX_op_nop); 969 970 rl = args[0]; 971 rh = args[1]; 972 s->gen_opc_buf[op_index] = INDEX_op_movi_i32; 973 s->gen_opc_buf[++op_index] = INDEX_op_movi_i32; 974 tcg_opt_gen_movi(&gen_args[0], rl, (uint32_t)r); 975 tcg_opt_gen_movi(&gen_args[2], rh, (uint32_t)(r >> 32)); 976 gen_args += 4; 977 args += 4; 978 break; 979 } 980 goto do_default; 981 982 case INDEX_op_brcond2_i32: 983 tmp = do_constant_folding_cond2(&args[0], &args[2], args[4]); 984 if (tmp != 2) { 985 if (tmp) { 986 reset_all_temps(nb_temps); 987 s->gen_opc_buf[op_index] = INDEX_op_br; 988 gen_args[0] = args[5]; 989 gen_args += 1; 990 } else { 991 s->gen_opc_buf[op_index] = INDEX_op_nop; 992 } 993 } else if ((args[4] == TCG_COND_LT || args[4] == TCG_COND_GE) 994 && temps[args[2]].state == TCG_TEMP_CONST 995 && temps[args[3]].state == TCG_TEMP_CONST 996 && temps[args[2]].val == 0 997 && temps[args[3]].val == 0) { 998 /* Simplify LT/GE comparisons vs zero to a single compare 999 vs the high word of the input. */ 1000 reset_all_temps(nb_temps); 1001 s->gen_opc_buf[op_index] = INDEX_op_brcond_i32; 1002 gen_args[0] = args[1]; 1003 gen_args[1] = args[3]; 1004 gen_args[2] = args[4]; 1005 gen_args[3] = args[5]; 1006 gen_args += 4; 1007 } else { 1008 goto do_default; 1009 } 1010 args += 6; 1011 break; 1012 1013 case INDEX_op_setcond2_i32: 1014 tmp = do_constant_folding_cond2(&args[1], &args[3], args[5]); 1015 if (tmp != 2) { 1016 s->gen_opc_buf[op_index] = INDEX_op_movi_i32; 1017 tcg_opt_gen_movi(gen_args, args[0], tmp); 1018 gen_args += 2; 1019 } else if ((args[5] == TCG_COND_LT || args[5] == TCG_COND_GE) 1020 && temps[args[3]].state == TCG_TEMP_CONST 1021 && temps[args[4]].state == TCG_TEMP_CONST 1022 && temps[args[3]].val == 0 1023 && temps[args[4]].val == 0) { 1024 /* Simplify LT/GE comparisons vs zero to a single compare 1025 vs the high word of the input. */ 1026 s->gen_opc_buf[op_index] = INDEX_op_setcond_i32; 1027 gen_args[0] = args[0]; 1028 gen_args[1] = args[2]; 1029 gen_args[2] = args[4]; 1030 gen_args[3] = args[5]; 1031 gen_args += 4; 1032 } else { 1033 goto do_default; 1034 } 1035 args += 6; 1036 break; 1037 1038 case INDEX_op_call: 1039 nb_call_args = (args[0] >> 16) + (args[0] & 0xffff); 1040 if (!(args[nb_call_args + 1] & (TCG_CALL_NO_READ_GLOBALS | 1041 TCG_CALL_NO_WRITE_GLOBALS))) { 1042 for (i = 0; i < nb_globals; i++) { 1043 reset_temp(i); 1044 } 1045 } 1046 for (i = 0; i < (args[0] >> 16); i++) { 1047 reset_temp(args[i + 1]); 1048 } 1049 i = nb_call_args + 3; 1050 while (i) { 1051 *gen_args = *args; 1052 args++; 1053 gen_args++; 1054 i--; 1055 } 1056 break; 1057 1058 default: 1059 do_default: 1060 /* Default case: we know nothing about operation (or were unable 1061 to compute the operation result) so no propagation is done. 1062 We trash everything if the operation is the end of a basic 1063 block, otherwise we only trash the output args. "mask" is 1064 the non-zero bits mask for the first output arg. */ 1065 if (def->flags & TCG_OPF_BB_END) { 1066 reset_all_temps(nb_temps); 1067 } else { 1068 for (i = 0; i < def->nb_oargs; i++) { 1069 reset_temp(args[i]); 1070 } 1071 } 1072 for (i = 0; i < def->nb_args; i++) { 1073 gen_args[i] = args[i]; 1074 } 1075 args += def->nb_args; 1076 gen_args += def->nb_args; 1077 break; 1078 } 1079 } 1080 1081 return gen_args; 1082 } 1083 1084 TCGArg *tcg_optimize(TCGContext *s, uint16_t *tcg_opc_ptr, 1085 TCGArg *args, TCGOpDef *tcg_op_defs) 1086 { 1087 TCGArg *res; 1088 res = tcg_constant_folding(s, tcg_opc_ptr, args, tcg_op_defs); 1089 return res; 1090 } 1091