1 /* LRA (local register allocator) driver and LRA utilities. 2 Copyright (C) 2010-2020 Free Software Foundation, Inc. 3 Contributed by Vladimir Makarov <vmakarov@redhat.com>. 4 5 This file is part of GCC. 6 7 GCC is free software; you can redistribute it and/or modify it under 8 the terms of the GNU General Public License as published by the Free 9 Software Foundation; either version 3, or (at your option) any later 10 version. 11 12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY 13 WARRANTY; without even the implied warranty of MERCHANTABILITY or 14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 15 for more details. 16 17 You should have received a copy of the GNU General Public License 18 along with GCC; see the file COPYING3. If not see 19 <http://www.gnu.org/licenses/>. */ 20 21 22 /* The Local Register Allocator (LRA) is a replacement of former 23 reload pass. It is focused to simplify code solving the reload 24 pass tasks, to make the code maintenance easier, and to implement new 25 perspective optimizations. 26 27 The major LRA design solutions are: 28 o division small manageable, separated sub-tasks 29 o reflection of all transformations and decisions in RTL as more 30 as possible 31 o insn constraints as a primary source of the info (minimizing 32 number of target-depended macros/hooks) 33 34 In brief LRA works by iterative insn process with the final goal is 35 to satisfy all insn and address constraints: 36 o New reload insns (in brief reloads) and reload pseudos might be 37 generated; 38 o Some pseudos might be spilled to assign hard registers to 39 new reload pseudos; 40 o Recalculating spilled pseudo values (rematerialization); 41 o Changing spilled pseudos to stack memory or their equivalences; 42 o Allocation stack memory changes the address displacement and 43 new iteration is needed. 44 45 Here is block diagram of LRA passes: 46 47 ------------------------ 48 --------------- | Undo inheritance for | --------------- 49 | Memory-memory | | spilled pseudos, | | New (and old) | 50 | move coalesce |<---| splits for pseudos got |<-- | pseudos | 51 --------------- | the same hard regs, | | assignment | 52 Start | | and optional reloads | --------------- 53 | | ------------------------ ^ 54 V | ---------------- | 55 ----------- V | Update virtual | | 56 | Remove |----> ------------>| register | | 57 | scratches | ^ | displacements | | 58 ----------- | ---------------- | 59 | | | 60 | V New | 61 | ------------ pseudos ------------------- 62 | |Constraints:| or insns | Inheritance/split | 63 | | RTL |--------->| transformations | 64 | | transfor- | | in EBB scope | 65 | substi- | mations | ------------------- 66 | tutions ------------ 67 | | No change 68 ---------------- V 69 | Spilled pseudo | ------------------- 70 | to memory |<----| Rematerialization | 71 | substitution | ------------------- 72 ---------------- 73 | No susbtitions 74 V 75 ------------------------- 76 | Hard regs substitution, | 77 | devirtalization, and |------> Finish 78 | restoring scratches got | 79 | memory | 80 ------------------------- 81 82 To speed up the process: 83 o We process only insns affected by changes on previous 84 iterations; 85 o We don't use DFA-infrastructure because it results in much slower 86 compiler speed than a special IR described below does; 87 o We use a special insn representation for quick access to insn 88 info which is always *synchronized* with the current RTL; 89 o Insn IR is minimized by memory. It is divided on three parts: 90 o one specific for each insn in RTL (only operand locations); 91 o one common for all insns in RTL with the same insn code 92 (different operand attributes from machine descriptions); 93 o one oriented for maintenance of live info (list of pseudos). 94 o Pseudo data: 95 o all insns where the pseudo is referenced; 96 o live info (conflicting hard regs, live ranges, # of 97 references etc); 98 o data used for assigning (preferred hard regs, costs etc). 99 100 This file contains LRA driver, LRA utility functions and data, and 101 code for dealing with scratches. */ 102 103 #include "config.h" 104 #include "system.h" 105 #include "coretypes.h" 106 #include "backend.h" 107 #include "target.h" 108 #include "rtl.h" 109 #include "tree.h" 110 #include "predict.h" 111 #include "df.h" 112 #include "memmodel.h" 113 #include "tm_p.h" 114 #include "optabs.h" 115 #include "regs.h" 116 #include "ira.h" 117 #include "recog.h" 118 #include "expr.h" 119 #include "cfgrtl.h" 120 #include "cfgbuild.h" 121 #include "lra.h" 122 #include "lra-int.h" 123 #include "print-rtl.h" 124 #include "function-abi.h" 125 126 /* Dump bitmap SET with TITLE and BB INDEX. */ 127 void 128 lra_dump_bitmap_with_title (const char *title, bitmap set, int index) 129 { 130 unsigned int i; 131 int count; 132 bitmap_iterator bi; 133 static const int max_nums_on_line = 10; 134 135 if (bitmap_empty_p (set)) 136 return; 137 fprintf (lra_dump_file, " %s %d:", title, index); 138 fprintf (lra_dump_file, "\n"); 139 count = max_nums_on_line + 1; 140 EXECUTE_IF_SET_IN_BITMAP (set, 0, i, bi) 141 { 142 if (count > max_nums_on_line) 143 { 144 fprintf (lra_dump_file, "\n "); 145 count = 0; 146 } 147 fprintf (lra_dump_file, " %4u", i); 148 count++; 149 } 150 fprintf (lra_dump_file, "\n"); 151 } 152 153 /* Hard registers currently not available for allocation. It can 154 changed after some hard registers become not eliminable. */ 155 HARD_REG_SET lra_no_alloc_regs; 156 157 static int get_new_reg_value (void); 158 static void expand_reg_info (void); 159 static void invalidate_insn_recog_data (int); 160 static int get_insn_freq (rtx_insn *); 161 static void invalidate_insn_data_regno_info (lra_insn_recog_data_t, 162 rtx_insn *, int); 163 static void remove_scratches_1 (rtx_insn *); 164 165 /* Expand all regno related info needed for LRA. */ 166 static void 167 expand_reg_data (int old) 168 { 169 resize_reg_info (); 170 expand_reg_info (); 171 ira_expand_reg_equiv (); 172 for (int i = (int) max_reg_num () - 1; i >= old; i--) 173 lra_change_class (i, ALL_REGS, " Set", true); 174 } 175 176 /* Create and return a new reg of ORIGINAL mode. If ORIGINAL is NULL 177 or of VOIDmode, use MD_MODE for the new reg. Initialize its 178 register class to RCLASS. Print message about assigning class 179 RCLASS containing new register name TITLE unless it is NULL. Use 180 attributes of ORIGINAL if it is a register. The created register 181 will have unique held value. */ 182 rtx 183 lra_create_new_reg_with_unique_value (machine_mode md_mode, rtx original, 184 enum reg_class rclass, const char *title) 185 { 186 machine_mode mode; 187 rtx new_reg; 188 189 if (original == NULL_RTX || (mode = GET_MODE (original)) == VOIDmode) 190 mode = md_mode; 191 lra_assert (mode != VOIDmode); 192 new_reg = gen_reg_rtx (mode); 193 if (original == NULL_RTX || ! REG_P (original)) 194 { 195 if (lra_dump_file != NULL) 196 fprintf (lra_dump_file, " Creating newreg=%i", REGNO (new_reg)); 197 } 198 else 199 { 200 if (ORIGINAL_REGNO (original) >= FIRST_PSEUDO_REGISTER) 201 ORIGINAL_REGNO (new_reg) = ORIGINAL_REGNO (original); 202 REG_USERVAR_P (new_reg) = REG_USERVAR_P (original); 203 REG_POINTER (new_reg) = REG_POINTER (original); 204 REG_ATTRS (new_reg) = REG_ATTRS (original); 205 if (lra_dump_file != NULL) 206 fprintf (lra_dump_file, " Creating newreg=%i from oldreg=%i", 207 REGNO (new_reg), REGNO (original)); 208 } 209 if (lra_dump_file != NULL) 210 { 211 if (title != NULL) 212 fprintf (lra_dump_file, ", assigning class %s to%s%s r%d", 213 reg_class_names[rclass], *title == '\0' ? "" : " ", 214 title, REGNO (new_reg)); 215 fprintf (lra_dump_file, "\n"); 216 } 217 expand_reg_data (max_reg_num ()); 218 setup_reg_classes (REGNO (new_reg), rclass, NO_REGS, rclass); 219 return new_reg; 220 } 221 222 /* Analogous to the previous function but also inherits value of 223 ORIGINAL. */ 224 rtx 225 lra_create_new_reg (machine_mode md_mode, rtx original, 226 enum reg_class rclass, const char *title) 227 { 228 rtx new_reg; 229 230 new_reg 231 = lra_create_new_reg_with_unique_value (md_mode, original, rclass, title); 232 if (original != NULL_RTX && REG_P (original)) 233 lra_assign_reg_val (REGNO (original), REGNO (new_reg)); 234 return new_reg; 235 } 236 237 /* Set up for REGNO unique hold value. */ 238 void 239 lra_set_regno_unique_value (int regno) 240 { 241 lra_reg_info[regno].val = get_new_reg_value (); 242 } 243 244 /* Invalidate INSN related info used by LRA. The info should never be 245 used after that. */ 246 void 247 lra_invalidate_insn_data (rtx_insn *insn) 248 { 249 lra_invalidate_insn_regno_info (insn); 250 invalidate_insn_recog_data (INSN_UID (insn)); 251 } 252 253 /* Mark INSN deleted and invalidate the insn related info used by 254 LRA. */ 255 void 256 lra_set_insn_deleted (rtx_insn *insn) 257 { 258 lra_invalidate_insn_data (insn); 259 SET_INSN_DELETED (insn); 260 } 261 262 /* Delete an unneeded INSN and any previous insns who sole purpose is 263 loading data that is dead in INSN. */ 264 void 265 lra_delete_dead_insn (rtx_insn *insn) 266 { 267 rtx_insn *prev = prev_real_insn (insn); 268 rtx prev_dest; 269 270 /* If the previous insn sets a register that dies in our insn, 271 delete it too. */ 272 if (prev && GET_CODE (PATTERN (prev)) == SET 273 && (prev_dest = SET_DEST (PATTERN (prev)), REG_P (prev_dest)) 274 && reg_mentioned_p (prev_dest, PATTERN (insn)) 275 && find_regno_note (insn, REG_DEAD, REGNO (prev_dest)) 276 && ! side_effects_p (SET_SRC (PATTERN (prev)))) 277 lra_delete_dead_insn (prev); 278 279 lra_set_insn_deleted (insn); 280 } 281 282 /* Emit insn x = y + z. Return NULL if we failed to do it. 283 Otherwise, return the insn. We don't use gen_add3_insn as it might 284 clobber CC. */ 285 static rtx_insn * 286 emit_add3_insn (rtx x, rtx y, rtx z) 287 { 288 rtx_insn *last; 289 290 last = get_last_insn (); 291 292 if (have_addptr3_insn (x, y, z)) 293 { 294 rtx_insn *insn = gen_addptr3_insn (x, y, z); 295 296 /* If the target provides an "addptr" pattern it hopefully does 297 for a reason. So falling back to the normal add would be 298 a bug. */ 299 lra_assert (insn != NULL_RTX); 300 emit_insn (insn); 301 return insn; 302 } 303 304 rtx_insn *insn = emit_insn (gen_rtx_SET (x, gen_rtx_PLUS (GET_MODE (y), 305 y, z))); 306 if (recog_memoized (insn) < 0) 307 { 308 delete_insns_since (last); 309 insn = NULL; 310 } 311 return insn; 312 } 313 314 /* Emit insn x = x + y. Return the insn. We use gen_add2_insn as the 315 last resort. */ 316 static rtx_insn * 317 emit_add2_insn (rtx x, rtx y) 318 { 319 rtx_insn *insn = emit_add3_insn (x, x, y); 320 if (insn == NULL_RTX) 321 { 322 insn = gen_add2_insn (x, y); 323 if (insn != NULL_RTX) 324 emit_insn (insn); 325 } 326 return insn; 327 } 328 329 /* Target checks operands through operand predicates to recognize an 330 insn. We should have a special precaution to generate add insns 331 which are frequent results of elimination. 332 333 Emit insns for x = y + z. X can be used to store intermediate 334 values and should be not in Y and Z when we use X to store an 335 intermediate value. Y + Z should form [base] [+ index[ * scale]] [ 336 + disp] where base and index are registers, disp and scale are 337 constants. Y should contain base if it is present, Z should 338 contain disp if any. index[*scale] can be part of Y or Z. */ 339 void 340 lra_emit_add (rtx x, rtx y, rtx z) 341 { 342 int old; 343 rtx_insn *last; 344 rtx a1, a2, base, index, disp, scale, index_scale; 345 bool ok_p; 346 347 rtx_insn *add3_insn = emit_add3_insn (x, y, z); 348 old = max_reg_num (); 349 if (add3_insn != NULL) 350 ; 351 else 352 { 353 disp = a2 = NULL_RTX; 354 if (GET_CODE (y) == PLUS) 355 { 356 a1 = XEXP (y, 0); 357 a2 = XEXP (y, 1); 358 disp = z; 359 } 360 else 361 { 362 a1 = y; 363 if (CONSTANT_P (z)) 364 disp = z; 365 else 366 a2 = z; 367 } 368 index_scale = scale = NULL_RTX; 369 if (GET_CODE (a1) == MULT) 370 { 371 index_scale = a1; 372 index = XEXP (a1, 0); 373 scale = XEXP (a1, 1); 374 base = a2; 375 } 376 else if (a2 != NULL_RTX && GET_CODE (a2) == MULT) 377 { 378 index_scale = a2; 379 index = XEXP (a2, 0); 380 scale = XEXP (a2, 1); 381 base = a1; 382 } 383 else 384 { 385 base = a1; 386 index = a2; 387 } 388 if ((base != NULL_RTX && ! (REG_P (base) || GET_CODE (base) == SUBREG)) 389 || (index != NULL_RTX 390 && ! (REG_P (index) || GET_CODE (index) == SUBREG)) 391 || (disp != NULL_RTX && ! CONSTANT_P (disp)) 392 || (scale != NULL_RTX && ! CONSTANT_P (scale))) 393 { 394 /* Probably we have no 3 op add. Last chance is to use 2-op 395 add insn. To succeed, don't move Z to X as an address 396 segment always comes in Y. Otherwise, we might fail when 397 adding the address segment to register. */ 398 lra_assert (x != y && x != z); 399 emit_move_insn (x, y); 400 rtx_insn *insn = emit_add2_insn (x, z); 401 lra_assert (insn != NULL_RTX); 402 } 403 else 404 { 405 if (index_scale == NULL_RTX) 406 index_scale = index; 407 if (disp == NULL_RTX) 408 { 409 /* Generate x = index_scale; x = x + base. */ 410 lra_assert (index_scale != NULL_RTX && base != NULL_RTX); 411 emit_move_insn (x, index_scale); 412 rtx_insn *insn = emit_add2_insn (x, base); 413 lra_assert (insn != NULL_RTX); 414 } 415 else if (scale == NULL_RTX) 416 { 417 /* Try x = base + disp. */ 418 lra_assert (base != NULL_RTX); 419 last = get_last_insn (); 420 rtx_insn *move_insn = 421 emit_move_insn (x, gen_rtx_PLUS (GET_MODE (base), base, disp)); 422 if (recog_memoized (move_insn) < 0) 423 { 424 delete_insns_since (last); 425 /* Generate x = disp; x = x + base. */ 426 emit_move_insn (x, disp); 427 rtx_insn *add2_insn = emit_add2_insn (x, base); 428 lra_assert (add2_insn != NULL_RTX); 429 } 430 /* Generate x = x + index. */ 431 if (index != NULL_RTX) 432 { 433 rtx_insn *insn = emit_add2_insn (x, index); 434 lra_assert (insn != NULL_RTX); 435 } 436 } 437 else 438 { 439 /* Try x = index_scale; x = x + disp; x = x + base. */ 440 last = get_last_insn (); 441 rtx_insn *move_insn = emit_move_insn (x, index_scale); 442 ok_p = false; 443 if (recog_memoized (move_insn) >= 0) 444 { 445 rtx_insn *insn = emit_add2_insn (x, disp); 446 if (insn != NULL_RTX) 447 { 448 if (base == NULL_RTX) 449 ok_p = true; 450 else 451 { 452 insn = emit_add2_insn (x, base); 453 if (insn != NULL_RTX) 454 ok_p = true; 455 } 456 } 457 } 458 if (! ok_p) 459 { 460 rtx_insn *insn; 461 462 delete_insns_since (last); 463 /* Generate x = disp; x = x + base; x = x + index_scale. */ 464 emit_move_insn (x, disp); 465 if (base != NULL_RTX) 466 { 467 insn = emit_add2_insn (x, base); 468 lra_assert (insn != NULL_RTX); 469 } 470 insn = emit_add2_insn (x, index_scale); 471 lra_assert (insn != NULL_RTX); 472 } 473 } 474 } 475 } 476 /* Functions emit_... can create pseudos -- so expand the pseudo 477 data. */ 478 if (old != max_reg_num ()) 479 expand_reg_data (old); 480 } 481 482 /* The number of emitted reload insns so far. */ 483 int lra_curr_reload_num; 484 485 /* Emit x := y, processing special case when y = u + v or y = u + v * 486 scale + w through emit_add (Y can be an address which is base + 487 index reg * scale + displacement in general case). X may be used 488 as intermediate result therefore it should be not in Y. */ 489 void 490 lra_emit_move (rtx x, rtx y) 491 { 492 int old; 493 rtx_insn *insn; 494 495 if (GET_CODE (y) != PLUS) 496 { 497 if (rtx_equal_p (x, y)) 498 return; 499 old = max_reg_num (); 500 501 insn = (GET_CODE (x) != STRICT_LOW_PART 502 ? emit_move_insn (x, y) : emit_insn (gen_rtx_SET (x, y))); 503 /* The move pattern may require scratch registers, so convert them 504 into real registers now. */ 505 if (insn != NULL_RTX) 506 remove_scratches_1 (insn); 507 if (REG_P (x)) 508 lra_reg_info[ORIGINAL_REGNO (x)].last_reload = ++lra_curr_reload_num; 509 /* Function emit_move can create pseudos -- so expand the pseudo 510 data. */ 511 if (old != max_reg_num ()) 512 expand_reg_data (old); 513 return; 514 } 515 lra_emit_add (x, XEXP (y, 0), XEXP (y, 1)); 516 } 517 518 /* Update insn operands which are duplication of operands whose 519 numbers are in array of NOPS (with end marker -1). The insn is 520 represented by its LRA internal representation ID. */ 521 void 522 lra_update_dups (lra_insn_recog_data_t id, signed char *nops) 523 { 524 int i, j, nop; 525 struct lra_static_insn_data *static_id = id->insn_static_data; 526 527 for (i = 0; i < static_id->n_dups; i++) 528 for (j = 0; (nop = nops[j]) >= 0; j++) 529 if (static_id->dup_num[i] == nop) 530 *id->dup_loc[i] = *id->operand_loc[nop]; 531 } 532 533 534 535 /* This page contains code dealing with info about registers in the 536 insns. */ 537 538 /* Pools for insn reg info. */ 539 object_allocator<lra_insn_reg> lra_insn_reg_pool ("insn regs"); 540 541 /* Create LRA insn related info about a reference to REGNO in INSN 542 with TYPE (in/out/inout), biggest reference mode MODE, flag that it 543 is reference through subreg (SUBREG_P), and reference to the next 544 insn reg info (NEXT). If REGNO can be early clobbered, 545 alternatives in which it can be early clobbered are given by 546 EARLY_CLOBBER_ALTS. */ 547 static struct lra_insn_reg * 548 new_insn_reg (rtx_insn *insn, int regno, enum op_type type, 549 machine_mode mode, bool subreg_p, 550 alternative_mask early_clobber_alts, 551 struct lra_insn_reg *next) 552 { 553 lra_insn_reg *ir = lra_insn_reg_pool.allocate (); 554 ir->type = type; 555 ir->biggest_mode = mode; 556 if (NONDEBUG_INSN_P (insn) 557 && partial_subreg_p (lra_reg_info[regno].biggest_mode, mode)) 558 lra_reg_info[regno].biggest_mode = mode; 559 ir->subreg_p = subreg_p; 560 ir->early_clobber_alts = early_clobber_alts; 561 ir->regno = regno; 562 ir->next = next; 563 return ir; 564 } 565 566 /* Free insn reg info list IR. */ 567 static void 568 free_insn_regs (struct lra_insn_reg *ir) 569 { 570 struct lra_insn_reg *next_ir; 571 572 for (; ir != NULL; ir = next_ir) 573 { 574 next_ir = ir->next; 575 lra_insn_reg_pool.remove (ir); 576 } 577 } 578 579 /* Finish pool for insn reg info. */ 580 static void 581 finish_insn_regs (void) 582 { 583 lra_insn_reg_pool.release (); 584 } 585 586 587 588 /* This page contains code dealing LRA insn info (or in other words 589 LRA internal insn representation). */ 590 591 /* Map INSN_CODE -> the static insn data. This info is valid during 592 all translation unit. */ 593 struct lra_static_insn_data *insn_code_data[NUM_INSN_CODES]; 594 595 /* Debug insns are represented as a special insn with one input 596 operand which is RTL expression in var_location. */ 597 598 /* The following data are used as static insn operand data for all 599 debug insns. If structure lra_operand_data is changed, the 600 initializer should be changed too. */ 601 static struct lra_operand_data debug_operand_data = 602 { 603 NULL, /* alternative */ 604 0, /* early_clobber_alts */ 605 E_VOIDmode, /* We are not interesting in the operand mode. */ 606 OP_IN, 607 0, 0, 0 608 }; 609 610 /* The following data are used as static insn data for all debug 611 bind insns. If structure lra_static_insn_data is changed, the 612 initializer should be changed too. */ 613 static struct lra_static_insn_data debug_bind_static_data = 614 { 615 &debug_operand_data, 616 0, /* Duplication operands #. */ 617 -1, /* Commutative operand #. */ 618 1, /* Operands #. There is only one operand which is debug RTL 619 expression. */ 620 0, /* Duplications #. */ 621 0, /* Alternatives #. We are not interesting in alternatives 622 because we does not proceed debug_insns for reloads. */ 623 NULL, /* Hard registers referenced in machine description. */ 624 NULL /* Descriptions of operands in alternatives. */ 625 }; 626 627 /* The following data are used as static insn data for all debug 628 marker insns. If structure lra_static_insn_data is changed, the 629 initializer should be changed too. */ 630 static struct lra_static_insn_data debug_marker_static_data = 631 { 632 &debug_operand_data, 633 0, /* Duplication operands #. */ 634 -1, /* Commutative operand #. */ 635 0, /* Operands #. There isn't any operand. */ 636 0, /* Duplications #. */ 637 0, /* Alternatives #. We are not interesting in alternatives 638 because we does not proceed debug_insns for reloads. */ 639 NULL, /* Hard registers referenced in machine description. */ 640 NULL /* Descriptions of operands in alternatives. */ 641 }; 642 643 /* Called once per compiler work to initialize some LRA data related 644 to insns. */ 645 static void 646 init_insn_code_data_once (void) 647 { 648 memset (insn_code_data, 0, sizeof (insn_code_data)); 649 } 650 651 /* Called once per compiler work to finalize some LRA data related to 652 insns. */ 653 static void 654 finish_insn_code_data_once (void) 655 { 656 for (unsigned int i = 0; i < NUM_INSN_CODES; i++) 657 { 658 if (insn_code_data[i] != NULL) 659 { 660 free (insn_code_data[i]); 661 insn_code_data[i] = NULL; 662 } 663 } 664 } 665 666 /* Return static insn data, allocate and setup if necessary. Although 667 dup_num is static data (it depends only on icode), to set it up we 668 need to extract insn first. So recog_data should be valid for 669 normal insn (ICODE >= 0) before the call. */ 670 static struct lra_static_insn_data * 671 get_static_insn_data (int icode, int nop, int ndup, int nalt) 672 { 673 struct lra_static_insn_data *data; 674 size_t n_bytes; 675 676 lra_assert (icode < (int) NUM_INSN_CODES); 677 if (icode >= 0 && (data = insn_code_data[icode]) != NULL) 678 return data; 679 lra_assert (nop >= 0 && ndup >= 0 && nalt >= 0); 680 n_bytes = sizeof (struct lra_static_insn_data) 681 + sizeof (struct lra_operand_data) * nop 682 + sizeof (int) * ndup; 683 data = XNEWVAR (struct lra_static_insn_data, n_bytes); 684 data->operand_alternative = NULL; 685 data->n_operands = nop; 686 data->n_dups = ndup; 687 data->n_alternatives = nalt; 688 data->operand = ((struct lra_operand_data *) 689 ((char *) data + sizeof (struct lra_static_insn_data))); 690 data->dup_num = ((int *) ((char *) data->operand 691 + sizeof (struct lra_operand_data) * nop)); 692 if (icode >= 0) 693 { 694 int i; 695 696 insn_code_data[icode] = data; 697 for (i = 0; i < nop; i++) 698 { 699 data->operand[i].constraint 700 = insn_data[icode].operand[i].constraint; 701 data->operand[i].mode = insn_data[icode].operand[i].mode; 702 data->operand[i].strict_low = insn_data[icode].operand[i].strict_low; 703 data->operand[i].is_operator 704 = insn_data[icode].operand[i].is_operator; 705 data->operand[i].type 706 = (data->operand[i].constraint[0] == '=' ? OP_OUT 707 : data->operand[i].constraint[0] == '+' ? OP_INOUT 708 : OP_IN); 709 data->operand[i].is_address = false; 710 } 711 for (i = 0; i < ndup; i++) 712 data->dup_num[i] = recog_data.dup_num[i]; 713 } 714 return data; 715 } 716 717 /* The current length of the following array. */ 718 int lra_insn_recog_data_len; 719 720 /* Map INSN_UID -> the insn recog data (NULL if unknown). */ 721 lra_insn_recog_data_t *lra_insn_recog_data; 722 723 /* Alloc pool we allocate entries for lra_insn_recog_data from. */ 724 static object_allocator<class lra_insn_recog_data> 725 lra_insn_recog_data_pool ("insn recog data pool"); 726 727 /* Initialize LRA data about insns. */ 728 static void 729 init_insn_recog_data (void) 730 { 731 lra_insn_recog_data_len = 0; 732 lra_insn_recog_data = NULL; 733 } 734 735 /* Expand, if necessary, LRA data about insns. */ 736 static void 737 check_and_expand_insn_recog_data (int index) 738 { 739 int i, old; 740 741 if (lra_insn_recog_data_len > index) 742 return; 743 old = lra_insn_recog_data_len; 744 lra_insn_recog_data_len = index * 3 / 2 + 1; 745 lra_insn_recog_data = XRESIZEVEC (lra_insn_recog_data_t, 746 lra_insn_recog_data, 747 lra_insn_recog_data_len); 748 for (i = old; i < lra_insn_recog_data_len; i++) 749 lra_insn_recog_data[i] = NULL; 750 } 751 752 /* Finish LRA DATA about insn. */ 753 static void 754 free_insn_recog_data (lra_insn_recog_data_t data) 755 { 756 if (data->operand_loc != NULL) 757 free (data->operand_loc); 758 if (data->dup_loc != NULL) 759 free (data->dup_loc); 760 if (data->arg_hard_regs != NULL) 761 free (data->arg_hard_regs); 762 if (data->icode < 0 && NONDEBUG_INSN_P (data->insn)) 763 { 764 if (data->insn_static_data->operand_alternative != NULL) 765 free (const_cast <operand_alternative *> 766 (data->insn_static_data->operand_alternative)); 767 free_insn_regs (data->insn_static_data->hard_regs); 768 free (data->insn_static_data); 769 } 770 free_insn_regs (data->regs); 771 data->regs = NULL; 772 lra_insn_recog_data_pool.remove (data); 773 } 774 775 /* Pools for copies. */ 776 static object_allocator<lra_copy> lra_copy_pool ("lra copies"); 777 778 /* Finish LRA data about all insns. */ 779 static void 780 finish_insn_recog_data (void) 781 { 782 int i; 783 lra_insn_recog_data_t data; 784 785 for (i = 0; i < lra_insn_recog_data_len; i++) 786 if ((data = lra_insn_recog_data[i]) != NULL) 787 free_insn_recog_data (data); 788 finish_insn_regs (); 789 lra_copy_pool.release (); 790 lra_insn_reg_pool.release (); 791 lra_insn_recog_data_pool.release (); 792 free (lra_insn_recog_data); 793 } 794 795 /* Setup info about operands in alternatives of LRA DATA of insn. */ 796 static void 797 setup_operand_alternative (lra_insn_recog_data_t data, 798 const operand_alternative *op_alt) 799 { 800 int i, j, nop, nalt; 801 int icode = data->icode; 802 struct lra_static_insn_data *static_data = data->insn_static_data; 803 804 static_data->commutative = -1; 805 nop = static_data->n_operands; 806 nalt = static_data->n_alternatives; 807 static_data->operand_alternative = op_alt; 808 for (i = 0; i < nop; i++) 809 { 810 static_data->operand[i].early_clobber_alts = 0; 811 static_data->operand[i].is_address = false; 812 if (static_data->operand[i].constraint[0] == '%') 813 { 814 /* We currently only support one commutative pair of operands. */ 815 if (static_data->commutative < 0) 816 static_data->commutative = i; 817 else 818 lra_assert (icode < 0); /* Asm */ 819 /* The last operand should not be marked commutative. */ 820 lra_assert (i != nop - 1); 821 } 822 } 823 for (j = 0; j < nalt; j++) 824 for (i = 0; i < nop; i++, op_alt++) 825 { 826 if (op_alt->earlyclobber) 827 static_data->operand[i].early_clobber_alts |= (alternative_mask) 1 << j; 828 static_data->operand[i].is_address |= op_alt->is_address; 829 } 830 } 831 832 /* Recursively process X and collect info about registers, which are 833 not the insn operands, in X with TYPE (in/out/inout) and flag that 834 it is early clobbered in the insn (EARLY_CLOBBER) and add the info 835 to LIST. X is a part of insn given by DATA. Return the result 836 list. */ 837 static struct lra_insn_reg * 838 collect_non_operand_hard_regs (rtx_insn *insn, rtx *x, 839 lra_insn_recog_data_t data, 840 struct lra_insn_reg *list, 841 enum op_type type, bool early_clobber) 842 { 843 int i, j, regno, last; 844 bool subreg_p; 845 machine_mode mode; 846 struct lra_insn_reg *curr; 847 rtx op = *x; 848 enum rtx_code code = GET_CODE (op); 849 const char *fmt = GET_RTX_FORMAT (code); 850 851 for (i = 0; i < data->insn_static_data->n_operands; i++) 852 if (! data->insn_static_data->operand[i].is_operator 853 && x == data->operand_loc[i]) 854 /* It is an operand loc. Stop here. */ 855 return list; 856 for (i = 0; i < data->insn_static_data->n_dups; i++) 857 if (x == data->dup_loc[i]) 858 /* It is a dup loc. Stop here. */ 859 return list; 860 mode = GET_MODE (op); 861 subreg_p = false; 862 if (code == SUBREG) 863 { 864 mode = wider_subreg_mode (op); 865 if (read_modify_subreg_p (op)) 866 subreg_p = true; 867 op = SUBREG_REG (op); 868 code = GET_CODE (op); 869 } 870 if (REG_P (op)) 871 { 872 if ((regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER) 873 return list; 874 /* Process all regs even unallocatable ones as we need info 875 about all regs for rematerialization pass. */ 876 for (last = end_hard_regno (mode, regno); regno < last; regno++) 877 { 878 for (curr = list; curr != NULL; curr = curr->next) 879 if (curr->regno == regno && curr->subreg_p == subreg_p 880 && curr->biggest_mode == mode) 881 { 882 if (curr->type != type) 883 curr->type = OP_INOUT; 884 if (early_clobber) 885 curr->early_clobber_alts = ALL_ALTERNATIVES; 886 break; 887 } 888 if (curr == NULL) 889 { 890 /* This is a new hard regno or the info cannot be 891 integrated into the found structure. */ 892 #ifdef STACK_REGS 893 early_clobber 894 = (early_clobber 895 /* This clobber is to inform popping floating 896 point stack only. */ 897 && ! (FIRST_STACK_REG <= regno 898 && regno <= LAST_STACK_REG)); 899 #endif 900 list = new_insn_reg (data->insn, regno, type, mode, subreg_p, 901 early_clobber ? ALL_ALTERNATIVES : 0, list); 902 } 903 } 904 return list; 905 } 906 switch (code) 907 { 908 case SET: 909 list = collect_non_operand_hard_regs (insn, &SET_DEST (op), data, 910 list, OP_OUT, false); 911 list = collect_non_operand_hard_regs (insn, &SET_SRC (op), data, 912 list, OP_IN, false); 913 break; 914 case CLOBBER: 915 /* We treat clobber of non-operand hard registers as early clobber. */ 916 list = collect_non_operand_hard_regs (insn, &XEXP (op, 0), data, 917 list, OP_OUT, true); 918 break; 919 case PRE_INC: case PRE_DEC: case POST_INC: case POST_DEC: 920 list = collect_non_operand_hard_regs (insn, &XEXP (op, 0), data, 921 list, OP_INOUT, false); 922 break; 923 case PRE_MODIFY: case POST_MODIFY: 924 list = collect_non_operand_hard_regs (insn, &XEXP (op, 0), data, 925 list, OP_INOUT, false); 926 list = collect_non_operand_hard_regs (insn, &XEXP (op, 1), data, 927 list, OP_IN, false); 928 break; 929 default: 930 fmt = GET_RTX_FORMAT (code); 931 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) 932 { 933 if (fmt[i] == 'e') 934 list = collect_non_operand_hard_regs (insn, &XEXP (op, i), data, 935 list, OP_IN, false); 936 else if (fmt[i] == 'E') 937 for (j = XVECLEN (op, i) - 1; j >= 0; j--) 938 list = collect_non_operand_hard_regs (insn, &XVECEXP (op, i, j), 939 data, list, OP_IN, false); 940 } 941 } 942 return list; 943 } 944 945 /* Set up and return info about INSN. Set up the info if it is not set up 946 yet. */ 947 lra_insn_recog_data_t 948 lra_set_insn_recog_data (rtx_insn *insn) 949 { 950 lra_insn_recog_data_t data; 951 int i, n, icode; 952 rtx **locs; 953 unsigned int uid = INSN_UID (insn); 954 struct lra_static_insn_data *insn_static_data; 955 956 check_and_expand_insn_recog_data (uid); 957 if (DEBUG_INSN_P (insn)) 958 icode = -1; 959 else 960 { 961 icode = INSN_CODE (insn); 962 if (icode < 0) 963 /* It might be a new simple insn which is not recognized yet. */ 964 INSN_CODE (insn) = icode = recog_memoized (insn); 965 } 966 data = lra_insn_recog_data_pool.allocate (); 967 lra_insn_recog_data[uid] = data; 968 data->insn = insn; 969 data->used_insn_alternative = LRA_UNKNOWN_ALT; 970 data->icode = icode; 971 data->regs = NULL; 972 if (DEBUG_INSN_P (insn)) 973 { 974 data->dup_loc = NULL; 975 data->arg_hard_regs = NULL; 976 data->preferred_alternatives = ALL_ALTERNATIVES; 977 if (DEBUG_BIND_INSN_P (insn)) 978 { 979 data->insn_static_data = &debug_bind_static_data; 980 data->operand_loc = XNEWVEC (rtx *, 1); 981 data->operand_loc[0] = &INSN_VAR_LOCATION_LOC (insn); 982 } 983 else if (DEBUG_MARKER_INSN_P (insn)) 984 { 985 data->insn_static_data = &debug_marker_static_data; 986 data->operand_loc = NULL; 987 } 988 return data; 989 } 990 if (icode < 0) 991 { 992 int nop, nalt; 993 machine_mode operand_mode[MAX_RECOG_OPERANDS]; 994 const char *constraints[MAX_RECOG_OPERANDS]; 995 996 nop = asm_noperands (PATTERN (insn)); 997 data->operand_loc = data->dup_loc = NULL; 998 nalt = 1; 999 if (nop < 0) 1000 { 1001 /* It is a special insn like USE or CLOBBER. We should 1002 recognize any regular insn otherwise LRA can do nothing 1003 with this insn. */ 1004 gcc_assert (GET_CODE (PATTERN (insn)) == USE 1005 || GET_CODE (PATTERN (insn)) == CLOBBER 1006 || GET_CODE (PATTERN (insn)) == ASM_INPUT); 1007 data->insn_static_data = insn_static_data 1008 = get_static_insn_data (-1, 0, 0, nalt); 1009 } 1010 else 1011 { 1012 /* expand_asm_operands makes sure there aren't too many 1013 operands. */ 1014 lra_assert (nop <= MAX_RECOG_OPERANDS); 1015 if (nop != 0) 1016 data->operand_loc = XNEWVEC (rtx *, nop); 1017 /* Now get the operand values and constraints out of the 1018 insn. */ 1019 decode_asm_operands (PATTERN (insn), NULL, 1020 data->operand_loc, 1021 constraints, operand_mode, NULL); 1022 if (nop > 0) 1023 for (const char *p =constraints[0]; *p; p++) 1024 nalt += *p == ','; 1025 data->insn_static_data = insn_static_data 1026 = get_static_insn_data (-1, nop, 0, nalt); 1027 for (i = 0; i < nop; i++) 1028 { 1029 insn_static_data->operand[i].mode = operand_mode[i]; 1030 insn_static_data->operand[i].constraint = constraints[i]; 1031 insn_static_data->operand[i].strict_low = false; 1032 insn_static_data->operand[i].is_operator = false; 1033 insn_static_data->operand[i].is_address = false; 1034 } 1035 } 1036 for (i = 0; i < insn_static_data->n_operands; i++) 1037 insn_static_data->operand[i].type 1038 = (insn_static_data->operand[i].constraint[0] == '=' ? OP_OUT 1039 : insn_static_data->operand[i].constraint[0] == '+' ? OP_INOUT 1040 : OP_IN); 1041 data->preferred_alternatives = ALL_ALTERNATIVES; 1042 if (nop > 0) 1043 { 1044 operand_alternative *op_alt = XCNEWVEC (operand_alternative, 1045 nalt * nop); 1046 preprocess_constraints (nop, nalt, constraints, op_alt, 1047 data->operand_loc); 1048 setup_operand_alternative (data, op_alt); 1049 } 1050 } 1051 else 1052 { 1053 insn_extract (insn); 1054 data->insn_static_data = insn_static_data 1055 = get_static_insn_data (icode, insn_data[icode].n_operands, 1056 insn_data[icode].n_dups, 1057 insn_data[icode].n_alternatives); 1058 n = insn_static_data->n_operands; 1059 if (n == 0) 1060 locs = NULL; 1061 else 1062 { 1063 locs = XNEWVEC (rtx *, n); 1064 memcpy (locs, recog_data.operand_loc, n * sizeof (rtx *)); 1065 } 1066 data->operand_loc = locs; 1067 n = insn_static_data->n_dups; 1068 if (n == 0) 1069 locs = NULL; 1070 else 1071 { 1072 locs = XNEWVEC (rtx *, n); 1073 memcpy (locs, recog_data.dup_loc, n * sizeof (rtx *)); 1074 } 1075 data->dup_loc = locs; 1076 data->preferred_alternatives = get_preferred_alternatives (insn); 1077 const operand_alternative *op_alt = preprocess_insn_constraints (icode); 1078 if (!insn_static_data->operand_alternative) 1079 setup_operand_alternative (data, op_alt); 1080 else if (op_alt != insn_static_data->operand_alternative) 1081 insn_static_data->operand_alternative = op_alt; 1082 } 1083 if (GET_CODE (PATTERN (insn)) == CLOBBER || GET_CODE (PATTERN (insn)) == USE) 1084 insn_static_data->hard_regs = NULL; 1085 else 1086 insn_static_data->hard_regs 1087 = collect_non_operand_hard_regs (insn, &PATTERN (insn), data, 1088 NULL, OP_IN, false); 1089 data->arg_hard_regs = NULL; 1090 if (CALL_P (insn)) 1091 { 1092 bool use_p; 1093 rtx link; 1094 int n_hard_regs, regno, arg_hard_regs[FIRST_PSEUDO_REGISTER]; 1095 1096 n_hard_regs = 0; 1097 /* Finding implicit hard register usage. We believe it will be 1098 not changed whatever transformations are used. Call insns 1099 are such example. */ 1100 for (link = CALL_INSN_FUNCTION_USAGE (insn); 1101 link != NULL_RTX; 1102 link = XEXP (link, 1)) 1103 if (((use_p = GET_CODE (XEXP (link, 0)) == USE) 1104 || GET_CODE (XEXP (link, 0)) == CLOBBER) 1105 && REG_P (XEXP (XEXP (link, 0), 0))) 1106 { 1107 regno = REGNO (XEXP (XEXP (link, 0), 0)); 1108 lra_assert (regno < FIRST_PSEUDO_REGISTER); 1109 /* It is an argument register. */ 1110 for (i = REG_NREGS (XEXP (XEXP (link, 0), 0)) - 1; i >= 0; i--) 1111 arg_hard_regs[n_hard_regs++] 1112 = regno + i + (use_p ? 0 : FIRST_PSEUDO_REGISTER); 1113 } 1114 1115 if (n_hard_regs != 0) 1116 { 1117 arg_hard_regs[n_hard_regs++] = -1; 1118 data->arg_hard_regs = XNEWVEC (int, n_hard_regs); 1119 memcpy (data->arg_hard_regs, arg_hard_regs, 1120 sizeof (int) * n_hard_regs); 1121 } 1122 } 1123 /* Some output operand can be recognized only from the context not 1124 from the constraints which are empty in this case. Call insn may 1125 contain a hard register in set destination with empty constraint 1126 and extract_insn treats them as an input. */ 1127 for (i = 0; i < insn_static_data->n_operands; i++) 1128 { 1129 int j; 1130 rtx pat, set; 1131 struct lra_operand_data *operand = &insn_static_data->operand[i]; 1132 1133 /* ??? Should we treat 'X' the same way. It looks to me that 1134 'X' means anything and empty constraint means we do not 1135 care. */ 1136 if (operand->type != OP_IN || *operand->constraint != '\0' 1137 || operand->is_operator) 1138 continue; 1139 pat = PATTERN (insn); 1140 if (GET_CODE (pat) == SET) 1141 { 1142 if (data->operand_loc[i] != &SET_DEST (pat)) 1143 continue; 1144 } 1145 else if (GET_CODE (pat) == PARALLEL) 1146 { 1147 for (j = XVECLEN (pat, 0) - 1; j >= 0; j--) 1148 { 1149 set = XVECEXP (PATTERN (insn), 0, j); 1150 if (GET_CODE (set) == SET 1151 && &SET_DEST (set) == data->operand_loc[i]) 1152 break; 1153 } 1154 if (j < 0) 1155 continue; 1156 } 1157 else 1158 continue; 1159 operand->type = OP_OUT; 1160 } 1161 return data; 1162 } 1163 1164 /* Return info about insn give by UID. The info should be already set 1165 up. */ 1166 static lra_insn_recog_data_t 1167 get_insn_recog_data_by_uid (int uid) 1168 { 1169 lra_insn_recog_data_t data; 1170 1171 data = lra_insn_recog_data[uid]; 1172 lra_assert (data != NULL); 1173 return data; 1174 } 1175 1176 /* Invalidate all info about insn given by its UID. */ 1177 static void 1178 invalidate_insn_recog_data (int uid) 1179 { 1180 lra_insn_recog_data_t data; 1181 1182 data = lra_insn_recog_data[uid]; 1183 lra_assert (data != NULL); 1184 free_insn_recog_data (data); 1185 lra_insn_recog_data[uid] = NULL; 1186 } 1187 1188 /* Update all the insn info about INSN. It is usually called when 1189 something in the insn was changed. Return the updated info. */ 1190 lra_insn_recog_data_t 1191 lra_update_insn_recog_data (rtx_insn *insn) 1192 { 1193 lra_insn_recog_data_t data; 1194 int n; 1195 unsigned int uid = INSN_UID (insn); 1196 struct lra_static_insn_data *insn_static_data; 1197 poly_int64 sp_offset = 0; 1198 1199 check_and_expand_insn_recog_data (uid); 1200 if ((data = lra_insn_recog_data[uid]) != NULL 1201 && data->icode != INSN_CODE (insn)) 1202 { 1203 sp_offset = data->sp_offset; 1204 invalidate_insn_data_regno_info (data, insn, get_insn_freq (insn)); 1205 invalidate_insn_recog_data (uid); 1206 data = NULL; 1207 } 1208 if (data == NULL) 1209 { 1210 data = lra_get_insn_recog_data (insn); 1211 /* Initiate or restore SP offset. */ 1212 data->sp_offset = sp_offset; 1213 return data; 1214 } 1215 insn_static_data = data->insn_static_data; 1216 data->used_insn_alternative = LRA_UNKNOWN_ALT; 1217 if (DEBUG_INSN_P (insn)) 1218 return data; 1219 if (data->icode < 0) 1220 { 1221 int nop; 1222 machine_mode operand_mode[MAX_RECOG_OPERANDS]; 1223 const char *constraints[MAX_RECOG_OPERANDS]; 1224 1225 nop = asm_noperands (PATTERN (insn)); 1226 if (nop >= 0) 1227 { 1228 lra_assert (nop == data->insn_static_data->n_operands); 1229 /* Now get the operand values and constraints out of the 1230 insn. */ 1231 decode_asm_operands (PATTERN (insn), NULL, 1232 data->operand_loc, 1233 constraints, operand_mode, NULL); 1234 1235 if (flag_checking) 1236 for (int i = 0; i < nop; i++) 1237 lra_assert 1238 (insn_static_data->operand[i].mode == operand_mode[i] 1239 && insn_static_data->operand[i].constraint == constraints[i] 1240 && ! insn_static_data->operand[i].is_operator); 1241 } 1242 1243 if (flag_checking) 1244 for (int i = 0; i < insn_static_data->n_operands; i++) 1245 lra_assert 1246 (insn_static_data->operand[i].type 1247 == (insn_static_data->operand[i].constraint[0] == '=' ? OP_OUT 1248 : insn_static_data->operand[i].constraint[0] == '+' ? OP_INOUT 1249 : OP_IN)); 1250 } 1251 else 1252 { 1253 insn_extract (insn); 1254 n = insn_static_data->n_operands; 1255 if (n != 0) 1256 memcpy (data->operand_loc, recog_data.operand_loc, n * sizeof (rtx *)); 1257 n = insn_static_data->n_dups; 1258 if (n != 0) 1259 memcpy (data->dup_loc, recog_data.dup_loc, n * sizeof (rtx *)); 1260 lra_assert (check_bool_attrs (insn)); 1261 } 1262 return data; 1263 } 1264 1265 /* Set up that INSN is using alternative ALT now. */ 1266 void 1267 lra_set_used_insn_alternative (rtx_insn *insn, int alt) 1268 { 1269 lra_insn_recog_data_t data; 1270 1271 data = lra_get_insn_recog_data (insn); 1272 data->used_insn_alternative = alt; 1273 } 1274 1275 /* Set up that insn with UID is using alternative ALT now. The insn 1276 info should be already set up. */ 1277 void 1278 lra_set_used_insn_alternative_by_uid (int uid, int alt) 1279 { 1280 lra_insn_recog_data_t data; 1281 1282 check_and_expand_insn_recog_data (uid); 1283 data = lra_insn_recog_data[uid]; 1284 lra_assert (data != NULL); 1285 data->used_insn_alternative = alt; 1286 } 1287 1288 1289 1290 /* This page contains code dealing with common register info and 1291 pseudo copies. */ 1292 1293 /* The size of the following array. */ 1294 static int reg_info_size; 1295 /* Common info about each register. */ 1296 class lra_reg *lra_reg_info; 1297 1298 HARD_REG_SET hard_regs_spilled_into; 1299 1300 /* Last register value. */ 1301 static int last_reg_value; 1302 1303 /* Return new register value. */ 1304 static int 1305 get_new_reg_value (void) 1306 { 1307 return ++last_reg_value; 1308 } 1309 1310 /* Vec referring to pseudo copies. */ 1311 static vec<lra_copy_t> copy_vec; 1312 1313 /* Initialize I-th element of lra_reg_info. */ 1314 static inline void 1315 initialize_lra_reg_info_element (int i) 1316 { 1317 bitmap_initialize (&lra_reg_info[i].insn_bitmap, ®_obstack); 1318 #ifdef STACK_REGS 1319 lra_reg_info[i].no_stack_p = false; 1320 #endif 1321 CLEAR_HARD_REG_SET (lra_reg_info[i].conflict_hard_regs); 1322 lra_reg_info[i].preferred_hard_regno1 = -1; 1323 lra_reg_info[i].preferred_hard_regno2 = -1; 1324 lra_reg_info[i].preferred_hard_regno_profit1 = 0; 1325 lra_reg_info[i].preferred_hard_regno_profit2 = 0; 1326 lra_reg_info[i].biggest_mode = VOIDmode; 1327 lra_reg_info[i].live_ranges = NULL; 1328 lra_reg_info[i].nrefs = lra_reg_info[i].freq = 0; 1329 lra_reg_info[i].last_reload = 0; 1330 lra_reg_info[i].restore_rtx = NULL_RTX; 1331 lra_reg_info[i].val = get_new_reg_value (); 1332 lra_reg_info[i].offset = 0; 1333 lra_reg_info[i].copies = NULL; 1334 } 1335 1336 /* Initialize common reg info and copies. */ 1337 static void 1338 init_reg_info (void) 1339 { 1340 int i; 1341 1342 last_reg_value = 0; 1343 reg_info_size = max_reg_num () * 3 / 2 + 1; 1344 lra_reg_info = XNEWVEC (class lra_reg, reg_info_size); 1345 for (i = 0; i < reg_info_size; i++) 1346 initialize_lra_reg_info_element (i); 1347 copy_vec.truncate (0); 1348 CLEAR_HARD_REG_SET (hard_regs_spilled_into); 1349 } 1350 1351 1352 /* Finish common reg info and copies. */ 1353 static void 1354 finish_reg_info (void) 1355 { 1356 int i; 1357 1358 for (i = 0; i < reg_info_size; i++) 1359 bitmap_clear (&lra_reg_info[i].insn_bitmap); 1360 free (lra_reg_info); 1361 reg_info_size = 0; 1362 } 1363 1364 /* Expand common reg info if it is necessary. */ 1365 static void 1366 expand_reg_info (void) 1367 { 1368 int i, old = reg_info_size; 1369 1370 if (reg_info_size > max_reg_num ()) 1371 return; 1372 reg_info_size = max_reg_num () * 3 / 2 + 1; 1373 lra_reg_info = XRESIZEVEC (class lra_reg, lra_reg_info, reg_info_size); 1374 for (i = old; i < reg_info_size; i++) 1375 initialize_lra_reg_info_element (i); 1376 } 1377 1378 /* Free all copies. */ 1379 void 1380 lra_free_copies (void) 1381 { 1382 lra_copy_t cp; 1383 1384 while (copy_vec.length () != 0) 1385 { 1386 cp = copy_vec.pop (); 1387 lra_reg_info[cp->regno1].copies = lra_reg_info[cp->regno2].copies = NULL; 1388 lra_copy_pool.remove (cp); 1389 } 1390 } 1391 1392 /* Create copy of two pseudos REGNO1 and REGNO2. The copy execution 1393 frequency is FREQ. */ 1394 void 1395 lra_create_copy (int regno1, int regno2, int freq) 1396 { 1397 bool regno1_dest_p; 1398 lra_copy_t cp; 1399 1400 lra_assert (regno1 != regno2); 1401 regno1_dest_p = true; 1402 if (regno1 > regno2) 1403 { 1404 std::swap (regno1, regno2); 1405 regno1_dest_p = false; 1406 } 1407 cp = lra_copy_pool.allocate (); 1408 copy_vec.safe_push (cp); 1409 cp->regno1_dest_p = regno1_dest_p; 1410 cp->freq = freq; 1411 cp->regno1 = regno1; 1412 cp->regno2 = regno2; 1413 cp->regno1_next = lra_reg_info[regno1].copies; 1414 lra_reg_info[regno1].copies = cp; 1415 cp->regno2_next = lra_reg_info[regno2].copies; 1416 lra_reg_info[regno2].copies = cp; 1417 if (lra_dump_file != NULL) 1418 fprintf (lra_dump_file, " Creating copy r%d%sr%d@%d\n", 1419 regno1, regno1_dest_p ? "<-" : "->", regno2, freq); 1420 } 1421 1422 /* Return N-th (0, 1, ...) copy. If there is no copy, return 1423 NULL. */ 1424 lra_copy_t 1425 lra_get_copy (int n) 1426 { 1427 if (n >= (int) copy_vec.length ()) 1428 return NULL; 1429 return copy_vec[n]; 1430 } 1431 1432 1433 1434 /* This page contains code dealing with info about registers in 1435 insns. */ 1436 1437 /* Process X of INSN recursively and add info (operand type is given 1438 by TYPE) about registers in X to the insn DATA. If X can be early 1439 clobbered, alternatives in which it can be early clobbered are given 1440 by EARLY_CLOBBER_ALTS. */ 1441 static void 1442 add_regs_to_insn_regno_info (lra_insn_recog_data_t data, rtx x, 1443 rtx_insn *insn, enum op_type type, 1444 alternative_mask early_clobber_alts) 1445 { 1446 int i, j, regno; 1447 bool subreg_p; 1448 machine_mode mode; 1449 const char *fmt; 1450 enum rtx_code code; 1451 struct lra_insn_reg *curr; 1452 1453 code = GET_CODE (x); 1454 mode = GET_MODE (x); 1455 subreg_p = false; 1456 if (GET_CODE (x) == SUBREG) 1457 { 1458 mode = wider_subreg_mode (x); 1459 if (read_modify_subreg_p (x)) 1460 subreg_p = true; 1461 x = SUBREG_REG (x); 1462 code = GET_CODE (x); 1463 } 1464 if (REG_P (x)) 1465 { 1466 regno = REGNO (x); 1467 /* Process all regs even unallocatable ones as we need info about 1468 all regs for rematerialization pass. */ 1469 expand_reg_info (); 1470 if (bitmap_set_bit (&lra_reg_info[regno].insn_bitmap, INSN_UID (insn))) 1471 { 1472 data->regs = new_insn_reg (data->insn, regno, type, mode, subreg_p, 1473 early_clobber_alts, data->regs); 1474 return; 1475 } 1476 else 1477 { 1478 for (curr = data->regs; curr != NULL; curr = curr->next) 1479 if (curr->regno == regno) 1480 { 1481 if (curr->subreg_p != subreg_p || curr->biggest_mode != mode) 1482 /* The info cannot be integrated into the found 1483 structure. */ 1484 data->regs = new_insn_reg (data->insn, regno, type, mode, 1485 subreg_p, early_clobber_alts, 1486 data->regs); 1487 else 1488 { 1489 if (curr->type != type) 1490 curr->type = OP_INOUT; 1491 curr->early_clobber_alts |= early_clobber_alts; 1492 } 1493 return; 1494 } 1495 gcc_unreachable (); 1496 } 1497 } 1498 1499 switch (code) 1500 { 1501 case SET: 1502 add_regs_to_insn_regno_info (data, SET_DEST (x), insn, OP_OUT, 0); 1503 add_regs_to_insn_regno_info (data, SET_SRC (x), insn, OP_IN, 0); 1504 break; 1505 case CLOBBER: 1506 /* We treat clobber of non-operand hard registers as early 1507 clobber. */ 1508 add_regs_to_insn_regno_info (data, XEXP (x, 0), insn, OP_OUT, 1509 ALL_ALTERNATIVES); 1510 break; 1511 case PRE_INC: case PRE_DEC: case POST_INC: case POST_DEC: 1512 add_regs_to_insn_regno_info (data, XEXP (x, 0), insn, OP_INOUT, 0); 1513 break; 1514 case PRE_MODIFY: case POST_MODIFY: 1515 add_regs_to_insn_regno_info (data, XEXP (x, 0), insn, OP_INOUT, 0); 1516 add_regs_to_insn_regno_info (data, XEXP (x, 1), insn, OP_IN, 0); 1517 break; 1518 default: 1519 if ((code != PARALLEL && code != EXPR_LIST) || type != OP_OUT) 1520 /* Some targets place small structures in registers for return 1521 values of functions, and those registers are wrapped in 1522 PARALLEL that we may see as the destination of a SET. Here 1523 is an example: 1524 1525 (call_insn 13 12 14 2 (set (parallel:BLK [ 1526 (expr_list:REG_DEP_TRUE (reg:DI 0 ax) 1527 (const_int 0 [0])) 1528 (expr_list:REG_DEP_TRUE (reg:DI 1 dx) 1529 (const_int 8 [0x8])) 1530 ]) 1531 (call (mem:QI (symbol_ref:DI (... */ 1532 type = OP_IN; 1533 fmt = GET_RTX_FORMAT (code); 1534 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) 1535 { 1536 if (fmt[i] == 'e') 1537 add_regs_to_insn_regno_info (data, XEXP (x, i), insn, type, 0); 1538 else if (fmt[i] == 'E') 1539 { 1540 for (j = XVECLEN (x, i) - 1; j >= 0; j--) 1541 add_regs_to_insn_regno_info (data, XVECEXP (x, i, j), insn, 1542 type, 0); 1543 } 1544 } 1545 } 1546 } 1547 1548 /* Return execution frequency of INSN. */ 1549 static int 1550 get_insn_freq (rtx_insn *insn) 1551 { 1552 basic_block bb = BLOCK_FOR_INSN (insn); 1553 1554 gcc_checking_assert (bb != NULL); 1555 return REG_FREQ_FROM_BB (bb); 1556 } 1557 1558 /* Invalidate all reg info of INSN with DATA and execution frequency 1559 FREQ. Update common info about the invalidated registers. */ 1560 static void 1561 invalidate_insn_data_regno_info (lra_insn_recog_data_t data, rtx_insn *insn, 1562 int freq) 1563 { 1564 int uid; 1565 bool debug_p; 1566 unsigned int i; 1567 struct lra_insn_reg *ir, *next_ir; 1568 1569 uid = INSN_UID (insn); 1570 debug_p = DEBUG_INSN_P (insn); 1571 for (ir = data->regs; ir != NULL; ir = next_ir) 1572 { 1573 i = ir->regno; 1574 next_ir = ir->next; 1575 lra_insn_reg_pool.remove (ir); 1576 bitmap_clear_bit (&lra_reg_info[i].insn_bitmap, uid); 1577 if (i >= FIRST_PSEUDO_REGISTER && ! debug_p) 1578 { 1579 lra_reg_info[i].nrefs--; 1580 lra_reg_info[i].freq -= freq; 1581 lra_assert (lra_reg_info[i].nrefs >= 0 && lra_reg_info[i].freq >= 0); 1582 } 1583 } 1584 data->regs = NULL; 1585 } 1586 1587 /* Invalidate all reg info of INSN. Update common info about the 1588 invalidated registers. */ 1589 void 1590 lra_invalidate_insn_regno_info (rtx_insn *insn) 1591 { 1592 invalidate_insn_data_regno_info (lra_get_insn_recog_data (insn), insn, 1593 get_insn_freq (insn)); 1594 } 1595 1596 /* Update common reg info from reg info of insn given by its DATA and 1597 execution frequency FREQ. */ 1598 static void 1599 setup_insn_reg_info (lra_insn_recog_data_t data, int freq) 1600 { 1601 unsigned int i; 1602 struct lra_insn_reg *ir; 1603 1604 for (ir = data->regs; ir != NULL; ir = ir->next) 1605 if ((i = ir->regno) >= FIRST_PSEUDO_REGISTER) 1606 { 1607 lra_reg_info[i].nrefs++; 1608 lra_reg_info[i].freq += freq; 1609 } 1610 } 1611 1612 /* Set up insn reg info of INSN. Update common reg info from reg info 1613 of INSN. */ 1614 void 1615 lra_update_insn_regno_info (rtx_insn *insn) 1616 { 1617 int i, freq; 1618 lra_insn_recog_data_t data; 1619 struct lra_static_insn_data *static_data; 1620 enum rtx_code code; 1621 rtx link; 1622 1623 if (! INSN_P (insn)) 1624 return; 1625 data = lra_get_insn_recog_data (insn); 1626 static_data = data->insn_static_data; 1627 freq = NONDEBUG_INSN_P (insn) ? get_insn_freq (insn) : 0; 1628 invalidate_insn_data_regno_info (data, insn, freq); 1629 for (i = static_data->n_operands - 1; i >= 0; i--) 1630 add_regs_to_insn_regno_info (data, *data->operand_loc[i], insn, 1631 static_data->operand[i].type, 1632 static_data->operand[i].early_clobber_alts); 1633 if ((code = GET_CODE (PATTERN (insn))) == CLOBBER || code == USE) 1634 add_regs_to_insn_regno_info (data, XEXP (PATTERN (insn), 0), insn, 1635 code == USE ? OP_IN : OP_OUT, 0); 1636 if (CALL_P (insn)) 1637 /* On some targets call insns can refer to pseudos in memory in 1638 CALL_INSN_FUNCTION_USAGE list. Process them in order to 1639 consider their occurrences in calls for different 1640 transformations (e.g. inheritance) with given pseudos. */ 1641 for (link = CALL_INSN_FUNCTION_USAGE (insn); 1642 link != NULL_RTX; 1643 link = XEXP (link, 1)) 1644 { 1645 code = GET_CODE (XEXP (link, 0)); 1646 if ((code == USE || code == CLOBBER) 1647 && MEM_P (XEXP (XEXP (link, 0), 0))) 1648 add_regs_to_insn_regno_info (data, XEXP (XEXP (link, 0), 0), insn, 1649 code == USE ? OP_IN : OP_OUT, 0); 1650 } 1651 if (NONDEBUG_INSN_P (insn)) 1652 setup_insn_reg_info (data, freq); 1653 } 1654 1655 /* Return reg info of insn given by it UID. */ 1656 struct lra_insn_reg * 1657 lra_get_insn_regs (int uid) 1658 { 1659 lra_insn_recog_data_t data; 1660 1661 data = get_insn_recog_data_by_uid (uid); 1662 return data->regs; 1663 } 1664 1665 1666 1667 /* Recursive hash function for RTL X. */ 1668 hashval_t 1669 lra_rtx_hash (rtx x) 1670 { 1671 int i, j; 1672 enum rtx_code code; 1673 const char *fmt; 1674 hashval_t val = 0; 1675 1676 if (x == 0) 1677 return val; 1678 1679 code = GET_CODE (x); 1680 val += (int) code + 4095; 1681 1682 /* Some RTL can be compared nonrecursively. */ 1683 switch (code) 1684 { 1685 case REG: 1686 return val + REGNO (x); 1687 1688 case LABEL_REF: 1689 return iterative_hash_object (XEXP (x, 0), val); 1690 1691 case SYMBOL_REF: 1692 return iterative_hash_object (XSTR (x, 0), val); 1693 1694 case SCRATCH: 1695 case CONST_DOUBLE: 1696 case CONST_VECTOR: 1697 return val; 1698 1699 case CONST_INT: 1700 return val + UINTVAL (x); 1701 1702 default: 1703 break; 1704 } 1705 1706 /* Hash the elements. */ 1707 fmt = GET_RTX_FORMAT (code); 1708 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) 1709 { 1710 switch (fmt[i]) 1711 { 1712 case 'w': 1713 val += XWINT (x, i); 1714 break; 1715 1716 case 'n': 1717 case 'i': 1718 val += XINT (x, i); 1719 break; 1720 1721 case 'V': 1722 case 'E': 1723 val += XVECLEN (x, i); 1724 1725 for (j = 0; j < XVECLEN (x, i); j++) 1726 val += lra_rtx_hash (XVECEXP (x, i, j)); 1727 break; 1728 1729 case 'e': 1730 val += lra_rtx_hash (XEXP (x, i)); 1731 break; 1732 1733 case 'S': 1734 case 's': 1735 val += htab_hash_string (XSTR (x, i)); 1736 break; 1737 1738 case 'u': 1739 case '0': 1740 case 't': 1741 break; 1742 1743 /* It is believed that rtx's at this level will never 1744 contain anything but integers and other rtx's, except for 1745 within LABEL_REFs and SYMBOL_REFs. */ 1746 default: 1747 abort (); 1748 } 1749 } 1750 return val; 1751 } 1752 1753 1754 1755 /* This page contains code dealing with stack of the insns which 1756 should be processed by the next constraint pass. */ 1757 1758 /* Bitmap used to put an insn on the stack only in one exemplar. */ 1759 static sbitmap lra_constraint_insn_stack_bitmap; 1760 1761 /* The stack itself. */ 1762 vec<rtx_insn *> lra_constraint_insn_stack; 1763 1764 /* Put INSN on the stack. If ALWAYS_UPDATE is true, always update the reg 1765 info for INSN, otherwise only update it if INSN is not already on the 1766 stack. */ 1767 static inline void 1768 lra_push_insn_1 (rtx_insn *insn, bool always_update) 1769 { 1770 unsigned int uid = INSN_UID (insn); 1771 if (always_update) 1772 lra_update_insn_regno_info (insn); 1773 if (uid >= SBITMAP_SIZE (lra_constraint_insn_stack_bitmap)) 1774 lra_constraint_insn_stack_bitmap = 1775 sbitmap_resize (lra_constraint_insn_stack_bitmap, 3 * uid / 2, 0); 1776 if (bitmap_bit_p (lra_constraint_insn_stack_bitmap, uid)) 1777 return; 1778 bitmap_set_bit (lra_constraint_insn_stack_bitmap, uid); 1779 if (! always_update) 1780 lra_update_insn_regno_info (insn); 1781 lra_constraint_insn_stack.safe_push (insn); 1782 } 1783 1784 /* Put INSN on the stack. */ 1785 void 1786 lra_push_insn (rtx_insn *insn) 1787 { 1788 lra_push_insn_1 (insn, false); 1789 } 1790 1791 /* Put INSN on the stack and update its reg info. */ 1792 void 1793 lra_push_insn_and_update_insn_regno_info (rtx_insn *insn) 1794 { 1795 lra_push_insn_1 (insn, true); 1796 } 1797 1798 /* Put insn with UID on the stack. */ 1799 void 1800 lra_push_insn_by_uid (unsigned int uid) 1801 { 1802 lra_push_insn (lra_insn_recog_data[uid]->insn); 1803 } 1804 1805 /* Take the last-inserted insns off the stack and return it. */ 1806 rtx_insn * 1807 lra_pop_insn (void) 1808 { 1809 rtx_insn *insn = lra_constraint_insn_stack.pop (); 1810 bitmap_clear_bit (lra_constraint_insn_stack_bitmap, INSN_UID (insn)); 1811 return insn; 1812 } 1813 1814 /* Return the current size of the insn stack. */ 1815 unsigned int 1816 lra_insn_stack_length (void) 1817 { 1818 return lra_constraint_insn_stack.length (); 1819 } 1820 1821 /* Push insns FROM to TO (excluding it) going in reverse order. */ 1822 static void 1823 push_insns (rtx_insn *from, rtx_insn *to) 1824 { 1825 rtx_insn *insn; 1826 1827 if (from == NULL_RTX) 1828 return; 1829 for (insn = from; insn != to; insn = PREV_INSN (insn)) 1830 if (INSN_P (insn)) 1831 lra_push_insn (insn); 1832 } 1833 1834 /* Set up sp offset for insn in range [FROM, LAST]. The offset is 1835 taken from the next BB insn after LAST or zero if there in such 1836 insn. */ 1837 static void 1838 setup_sp_offset (rtx_insn *from, rtx_insn *last) 1839 { 1840 rtx_insn *before = next_nonnote_nondebug_insn_bb (last); 1841 poly_int64 offset = (before == NULL_RTX || ! INSN_P (before) 1842 ? 0 : lra_get_insn_recog_data (before)->sp_offset); 1843 1844 for (rtx_insn *insn = from; insn != NEXT_INSN (last); insn = NEXT_INSN (insn)) 1845 lra_get_insn_recog_data (insn)->sp_offset = offset; 1846 } 1847 1848 /* Emit insns BEFORE before INSN and insns AFTER after INSN. Put the 1849 insns onto the stack. Print about emitting the insns with 1850 TITLE. */ 1851 void 1852 lra_process_new_insns (rtx_insn *insn, rtx_insn *before, rtx_insn *after, 1853 const char *title) 1854 { 1855 rtx_insn *last; 1856 1857 if (before == NULL_RTX && after == NULL_RTX) 1858 return; 1859 if (lra_dump_file != NULL) 1860 { 1861 dump_insn_slim (lra_dump_file, insn); 1862 if (before != NULL_RTX) 1863 { 1864 fprintf (lra_dump_file," %s before:\n", title); 1865 dump_rtl_slim (lra_dump_file, before, NULL, -1, 0); 1866 } 1867 if (after != NULL_RTX) 1868 { 1869 fprintf (lra_dump_file, " %s after:\n", title); 1870 dump_rtl_slim (lra_dump_file, after, NULL, -1, 0); 1871 } 1872 fprintf (lra_dump_file, "\n"); 1873 } 1874 if (before != NULL_RTX) 1875 { 1876 if (cfun->can_throw_non_call_exceptions) 1877 copy_reg_eh_region_note_forward (insn, before, NULL); 1878 emit_insn_before (before, insn); 1879 push_insns (PREV_INSN (insn), PREV_INSN (before)); 1880 setup_sp_offset (before, PREV_INSN (insn)); 1881 } 1882 if (after != NULL_RTX) 1883 { 1884 if (cfun->can_throw_non_call_exceptions) 1885 copy_reg_eh_region_note_forward (insn, after, NULL); 1886 for (last = after; NEXT_INSN (last) != NULL_RTX; last = NEXT_INSN (last)) 1887 ; 1888 emit_insn_after (after, insn); 1889 push_insns (last, insn); 1890 setup_sp_offset (after, last); 1891 } 1892 if (cfun->can_throw_non_call_exceptions) 1893 { 1894 rtx note = find_reg_note (insn, REG_EH_REGION, NULL_RTX); 1895 if (note && !insn_could_throw_p (insn)) 1896 remove_note (insn, note); 1897 } 1898 } 1899 1900 1901 /* Replace all references to register OLD_REGNO in *LOC with pseudo 1902 register NEW_REG. Try to simplify subreg of constant if SUBREG_P. 1903 DEBUG_P is if LOC is within a DEBUG_INSN. Return true if any 1904 change was made. */ 1905 bool 1906 lra_substitute_pseudo (rtx *loc, int old_regno, rtx new_reg, bool subreg_p, 1907 bool debug_p) 1908 { 1909 rtx x = *loc; 1910 bool result = false; 1911 enum rtx_code code; 1912 const char *fmt; 1913 int i, j; 1914 1915 if (x == NULL_RTX) 1916 return false; 1917 1918 code = GET_CODE (x); 1919 if (code == SUBREG && subreg_p) 1920 { 1921 rtx subst, inner = SUBREG_REG (x); 1922 /* Transform subreg of constant while we still have inner mode 1923 of the subreg. The subreg internal should not be an insn 1924 operand. */ 1925 if (REG_P (inner) && (int) REGNO (inner) == old_regno 1926 && CONSTANT_P (new_reg) 1927 && (subst = simplify_subreg (GET_MODE (x), new_reg, GET_MODE (inner), 1928 SUBREG_BYTE (x))) != NULL_RTX) 1929 { 1930 *loc = subst; 1931 return true; 1932 } 1933 1934 } 1935 else if (code == REG && (int) REGNO (x) == old_regno) 1936 { 1937 machine_mode mode = GET_MODE (x); 1938 machine_mode inner_mode = GET_MODE (new_reg); 1939 1940 if (mode != inner_mode 1941 && ! (CONST_SCALAR_INT_P (new_reg) && SCALAR_INT_MODE_P (mode))) 1942 { 1943 poly_uint64 offset = 0; 1944 if (partial_subreg_p (mode, inner_mode) 1945 && SCALAR_INT_MODE_P (inner_mode)) 1946 offset = subreg_lowpart_offset (mode, inner_mode); 1947 if (debug_p) 1948 new_reg = gen_rtx_raw_SUBREG (mode, new_reg, offset); 1949 else 1950 new_reg = gen_rtx_SUBREG (mode, new_reg, offset); 1951 } 1952 *loc = new_reg; 1953 return true; 1954 } 1955 1956 /* Scan all the operand sub-expressions. */ 1957 fmt = GET_RTX_FORMAT (code); 1958 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) 1959 { 1960 if (fmt[i] == 'e') 1961 { 1962 if (lra_substitute_pseudo (&XEXP (x, i), old_regno, 1963 new_reg, subreg_p, debug_p)) 1964 result = true; 1965 } 1966 else if (fmt[i] == 'E') 1967 { 1968 for (j = XVECLEN (x, i) - 1; j >= 0; j--) 1969 if (lra_substitute_pseudo (&XVECEXP (x, i, j), old_regno, 1970 new_reg, subreg_p, debug_p)) 1971 result = true; 1972 } 1973 } 1974 return result; 1975 } 1976 1977 /* Call lra_substitute_pseudo within an insn. Try to simplify subreg 1978 of constant if SUBREG_P. This won't update the insn ptr, just the 1979 contents of the insn. */ 1980 bool 1981 lra_substitute_pseudo_within_insn (rtx_insn *insn, int old_regno, 1982 rtx new_reg, bool subreg_p) 1983 { 1984 rtx loc = insn; 1985 return lra_substitute_pseudo (&loc, old_regno, new_reg, subreg_p, 1986 DEBUG_INSN_P (insn)); 1987 } 1988 1989 1990 1991 /* This page contains code dealing with scratches (changing them onto 1992 pseudos and restoring them from the pseudos). 1993 1994 We change scratches into pseudos at the beginning of LRA to 1995 simplify dealing with them (conflicts, hard register assignments). 1996 1997 If the pseudo denoting scratch was spilled it means that we do need 1998 a hard register for it. Such pseudos are transformed back to 1999 scratches at the end of LRA. */ 2000 2001 /* Description of location of a former scratch operand. */ 2002 struct sloc 2003 { 2004 rtx_insn *insn; /* Insn where the scratch was. */ 2005 int nop; /* Number of the operand which was a scratch. */ 2006 int icode; /* Original icode from which scratch was removed. */ 2007 }; 2008 2009 typedef struct sloc *sloc_t; 2010 2011 /* Locations of the former scratches. */ 2012 static vec<sloc_t> scratches; 2013 2014 /* Bitmap of scratch regnos. */ 2015 static bitmap_head scratch_bitmap; 2016 2017 /* Bitmap of scratch operands. */ 2018 static bitmap_head scratch_operand_bitmap; 2019 2020 /* Return true if pseudo REGNO is made of SCRATCH. */ 2021 bool 2022 lra_former_scratch_p (int regno) 2023 { 2024 return bitmap_bit_p (&scratch_bitmap, regno); 2025 } 2026 2027 /* Return true if the operand NOP of INSN is a former scratch. */ 2028 bool 2029 lra_former_scratch_operand_p (rtx_insn *insn, int nop) 2030 { 2031 return bitmap_bit_p (&scratch_operand_bitmap, 2032 INSN_UID (insn) * MAX_RECOG_OPERANDS + nop) != 0; 2033 } 2034 2035 /* Register operand NOP in INSN as a former scratch. It will be 2036 changed to scratch back, if it is necessary, at the LRA end. */ 2037 void 2038 lra_register_new_scratch_op (rtx_insn *insn, int nop, int icode) 2039 { 2040 lra_insn_recog_data_t id = lra_get_insn_recog_data (insn); 2041 rtx op = *id->operand_loc[nop]; 2042 sloc_t loc = XNEW (struct sloc); 2043 lra_assert (REG_P (op)); 2044 loc->insn = insn; 2045 loc->nop = nop; 2046 loc->icode = icode; 2047 scratches.safe_push (loc); 2048 bitmap_set_bit (&scratch_bitmap, REGNO (op)); 2049 bitmap_set_bit (&scratch_operand_bitmap, 2050 INSN_UID (insn) * MAX_RECOG_OPERANDS + nop); 2051 add_reg_note (insn, REG_UNUSED, op); 2052 } 2053 2054 /* Change INSN's scratches into pseudos and save their location. */ 2055 static void 2056 remove_scratches_1 (rtx_insn *insn) 2057 { 2058 int i; 2059 bool insn_changed_p; 2060 rtx reg; 2061 lra_insn_recog_data_t id; 2062 struct lra_static_insn_data *static_id; 2063 2064 id = lra_get_insn_recog_data (insn); 2065 static_id = id->insn_static_data; 2066 insn_changed_p = false; 2067 for (i = 0; i < static_id->n_operands; i++) 2068 if (GET_CODE (*id->operand_loc[i]) == SCRATCH 2069 && GET_MODE (*id->operand_loc[i]) != VOIDmode) 2070 { 2071 insn_changed_p = true; 2072 *id->operand_loc[i] = reg 2073 = lra_create_new_reg (static_id->operand[i].mode, 2074 *id->operand_loc[i], ALL_REGS, NULL); 2075 lra_register_new_scratch_op (insn, i, id->icode); 2076 if (lra_dump_file != NULL) 2077 fprintf (lra_dump_file, 2078 "Removing SCRATCH in insn #%u (nop %d)\n", 2079 INSN_UID (insn), i); 2080 } 2081 if (insn_changed_p) 2082 /* Because we might use DF right after caller-saves sub-pass 2083 we need to keep DF info up to date. */ 2084 df_insn_rescan (insn); 2085 } 2086 2087 /* Change scratches into pseudos and save their location. */ 2088 static void 2089 remove_scratches (void) 2090 { 2091 basic_block bb; 2092 rtx_insn *insn; 2093 2094 scratches.create (get_max_uid ()); 2095 bitmap_initialize (&scratch_bitmap, ®_obstack); 2096 bitmap_initialize (&scratch_operand_bitmap, ®_obstack); 2097 FOR_EACH_BB_FN (bb, cfun) 2098 FOR_BB_INSNS (bb, insn) 2099 if (INSN_P (insn)) 2100 remove_scratches_1 (insn); 2101 } 2102 2103 /* Changes pseudos created by function remove_scratches onto scratches. */ 2104 static void 2105 restore_scratches (void) 2106 { 2107 int regno; 2108 unsigned i; 2109 sloc_t loc; 2110 rtx_insn *last = NULL; 2111 lra_insn_recog_data_t id = NULL; 2112 2113 for (i = 0; scratches.iterate (i, &loc); i++) 2114 { 2115 /* Ignore already deleted insns. */ 2116 if (NOTE_P (loc->insn) 2117 && NOTE_KIND (loc->insn) == NOTE_INSN_DELETED) 2118 continue; 2119 if (last != loc->insn) 2120 { 2121 last = loc->insn; 2122 id = lra_get_insn_recog_data (last); 2123 } 2124 if (loc->icode != id->icode) 2125 { 2126 /* The icode doesn't match, which means the insn has been modified 2127 (e.g. register elimination). The scratch cannot be restored. */ 2128 continue; 2129 } 2130 if (REG_P (*id->operand_loc[loc->nop]) 2131 && ((regno = REGNO (*id->operand_loc[loc->nop])) 2132 >= FIRST_PSEUDO_REGISTER) 2133 && lra_get_regno_hard_regno (regno) < 0) 2134 { 2135 /* It should be only case when scratch register with chosen 2136 constraint 'X' did not get memory or hard register. */ 2137 lra_assert (lra_former_scratch_p (regno)); 2138 *id->operand_loc[loc->nop] 2139 = gen_rtx_SCRATCH (GET_MODE (*id->operand_loc[loc->nop])); 2140 lra_update_dup (id, loc->nop); 2141 if (lra_dump_file != NULL) 2142 fprintf (lra_dump_file, "Restoring SCRATCH in insn #%u(nop %d)\n", 2143 INSN_UID (loc->insn), loc->nop); 2144 } 2145 } 2146 for (i = 0; scratches.iterate (i, &loc); i++) 2147 free (loc); 2148 scratches.release (); 2149 bitmap_clear (&scratch_bitmap); 2150 bitmap_clear (&scratch_operand_bitmap); 2151 } 2152 2153 2154 2155 /* Function checks RTL for correctness. If FINAL_P is true, it is 2156 done at the end of LRA and the check is more rigorous. */ 2157 static void 2158 check_rtl (bool final_p) 2159 { 2160 basic_block bb; 2161 rtx_insn *insn; 2162 2163 lra_assert (! final_p || reload_completed); 2164 FOR_EACH_BB_FN (bb, cfun) 2165 FOR_BB_INSNS (bb, insn) 2166 if (NONDEBUG_INSN_P (insn) 2167 && GET_CODE (PATTERN (insn)) != USE 2168 && GET_CODE (PATTERN (insn)) != CLOBBER 2169 && GET_CODE (PATTERN (insn)) != ASM_INPUT) 2170 { 2171 if (final_p) 2172 { 2173 extract_constrain_insn (insn); 2174 continue; 2175 } 2176 /* LRA code is based on assumption that all addresses can be 2177 correctly decomposed. LRA can generate reloads for 2178 decomposable addresses. The decomposition code checks the 2179 correctness of the addresses. So we don't need to check 2180 the addresses here. Don't call insn_invalid_p here, it can 2181 change the code at this stage. */ 2182 if (recog_memoized (insn) < 0 && asm_noperands (PATTERN (insn)) < 0) 2183 fatal_insn_not_found (insn); 2184 } 2185 } 2186 2187 /* Determine if the current function has an exception receiver block 2188 that reaches the exit block via non-exceptional edges */ 2189 static bool 2190 has_nonexceptional_receiver (void) 2191 { 2192 edge e; 2193 edge_iterator ei; 2194 basic_block *tos, *worklist, bb; 2195 2196 /* If we're not optimizing, then just err on the safe side. */ 2197 if (!optimize) 2198 return true; 2199 2200 /* First determine which blocks can reach exit via normal paths. */ 2201 tos = worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun) + 1); 2202 2203 FOR_EACH_BB_FN (bb, cfun) 2204 bb->flags &= ~BB_REACHABLE; 2205 2206 /* Place the exit block on our worklist. */ 2207 EXIT_BLOCK_PTR_FOR_FN (cfun)->flags |= BB_REACHABLE; 2208 *tos++ = EXIT_BLOCK_PTR_FOR_FN (cfun); 2209 2210 /* Iterate: find everything reachable from what we've already seen. */ 2211 while (tos != worklist) 2212 { 2213 bb = *--tos; 2214 2215 FOR_EACH_EDGE (e, ei, bb->preds) 2216 if (e->flags & EDGE_ABNORMAL) 2217 { 2218 free (worklist); 2219 return true; 2220 } 2221 else 2222 { 2223 basic_block src = e->src; 2224 2225 if (!(src->flags & BB_REACHABLE)) 2226 { 2227 src->flags |= BB_REACHABLE; 2228 *tos++ = src; 2229 } 2230 } 2231 } 2232 free (worklist); 2233 /* No exceptional block reached exit unexceptionally. */ 2234 return false; 2235 } 2236 2237 2238 /* Process recursively X of INSN and add REG_INC notes if necessary. */ 2239 static void 2240 add_auto_inc_notes (rtx_insn *insn, rtx x) 2241 { 2242 enum rtx_code code = GET_CODE (x); 2243 const char *fmt; 2244 int i, j; 2245 2246 if (code == MEM && auto_inc_p (XEXP (x, 0))) 2247 { 2248 add_reg_note (insn, REG_INC, XEXP (XEXP (x, 0), 0)); 2249 return; 2250 } 2251 2252 /* Scan all X sub-expressions. */ 2253 fmt = GET_RTX_FORMAT (code); 2254 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) 2255 { 2256 if (fmt[i] == 'e') 2257 add_auto_inc_notes (insn, XEXP (x, i)); 2258 else if (fmt[i] == 'E') 2259 for (j = XVECLEN (x, i) - 1; j >= 0; j--) 2260 add_auto_inc_notes (insn, XVECEXP (x, i, j)); 2261 } 2262 } 2263 2264 2265 /* Remove all REG_DEAD and REG_UNUSED notes and regenerate REG_INC. 2266 We change pseudos by hard registers without notification of DF and 2267 that can make the notes obsolete. DF-infrastructure does not deal 2268 with REG_INC notes -- so we should regenerate them here. */ 2269 static void 2270 update_inc_notes (void) 2271 { 2272 rtx *pnote; 2273 basic_block bb; 2274 rtx_insn *insn; 2275 2276 FOR_EACH_BB_FN (bb, cfun) 2277 FOR_BB_INSNS (bb, insn) 2278 if (NONDEBUG_INSN_P (insn)) 2279 { 2280 pnote = ®_NOTES (insn); 2281 while (*pnote != 0) 2282 { 2283 if (REG_NOTE_KIND (*pnote) == REG_DEAD 2284 || REG_NOTE_KIND (*pnote) == REG_UNUSED 2285 || REG_NOTE_KIND (*pnote) == REG_INC) 2286 *pnote = XEXP (*pnote, 1); 2287 else 2288 pnote = &XEXP (*pnote, 1); 2289 } 2290 2291 if (AUTO_INC_DEC) 2292 add_auto_inc_notes (insn, PATTERN (insn)); 2293 } 2294 } 2295 2296 /* Set to 1 while in lra. */ 2297 int lra_in_progress; 2298 2299 /* Start of pseudo regnos before the LRA. */ 2300 int lra_new_regno_start; 2301 2302 /* Start of reload pseudo regnos before the new spill pass. */ 2303 int lra_constraint_new_regno_start; 2304 2305 /* Avoid spilling pseudos with regno more than the following value if 2306 it is possible. */ 2307 int lra_bad_spill_regno_start; 2308 2309 /* A pseudo of Pmode. */ 2310 rtx lra_pmode_pseudo; 2311 2312 /* Inheritance pseudo regnos before the new spill pass. */ 2313 bitmap_head lra_inheritance_pseudos; 2314 2315 /* Split regnos before the new spill pass. */ 2316 bitmap_head lra_split_regs; 2317 2318 /* Reload pseudo regnos before the new assignment pass which still can 2319 be spilled after the assignment pass as memory is also accepted in 2320 insns for the reload pseudos. */ 2321 bitmap_head lra_optional_reload_pseudos; 2322 2323 /* Pseudo regnos used for subreg reloads before the new assignment 2324 pass. Such pseudos still can be spilled after the assignment 2325 pass. */ 2326 bitmap_head lra_subreg_reload_pseudos; 2327 2328 /* File used for output of LRA debug information. */ 2329 FILE *lra_dump_file; 2330 2331 /* True if we found an asm error. */ 2332 bool lra_asm_error_p; 2333 2334 /* True if we should try spill into registers of different classes 2335 instead of memory. */ 2336 bool lra_reg_spill_p; 2337 2338 /* Set up value LRA_REG_SPILL_P. */ 2339 static void 2340 setup_reg_spill_flag (void) 2341 { 2342 int cl, mode; 2343 2344 if (targetm.spill_class != NULL) 2345 for (cl = 0; cl < (int) LIM_REG_CLASSES; cl++) 2346 for (mode = 0; mode < MAX_MACHINE_MODE; mode++) 2347 if (targetm.spill_class ((enum reg_class) cl, 2348 (machine_mode) mode) != NO_REGS) 2349 { 2350 lra_reg_spill_p = true; 2351 return; 2352 } 2353 lra_reg_spill_p = false; 2354 } 2355 2356 /* True if the current function is too big to use regular algorithms 2357 in LRA. In other words, we should use simpler and faster algorithms 2358 in LRA. It also means we should not worry about generation code 2359 for caller saves. The value is set up in IRA. */ 2360 bool lra_simple_p; 2361 2362 /* Major LRA entry function. F is a file should be used to dump LRA 2363 debug info. */ 2364 void 2365 lra (FILE *f) 2366 { 2367 int i; 2368 bool live_p, inserted_p; 2369 2370 lra_dump_file = f; 2371 lra_asm_error_p = false; 2372 lra_pmode_pseudo = gen_reg_rtx (Pmode); 2373 2374 timevar_push (TV_LRA); 2375 2376 /* Make sure that the last insn is a note. Some subsequent passes 2377 need it. */ 2378 emit_note (NOTE_INSN_DELETED); 2379 2380 lra_no_alloc_regs = ira_no_alloc_regs; 2381 2382 init_reg_info (); 2383 expand_reg_info (); 2384 2385 init_insn_recog_data (); 2386 2387 /* Some quick check on RTL generated by previous passes. */ 2388 if (flag_checking) 2389 check_rtl (false); 2390 2391 lra_in_progress = 1; 2392 2393 lra_live_range_iter = lra_coalesce_iter = lra_constraint_iter = 0; 2394 lra_assignment_iter = lra_assignment_iter_after_spill = 0; 2395 lra_inheritance_iter = lra_undo_inheritance_iter = 0; 2396 lra_rematerialization_iter = 0; 2397 2398 setup_reg_spill_flag (); 2399 2400 /* Function remove_scratches can creates new pseudos for clobbers -- 2401 so set up lra_constraint_new_regno_start before its call to 2402 permit changing reg classes for pseudos created by this 2403 simplification. */ 2404 lra_constraint_new_regno_start = lra_new_regno_start = max_reg_num (); 2405 lra_bad_spill_regno_start = INT_MAX; 2406 remove_scratches (); 2407 2408 /* A function that has a non-local label that can reach the exit 2409 block via non-exceptional paths must save all call-saved 2410 registers. */ 2411 if (cfun->has_nonlocal_label && has_nonexceptional_receiver ()) 2412 crtl->saves_all_registers = 1; 2413 2414 if (crtl->saves_all_registers) 2415 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) 2416 if (!crtl->abi->clobbers_full_reg_p (i) 2417 && !fixed_regs[i] 2418 && !LOCAL_REGNO (i)) 2419 df_set_regs_ever_live (i, true); 2420 2421 /* We don't DF from now and avoid its using because it is to 2422 expensive when a lot of RTL changes are made. */ 2423 df_set_flags (DF_NO_INSN_RESCAN); 2424 lra_constraint_insn_stack.create (get_max_uid ()); 2425 lra_constraint_insn_stack_bitmap = sbitmap_alloc (get_max_uid ()); 2426 bitmap_clear (lra_constraint_insn_stack_bitmap); 2427 lra_live_ranges_init (); 2428 lra_constraints_init (); 2429 lra_curr_reload_num = 0; 2430 push_insns (get_last_insn (), NULL); 2431 /* It is needed for the 1st coalescing. */ 2432 bitmap_initialize (&lra_inheritance_pseudos, ®_obstack); 2433 bitmap_initialize (&lra_split_regs, ®_obstack); 2434 bitmap_initialize (&lra_optional_reload_pseudos, ®_obstack); 2435 bitmap_initialize (&lra_subreg_reload_pseudos, ®_obstack); 2436 live_p = false; 2437 if (maybe_ne (get_frame_size (), 0) && crtl->stack_alignment_needed) 2438 /* If we have a stack frame, we must align it now. The stack size 2439 may be a part of the offset computation for register 2440 elimination. */ 2441 assign_stack_local (BLKmode, 0, crtl->stack_alignment_needed); 2442 lra_init_equiv (); 2443 for (;;) 2444 { 2445 for (;;) 2446 { 2447 bool reloads_p = lra_constraints (lra_constraint_iter == 0); 2448 /* Constraint transformations may result in that eliminable 2449 hard regs become uneliminable and pseudos which use them 2450 should be spilled. It is better to do it before pseudo 2451 assignments. 2452 2453 For example, rs6000 can make 2454 RS6000_PIC_OFFSET_TABLE_REGNUM uneliminable if we started 2455 to use a constant pool. */ 2456 lra_eliminate (false, false); 2457 /* We should try to assign hard registers to scratches even 2458 if there were no RTL transformations in lra_constraints. 2459 Also we should check IRA assignments on the first 2460 iteration as they can be wrong because of early clobbers 2461 operands which are ignored in IRA. */ 2462 if (! reloads_p && lra_constraint_iter > 1) 2463 { 2464 /* Stack is not empty here only when there are changes 2465 during the elimination sub-pass. */ 2466 if (bitmap_empty_p (lra_constraint_insn_stack_bitmap)) 2467 break; 2468 else 2469 /* If there are no reloads but changing due 2470 elimination, restart the constraint sub-pass 2471 first. */ 2472 continue; 2473 } 2474 /* Do inheritance only for regular algorithms. */ 2475 if (! lra_simple_p) 2476 lra_inheritance (); 2477 if (live_p) 2478 lra_clear_live_ranges (); 2479 bool fails_p; 2480 do 2481 { 2482 /* We need live ranges for lra_assign -- so build them. 2483 But don't remove dead insns or change global live 2484 info as we can undo inheritance transformations after 2485 inheritance pseudo assigning. */ 2486 lra_create_live_ranges (true, !lra_simple_p); 2487 live_p = true; 2488 /* If we don't spill non-reload and non-inheritance 2489 pseudos, there is no sense to run memory-memory move 2490 coalescing. If inheritance pseudos were spilled, the 2491 memory-memory moves involving them will be removed by 2492 pass undoing inheritance. */ 2493 if (lra_simple_p) 2494 lra_assign (fails_p); 2495 else 2496 { 2497 bool spill_p = !lra_assign (fails_p); 2498 2499 if (lra_undo_inheritance ()) 2500 live_p = false; 2501 if (spill_p && ! fails_p) 2502 { 2503 if (! live_p) 2504 { 2505 lra_create_live_ranges (true, true); 2506 live_p = true; 2507 } 2508 if (lra_coalesce ()) 2509 live_p = false; 2510 } 2511 if (! live_p) 2512 lra_clear_live_ranges (); 2513 } 2514 if (fails_p) 2515 { 2516 /* It is a very rare case. It is the last hope to 2517 split a hard regno live range for a reload 2518 pseudo. */ 2519 if (live_p) 2520 lra_clear_live_ranges (); 2521 live_p = false; 2522 if (! lra_split_hard_reg_for ()) 2523 break; 2524 } 2525 } 2526 while (fails_p); 2527 if (! live_p) { 2528 /* We need the correct reg notes for work of constraint sub-pass. */ 2529 lra_create_live_ranges (true, true); 2530 live_p = true; 2531 } 2532 } 2533 /* Don't clear optional reloads bitmap until all constraints are 2534 satisfied as we need to differ them from regular reloads. */ 2535 bitmap_clear (&lra_optional_reload_pseudos); 2536 bitmap_clear (&lra_subreg_reload_pseudos); 2537 bitmap_clear (&lra_inheritance_pseudos); 2538 bitmap_clear (&lra_split_regs); 2539 if (! live_p) 2540 { 2541 /* We need full live info for spilling pseudos into 2542 registers instead of memory. */ 2543 lra_create_live_ranges (lra_reg_spill_p, true); 2544 live_p = true; 2545 } 2546 /* We should check necessity for spilling here as the above live 2547 range pass can remove spilled pseudos. */ 2548 if (! lra_need_for_spills_p ()) 2549 break; 2550 /* Now we know what pseudos should be spilled. Try to 2551 rematerialize them first. */ 2552 if (lra_remat ()) 2553 { 2554 /* We need full live info -- see the comment above. */ 2555 lra_create_live_ranges (lra_reg_spill_p, true); 2556 live_p = true; 2557 if (! lra_need_for_spills_p ()) 2558 { 2559 if (lra_need_for_scratch_reg_p ()) 2560 continue; 2561 break; 2562 } 2563 } 2564 lra_spill (); 2565 /* Assignment of stack slots changes elimination offsets for 2566 some eliminations. So update the offsets here. */ 2567 lra_eliminate (false, false); 2568 lra_constraint_new_regno_start = max_reg_num (); 2569 if (lra_bad_spill_regno_start == INT_MAX 2570 && lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES 2571 && lra_rematerialization_iter > LRA_MAX_REMATERIALIZATION_PASSES) 2572 /* After switching off inheritance and rematerialization 2573 passes, avoid spilling reload pseudos will be created to 2574 prevent LRA cycling in some complicated cases. */ 2575 lra_bad_spill_regno_start = lra_constraint_new_regno_start; 2576 lra_assignment_iter_after_spill = 0; 2577 } 2578 restore_scratches (); 2579 lra_eliminate (true, false); 2580 lra_final_code_change (); 2581 lra_in_progress = 0; 2582 if (live_p) 2583 lra_clear_live_ranges (); 2584 lra_live_ranges_finish (); 2585 lra_constraints_finish (); 2586 finish_reg_info (); 2587 sbitmap_free (lra_constraint_insn_stack_bitmap); 2588 lra_constraint_insn_stack.release (); 2589 finish_insn_recog_data (); 2590 regstat_free_n_sets_and_refs (); 2591 regstat_free_ri (); 2592 reload_completed = 1; 2593 update_inc_notes (); 2594 2595 inserted_p = fixup_abnormal_edges (); 2596 2597 /* We've possibly turned single trapping insn into multiple ones. */ 2598 if (cfun->can_throw_non_call_exceptions) 2599 { 2600 auto_sbitmap blocks (last_basic_block_for_fn (cfun)); 2601 bitmap_ones (blocks); 2602 find_many_sub_basic_blocks (blocks); 2603 } 2604 2605 if (inserted_p) 2606 commit_edge_insertions (); 2607 2608 /* Replacing pseudos with their memory equivalents might have 2609 created shared rtx. Subsequent passes would get confused 2610 by this, so unshare everything here. */ 2611 unshare_all_rtl_again (get_insns ()); 2612 2613 if (flag_checking) 2614 check_rtl (true); 2615 2616 timevar_pop (TV_LRA); 2617 } 2618 2619 /* Called once per compiler to initialize LRA data once. */ 2620 void 2621 lra_init_once (void) 2622 { 2623 init_insn_code_data_once (); 2624 } 2625 2626 /* Called once per compiler to finish LRA data which are initialize 2627 once. */ 2628 void 2629 lra_finish_once (void) 2630 { 2631 finish_insn_code_data_once (); 2632 } 2633