1 /* Code for RTL transformations to satisfy insn constraints. 2 Copyright (C) 2010-2018 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 /* This file contains code for 3 passes: constraint pass, 23 inheritance/split pass, and pass for undoing failed inheritance and 24 split. 25 26 The major goal of constraint pass is to transform RTL to satisfy 27 insn and address constraints by: 28 o choosing insn alternatives; 29 o generating *reload insns* (or reloads in brief) and *reload 30 pseudos* which will get necessary hard registers later; 31 o substituting pseudos with equivalent values and removing the 32 instructions that initialized those pseudos. 33 34 The constraint pass has biggest and most complicated code in LRA. 35 There are a lot of important details like: 36 o reuse of input reload pseudos to simplify reload pseudo 37 allocations; 38 o some heuristics to choose insn alternative to improve the 39 inheritance; 40 o early clobbers etc. 41 42 The pass is mimicking former reload pass in alternative choosing 43 because the reload pass is oriented to current machine description 44 model. It might be changed if the machine description model is 45 changed. 46 47 There is special code for preventing all LRA and this pass cycling 48 in case of bugs. 49 50 On the first iteration of the pass we process every instruction and 51 choose an alternative for each one. On subsequent iterations we try 52 to avoid reprocessing instructions if we can be sure that the old 53 choice is still valid. 54 55 The inheritance/spilt pass is to transform code to achieve 56 ineheritance and live range splitting. It is done on backward 57 traversal of EBBs. 58 59 The inheritance optimization goal is to reuse values in hard 60 registers. There is analogous optimization in old reload pass. The 61 inheritance is achieved by following transformation: 62 63 reload_p1 <- p reload_p1 <- p 64 ... new_p <- reload_p1 65 ... => ... 66 reload_p2 <- p reload_p2 <- new_p 67 68 where p is spilled and not changed between the insns. Reload_p1 is 69 also called *original pseudo* and new_p is called *inheritance 70 pseudo*. 71 72 The subsequent assignment pass will try to assign the same (or 73 another if it is not possible) hard register to new_p as to 74 reload_p1 or reload_p2. 75 76 If the assignment pass fails to assign a hard register to new_p, 77 this file will undo the inheritance and restore the original code. 78 This is because implementing the above sequence with a spilled 79 new_p would make the code much worse. The inheritance is done in 80 EBB scope. The above is just a simplified example to get an idea 81 of the inheritance as the inheritance is also done for non-reload 82 insns. 83 84 Splitting (transformation) is also done in EBB scope on the same 85 pass as the inheritance: 86 87 r <- ... or ... <- r r <- ... or ... <- r 88 ... s <- r (new insn -- save) 89 ... => 90 ... r <- s (new insn -- restore) 91 ... <- r ... <- r 92 93 The *split pseudo* s is assigned to the hard register of the 94 original pseudo or hard register r. 95 96 Splitting is done: 97 o In EBBs with high register pressure for global pseudos (living 98 in at least 2 BBs) and assigned to hard registers when there 99 are more one reloads needing the hard registers; 100 o for pseudos needing save/restore code around calls. 101 102 If the split pseudo still has the same hard register as the 103 original pseudo after the subsequent assignment pass or the 104 original pseudo was split, the opposite transformation is done on 105 the same pass for undoing inheritance. */ 106 107 #undef REG_OK_STRICT 108 109 #include "config.h" 110 #include "system.h" 111 #include "coretypes.h" 112 #include "backend.h" 113 #include "target.h" 114 #include "rtl.h" 115 #include "tree.h" 116 #include "predict.h" 117 #include "df.h" 118 #include "memmodel.h" 119 #include "tm_p.h" 120 #include "expmed.h" 121 #include "optabs.h" 122 #include "regs.h" 123 #include "ira.h" 124 #include "recog.h" 125 #include "output.h" 126 #include "addresses.h" 127 #include "expr.h" 128 #include "cfgrtl.h" 129 #include "rtl-error.h" 130 #include "params.h" 131 #include "lra.h" 132 #include "lra-int.h" 133 #include "print-rtl.h" 134 135 /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current 136 insn. Remember that LRA_CURR_RELOAD_NUM is the number of emitted 137 reload insns. */ 138 static int bb_reload_num; 139 140 /* The current insn being processed and corresponding its single set 141 (NULL otherwise), its data (basic block, the insn data, the insn 142 static data, and the mode of each operand). */ 143 static rtx_insn *curr_insn; 144 static rtx curr_insn_set; 145 static basic_block curr_bb; 146 static lra_insn_recog_data_t curr_id; 147 static struct lra_static_insn_data *curr_static_id; 148 static machine_mode curr_operand_mode[MAX_RECOG_OPERANDS]; 149 /* Mode of the register substituted by its equivalence with VOIDmode 150 (e.g. constant) and whose subreg is given operand of the current 151 insn. VOIDmode in all other cases. */ 152 static machine_mode original_subreg_reg_mode[MAX_RECOG_OPERANDS]; 153 154 155 156 /* Start numbers for new registers and insns at the current constraints 157 pass start. */ 158 static int new_regno_start; 159 static int new_insn_uid_start; 160 161 /* If LOC is nonnull, strip any outer subreg from it. */ 162 static inline rtx * 163 strip_subreg (rtx *loc) 164 { 165 return loc && GET_CODE (*loc) == SUBREG ? &SUBREG_REG (*loc) : loc; 166 } 167 168 /* Return hard regno of REGNO or if it is was not assigned to a hard 169 register, use a hard register from its allocno class. */ 170 static int 171 get_try_hard_regno (int regno) 172 { 173 int hard_regno; 174 enum reg_class rclass; 175 176 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER) 177 hard_regno = lra_get_regno_hard_regno (regno); 178 if (hard_regno >= 0) 179 return hard_regno; 180 rclass = lra_get_allocno_class (regno); 181 if (rclass == NO_REGS) 182 return -1; 183 return ira_class_hard_regs[rclass][0]; 184 } 185 186 /* Return the hard regno of X after removing its subreg. If X is not 187 a register or a subreg of a register, return -1. If X is a pseudo, 188 use its assignment. If FINAL_P return the final hard regno which will 189 be after elimination. */ 190 static int 191 get_hard_regno (rtx x, bool final_p) 192 { 193 rtx reg; 194 int hard_regno; 195 196 reg = x; 197 if (SUBREG_P (x)) 198 reg = SUBREG_REG (x); 199 if (! REG_P (reg)) 200 return -1; 201 if (! HARD_REGISTER_NUM_P (hard_regno = REGNO (reg))) 202 hard_regno = lra_get_regno_hard_regno (hard_regno); 203 if (hard_regno < 0) 204 return -1; 205 if (final_p) 206 hard_regno = lra_get_elimination_hard_regno (hard_regno); 207 if (SUBREG_P (x)) 208 hard_regno += subreg_regno_offset (hard_regno, GET_MODE (reg), 209 SUBREG_BYTE (x), GET_MODE (x)); 210 return hard_regno; 211 } 212 213 /* If REGNO is a hard register or has been allocated a hard register, 214 return the class of that register. If REGNO is a reload pseudo 215 created by the current constraints pass, return its allocno class. 216 Return NO_REGS otherwise. */ 217 static enum reg_class 218 get_reg_class (int regno) 219 { 220 int hard_regno; 221 222 if (! HARD_REGISTER_NUM_P (hard_regno = regno)) 223 hard_regno = lra_get_regno_hard_regno (regno); 224 if (hard_regno >= 0) 225 { 226 hard_regno = lra_get_elimination_hard_regno (hard_regno); 227 return REGNO_REG_CLASS (hard_regno); 228 } 229 if (regno >= new_regno_start) 230 return lra_get_allocno_class (regno); 231 return NO_REGS; 232 } 233 234 /* Return true if REG satisfies (or will satisfy) reg class constraint 235 CL. Use elimination first if REG is a hard register. If REG is a 236 reload pseudo created by this constraints pass, assume that it will 237 be allocated a hard register from its allocno class, but allow that 238 class to be narrowed to CL if it is currently a superset of CL. 239 240 If NEW_CLASS is nonnull, set *NEW_CLASS to the new allocno class of 241 REGNO (reg), or NO_REGS if no change in its class was needed. */ 242 static bool 243 in_class_p (rtx reg, enum reg_class cl, enum reg_class *new_class) 244 { 245 enum reg_class rclass, common_class; 246 machine_mode reg_mode; 247 int class_size, hard_regno, nregs, i, j; 248 int regno = REGNO (reg); 249 250 if (new_class != NULL) 251 *new_class = NO_REGS; 252 if (regno < FIRST_PSEUDO_REGISTER) 253 { 254 rtx final_reg = reg; 255 rtx *final_loc = &final_reg; 256 257 lra_eliminate_reg_if_possible (final_loc); 258 return TEST_HARD_REG_BIT (reg_class_contents[cl], REGNO (*final_loc)); 259 } 260 reg_mode = GET_MODE (reg); 261 rclass = get_reg_class (regno); 262 if (regno < new_regno_start 263 /* Do not allow the constraints for reload instructions to 264 influence the classes of new pseudos. These reloads are 265 typically moves that have many alternatives, and restricting 266 reload pseudos for one alternative may lead to situations 267 where other reload pseudos are no longer allocatable. */ 268 || (INSN_UID (curr_insn) >= new_insn_uid_start 269 && curr_insn_set != NULL 270 && ((OBJECT_P (SET_SRC (curr_insn_set)) 271 && ! CONSTANT_P (SET_SRC (curr_insn_set))) 272 || (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG 273 && OBJECT_P (SUBREG_REG (SET_SRC (curr_insn_set))) 274 && ! CONSTANT_P (SUBREG_REG (SET_SRC (curr_insn_set))))))) 275 /* When we don't know what class will be used finally for reload 276 pseudos, we use ALL_REGS. */ 277 return ((regno >= new_regno_start && rclass == ALL_REGS) 278 || (rclass != NO_REGS && ira_class_subset_p[rclass][cl] 279 && ! hard_reg_set_subset_p (reg_class_contents[cl], 280 lra_no_alloc_regs))); 281 else 282 { 283 common_class = ira_reg_class_subset[rclass][cl]; 284 if (new_class != NULL) 285 *new_class = common_class; 286 if (hard_reg_set_subset_p (reg_class_contents[common_class], 287 lra_no_alloc_regs)) 288 return false; 289 /* Check that there are enough allocatable regs. */ 290 class_size = ira_class_hard_regs_num[common_class]; 291 for (i = 0; i < class_size; i++) 292 { 293 hard_regno = ira_class_hard_regs[common_class][i]; 294 nregs = hard_regno_nregs (hard_regno, reg_mode); 295 if (nregs == 1) 296 return true; 297 for (j = 0; j < nregs; j++) 298 if (TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno + j) 299 || ! TEST_HARD_REG_BIT (reg_class_contents[common_class], 300 hard_regno + j)) 301 break; 302 if (j >= nregs) 303 return true; 304 } 305 return false; 306 } 307 } 308 309 /* Return true if REGNO satisfies a memory constraint. */ 310 static bool 311 in_mem_p (int regno) 312 { 313 return get_reg_class (regno) == NO_REGS; 314 } 315 316 /* Return 1 if ADDR is a valid memory address for mode MODE in address 317 space AS, and check that each pseudo has the proper kind of hard 318 reg. */ 319 static int 320 valid_address_p (machine_mode mode ATTRIBUTE_UNUSED, 321 rtx addr, addr_space_t as) 322 { 323 #ifdef GO_IF_LEGITIMATE_ADDRESS 324 lra_assert (ADDR_SPACE_GENERIC_P (as)); 325 GO_IF_LEGITIMATE_ADDRESS (mode, addr, win); 326 return 0; 327 328 win: 329 return 1; 330 #else 331 return targetm.addr_space.legitimate_address_p (mode, addr, 0, as); 332 #endif 333 } 334 335 namespace { 336 /* Temporarily eliminates registers in an address (for the lifetime of 337 the object). */ 338 class address_eliminator { 339 public: 340 address_eliminator (struct address_info *ad); 341 ~address_eliminator (); 342 343 private: 344 struct address_info *m_ad; 345 rtx *m_base_loc; 346 rtx m_base_reg; 347 rtx *m_index_loc; 348 rtx m_index_reg; 349 }; 350 } 351 352 address_eliminator::address_eliminator (struct address_info *ad) 353 : m_ad (ad), 354 m_base_loc (strip_subreg (ad->base_term)), 355 m_base_reg (NULL_RTX), 356 m_index_loc (strip_subreg (ad->index_term)), 357 m_index_reg (NULL_RTX) 358 { 359 if (m_base_loc != NULL) 360 { 361 m_base_reg = *m_base_loc; 362 lra_eliminate_reg_if_possible (m_base_loc); 363 if (m_ad->base_term2 != NULL) 364 *m_ad->base_term2 = *m_ad->base_term; 365 } 366 if (m_index_loc != NULL) 367 { 368 m_index_reg = *m_index_loc; 369 lra_eliminate_reg_if_possible (m_index_loc); 370 } 371 } 372 373 address_eliminator::~address_eliminator () 374 { 375 if (m_base_loc && *m_base_loc != m_base_reg) 376 { 377 *m_base_loc = m_base_reg; 378 if (m_ad->base_term2 != NULL) 379 *m_ad->base_term2 = *m_ad->base_term; 380 } 381 if (m_index_loc && *m_index_loc != m_index_reg) 382 *m_index_loc = m_index_reg; 383 } 384 385 /* Return true if the eliminated form of AD is a legitimate target address. */ 386 static bool 387 valid_address_p (struct address_info *ad) 388 { 389 address_eliminator eliminator (ad); 390 return valid_address_p (ad->mode, *ad->outer, ad->as); 391 } 392 393 /* Return true if the eliminated form of memory reference OP satisfies 394 extra (special) memory constraint CONSTRAINT. */ 395 static bool 396 satisfies_memory_constraint_p (rtx op, enum constraint_num constraint) 397 { 398 struct address_info ad; 399 400 decompose_mem_address (&ad, op); 401 address_eliminator eliminator (&ad); 402 return constraint_satisfied_p (op, constraint); 403 } 404 405 /* Return true if the eliminated form of address AD satisfies extra 406 address constraint CONSTRAINT. */ 407 static bool 408 satisfies_address_constraint_p (struct address_info *ad, 409 enum constraint_num constraint) 410 { 411 address_eliminator eliminator (ad); 412 return constraint_satisfied_p (*ad->outer, constraint); 413 } 414 415 /* Return true if the eliminated form of address OP satisfies extra 416 address constraint CONSTRAINT. */ 417 static bool 418 satisfies_address_constraint_p (rtx op, enum constraint_num constraint) 419 { 420 struct address_info ad; 421 422 decompose_lea_address (&ad, &op); 423 return satisfies_address_constraint_p (&ad, constraint); 424 } 425 426 /* Initiate equivalences for LRA. As we keep original equivalences 427 before any elimination, we need to make copies otherwise any change 428 in insns might change the equivalences. */ 429 void 430 lra_init_equiv (void) 431 { 432 ira_expand_reg_equiv (); 433 for (int i = FIRST_PSEUDO_REGISTER; i < max_reg_num (); i++) 434 { 435 rtx res; 436 437 if ((res = ira_reg_equiv[i].memory) != NULL_RTX) 438 ira_reg_equiv[i].memory = copy_rtx (res); 439 if ((res = ira_reg_equiv[i].invariant) != NULL_RTX) 440 ira_reg_equiv[i].invariant = copy_rtx (res); 441 } 442 } 443 444 static rtx loc_equivalence_callback (rtx, const_rtx, void *); 445 446 /* Update equivalence for REGNO. We need to this as the equivalence 447 might contain other pseudos which are changed by their 448 equivalences. */ 449 static void 450 update_equiv (int regno) 451 { 452 rtx x; 453 454 if ((x = ira_reg_equiv[regno].memory) != NULL_RTX) 455 ira_reg_equiv[regno].memory 456 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback, 457 NULL_RTX); 458 if ((x = ira_reg_equiv[regno].invariant) != NULL_RTX) 459 ira_reg_equiv[regno].invariant 460 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback, 461 NULL_RTX); 462 } 463 464 /* If we have decided to substitute X with another value, return that 465 value, otherwise return X. */ 466 static rtx 467 get_equiv (rtx x) 468 { 469 int regno; 470 rtx res; 471 472 if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER 473 || ! ira_reg_equiv[regno].defined_p 474 || ! ira_reg_equiv[regno].profitable_p 475 || lra_get_regno_hard_regno (regno) >= 0) 476 return x; 477 if ((res = ira_reg_equiv[regno].memory) != NULL_RTX) 478 { 479 if (targetm.cannot_substitute_mem_equiv_p (res)) 480 return x; 481 return res; 482 } 483 if ((res = ira_reg_equiv[regno].constant) != NULL_RTX) 484 return res; 485 if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX) 486 return res; 487 gcc_unreachable (); 488 } 489 490 /* If we have decided to substitute X with the equivalent value, 491 return that value after elimination for INSN, otherwise return 492 X. */ 493 static rtx 494 get_equiv_with_elimination (rtx x, rtx_insn *insn) 495 { 496 rtx res = get_equiv (x); 497 498 if (x == res || CONSTANT_P (res)) 499 return res; 500 return lra_eliminate_regs_1 (insn, res, GET_MODE (res), 501 false, false, 0, true); 502 } 503 504 /* Set up curr_operand_mode. */ 505 static void 506 init_curr_operand_mode (void) 507 { 508 int nop = curr_static_id->n_operands; 509 for (int i = 0; i < nop; i++) 510 { 511 machine_mode mode = GET_MODE (*curr_id->operand_loc[i]); 512 if (mode == VOIDmode) 513 { 514 /* The .md mode for address operands is the mode of the 515 addressed value rather than the mode of the address itself. */ 516 if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address) 517 mode = Pmode; 518 else 519 mode = curr_static_id->operand[i].mode; 520 } 521 curr_operand_mode[i] = mode; 522 } 523 } 524 525 526 527 /* The page contains code to reuse input reloads. */ 528 529 /* Structure describes input reload of the current insns. */ 530 struct input_reload 531 { 532 /* True for input reload of matched operands. */ 533 bool match_p; 534 /* Reloaded value. */ 535 rtx input; 536 /* Reload pseudo used. */ 537 rtx reg; 538 }; 539 540 /* The number of elements in the following array. */ 541 static int curr_insn_input_reloads_num; 542 /* Array containing info about input reloads. It is used to find the 543 same input reload and reuse the reload pseudo in this case. */ 544 static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS]; 545 546 /* Initiate data concerning reuse of input reloads for the current 547 insn. */ 548 static void 549 init_curr_insn_input_reloads (void) 550 { 551 curr_insn_input_reloads_num = 0; 552 } 553 554 /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse already 555 created input reload pseudo (only if TYPE is not OP_OUT). Don't 556 reuse pseudo if IN_SUBREG_P is true and the reused pseudo should be 557 wrapped up in SUBREG. The result pseudo is returned through 558 RESULT_REG. Return TRUE if we created a new pseudo, FALSE if we 559 reused the already created input reload pseudo. Use TITLE to 560 describe new registers for debug purposes. */ 561 static bool 562 get_reload_reg (enum op_type type, machine_mode mode, rtx original, 563 enum reg_class rclass, bool in_subreg_p, 564 const char *title, rtx *result_reg) 565 { 566 int i, regno; 567 enum reg_class new_class; 568 bool unique_p = false; 569 570 if (type == OP_OUT) 571 { 572 *result_reg 573 = lra_create_new_reg_with_unique_value (mode, original, rclass, title); 574 return true; 575 } 576 /* Prevent reuse value of expression with side effects, 577 e.g. volatile memory. */ 578 if (! side_effects_p (original)) 579 for (i = 0; i < curr_insn_input_reloads_num; i++) 580 { 581 if (! curr_insn_input_reloads[i].match_p 582 && rtx_equal_p (curr_insn_input_reloads[i].input, original) 583 && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class)) 584 { 585 rtx reg = curr_insn_input_reloads[i].reg; 586 regno = REGNO (reg); 587 /* If input is equal to original and both are VOIDmode, 588 GET_MODE (reg) might be still different from mode. 589 Ensure we don't return *result_reg with wrong mode. */ 590 if (GET_MODE (reg) != mode) 591 { 592 if (in_subreg_p) 593 continue; 594 if (maybe_lt (GET_MODE_SIZE (GET_MODE (reg)), 595 GET_MODE_SIZE (mode))) 596 continue; 597 reg = lowpart_subreg (mode, reg, GET_MODE (reg)); 598 if (reg == NULL_RTX || GET_CODE (reg) != SUBREG) 599 continue; 600 } 601 *result_reg = reg; 602 if (lra_dump_file != NULL) 603 { 604 fprintf (lra_dump_file, " Reuse r%d for reload ", regno); 605 dump_value_slim (lra_dump_file, original, 1); 606 } 607 if (new_class != lra_get_allocno_class (regno)) 608 lra_change_class (regno, new_class, ", change to", false); 609 if (lra_dump_file != NULL) 610 fprintf (lra_dump_file, "\n"); 611 return false; 612 } 613 /* If we have an input reload with a different mode, make sure it 614 will get a different hard reg. */ 615 else if (REG_P (original) 616 && REG_P (curr_insn_input_reloads[i].input) 617 && REGNO (original) == REGNO (curr_insn_input_reloads[i].input) 618 && (GET_MODE (original) 619 != GET_MODE (curr_insn_input_reloads[i].input))) 620 unique_p = true; 621 } 622 *result_reg = (unique_p 623 ? lra_create_new_reg_with_unique_value 624 : lra_create_new_reg) (mode, original, rclass, title); 625 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS); 626 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original; 627 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = false; 628 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg; 629 return true; 630 } 631 632 633 634 /* The page contains code to extract memory address parts. */ 635 636 /* Wrapper around REGNO_OK_FOR_INDEX_P, to allow pseudos. */ 637 static inline bool 638 ok_for_index_p_nonstrict (rtx reg) 639 { 640 unsigned regno = REGNO (reg); 641 642 return regno >= FIRST_PSEUDO_REGISTER || REGNO_OK_FOR_INDEX_P (regno); 643 } 644 645 /* A version of regno_ok_for_base_p for use here, when all pseudos 646 should count as OK. Arguments as for regno_ok_for_base_p. */ 647 static inline bool 648 ok_for_base_p_nonstrict (rtx reg, machine_mode mode, addr_space_t as, 649 enum rtx_code outer_code, enum rtx_code index_code) 650 { 651 unsigned regno = REGNO (reg); 652 653 if (regno >= FIRST_PSEUDO_REGISTER) 654 return true; 655 return ok_for_base_p_1 (regno, mode, as, outer_code, index_code); 656 } 657 658 659 660 /* The page contains major code to choose the current insn alternative 661 and generate reloads for it. */ 662 663 /* Return the offset from REGNO of the least significant register 664 in (reg:MODE REGNO). 665 666 This function is used to tell whether two registers satisfy 667 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if: 668 669 REGNO1 + lra_constraint_offset (REGNO1, MODE1) 670 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */ 671 int 672 lra_constraint_offset (int regno, machine_mode mode) 673 { 674 lra_assert (regno < FIRST_PSEUDO_REGISTER); 675 676 scalar_int_mode int_mode; 677 if (WORDS_BIG_ENDIAN 678 && is_a <scalar_int_mode> (mode, &int_mode) 679 && GET_MODE_SIZE (int_mode) > UNITS_PER_WORD) 680 return hard_regno_nregs (regno, mode) - 1; 681 return 0; 682 } 683 684 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match 685 if they are the same hard reg, and has special hacks for 686 auto-increment and auto-decrement. This is specifically intended for 687 process_alt_operands to use in determining whether two operands 688 match. X is the operand whose number is the lower of the two. 689 690 It is supposed that X is the output operand and Y is the input 691 operand. Y_HARD_REGNO is the final hard regno of register Y or 692 register in subreg Y as we know it now. Otherwise, it is a 693 negative value. */ 694 static bool 695 operands_match_p (rtx x, rtx y, int y_hard_regno) 696 { 697 int i; 698 RTX_CODE code = GET_CODE (x); 699 const char *fmt; 700 701 if (x == y) 702 return true; 703 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x)))) 704 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))))) 705 { 706 int j; 707 708 i = get_hard_regno (x, false); 709 if (i < 0) 710 goto slow; 711 712 if ((j = y_hard_regno) < 0) 713 goto slow; 714 715 i += lra_constraint_offset (i, GET_MODE (x)); 716 j += lra_constraint_offset (j, GET_MODE (y)); 717 718 return i == j; 719 } 720 721 /* If two operands must match, because they are really a single 722 operand of an assembler insn, then two post-increments are invalid 723 because the assembler insn would increment only once. On the 724 other hand, a post-increment matches ordinary indexing if the 725 post-increment is the output operand. */ 726 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY) 727 return operands_match_p (XEXP (x, 0), y, y_hard_regno); 728 729 /* Two pre-increments are invalid because the assembler insn would 730 increment only once. On the other hand, a pre-increment matches 731 ordinary indexing if the pre-increment is the input operand. */ 732 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC 733 || GET_CODE (y) == PRE_MODIFY) 734 return operands_match_p (x, XEXP (y, 0), -1); 735 736 slow: 737 738 if (code == REG && REG_P (y)) 739 return REGNO (x) == REGNO (y); 740 741 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)) 742 && x == SUBREG_REG (y)) 743 return true; 744 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x)) 745 && SUBREG_REG (x) == y) 746 return true; 747 748 /* Now we have disposed of all the cases in which different rtx 749 codes can match. */ 750 if (code != GET_CODE (y)) 751 return false; 752 753 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */ 754 if (GET_MODE (x) != GET_MODE (y)) 755 return false; 756 757 switch (code) 758 { 759 CASE_CONST_UNIQUE: 760 return false; 761 762 case LABEL_REF: 763 return label_ref_label (x) == label_ref_label (y); 764 case SYMBOL_REF: 765 return XSTR (x, 0) == XSTR (y, 0); 766 767 default: 768 break; 769 } 770 771 /* Compare the elements. If any pair of corresponding elements fail 772 to match, return false for the whole things. */ 773 774 fmt = GET_RTX_FORMAT (code); 775 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) 776 { 777 int val, j; 778 switch (fmt[i]) 779 { 780 case 'w': 781 if (XWINT (x, i) != XWINT (y, i)) 782 return false; 783 break; 784 785 case 'i': 786 if (XINT (x, i) != XINT (y, i)) 787 return false; 788 break; 789 790 case 'p': 791 if (maybe_ne (SUBREG_BYTE (x), SUBREG_BYTE (y))) 792 return false; 793 break; 794 795 case 'e': 796 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1); 797 if (val == 0) 798 return false; 799 break; 800 801 case '0': 802 break; 803 804 case 'E': 805 if (XVECLEN (x, i) != XVECLEN (y, i)) 806 return false; 807 for (j = XVECLEN (x, i) - 1; j >= 0; --j) 808 { 809 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1); 810 if (val == 0) 811 return false; 812 } 813 break; 814 815 /* It is believed that rtx's at this level will never 816 contain anything but integers and other rtx's, except for 817 within LABEL_REFs and SYMBOL_REFs. */ 818 default: 819 gcc_unreachable (); 820 } 821 } 822 return true; 823 } 824 825 /* True if X is a constant that can be forced into the constant pool. 826 MODE is the mode of the operand, or VOIDmode if not known. */ 827 #define CONST_POOL_OK_P(MODE, X) \ 828 ((MODE) != VOIDmode \ 829 && CONSTANT_P (X) \ 830 && GET_CODE (X) != HIGH \ 831 && GET_MODE_SIZE (MODE).is_constant () \ 832 && !targetm.cannot_force_const_mem (MODE, X)) 833 834 /* True if C is a non-empty register class that has too few registers 835 to be safely used as a reload target class. */ 836 #define SMALL_REGISTER_CLASS_P(C) \ 837 (ira_class_hard_regs_num [(C)] == 1 \ 838 || (ira_class_hard_regs_num [(C)] >= 1 \ 839 && targetm.class_likely_spilled_p (C))) 840 841 /* If REG is a reload pseudo, try to make its class satisfying CL. */ 842 static void 843 narrow_reload_pseudo_class (rtx reg, enum reg_class cl) 844 { 845 enum reg_class rclass; 846 847 /* Do not make more accurate class from reloads generated. They are 848 mostly moves with a lot of constraints. Making more accurate 849 class may results in very narrow class and impossibility of find 850 registers for several reloads of one insn. */ 851 if (INSN_UID (curr_insn) >= new_insn_uid_start) 852 return; 853 if (GET_CODE (reg) == SUBREG) 854 reg = SUBREG_REG (reg); 855 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start) 856 return; 857 if (in_class_p (reg, cl, &rclass) && rclass != cl) 858 lra_change_class (REGNO (reg), rclass, " Change to", true); 859 } 860 861 /* Searches X for any reference to a reg with the same value as REGNO, 862 returning the rtx of the reference found if any. Otherwise, 863 returns NULL_RTX. */ 864 static rtx 865 regno_val_use_in (unsigned int regno, rtx x) 866 { 867 const char *fmt; 868 int i, j; 869 rtx tem; 870 871 if (REG_P (x) && lra_reg_info[REGNO (x)].val == lra_reg_info[regno].val) 872 return x; 873 874 fmt = GET_RTX_FORMAT (GET_CODE (x)); 875 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) 876 { 877 if (fmt[i] == 'e') 878 { 879 if ((tem = regno_val_use_in (regno, XEXP (x, i)))) 880 return tem; 881 } 882 else if (fmt[i] == 'E') 883 for (j = XVECLEN (x, i) - 1; j >= 0; j--) 884 if ((tem = regno_val_use_in (regno , XVECEXP (x, i, j)))) 885 return tem; 886 } 887 888 return NULL_RTX; 889 } 890 891 /* Return true if all current insn non-output operands except INS (it 892 has a negaitve end marker) do not use pseudos with the same value 893 as REGNO. */ 894 static bool 895 check_conflict_input_operands (int regno, signed char *ins) 896 { 897 int in; 898 int n_operands = curr_static_id->n_operands; 899 900 for (int nop = 0; nop < n_operands; nop++) 901 if (! curr_static_id->operand[nop].is_operator 902 && curr_static_id->operand[nop].type != OP_OUT) 903 { 904 for (int i = 0; (in = ins[i]) >= 0; i++) 905 if (in == nop) 906 break; 907 if (in < 0 908 && regno_val_use_in (regno, *curr_id->operand_loc[nop]) != NULL_RTX) 909 return false; 910 } 911 return true; 912 } 913 914 /* Generate reloads for matching OUT and INS (array of input operand 915 numbers with end marker -1) with reg class GOAL_CLASS, considering 916 output operands OUTS (similar array to INS) needing to be in different 917 registers. Add input and output reloads correspondingly to the lists 918 *BEFORE and *AFTER. OUT might be negative. In this case we generate 919 input reloads for matched input operands INS. EARLY_CLOBBER_P is a flag 920 that the output operand is early clobbered for chosen alternative. */ 921 static void 922 match_reload (signed char out, signed char *ins, signed char *outs, 923 enum reg_class goal_class, rtx_insn **before, 924 rtx_insn **after, bool early_clobber_p) 925 { 926 bool out_conflict; 927 int i, in; 928 rtx new_in_reg, new_out_reg, reg; 929 machine_mode inmode, outmode; 930 rtx in_rtx = *curr_id->operand_loc[ins[0]]; 931 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out]; 932 933 inmode = curr_operand_mode[ins[0]]; 934 outmode = out < 0 ? inmode : curr_operand_mode[out]; 935 push_to_sequence (*before); 936 if (inmode != outmode) 937 { 938 /* process_alt_operands has already checked that the mode sizes 939 are ordered. */ 940 if (partial_subreg_p (outmode, inmode)) 941 { 942 reg = new_in_reg 943 = lra_create_new_reg_with_unique_value (inmode, in_rtx, 944 goal_class, ""); 945 new_out_reg = gen_lowpart_SUBREG (outmode, reg); 946 LRA_SUBREG_P (new_out_reg) = 1; 947 /* If the input reg is dying here, we can use the same hard 948 register for REG and IN_RTX. We do it only for original 949 pseudos as reload pseudos can die although original 950 pseudos still live where reload pseudos dies. */ 951 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start 952 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx)) 953 && (!early_clobber_p 954 || check_conflict_input_operands(REGNO (in_rtx), ins))) 955 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg)); 956 } 957 else 958 { 959 reg = new_out_reg 960 = lra_create_new_reg_with_unique_value (outmode, out_rtx, 961 goal_class, ""); 962 new_in_reg = gen_lowpart_SUBREG (inmode, reg); 963 /* NEW_IN_REG is non-paradoxical subreg. We don't want 964 NEW_OUT_REG living above. We add clobber clause for 965 this. This is just a temporary clobber. We can remove 966 it at the end of LRA work. */ 967 rtx_insn *clobber = emit_clobber (new_out_reg); 968 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1; 969 LRA_SUBREG_P (new_in_reg) = 1; 970 if (GET_CODE (in_rtx) == SUBREG) 971 { 972 rtx subreg_reg = SUBREG_REG (in_rtx); 973 974 /* If SUBREG_REG is dying here and sub-registers IN_RTX 975 and NEW_IN_REG are similar, we can use the same hard 976 register for REG and SUBREG_REG. */ 977 if (REG_P (subreg_reg) 978 && (int) REGNO (subreg_reg) < lra_new_regno_start 979 && GET_MODE (subreg_reg) == outmode 980 && known_eq (SUBREG_BYTE (in_rtx), SUBREG_BYTE (new_in_reg)) 981 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg)) 982 && (! early_clobber_p 983 || check_conflict_input_operands (REGNO (subreg_reg), 984 ins))) 985 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg)); 986 } 987 } 988 } 989 else 990 { 991 /* Pseudos have values -- see comments for lra_reg_info. 992 Different pseudos with the same value do not conflict even if 993 they live in the same place. When we create a pseudo we 994 assign value of original pseudo (if any) from which we 995 created the new pseudo. If we create the pseudo from the 996 input pseudo, the new pseudo will have no conflict with the 997 input pseudo which is wrong when the input pseudo lives after 998 the insn and as the new pseudo value is changed by the insn 999 output. Therefore we create the new pseudo from the output 1000 except the case when we have single matched dying input 1001 pseudo. 1002 1003 We cannot reuse the current output register because we might 1004 have a situation like "a <- a op b", where the constraints 1005 force the second input operand ("b") to match the output 1006 operand ("a"). "b" must then be copied into a new register 1007 so that it doesn't clobber the current value of "a". 1008 1009 We can not use the same value if the output pseudo is 1010 early clobbered or the input pseudo is mentioned in the 1011 output, e.g. as an address part in memory, because 1012 output reload will actually extend the pseudo liveness. 1013 We don't care about eliminable hard regs here as we are 1014 interesting only in pseudos. */ 1015 1016 /* Matching input's register value is the same as one of the other 1017 output operand. Output operands in a parallel insn must be in 1018 different registers. */ 1019 out_conflict = false; 1020 if (REG_P (in_rtx)) 1021 { 1022 for (i = 0; outs[i] >= 0; i++) 1023 { 1024 rtx other_out_rtx = *curr_id->operand_loc[outs[i]]; 1025 if (REG_P (other_out_rtx) 1026 && (regno_val_use_in (REGNO (in_rtx), other_out_rtx) 1027 != NULL_RTX)) 1028 { 1029 out_conflict = true; 1030 break; 1031 } 1032 } 1033 } 1034 1035 new_in_reg = new_out_reg 1036 = (! early_clobber_p && ins[1] < 0 && REG_P (in_rtx) 1037 && (int) REGNO (in_rtx) < lra_new_regno_start 1038 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx)) 1039 && (! early_clobber_p 1040 || check_conflict_input_operands (REGNO (in_rtx), ins)) 1041 && (out < 0 1042 || regno_val_use_in (REGNO (in_rtx), out_rtx) == NULL_RTX) 1043 && !out_conflict 1044 ? lra_create_new_reg (inmode, in_rtx, goal_class, "") 1045 : lra_create_new_reg_with_unique_value (outmode, out_rtx, 1046 goal_class, "")); 1047 } 1048 /* In operand can be got from transformations before processing insn 1049 constraints. One example of such transformations is subreg 1050 reloading (see function simplify_operand_subreg). The new 1051 pseudos created by the transformations might have inaccurate 1052 class (ALL_REGS) and we should make their classes more 1053 accurate. */ 1054 narrow_reload_pseudo_class (in_rtx, goal_class); 1055 lra_emit_move (copy_rtx (new_in_reg), in_rtx); 1056 *before = get_insns (); 1057 end_sequence (); 1058 /* Add the new pseudo to consider values of subsequent input reload 1059 pseudos. */ 1060 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS); 1061 curr_insn_input_reloads[curr_insn_input_reloads_num].input = in_rtx; 1062 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = true; 1063 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = new_in_reg; 1064 for (i = 0; (in = ins[i]) >= 0; i++) 1065 { 1066 lra_assert 1067 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode 1068 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in])); 1069 *curr_id->operand_loc[in] = new_in_reg; 1070 } 1071 lra_update_dups (curr_id, ins); 1072 if (out < 0) 1073 return; 1074 /* See a comment for the input operand above. */ 1075 narrow_reload_pseudo_class (out_rtx, goal_class); 1076 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX) 1077 { 1078 start_sequence (); 1079 lra_emit_move (out_rtx, copy_rtx (new_out_reg)); 1080 emit_insn (*after); 1081 *after = get_insns (); 1082 end_sequence (); 1083 } 1084 *curr_id->operand_loc[out] = new_out_reg; 1085 lra_update_dup (curr_id, out); 1086 } 1087 1088 /* Return register class which is union of all reg classes in insn 1089 constraint alternative string starting with P. */ 1090 static enum reg_class 1091 reg_class_from_constraints (const char *p) 1092 { 1093 int c, len; 1094 enum reg_class op_class = NO_REGS; 1095 1096 do 1097 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c) 1098 { 1099 case '#': 1100 case ',': 1101 return op_class; 1102 1103 case 'g': 1104 op_class = reg_class_subunion[op_class][GENERAL_REGS]; 1105 break; 1106 1107 default: 1108 enum constraint_num cn = lookup_constraint (p); 1109 enum reg_class cl = reg_class_for_constraint (cn); 1110 if (cl == NO_REGS) 1111 { 1112 if (insn_extra_address_constraint (cn)) 1113 op_class 1114 = (reg_class_subunion 1115 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC, 1116 ADDRESS, SCRATCH)]); 1117 break; 1118 } 1119 1120 op_class = reg_class_subunion[op_class][cl]; 1121 break; 1122 } 1123 while ((p += len), c); 1124 return op_class; 1125 } 1126 1127 /* If OP is a register, return the class of the register as per 1128 get_reg_class, otherwise return NO_REGS. */ 1129 static inline enum reg_class 1130 get_op_class (rtx op) 1131 { 1132 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS; 1133 } 1134 1135 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo 1136 otherwise. If modes of MEM_PSEUDO and VAL are different, use 1137 SUBREG for VAL to make them equal. */ 1138 static rtx_insn * 1139 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val) 1140 { 1141 if (GET_MODE (mem_pseudo) != GET_MODE (val)) 1142 { 1143 /* Usually size of mem_pseudo is greater than val size but in 1144 rare cases it can be less as it can be defined by target 1145 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */ 1146 if (! MEM_P (val)) 1147 { 1148 val = gen_lowpart_SUBREG (GET_MODE (mem_pseudo), 1149 GET_CODE (val) == SUBREG 1150 ? SUBREG_REG (val) : val); 1151 LRA_SUBREG_P (val) = 1; 1152 } 1153 else 1154 { 1155 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo); 1156 LRA_SUBREG_P (mem_pseudo) = 1; 1157 } 1158 } 1159 return to_p ? gen_move_insn (mem_pseudo, val) 1160 : gen_move_insn (val, mem_pseudo); 1161 } 1162 1163 /* Process a special case insn (register move), return true if we 1164 don't need to process it anymore. INSN should be a single set 1165 insn. Set up that RTL was changed through CHANGE_P and that hook 1166 TARGET_SECONDARY_MEMORY_NEEDED says to use secondary memory through 1167 SEC_MEM_P. */ 1168 static bool 1169 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED) 1170 { 1171 int sregno, dregno; 1172 rtx dest, src, dreg, sreg, new_reg, scratch_reg; 1173 rtx_insn *before; 1174 enum reg_class dclass, sclass, secondary_class; 1175 secondary_reload_info sri; 1176 1177 lra_assert (curr_insn_set != NULL_RTX); 1178 dreg = dest = SET_DEST (curr_insn_set); 1179 sreg = src = SET_SRC (curr_insn_set); 1180 if (GET_CODE (dest) == SUBREG) 1181 dreg = SUBREG_REG (dest); 1182 if (GET_CODE (src) == SUBREG) 1183 sreg = SUBREG_REG (src); 1184 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg))) 1185 return false; 1186 sclass = dclass = NO_REGS; 1187 if (REG_P (dreg)) 1188 dclass = get_reg_class (REGNO (dreg)); 1189 gcc_assert (dclass < LIM_REG_CLASSES); 1190 if (dclass == ALL_REGS) 1191 /* ALL_REGS is used for new pseudos created by transformations 1192 like reload of SUBREG_REG (see function 1193 simplify_operand_subreg). We don't know their class yet. We 1194 should figure out the class from processing the insn 1195 constraints not in this fast path function. Even if ALL_REGS 1196 were a right class for the pseudo, secondary_... hooks usually 1197 are not define for ALL_REGS. */ 1198 return false; 1199 if (REG_P (sreg)) 1200 sclass = get_reg_class (REGNO (sreg)); 1201 gcc_assert (sclass < LIM_REG_CLASSES); 1202 if (sclass == ALL_REGS) 1203 /* See comments above. */ 1204 return false; 1205 if (sclass == NO_REGS && dclass == NO_REGS) 1206 return false; 1207 if (targetm.secondary_memory_needed (GET_MODE (src), sclass, dclass) 1208 && ((sclass != NO_REGS && dclass != NO_REGS) 1209 || (GET_MODE (src) 1210 != targetm.secondary_memory_needed_mode (GET_MODE (src))))) 1211 { 1212 *sec_mem_p = true; 1213 return false; 1214 } 1215 if (! REG_P (dreg) || ! REG_P (sreg)) 1216 return false; 1217 sri.prev_sri = NULL; 1218 sri.icode = CODE_FOR_nothing; 1219 sri.extra_cost = 0; 1220 secondary_class = NO_REGS; 1221 /* Set up hard register for a reload pseudo for hook 1222 secondary_reload because some targets just ignore unassigned 1223 pseudos in the hook. */ 1224 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0) 1225 { 1226 dregno = REGNO (dreg); 1227 reg_renumber[dregno] = ira_class_hard_regs[dclass][0]; 1228 } 1229 else 1230 dregno = -1; 1231 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0) 1232 { 1233 sregno = REGNO (sreg); 1234 reg_renumber[sregno] = ira_class_hard_regs[sclass][0]; 1235 } 1236 else 1237 sregno = -1; 1238 if (sclass != NO_REGS) 1239 secondary_class 1240 = (enum reg_class) targetm.secondary_reload (false, dest, 1241 (reg_class_t) sclass, 1242 GET_MODE (src), &sri); 1243 if (sclass == NO_REGS 1244 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing) 1245 && dclass != NO_REGS)) 1246 { 1247 enum reg_class old_sclass = secondary_class; 1248 secondary_reload_info old_sri = sri; 1249 1250 sri.prev_sri = NULL; 1251 sri.icode = CODE_FOR_nothing; 1252 sri.extra_cost = 0; 1253 secondary_class 1254 = (enum reg_class) targetm.secondary_reload (true, src, 1255 (reg_class_t) dclass, 1256 GET_MODE (src), &sri); 1257 /* Check the target hook consistency. */ 1258 lra_assert 1259 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing) 1260 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing) 1261 || (secondary_class == old_sclass && sri.icode == old_sri.icode)); 1262 } 1263 if (sregno >= 0) 1264 reg_renumber [sregno] = -1; 1265 if (dregno >= 0) 1266 reg_renumber [dregno] = -1; 1267 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing) 1268 return false; 1269 *change_p = true; 1270 new_reg = NULL_RTX; 1271 if (secondary_class != NO_REGS) 1272 new_reg = lra_create_new_reg_with_unique_value (GET_MODE (src), NULL_RTX, 1273 secondary_class, 1274 "secondary"); 1275 start_sequence (); 1276 if (sri.icode == CODE_FOR_nothing) 1277 lra_emit_move (new_reg, src); 1278 else 1279 { 1280 enum reg_class scratch_class; 1281 1282 scratch_class = (reg_class_from_constraints 1283 (insn_data[sri.icode].operand[2].constraint)); 1284 scratch_reg = (lra_create_new_reg_with_unique_value 1285 (insn_data[sri.icode].operand[2].mode, NULL_RTX, 1286 scratch_class, "scratch")); 1287 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest, 1288 src, scratch_reg)); 1289 } 1290 before = get_insns (); 1291 end_sequence (); 1292 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move"); 1293 if (new_reg != NULL_RTX) 1294 SET_SRC (curr_insn_set) = new_reg; 1295 else 1296 { 1297 if (lra_dump_file != NULL) 1298 { 1299 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn)); 1300 dump_insn_slim (lra_dump_file, curr_insn); 1301 } 1302 lra_set_insn_deleted (curr_insn); 1303 return true; 1304 } 1305 return false; 1306 } 1307 1308 /* The following data describe the result of process_alt_operands. 1309 The data are used in curr_insn_transform to generate reloads. */ 1310 1311 /* The chosen reg classes which should be used for the corresponding 1312 operands. */ 1313 static enum reg_class goal_alt[MAX_RECOG_OPERANDS]; 1314 /* True if the operand should be the same as another operand and that 1315 other operand does not need a reload. */ 1316 static bool goal_alt_match_win[MAX_RECOG_OPERANDS]; 1317 /* True if the operand does not need a reload. */ 1318 static bool goal_alt_win[MAX_RECOG_OPERANDS]; 1319 /* True if the operand can be offsetable memory. */ 1320 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS]; 1321 /* The number of an operand to which given operand can be matched to. */ 1322 static int goal_alt_matches[MAX_RECOG_OPERANDS]; 1323 /* The number of elements in the following array. */ 1324 static int goal_alt_dont_inherit_ops_num; 1325 /* Numbers of operands whose reload pseudos should not be inherited. */ 1326 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS]; 1327 /* True if the insn commutative operands should be swapped. */ 1328 static bool goal_alt_swapped; 1329 /* The chosen insn alternative. */ 1330 static int goal_alt_number; 1331 1332 /* True if the corresponding operand is the result of an equivalence 1333 substitution. */ 1334 static bool equiv_substition_p[MAX_RECOG_OPERANDS]; 1335 1336 /* The following five variables are used to choose the best insn 1337 alternative. They reflect final characteristics of the best 1338 alternative. */ 1339 1340 /* Number of necessary reloads and overall cost reflecting the 1341 previous value and other unpleasantness of the best alternative. */ 1342 static int best_losers, best_overall; 1343 /* Overall number hard registers used for reloads. For example, on 1344 some targets we need 2 general registers to reload DFmode and only 1345 one floating point register. */ 1346 static int best_reload_nregs; 1347 /* Overall number reflecting distances of previous reloading the same 1348 value. The distances are counted from the current BB start. It is 1349 used to improve inheritance chances. */ 1350 static int best_reload_sum; 1351 1352 /* True if the current insn should have no correspondingly input or 1353 output reloads. */ 1354 static bool no_input_reloads_p, no_output_reloads_p; 1355 1356 /* True if we swapped the commutative operands in the current 1357 insn. */ 1358 static int curr_swapped; 1359 1360 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a 1361 register of class CL. Add any input reloads to list BEFORE. AFTER 1362 is nonnull if *LOC is an automodified value; handle that case by 1363 adding the required output reloads to list AFTER. Return true if 1364 the RTL was changed. 1365 1366 if CHECK_ONLY_P is true, check that the *LOC is a correct address 1367 register. Return false if the address register is correct. */ 1368 static bool 1369 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after, 1370 enum reg_class cl) 1371 { 1372 int regno; 1373 enum reg_class rclass, new_class; 1374 rtx reg; 1375 rtx new_reg; 1376 machine_mode mode; 1377 bool subreg_p, before_p = false; 1378 1379 subreg_p = GET_CODE (*loc) == SUBREG; 1380 if (subreg_p) 1381 { 1382 reg = SUBREG_REG (*loc); 1383 mode = GET_MODE (reg); 1384 1385 /* For mode with size bigger than ptr_mode, there unlikely to be "mov" 1386 between two registers with different classes, but there normally will 1387 be "mov" which transfers element of vector register into the general 1388 register, and this normally will be a subreg which should be reloaded 1389 as a whole. This is particularly likely to be triggered when 1390 -fno-split-wide-types specified. */ 1391 if (!REG_P (reg) 1392 || in_class_p (reg, cl, &new_class) 1393 || known_le (GET_MODE_SIZE (mode), GET_MODE_SIZE (ptr_mode))) 1394 loc = &SUBREG_REG (*loc); 1395 } 1396 1397 reg = *loc; 1398 mode = GET_MODE (reg); 1399 if (! REG_P (reg)) 1400 { 1401 if (check_only_p) 1402 return true; 1403 /* Always reload memory in an address even if the target supports 1404 such addresses. */ 1405 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address"); 1406 before_p = true; 1407 } 1408 else 1409 { 1410 regno = REGNO (reg); 1411 rclass = get_reg_class (regno); 1412 if (! check_only_p 1413 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg) 1414 { 1415 if (lra_dump_file != NULL) 1416 { 1417 fprintf (lra_dump_file, 1418 "Changing pseudo %d in address of insn %u on equiv ", 1419 REGNO (reg), INSN_UID (curr_insn)); 1420 dump_value_slim (lra_dump_file, *loc, 1); 1421 fprintf (lra_dump_file, "\n"); 1422 } 1423 *loc = copy_rtx (*loc); 1424 } 1425 if (*loc != reg || ! in_class_p (reg, cl, &new_class)) 1426 { 1427 if (check_only_p) 1428 return true; 1429 reg = *loc; 1430 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT, 1431 mode, reg, cl, subreg_p, "address", &new_reg)) 1432 before_p = true; 1433 } 1434 else if (new_class != NO_REGS && rclass != new_class) 1435 { 1436 if (check_only_p) 1437 return true; 1438 lra_change_class (regno, new_class, " Change to", true); 1439 return false; 1440 } 1441 else 1442 return false; 1443 } 1444 if (before_p) 1445 { 1446 push_to_sequence (*before); 1447 lra_emit_move (new_reg, reg); 1448 *before = get_insns (); 1449 end_sequence (); 1450 } 1451 *loc = new_reg; 1452 if (after != NULL) 1453 { 1454 start_sequence (); 1455 lra_emit_move (before_p ? copy_rtx (reg) : reg, new_reg); 1456 emit_insn (*after); 1457 *after = get_insns (); 1458 end_sequence (); 1459 } 1460 return true; 1461 } 1462 1463 /* Insert move insn in simplify_operand_subreg. BEFORE returns 1464 the insn to be inserted before curr insn. AFTER returns the 1465 the insn to be inserted after curr insn. ORIGREG and NEWREG 1466 are the original reg and new reg for reload. */ 1467 static void 1468 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg, 1469 rtx newreg) 1470 { 1471 if (before) 1472 { 1473 push_to_sequence (*before); 1474 lra_emit_move (newreg, origreg); 1475 *before = get_insns (); 1476 end_sequence (); 1477 } 1478 if (after) 1479 { 1480 start_sequence (); 1481 lra_emit_move (origreg, newreg); 1482 emit_insn (*after); 1483 *after = get_insns (); 1484 end_sequence (); 1485 } 1486 } 1487 1488 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as); 1489 static bool process_address (int, bool, rtx_insn **, rtx_insn **); 1490 1491 /* Make reloads for subreg in operand NOP with internal subreg mode 1492 REG_MODE, add new reloads for further processing. Return true if 1493 any change was done. */ 1494 static bool 1495 simplify_operand_subreg (int nop, machine_mode reg_mode) 1496 { 1497 int hard_regno; 1498 rtx_insn *before, *after; 1499 machine_mode mode, innermode; 1500 rtx reg, new_reg; 1501 rtx operand = *curr_id->operand_loc[nop]; 1502 enum reg_class regclass; 1503 enum op_type type; 1504 1505 before = after = NULL; 1506 1507 if (GET_CODE (operand) != SUBREG) 1508 return false; 1509 1510 mode = GET_MODE (operand); 1511 reg = SUBREG_REG (operand); 1512 innermode = GET_MODE (reg); 1513 type = curr_static_id->operand[nop].type; 1514 if (MEM_P (reg)) 1515 { 1516 const bool addr_was_valid 1517 = valid_address_p (innermode, XEXP (reg, 0), MEM_ADDR_SPACE (reg)); 1518 alter_subreg (curr_id->operand_loc[nop], false); 1519 rtx subst = *curr_id->operand_loc[nop]; 1520 lra_assert (MEM_P (subst)); 1521 1522 if (!addr_was_valid 1523 || valid_address_p (GET_MODE (subst), XEXP (subst, 0), 1524 MEM_ADDR_SPACE (subst)) 1525 || ((get_constraint_type (lookup_constraint 1526 (curr_static_id->operand[nop].constraint)) 1527 != CT_SPECIAL_MEMORY) 1528 /* We still can reload address and if the address is 1529 valid, we can remove subreg without reloading its 1530 inner memory. */ 1531 && valid_address_p (GET_MODE (subst), 1532 regno_reg_rtx 1533 [ira_class_hard_regs 1534 [base_reg_class (GET_MODE (subst), 1535 MEM_ADDR_SPACE (subst), 1536 ADDRESS, SCRATCH)][0]], 1537 MEM_ADDR_SPACE (subst)))) 1538 { 1539 /* If we change the address for a paradoxical subreg of memory, the 1540 new address might violate the necessary alignment or the access 1541 might be slow; take this into consideration. We need not worry 1542 about accesses beyond allocated memory for paradoxical memory 1543 subregs as we don't substitute such equiv memory (see processing 1544 equivalences in function lra_constraints) and because for spilled 1545 pseudos we allocate stack memory enough for the biggest 1546 corresponding paradoxical subreg. 1547 1548 However, do not blindly simplify a (subreg (mem ...)) for 1549 WORD_REGISTER_OPERATIONS targets as this may lead to loading junk 1550 data into a register when the inner is narrower than outer or 1551 missing important data from memory when the inner is wider than 1552 outer. This rule only applies to modes that are no wider than 1553 a word. */ 1554 if (!(maybe_ne (GET_MODE_PRECISION (mode), 1555 GET_MODE_PRECISION (innermode)) 1556 && known_le (GET_MODE_SIZE (mode), UNITS_PER_WORD) 1557 && known_le (GET_MODE_SIZE (innermode), UNITS_PER_WORD) 1558 && WORD_REGISTER_OPERATIONS) 1559 && (!(MEM_ALIGN (subst) < GET_MODE_ALIGNMENT (mode) 1560 && targetm.slow_unaligned_access (mode, MEM_ALIGN (subst))) 1561 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode) 1562 && targetm.slow_unaligned_access (innermode, 1563 MEM_ALIGN (reg))))) 1564 return true; 1565 1566 *curr_id->operand_loc[nop] = operand; 1567 1568 /* But if the address was not valid, we cannot reload the MEM without 1569 reloading the address first. */ 1570 if (!addr_was_valid) 1571 process_address (nop, false, &before, &after); 1572 1573 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */ 1574 enum reg_class rclass 1575 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS); 1576 if (get_reload_reg (curr_static_id->operand[nop].type, innermode, 1577 reg, rclass, TRUE, "slow mem", &new_reg)) 1578 { 1579 bool insert_before, insert_after; 1580 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg)); 1581 1582 insert_before = (type != OP_OUT 1583 || partial_subreg_p (mode, innermode)); 1584 insert_after = type != OP_IN; 1585 insert_move_for_subreg (insert_before ? &before : NULL, 1586 insert_after ? &after : NULL, 1587 reg, new_reg); 1588 } 1589 SUBREG_REG (operand) = new_reg; 1590 1591 /* Convert to MODE. */ 1592 reg = operand; 1593 rclass 1594 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS); 1595 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg, 1596 rclass, TRUE, "slow mem", &new_reg)) 1597 { 1598 bool insert_before, insert_after; 1599 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg)); 1600 1601 insert_before = type != OP_OUT; 1602 insert_after = type != OP_IN; 1603 insert_move_for_subreg (insert_before ? &before : NULL, 1604 insert_after ? &after : NULL, 1605 reg, new_reg); 1606 } 1607 *curr_id->operand_loc[nop] = new_reg; 1608 lra_process_new_insns (curr_insn, before, after, 1609 "Inserting slow mem reload"); 1610 return true; 1611 } 1612 1613 /* If the address was valid and became invalid, prefer to reload 1614 the memory. Typical case is when the index scale should 1615 correspond the memory. */ 1616 *curr_id->operand_loc[nop] = operand; 1617 /* Do not return false here as the MEM_P (reg) will be processed 1618 later in this function. */ 1619 } 1620 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER) 1621 { 1622 alter_subreg (curr_id->operand_loc[nop], false); 1623 return true; 1624 } 1625 else if (CONSTANT_P (reg)) 1626 { 1627 /* Try to simplify subreg of constant. It is usually result of 1628 equivalence substitution. */ 1629 if (innermode == VOIDmode 1630 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode) 1631 innermode = curr_static_id->operand[nop].mode; 1632 if ((new_reg = simplify_subreg (mode, reg, innermode, 1633 SUBREG_BYTE (operand))) != NULL_RTX) 1634 { 1635 *curr_id->operand_loc[nop] = new_reg; 1636 return true; 1637 } 1638 } 1639 /* Put constant into memory when we have mixed modes. It generates 1640 a better code in most cases as it does not need a secondary 1641 reload memory. It also prevents LRA looping when LRA is using 1642 secondary reload memory again and again. */ 1643 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg) 1644 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode)) 1645 { 1646 SUBREG_REG (operand) = force_const_mem (reg_mode, reg); 1647 alter_subreg (curr_id->operand_loc[nop], false); 1648 return true; 1649 } 1650 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or 1651 if there may be a problem accessing OPERAND in the outer 1652 mode. */ 1653 if ((REG_P (reg) 1654 && REGNO (reg) >= FIRST_PSEUDO_REGISTER 1655 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0 1656 /* Don't reload paradoxical subregs because we could be looping 1657 having repeatedly final regno out of hard regs range. */ 1658 && (hard_regno_nregs (hard_regno, innermode) 1659 >= hard_regno_nregs (hard_regno, mode)) 1660 && simplify_subreg_regno (hard_regno, innermode, 1661 SUBREG_BYTE (operand), mode) < 0 1662 /* Don't reload subreg for matching reload. It is actually 1663 valid subreg in LRA. */ 1664 && ! LRA_SUBREG_P (operand)) 1665 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg)) 1666 { 1667 enum reg_class rclass; 1668 1669 if (REG_P (reg)) 1670 /* There is a big probability that we will get the same class 1671 for the new pseudo and we will get the same insn which 1672 means infinite looping. So spill the new pseudo. */ 1673 rclass = NO_REGS; 1674 else 1675 /* The class will be defined later in curr_insn_transform. */ 1676 rclass 1677 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS); 1678 1679 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg, 1680 rclass, TRUE, "subreg reg", &new_reg)) 1681 { 1682 bool insert_before, insert_after; 1683 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg)); 1684 1685 insert_before = (type != OP_OUT 1686 || read_modify_subreg_p (operand)); 1687 insert_after = (type != OP_IN); 1688 insert_move_for_subreg (insert_before ? &before : NULL, 1689 insert_after ? &after : NULL, 1690 reg, new_reg); 1691 } 1692 SUBREG_REG (operand) = new_reg; 1693 lra_process_new_insns (curr_insn, before, after, 1694 "Inserting subreg reload"); 1695 return true; 1696 } 1697 /* Force a reload for a paradoxical subreg. For paradoxical subreg, 1698 IRA allocates hardreg to the inner pseudo reg according to its mode 1699 instead of the outermode, so the size of the hardreg may not be enough 1700 to contain the outermode operand, in that case we may need to insert 1701 reload for the reg. For the following two types of paradoxical subreg, 1702 we need to insert reload: 1703 1. If the op_type is OP_IN, and the hardreg could not be paired with 1704 other hardreg to contain the outermode operand 1705 (checked by in_hard_reg_set_p), we need to insert the reload. 1706 2. If the op_type is OP_OUT or OP_INOUT. 1707 1708 Here is a paradoxical subreg example showing how the reload is generated: 1709 1710 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ]) 1711 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64} 1712 1713 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example 1714 here, if reg107 is assigned to hardreg R15, because R15 is the last 1715 hardreg, compiler cannot find another hardreg to pair with R15 to 1716 contain TImode data. So we insert a TImode reload reg180 for it. 1717 After reload is inserted: 1718 1719 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0) 1720 (reg:DI 107 [ __comp ])) -1 1721 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ]) 1722 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64} 1723 1724 Two reload hard registers will be allocated to reg180 to save TImode data 1725 in LRA_assign. */ 1726 else if (REG_P (reg) 1727 && REGNO (reg) >= FIRST_PSEUDO_REGISTER 1728 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0 1729 && (hard_regno_nregs (hard_regno, innermode) 1730 < hard_regno_nregs (hard_regno, mode)) 1731 && (regclass = lra_get_allocno_class (REGNO (reg))) 1732 && (type != OP_IN 1733 || !in_hard_reg_set_p (reg_class_contents[regclass], 1734 mode, hard_regno))) 1735 { 1736 /* The class will be defined later in curr_insn_transform. */ 1737 enum reg_class rclass 1738 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS); 1739 1740 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg, 1741 rclass, TRUE, "paradoxical subreg", &new_reg)) 1742 { 1743 rtx subreg; 1744 bool insert_before, insert_after; 1745 1746 PUT_MODE (new_reg, mode); 1747 subreg = gen_lowpart_SUBREG (innermode, new_reg); 1748 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg)); 1749 1750 insert_before = (type != OP_OUT); 1751 insert_after = (type != OP_IN); 1752 insert_move_for_subreg (insert_before ? &before : NULL, 1753 insert_after ? &after : NULL, 1754 reg, subreg); 1755 } 1756 SUBREG_REG (operand) = new_reg; 1757 lra_process_new_insns (curr_insn, before, after, 1758 "Inserting paradoxical subreg reload"); 1759 return true; 1760 } 1761 return false; 1762 } 1763 1764 /* Return TRUE if X refers for a hard register from SET. */ 1765 static bool 1766 uses_hard_regs_p (rtx x, HARD_REG_SET set) 1767 { 1768 int i, j, x_hard_regno; 1769 machine_mode mode; 1770 const char *fmt; 1771 enum rtx_code code; 1772 1773 if (x == NULL_RTX) 1774 return false; 1775 code = GET_CODE (x); 1776 mode = GET_MODE (x); 1777 if (code == SUBREG) 1778 { 1779 mode = wider_subreg_mode (x); 1780 x = SUBREG_REG (x); 1781 code = GET_CODE (x); 1782 } 1783 1784 if (REG_P (x)) 1785 { 1786 x_hard_regno = get_hard_regno (x, true); 1787 return (x_hard_regno >= 0 1788 && overlaps_hard_reg_set_p (set, mode, x_hard_regno)); 1789 } 1790 if (MEM_P (x)) 1791 { 1792 struct address_info ad; 1793 1794 decompose_mem_address (&ad, x); 1795 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set)) 1796 return true; 1797 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set)) 1798 return true; 1799 } 1800 fmt = GET_RTX_FORMAT (code); 1801 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) 1802 { 1803 if (fmt[i] == 'e') 1804 { 1805 if (uses_hard_regs_p (XEXP (x, i), set)) 1806 return true; 1807 } 1808 else if (fmt[i] == 'E') 1809 { 1810 for (j = XVECLEN (x, i) - 1; j >= 0; j--) 1811 if (uses_hard_regs_p (XVECEXP (x, i, j), set)) 1812 return true; 1813 } 1814 } 1815 return false; 1816 } 1817 1818 /* Return true if OP is a spilled pseudo. */ 1819 static inline bool 1820 spilled_pseudo_p (rtx op) 1821 { 1822 return (REG_P (op) 1823 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op))); 1824 } 1825 1826 /* Return true if X is a general constant. */ 1827 static inline bool 1828 general_constant_p (rtx x) 1829 { 1830 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x)); 1831 } 1832 1833 static bool 1834 reg_in_class_p (rtx reg, enum reg_class cl) 1835 { 1836 if (cl == NO_REGS) 1837 return get_reg_class (REGNO (reg)) == NO_REGS; 1838 return in_class_p (reg, cl, NULL); 1839 } 1840 1841 /* Return true if SET of RCLASS contains no hard regs which can be 1842 used in MODE. */ 1843 static bool 1844 prohibited_class_reg_set_mode_p (enum reg_class rclass, 1845 HARD_REG_SET &set, 1846 machine_mode mode) 1847 { 1848 HARD_REG_SET temp; 1849 1850 lra_assert (hard_reg_set_subset_p (reg_class_contents[rclass], set)); 1851 COPY_HARD_REG_SET (temp, set); 1852 AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs); 1853 return (hard_reg_set_subset_p 1854 (temp, ira_prohibited_class_mode_regs[rclass][mode])); 1855 } 1856 1857 1858 /* Used to check validity info about small class input operands. It 1859 should be incremented at start of processing an insn 1860 alternative. */ 1861 static unsigned int curr_small_class_check = 0; 1862 1863 /* Update number of used inputs of class OP_CLASS for operand NOP. 1864 Return true if we have more such class operands than the number of 1865 available regs. */ 1866 static bool 1867 update_and_check_small_class_inputs (int nop, enum reg_class op_class) 1868 { 1869 static unsigned int small_class_check[LIM_REG_CLASSES]; 1870 static int small_class_input_nums[LIM_REG_CLASSES]; 1871 1872 if (SMALL_REGISTER_CLASS_P (op_class) 1873 /* We are interesting in classes became small because of fixing 1874 some hard regs, e.g. by an user through GCC options. */ 1875 && hard_reg_set_intersect_p (reg_class_contents[op_class], 1876 ira_no_alloc_regs) 1877 && (curr_static_id->operand[nop].type != OP_OUT 1878 || curr_static_id->operand[nop].early_clobber)) 1879 { 1880 if (small_class_check[op_class] == curr_small_class_check) 1881 small_class_input_nums[op_class]++; 1882 else 1883 { 1884 small_class_check[op_class] = curr_small_class_check; 1885 small_class_input_nums[op_class] = 1; 1886 } 1887 if (small_class_input_nums[op_class] > ira_class_hard_regs_num[op_class]) 1888 return true; 1889 } 1890 return false; 1891 } 1892 1893 /* Major function to choose the current insn alternative and what 1894 operands should be reloaded and how. If ONLY_ALTERNATIVE is not 1895 negative we should consider only this alternative. Return false if 1896 we can not choose the alternative or find how to reload the 1897 operands. */ 1898 static bool 1899 process_alt_operands (int only_alternative) 1900 { 1901 bool ok_p = false; 1902 int nop, overall, nalt; 1903 int n_alternatives = curr_static_id->n_alternatives; 1904 int n_operands = curr_static_id->n_operands; 1905 /* LOSERS counts the operands that don't fit this alternative and 1906 would require loading. */ 1907 int losers; 1908 int addr_losers; 1909 /* REJECT is a count of how undesirable this alternative says it is 1910 if any reloading is required. If the alternative matches exactly 1911 then REJECT is ignored, but otherwise it gets this much counted 1912 against it in addition to the reloading needed. */ 1913 int reject; 1914 /* This is defined by '!' or '?' alternative constraint and added to 1915 reject. But in some cases it can be ignored. */ 1916 int static_reject; 1917 int op_reject; 1918 /* The number of elements in the following array. */ 1919 int early_clobbered_regs_num; 1920 /* Numbers of operands which are early clobber registers. */ 1921 int early_clobbered_nops[MAX_RECOG_OPERANDS]; 1922 enum reg_class curr_alt[MAX_RECOG_OPERANDS]; 1923 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS]; 1924 bool curr_alt_match_win[MAX_RECOG_OPERANDS]; 1925 bool curr_alt_win[MAX_RECOG_OPERANDS]; 1926 bool curr_alt_offmemok[MAX_RECOG_OPERANDS]; 1927 int curr_alt_matches[MAX_RECOG_OPERANDS]; 1928 /* The number of elements in the following array. */ 1929 int curr_alt_dont_inherit_ops_num; 1930 /* Numbers of operands whose reload pseudos should not be inherited. */ 1931 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS]; 1932 rtx op; 1933 /* The register when the operand is a subreg of register, otherwise the 1934 operand itself. */ 1935 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS]; 1936 /* The register if the operand is a register or subreg of register, 1937 otherwise NULL. */ 1938 rtx operand_reg[MAX_RECOG_OPERANDS]; 1939 int hard_regno[MAX_RECOG_OPERANDS]; 1940 machine_mode biggest_mode[MAX_RECOG_OPERANDS]; 1941 int reload_nregs, reload_sum; 1942 bool costly_p; 1943 enum reg_class cl; 1944 1945 /* Calculate some data common for all alternatives to speed up the 1946 function. */ 1947 for (nop = 0; nop < n_operands; nop++) 1948 { 1949 rtx reg; 1950 1951 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop]; 1952 /* The real hard regno of the operand after the allocation. */ 1953 hard_regno[nop] = get_hard_regno (op, true); 1954 1955 operand_reg[nop] = reg = op; 1956 biggest_mode[nop] = GET_MODE (op); 1957 if (GET_CODE (op) == SUBREG) 1958 { 1959 biggest_mode[nop] = wider_subreg_mode (op); 1960 operand_reg[nop] = reg = SUBREG_REG (op); 1961 } 1962 if (! REG_P (reg)) 1963 operand_reg[nop] = NULL_RTX; 1964 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER 1965 || ((int) REGNO (reg) 1966 == lra_get_elimination_hard_regno (REGNO (reg)))) 1967 no_subreg_reg_operand[nop] = reg; 1968 else 1969 operand_reg[nop] = no_subreg_reg_operand[nop] 1970 /* Just use natural mode for elimination result. It should 1971 be enough for extra constraints hooks. */ 1972 = regno_reg_rtx[hard_regno[nop]]; 1973 } 1974 1975 /* The constraints are made of several alternatives. Each operand's 1976 constraint looks like foo,bar,... with commas separating the 1977 alternatives. The first alternatives for all operands go 1978 together, the second alternatives go together, etc. 1979 1980 First loop over alternatives. */ 1981 alternative_mask preferred = curr_id->preferred_alternatives; 1982 if (only_alternative >= 0) 1983 preferred &= ALTERNATIVE_BIT (only_alternative); 1984 1985 for (nalt = 0; nalt < n_alternatives; nalt++) 1986 { 1987 /* Loop over operands for one constraint alternative. */ 1988 if (!TEST_BIT (preferred, nalt)) 1989 continue; 1990 1991 curr_small_class_check++; 1992 overall = losers = addr_losers = 0; 1993 static_reject = reject = reload_nregs = reload_sum = 0; 1994 for (nop = 0; nop < n_operands; nop++) 1995 { 1996 int inc = (curr_static_id 1997 ->operand_alternative[nalt * n_operands + nop].reject); 1998 if (lra_dump_file != NULL && inc != 0) 1999 fprintf (lra_dump_file, 2000 " Staticly defined alt reject+=%d\n", inc); 2001 static_reject += inc; 2002 } 2003 reject += static_reject; 2004 early_clobbered_regs_num = 0; 2005 2006 for (nop = 0; nop < n_operands; nop++) 2007 { 2008 const char *p; 2009 char *end; 2010 int len, c, m, i, opalt_num, this_alternative_matches; 2011 bool win, did_match, offmemok, early_clobber_p; 2012 /* false => this operand can be reloaded somehow for this 2013 alternative. */ 2014 bool badop; 2015 /* true => this operand can be reloaded if the alternative 2016 allows regs. */ 2017 bool winreg; 2018 /* True if a constant forced into memory would be OK for 2019 this operand. */ 2020 bool constmemok; 2021 enum reg_class this_alternative, this_costly_alternative; 2022 HARD_REG_SET this_alternative_set, this_costly_alternative_set; 2023 bool this_alternative_match_win, this_alternative_win; 2024 bool this_alternative_offmemok; 2025 bool scratch_p; 2026 machine_mode mode; 2027 enum constraint_num cn; 2028 2029 opalt_num = nalt * n_operands + nop; 2030 if (curr_static_id->operand_alternative[opalt_num].anything_ok) 2031 { 2032 /* Fast track for no constraints at all. */ 2033 curr_alt[nop] = NO_REGS; 2034 CLEAR_HARD_REG_SET (curr_alt_set[nop]); 2035 curr_alt_win[nop] = true; 2036 curr_alt_match_win[nop] = false; 2037 curr_alt_offmemok[nop] = false; 2038 curr_alt_matches[nop] = -1; 2039 continue; 2040 } 2041 2042 op = no_subreg_reg_operand[nop]; 2043 mode = curr_operand_mode[nop]; 2044 2045 win = did_match = winreg = offmemok = constmemok = false; 2046 badop = true; 2047 2048 early_clobber_p = false; 2049 p = curr_static_id->operand_alternative[opalt_num].constraint; 2050 2051 this_costly_alternative = this_alternative = NO_REGS; 2052 /* We update set of possible hard regs besides its class 2053 because reg class might be inaccurate. For example, 2054 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM 2055 is translated in HI_REGS because classes are merged by 2056 pairs and there is no accurate intermediate class. */ 2057 CLEAR_HARD_REG_SET (this_alternative_set); 2058 CLEAR_HARD_REG_SET (this_costly_alternative_set); 2059 this_alternative_win = false; 2060 this_alternative_match_win = false; 2061 this_alternative_offmemok = false; 2062 this_alternative_matches = -1; 2063 2064 /* An empty constraint should be excluded by the fast 2065 track. */ 2066 lra_assert (*p != 0 && *p != ','); 2067 2068 op_reject = 0; 2069 /* Scan this alternative's specs for this operand; set WIN 2070 if the operand fits any letter in this alternative. 2071 Otherwise, clear BADOP if this operand could fit some 2072 letter after reloads, or set WINREG if this operand could 2073 fit after reloads provided the constraint allows some 2074 registers. */ 2075 costly_p = false; 2076 do 2077 { 2078 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c) 2079 { 2080 case '\0': 2081 len = 0; 2082 break; 2083 case ',': 2084 c = '\0'; 2085 break; 2086 2087 case '&': 2088 early_clobber_p = true; 2089 break; 2090 2091 case '$': 2092 op_reject += LRA_MAX_REJECT; 2093 break; 2094 case '^': 2095 op_reject += LRA_LOSER_COST_FACTOR; 2096 break; 2097 2098 case '#': 2099 /* Ignore rest of this alternative. */ 2100 c = '\0'; 2101 break; 2102 2103 case '0': case '1': case '2': case '3': case '4': 2104 case '5': case '6': case '7': case '8': case '9': 2105 { 2106 int m_hregno; 2107 bool match_p; 2108 2109 m = strtoul (p, &end, 10); 2110 p = end; 2111 len = 0; 2112 lra_assert (nop > m); 2113 2114 /* Reject matches if we don't know which operand is 2115 bigger. This situation would arguably be a bug in 2116 an .md pattern, but could also occur in a user asm. */ 2117 if (!ordered_p (GET_MODE_SIZE (biggest_mode[m]), 2118 GET_MODE_SIZE (biggest_mode[nop]))) 2119 break; 2120 2121 /* Don't match wrong asm insn operands for proper 2122 diagnostic later. */ 2123 if (INSN_CODE (curr_insn) < 0 2124 && (curr_operand_mode[m] == BLKmode 2125 || curr_operand_mode[nop] == BLKmode) 2126 && curr_operand_mode[m] != curr_operand_mode[nop]) 2127 break; 2128 2129 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false); 2130 /* We are supposed to match a previous operand. 2131 If we do, we win if that one did. If we do 2132 not, count both of the operands as losers. 2133 (This is too conservative, since most of the 2134 time only a single reload insn will be needed 2135 to make the two operands win. As a result, 2136 this alternative may be rejected when it is 2137 actually desirable.) */ 2138 match_p = false; 2139 if (operands_match_p (*curr_id->operand_loc[nop], 2140 *curr_id->operand_loc[m], m_hregno)) 2141 { 2142 /* We should reject matching of an early 2143 clobber operand if the matching operand is 2144 not dying in the insn. */ 2145 if (! curr_static_id->operand[m].early_clobber 2146 || operand_reg[nop] == NULL_RTX 2147 || (find_regno_note (curr_insn, REG_DEAD, 2148 REGNO (op)) 2149 || REGNO (op) == REGNO (operand_reg[m]))) 2150 match_p = true; 2151 } 2152 if (match_p) 2153 { 2154 /* If we are matching a non-offsettable 2155 address where an offsettable address was 2156 expected, then we must reject this 2157 combination, because we can't reload 2158 it. */ 2159 if (curr_alt_offmemok[m] 2160 && MEM_P (*curr_id->operand_loc[m]) 2161 && curr_alt[m] == NO_REGS && ! curr_alt_win[m]) 2162 continue; 2163 } 2164 else 2165 { 2166 /* Operands don't match. Both operands must 2167 allow a reload register, otherwise we 2168 cannot make them match. */ 2169 if (curr_alt[m] == NO_REGS) 2170 break; 2171 /* Retroactively mark the operand we had to 2172 match as a loser, if it wasn't already and 2173 it wasn't matched to a register constraint 2174 (e.g it might be matched by memory). */ 2175 if (curr_alt_win[m] 2176 && (operand_reg[m] == NULL_RTX 2177 || hard_regno[m] < 0)) 2178 { 2179 losers++; 2180 reload_nregs 2181 += (ira_reg_class_max_nregs[curr_alt[m]] 2182 [GET_MODE (*curr_id->operand_loc[m])]); 2183 } 2184 2185 /* Prefer matching earlyclobber alternative as 2186 it results in less hard regs required for 2187 the insn than a non-matching earlyclobber 2188 alternative. */ 2189 if (curr_static_id->operand[m].early_clobber) 2190 { 2191 if (lra_dump_file != NULL) 2192 fprintf 2193 (lra_dump_file, 2194 " %d Matching earlyclobber alt:" 2195 " reject--\n", 2196 nop); 2197 reject--; 2198 } 2199 /* Otherwise we prefer no matching 2200 alternatives because it gives more freedom 2201 in RA. */ 2202 else if (operand_reg[nop] == NULL_RTX 2203 || (find_regno_note (curr_insn, REG_DEAD, 2204 REGNO (operand_reg[nop])) 2205 == NULL_RTX)) 2206 { 2207 if (lra_dump_file != NULL) 2208 fprintf 2209 (lra_dump_file, 2210 " %d Matching alt: reject+=2\n", 2211 nop); 2212 reject += 2; 2213 } 2214 } 2215 /* If we have to reload this operand and some 2216 previous operand also had to match the same 2217 thing as this operand, we don't know how to do 2218 that. */ 2219 if (!match_p || !curr_alt_win[m]) 2220 { 2221 for (i = 0; i < nop; i++) 2222 if (curr_alt_matches[i] == m) 2223 break; 2224 if (i < nop) 2225 break; 2226 } 2227 else 2228 did_match = true; 2229 2230 this_alternative_matches = m; 2231 /* This can be fixed with reloads if the operand 2232 we are supposed to match can be fixed with 2233 reloads. */ 2234 badop = false; 2235 this_alternative = curr_alt[m]; 2236 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]); 2237 winreg = this_alternative != NO_REGS; 2238 break; 2239 } 2240 2241 case 'g': 2242 if (MEM_P (op) 2243 || general_constant_p (op) 2244 || spilled_pseudo_p (op)) 2245 win = true; 2246 cl = GENERAL_REGS; 2247 goto reg; 2248 2249 default: 2250 cn = lookup_constraint (p); 2251 switch (get_constraint_type (cn)) 2252 { 2253 case CT_REGISTER: 2254 cl = reg_class_for_constraint (cn); 2255 if (cl != NO_REGS) 2256 goto reg; 2257 break; 2258 2259 case CT_CONST_INT: 2260 if (CONST_INT_P (op) 2261 && insn_const_int_ok_for_constraint (INTVAL (op), cn)) 2262 win = true; 2263 break; 2264 2265 case CT_MEMORY: 2266 if (MEM_P (op) 2267 && satisfies_memory_constraint_p (op, cn)) 2268 win = true; 2269 else if (spilled_pseudo_p (op)) 2270 win = true; 2271 2272 /* If we didn't already win, we can reload constants 2273 via force_const_mem or put the pseudo value into 2274 memory, or make other memory by reloading the 2275 address like for 'o'. */ 2276 if (CONST_POOL_OK_P (mode, op) 2277 || MEM_P (op) || REG_P (op) 2278 /* We can restore the equiv insn by a 2279 reload. */ 2280 || equiv_substition_p[nop]) 2281 badop = false; 2282 constmemok = true; 2283 offmemok = true; 2284 break; 2285 2286 case CT_ADDRESS: 2287 /* An asm operand with an address constraint 2288 that doesn't satisfy address_operand has 2289 is_address cleared, so that we don't try to 2290 make a non-address fit. */ 2291 if (!curr_static_id->operand[nop].is_address) 2292 break; 2293 /* If we didn't already win, we can reload the address 2294 into a base register. */ 2295 if (satisfies_address_constraint_p (op, cn)) 2296 win = true; 2297 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC, 2298 ADDRESS, SCRATCH); 2299 badop = false; 2300 goto reg; 2301 2302 case CT_FIXED_FORM: 2303 if (constraint_satisfied_p (op, cn)) 2304 win = true; 2305 break; 2306 2307 case CT_SPECIAL_MEMORY: 2308 if (MEM_P (op) 2309 && satisfies_memory_constraint_p (op, cn)) 2310 win = true; 2311 else if (spilled_pseudo_p (op)) 2312 win = true; 2313 break; 2314 } 2315 break; 2316 2317 reg: 2318 this_alternative = reg_class_subunion[this_alternative][cl]; 2319 IOR_HARD_REG_SET (this_alternative_set, 2320 reg_class_contents[cl]); 2321 if (costly_p) 2322 { 2323 this_costly_alternative 2324 = reg_class_subunion[this_costly_alternative][cl]; 2325 IOR_HARD_REG_SET (this_costly_alternative_set, 2326 reg_class_contents[cl]); 2327 } 2328 if (mode == BLKmode) 2329 break; 2330 winreg = true; 2331 if (REG_P (op)) 2332 { 2333 if (hard_regno[nop] >= 0 2334 && in_hard_reg_set_p (this_alternative_set, 2335 mode, hard_regno[nop])) 2336 win = true; 2337 else if (hard_regno[nop] < 0 2338 && in_class_p (op, this_alternative, NULL)) 2339 win = true; 2340 } 2341 break; 2342 } 2343 if (c != ' ' && c != '\t') 2344 costly_p = c == '*'; 2345 } 2346 while ((p += len), c); 2347 2348 scratch_p = (operand_reg[nop] != NULL_RTX 2349 && lra_former_scratch_p (REGNO (operand_reg[nop]))); 2350 /* Record which operands fit this alternative. */ 2351 if (win) 2352 { 2353 this_alternative_win = true; 2354 if (operand_reg[nop] != NULL_RTX) 2355 { 2356 if (hard_regno[nop] >= 0) 2357 { 2358 if (in_hard_reg_set_p (this_costly_alternative_set, 2359 mode, hard_regno[nop])) 2360 { 2361 if (lra_dump_file != NULL) 2362 fprintf (lra_dump_file, 2363 " %d Costly set: reject++\n", 2364 nop); 2365 reject++; 2366 } 2367 } 2368 else 2369 { 2370 /* Prefer won reg to spilled pseudo under other 2371 equal conditions for possibe inheritance. */ 2372 if (! scratch_p) 2373 { 2374 if (lra_dump_file != NULL) 2375 fprintf 2376 (lra_dump_file, 2377 " %d Non pseudo reload: reject++\n", 2378 nop); 2379 reject++; 2380 } 2381 if (in_class_p (operand_reg[nop], 2382 this_costly_alternative, NULL)) 2383 { 2384 if (lra_dump_file != NULL) 2385 fprintf 2386 (lra_dump_file, 2387 " %d Non pseudo costly reload:" 2388 " reject++\n", 2389 nop); 2390 reject++; 2391 } 2392 } 2393 /* We simulate the behavior of old reload here. 2394 Although scratches need hard registers and it 2395 might result in spilling other pseudos, no reload 2396 insns are generated for the scratches. So it 2397 might cost something but probably less than old 2398 reload pass believes. */ 2399 if (scratch_p) 2400 { 2401 if (lra_dump_file != NULL) 2402 fprintf (lra_dump_file, 2403 " %d Scratch win: reject+=2\n", 2404 nop); 2405 reject += 2; 2406 } 2407 } 2408 } 2409 else if (did_match) 2410 this_alternative_match_win = true; 2411 else 2412 { 2413 int const_to_mem = 0; 2414 bool no_regs_p; 2415 2416 reject += op_reject; 2417 /* Never do output reload of stack pointer. It makes 2418 impossible to do elimination when SP is changed in 2419 RTL. */ 2420 if (op == stack_pointer_rtx && ! frame_pointer_needed 2421 && curr_static_id->operand[nop].type != OP_IN) 2422 goto fail; 2423 2424 /* If this alternative asks for a specific reg class, see if there 2425 is at least one allocatable register in that class. */ 2426 no_regs_p 2427 = (this_alternative == NO_REGS 2428 || (hard_reg_set_subset_p 2429 (reg_class_contents[this_alternative], 2430 lra_no_alloc_regs))); 2431 2432 /* For asms, verify that the class for this alternative is possible 2433 for the mode that is specified. */ 2434 if (!no_regs_p && INSN_CODE (curr_insn) < 0) 2435 { 2436 int i; 2437 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) 2438 if (targetm.hard_regno_mode_ok (i, mode) 2439 && in_hard_reg_set_p (reg_class_contents[this_alternative], 2440 mode, i)) 2441 break; 2442 if (i == FIRST_PSEUDO_REGISTER) 2443 winreg = false; 2444 } 2445 2446 /* If this operand accepts a register, and if the 2447 register class has at least one allocatable register, 2448 then this operand can be reloaded. */ 2449 if (winreg && !no_regs_p) 2450 badop = false; 2451 2452 if (badop) 2453 { 2454 if (lra_dump_file != NULL) 2455 fprintf (lra_dump_file, 2456 " alt=%d: Bad operand -- refuse\n", 2457 nalt); 2458 goto fail; 2459 } 2460 2461 if (this_alternative != NO_REGS) 2462 { 2463 HARD_REG_SET available_regs; 2464 2465 COPY_HARD_REG_SET (available_regs, 2466 reg_class_contents[this_alternative]); 2467 AND_COMPL_HARD_REG_SET 2468 (available_regs, 2469 ira_prohibited_class_mode_regs[this_alternative][mode]); 2470 AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs); 2471 if (hard_reg_set_empty_p (available_regs)) 2472 { 2473 /* There are no hard regs holding a value of given 2474 mode. */ 2475 if (offmemok) 2476 { 2477 this_alternative = NO_REGS; 2478 if (lra_dump_file != NULL) 2479 fprintf (lra_dump_file, 2480 " %d Using memory because of" 2481 " a bad mode: reject+=2\n", 2482 nop); 2483 reject += 2; 2484 } 2485 else 2486 { 2487 if (lra_dump_file != NULL) 2488 fprintf (lra_dump_file, 2489 " alt=%d: Wrong mode -- refuse\n", 2490 nalt); 2491 goto fail; 2492 } 2493 } 2494 } 2495 2496 /* If not assigned pseudo has a class which a subset of 2497 required reg class, it is a less costly alternative 2498 as the pseudo still can get a hard reg of necessary 2499 class. */ 2500 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0 2501 && (cl = get_reg_class (REGNO (op))) != NO_REGS 2502 && ira_class_subset_p[this_alternative][cl]) 2503 { 2504 if (lra_dump_file != NULL) 2505 fprintf 2506 (lra_dump_file, 2507 " %d Super set class reg: reject-=3\n", nop); 2508 reject -= 3; 2509 } 2510 2511 this_alternative_offmemok = offmemok; 2512 if (this_costly_alternative != NO_REGS) 2513 { 2514 if (lra_dump_file != NULL) 2515 fprintf (lra_dump_file, 2516 " %d Costly loser: reject++\n", nop); 2517 reject++; 2518 } 2519 /* If the operand is dying, has a matching constraint, 2520 and satisfies constraints of the matched operand 2521 which failed to satisfy the own constraints, most probably 2522 the reload for this operand will be gone. */ 2523 if (this_alternative_matches >= 0 2524 && !curr_alt_win[this_alternative_matches] 2525 && REG_P (op) 2526 && find_regno_note (curr_insn, REG_DEAD, REGNO (op)) 2527 && (hard_regno[nop] >= 0 2528 ? in_hard_reg_set_p (this_alternative_set, 2529 mode, hard_regno[nop]) 2530 : in_class_p (op, this_alternative, NULL))) 2531 { 2532 if (lra_dump_file != NULL) 2533 fprintf 2534 (lra_dump_file, 2535 " %d Dying matched operand reload: reject++\n", 2536 nop); 2537 reject++; 2538 } 2539 else 2540 { 2541 /* Strict_low_part requires to reload the register 2542 not the sub-register. In this case we should 2543 check that a final reload hard reg can hold the 2544 value mode. */ 2545 if (curr_static_id->operand[nop].strict_low 2546 && REG_P (op) 2547 && hard_regno[nop] < 0 2548 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG 2549 && ira_class_hard_regs_num[this_alternative] > 0 2550 && (!targetm.hard_regno_mode_ok 2551 (ira_class_hard_regs[this_alternative][0], 2552 GET_MODE (*curr_id->operand_loc[nop])))) 2553 { 2554 if (lra_dump_file != NULL) 2555 fprintf 2556 (lra_dump_file, 2557 " alt=%d: Strict low subreg reload -- refuse\n", 2558 nalt); 2559 goto fail; 2560 } 2561 losers++; 2562 } 2563 if (operand_reg[nop] != NULL_RTX 2564 /* Output operands and matched input operands are 2565 not inherited. The following conditions do not 2566 exactly describe the previous statement but they 2567 are pretty close. */ 2568 && curr_static_id->operand[nop].type != OP_OUT 2569 && (this_alternative_matches < 0 2570 || curr_static_id->operand[nop].type != OP_IN)) 2571 { 2572 int last_reload = (lra_reg_info[ORIGINAL_REGNO 2573 (operand_reg[nop])] 2574 .last_reload); 2575 2576 /* The value of reload_sum has sense only if we 2577 process insns in their order. It happens only on 2578 the first constraints sub-pass when we do most of 2579 reload work. */ 2580 if (lra_constraint_iter == 1 && last_reload > bb_reload_num) 2581 reload_sum += last_reload - bb_reload_num; 2582 } 2583 /* If this is a constant that is reloaded into the 2584 desired class by copying it to memory first, count 2585 that as another reload. This is consistent with 2586 other code and is required to avoid choosing another 2587 alternative when the constant is moved into memory. 2588 Note that the test here is precisely the same as in 2589 the code below that calls force_const_mem. */ 2590 if (CONST_POOL_OK_P (mode, op) 2591 && ((targetm.preferred_reload_class 2592 (op, this_alternative) == NO_REGS) 2593 || no_input_reloads_p)) 2594 { 2595 const_to_mem = 1; 2596 if (! no_regs_p) 2597 losers++; 2598 } 2599 2600 /* Alternative loses if it requires a type of reload not 2601 permitted for this insn. We can always reload 2602 objects with a REG_UNUSED note. */ 2603 if ((curr_static_id->operand[nop].type != OP_IN 2604 && no_output_reloads_p 2605 && ! find_reg_note (curr_insn, REG_UNUSED, op)) 2606 || (curr_static_id->operand[nop].type != OP_OUT 2607 && no_input_reloads_p && ! const_to_mem) 2608 || (this_alternative_matches >= 0 2609 && (no_input_reloads_p 2610 || (no_output_reloads_p 2611 && (curr_static_id->operand 2612 [this_alternative_matches].type != OP_IN) 2613 && ! find_reg_note (curr_insn, REG_UNUSED, 2614 no_subreg_reg_operand 2615 [this_alternative_matches]))))) 2616 { 2617 if (lra_dump_file != NULL) 2618 fprintf 2619 (lra_dump_file, 2620 " alt=%d: No input/otput reload -- refuse\n", 2621 nalt); 2622 goto fail; 2623 } 2624 2625 /* Alternative loses if it required class pseudo can not 2626 hold value of required mode. Such insns can be 2627 described by insn definitions with mode iterators. */ 2628 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode 2629 && ! hard_reg_set_empty_p (this_alternative_set) 2630 /* It is common practice for constraints to use a 2631 class which does not have actually enough regs to 2632 hold the value (e.g. x86 AREG for mode requiring 2633 more one general reg). Therefore we have 2 2634 conditions to check that the reload pseudo can 2635 not hold the mode value. */ 2636 && (!targetm.hard_regno_mode_ok 2637 (ira_class_hard_regs[this_alternative][0], 2638 GET_MODE (*curr_id->operand_loc[nop]))) 2639 /* The above condition is not enough as the first 2640 reg in ira_class_hard_regs can be not aligned for 2641 multi-words mode values. */ 2642 && (prohibited_class_reg_set_mode_p 2643 (this_alternative, this_alternative_set, 2644 GET_MODE (*curr_id->operand_loc[nop])))) 2645 { 2646 if (lra_dump_file != NULL) 2647 fprintf (lra_dump_file, 2648 " alt=%d: reload pseudo for op %d " 2649 " can not hold the mode value -- refuse\n", 2650 nalt, nop); 2651 goto fail; 2652 } 2653 2654 /* Check strong discouragement of reload of non-constant 2655 into class THIS_ALTERNATIVE. */ 2656 if (! CONSTANT_P (op) && ! no_regs_p 2657 && (targetm.preferred_reload_class 2658 (op, this_alternative) == NO_REGS 2659 || (curr_static_id->operand[nop].type == OP_OUT 2660 && (targetm.preferred_output_reload_class 2661 (op, this_alternative) == NO_REGS)))) 2662 { 2663 if (lra_dump_file != NULL) 2664 fprintf (lra_dump_file, 2665 " %d Non-prefered reload: reject+=%d\n", 2666 nop, LRA_MAX_REJECT); 2667 reject += LRA_MAX_REJECT; 2668 } 2669 2670 if (! (MEM_P (op) && offmemok) 2671 && ! (const_to_mem && constmemok)) 2672 { 2673 /* We prefer to reload pseudos over reloading other 2674 things, since such reloads may be able to be 2675 eliminated later. So bump REJECT in other cases. 2676 Don't do this in the case where we are forcing a 2677 constant into memory and it will then win since 2678 we don't want to have a different alternative 2679 match then. */ 2680 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER)) 2681 { 2682 if (lra_dump_file != NULL) 2683 fprintf 2684 (lra_dump_file, 2685 " %d Non-pseudo reload: reject+=2\n", 2686 nop); 2687 reject += 2; 2688 } 2689 2690 if (! no_regs_p) 2691 reload_nregs 2692 += ira_reg_class_max_nregs[this_alternative][mode]; 2693 2694 if (SMALL_REGISTER_CLASS_P (this_alternative)) 2695 { 2696 if (lra_dump_file != NULL) 2697 fprintf 2698 (lra_dump_file, 2699 " %d Small class reload: reject+=%d\n", 2700 nop, LRA_LOSER_COST_FACTOR / 2); 2701 reject += LRA_LOSER_COST_FACTOR / 2; 2702 } 2703 } 2704 2705 /* We are trying to spill pseudo into memory. It is 2706 usually more costly than moving to a hard register 2707 although it might takes the same number of 2708 reloads. 2709 2710 Non-pseudo spill may happen also. Suppose a target allows both 2711 register and memory in the operand constraint alternatives, 2712 then it's typical that an eliminable register has a substition 2713 of "base + offset" which can either be reloaded by a simple 2714 "new_reg <= base + offset" which will match the register 2715 constraint, or a similar reg addition followed by further spill 2716 to and reload from memory which will match the memory 2717 constraint, but this memory spill will be much more costly 2718 usually. 2719 2720 Code below increases the reject for both pseudo and non-pseudo 2721 spill. */ 2722 if (no_regs_p 2723 && !(MEM_P (op) && offmemok) 2724 && !(REG_P (op) && hard_regno[nop] < 0)) 2725 { 2726 if (lra_dump_file != NULL) 2727 fprintf 2728 (lra_dump_file, 2729 " %d Spill %spseudo into memory: reject+=3\n", 2730 nop, REG_P (op) ? "" : "Non-"); 2731 reject += 3; 2732 if (VECTOR_MODE_P (mode)) 2733 { 2734 /* Spilling vectors into memory is usually more 2735 costly as they contain big values. */ 2736 if (lra_dump_file != NULL) 2737 fprintf 2738 (lra_dump_file, 2739 " %d Spill vector pseudo: reject+=2\n", 2740 nop); 2741 reject += 2; 2742 } 2743 } 2744 2745 /* When we use an operand requiring memory in given 2746 alternative, the insn should write *and* read the 2747 value to/from memory it is costly in comparison with 2748 an insn alternative which does not use memory 2749 (e.g. register or immediate operand). We exclude 2750 memory operand for such case as we can satisfy the 2751 memory constraints by reloading address. */ 2752 if (no_regs_p && offmemok && !MEM_P (op)) 2753 { 2754 if (lra_dump_file != NULL) 2755 fprintf 2756 (lra_dump_file, 2757 " Using memory insn operand %d: reject+=3\n", 2758 nop); 2759 reject += 3; 2760 } 2761 2762 /* If reload requires moving value through secondary 2763 memory, it will need one more insn at least. */ 2764 if (this_alternative != NO_REGS 2765 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS 2766 && ((curr_static_id->operand[nop].type != OP_OUT 2767 && targetm.secondary_memory_needed (GET_MODE (op), cl, 2768 this_alternative)) 2769 || (curr_static_id->operand[nop].type != OP_IN 2770 && (targetm.secondary_memory_needed 2771 (GET_MODE (op), this_alternative, cl))))) 2772 losers++; 2773 2774 /* Input reloads can be inherited more often than output 2775 reloads can be removed, so penalize output 2776 reloads. */ 2777 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN) 2778 { 2779 if (lra_dump_file != NULL) 2780 fprintf 2781 (lra_dump_file, 2782 " %d Non input pseudo reload: reject++\n", 2783 nop); 2784 reject++; 2785 } 2786 2787 if (MEM_P (op) && offmemok) 2788 addr_losers++; 2789 else if (curr_static_id->operand[nop].type == OP_INOUT) 2790 { 2791 if (lra_dump_file != NULL) 2792 fprintf 2793 (lra_dump_file, 2794 " %d Input/Output reload: reject+=%d\n", 2795 nop, LRA_LOSER_COST_FACTOR); 2796 reject += LRA_LOSER_COST_FACTOR; 2797 } 2798 } 2799 2800 if (early_clobber_p && ! scratch_p) 2801 { 2802 if (lra_dump_file != NULL) 2803 fprintf (lra_dump_file, 2804 " %d Early clobber: reject++\n", nop); 2805 reject++; 2806 } 2807 /* ??? We check early clobbers after processing all operands 2808 (see loop below) and there we update the costs more. 2809 Should we update the cost (may be approximately) here 2810 because of early clobber register reloads or it is a rare 2811 or non-important thing to be worth to do it. */ 2812 overall = (losers * LRA_LOSER_COST_FACTOR + reject 2813 - (addr_losers == losers ? static_reject : 0)); 2814 if ((best_losers == 0 || losers != 0) && best_overall < overall) 2815 { 2816 if (lra_dump_file != NULL) 2817 fprintf (lra_dump_file, 2818 " alt=%d,overall=%d,losers=%d -- refuse\n", 2819 nalt, overall, losers); 2820 goto fail; 2821 } 2822 2823 if (update_and_check_small_class_inputs (nop, this_alternative)) 2824 { 2825 if (lra_dump_file != NULL) 2826 fprintf (lra_dump_file, 2827 " alt=%d, not enough small class regs -- refuse\n", 2828 nalt); 2829 goto fail; 2830 } 2831 curr_alt[nop] = this_alternative; 2832 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set); 2833 curr_alt_win[nop] = this_alternative_win; 2834 curr_alt_match_win[nop] = this_alternative_match_win; 2835 curr_alt_offmemok[nop] = this_alternative_offmemok; 2836 curr_alt_matches[nop] = this_alternative_matches; 2837 2838 if (this_alternative_matches >= 0 2839 && !did_match && !this_alternative_win) 2840 curr_alt_win[this_alternative_matches] = false; 2841 2842 if (early_clobber_p && operand_reg[nop] != NULL_RTX) 2843 early_clobbered_nops[early_clobbered_regs_num++] = nop; 2844 } 2845 2846 if (curr_insn_set != NULL_RTX && n_operands == 2 2847 /* Prevent processing non-move insns. */ 2848 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG 2849 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1]) 2850 && ((! curr_alt_win[0] && ! curr_alt_win[1] 2851 && REG_P (no_subreg_reg_operand[0]) 2852 && REG_P (no_subreg_reg_operand[1]) 2853 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1]) 2854 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))) 2855 || (! curr_alt_win[0] && curr_alt_win[1] 2856 && REG_P (no_subreg_reg_operand[1]) 2857 /* Check that we reload memory not the memory 2858 address. */ 2859 && ! (curr_alt_offmemok[0] 2860 && MEM_P (no_subreg_reg_operand[0])) 2861 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])) 2862 || (curr_alt_win[0] && ! curr_alt_win[1] 2863 && REG_P (no_subreg_reg_operand[0]) 2864 /* Check that we reload memory not the memory 2865 address. */ 2866 && ! (curr_alt_offmemok[1] 2867 && MEM_P (no_subreg_reg_operand[1])) 2868 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1]) 2869 && (! CONST_POOL_OK_P (curr_operand_mode[1], 2870 no_subreg_reg_operand[1]) 2871 || (targetm.preferred_reload_class 2872 (no_subreg_reg_operand[1], 2873 (enum reg_class) curr_alt[1]) != NO_REGS)) 2874 /* If it is a result of recent elimination in move 2875 insn we can transform it into an add still by 2876 using this alternative. */ 2877 && GET_CODE (no_subreg_reg_operand[1]) != PLUS 2878 /* Likewise if the source has been replaced with an 2879 equivalent value. This only happens once -- the reload 2880 will use the equivalent value instead of the register it 2881 replaces -- so there should be no danger of cycling. */ 2882 && !equiv_substition_p[1]))) 2883 { 2884 /* We have a move insn and a new reload insn will be similar 2885 to the current insn. We should avoid such situation as 2886 it results in LRA cycling. */ 2887 if (lra_dump_file != NULL) 2888 fprintf (lra_dump_file, 2889 " Cycle danger: overall += LRA_MAX_REJECT\n"); 2890 overall += LRA_MAX_REJECT; 2891 } 2892 ok_p = true; 2893 curr_alt_dont_inherit_ops_num = 0; 2894 for (nop = 0; nop < early_clobbered_regs_num; nop++) 2895 { 2896 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j; 2897 HARD_REG_SET temp_set; 2898 2899 i = early_clobbered_nops[nop]; 2900 if ((! curr_alt_win[i] && ! curr_alt_match_win[i]) 2901 || hard_regno[i] < 0) 2902 continue; 2903 lra_assert (operand_reg[i] != NULL_RTX); 2904 clobbered_hard_regno = hard_regno[i]; 2905 CLEAR_HARD_REG_SET (temp_set); 2906 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno); 2907 first_conflict_j = last_conflict_j = -1; 2908 for (j = 0; j < n_operands; j++) 2909 if (j == i 2910 /* We don't want process insides of match_operator and 2911 match_parallel because otherwise we would process 2912 their operands once again generating a wrong 2913 code. */ 2914 || curr_static_id->operand[j].is_operator) 2915 continue; 2916 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j]) 2917 || (curr_alt_matches[i] == j && curr_alt_match_win[i])) 2918 continue; 2919 /* If we don't reload j-th operand, check conflicts. */ 2920 else if ((curr_alt_win[j] || curr_alt_match_win[j]) 2921 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set)) 2922 { 2923 if (first_conflict_j < 0) 2924 first_conflict_j = j; 2925 last_conflict_j = j; 2926 } 2927 if (last_conflict_j < 0) 2928 continue; 2929 /* If earlyclobber operand conflicts with another 2930 non-matching operand which is actually the same register 2931 as the earlyclobber operand, it is better to reload the 2932 another operand as an operand matching the earlyclobber 2933 operand can be also the same. */ 2934 if (first_conflict_j == last_conflict_j 2935 && operand_reg[last_conflict_j] != NULL_RTX 2936 && ! curr_alt_match_win[last_conflict_j] 2937 && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j])) 2938 { 2939 curr_alt_win[last_conflict_j] = false; 2940 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] 2941 = last_conflict_j; 2942 losers++; 2943 /* Early clobber was already reflected in REJECT. */ 2944 lra_assert (reject > 0); 2945 if (lra_dump_file != NULL) 2946 fprintf 2947 (lra_dump_file, 2948 " %d Conflict early clobber reload: reject--\n", 2949 i); 2950 reject--; 2951 overall += LRA_LOSER_COST_FACTOR - 1; 2952 } 2953 else 2954 { 2955 /* We need to reload early clobbered register and the 2956 matched registers. */ 2957 for (j = 0; j < n_operands; j++) 2958 if (curr_alt_matches[j] == i) 2959 { 2960 curr_alt_match_win[j] = false; 2961 losers++; 2962 overall += LRA_LOSER_COST_FACTOR; 2963 } 2964 if (! curr_alt_match_win[i]) 2965 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i; 2966 else 2967 { 2968 /* Remember pseudos used for match reloads are never 2969 inherited. */ 2970 lra_assert (curr_alt_matches[i] >= 0); 2971 curr_alt_win[curr_alt_matches[i]] = false; 2972 } 2973 curr_alt_win[i] = curr_alt_match_win[i] = false; 2974 losers++; 2975 /* Early clobber was already reflected in REJECT. */ 2976 lra_assert (reject > 0); 2977 if (lra_dump_file != NULL) 2978 fprintf 2979 (lra_dump_file, 2980 " %d Matched conflict early clobber reloads: " 2981 "reject--\n", 2982 i); 2983 reject--; 2984 overall += LRA_LOSER_COST_FACTOR - 1; 2985 } 2986 } 2987 if (lra_dump_file != NULL) 2988 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n", 2989 nalt, overall, losers, reload_nregs); 2990 2991 /* If this alternative can be made to work by reloading, and it 2992 needs less reloading than the others checked so far, record 2993 it as the chosen goal for reloading. */ 2994 if ((best_losers != 0 && losers == 0) 2995 || (((best_losers == 0 && losers == 0) 2996 || (best_losers != 0 && losers != 0)) 2997 && (best_overall > overall 2998 || (best_overall == overall 2999 /* If the cost of the reloads is the same, 3000 prefer alternative which requires minimal 3001 number of reload regs. */ 3002 && (reload_nregs < best_reload_nregs 3003 || (reload_nregs == best_reload_nregs 3004 && (best_reload_sum < reload_sum 3005 || (best_reload_sum == reload_sum 3006 && nalt < goal_alt_number)))))))) 3007 { 3008 for (nop = 0; nop < n_operands; nop++) 3009 { 3010 goal_alt_win[nop] = curr_alt_win[nop]; 3011 goal_alt_match_win[nop] = curr_alt_match_win[nop]; 3012 goal_alt_matches[nop] = curr_alt_matches[nop]; 3013 goal_alt[nop] = curr_alt[nop]; 3014 goal_alt_offmemok[nop] = curr_alt_offmemok[nop]; 3015 } 3016 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num; 3017 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++) 3018 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop]; 3019 goal_alt_swapped = curr_swapped; 3020 best_overall = overall; 3021 best_losers = losers; 3022 best_reload_nregs = reload_nregs; 3023 best_reload_sum = reload_sum; 3024 goal_alt_number = nalt; 3025 } 3026 if (losers == 0) 3027 /* Everything is satisfied. Do not process alternatives 3028 anymore. */ 3029 break; 3030 fail: 3031 ; 3032 } 3033 return ok_p; 3034 } 3035 3036 /* Make reload base reg from address AD. */ 3037 static rtx 3038 base_to_reg (struct address_info *ad) 3039 { 3040 enum reg_class cl; 3041 int code = -1; 3042 rtx new_inner = NULL_RTX; 3043 rtx new_reg = NULL_RTX; 3044 rtx_insn *insn; 3045 rtx_insn *last_insn = get_last_insn(); 3046 3047 lra_assert (ad->disp == ad->disp_term); 3048 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code, 3049 get_index_code (ad)); 3050 new_reg = lra_create_new_reg (GET_MODE (*ad->base), NULL_RTX, 3051 cl, "base"); 3052 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg, 3053 ad->disp_term == NULL 3054 ? const0_rtx 3055 : *ad->disp_term); 3056 if (!valid_address_p (ad->mode, new_inner, ad->as)) 3057 return NULL_RTX; 3058 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base)); 3059 code = recog_memoized (insn); 3060 if (code < 0) 3061 { 3062 delete_insns_since (last_insn); 3063 return NULL_RTX; 3064 } 3065 3066 return new_inner; 3067 } 3068 3069 /* Make reload base reg + DISP from address AD. Return the new pseudo. */ 3070 static rtx 3071 base_plus_disp_to_reg (struct address_info *ad, rtx disp) 3072 { 3073 enum reg_class cl; 3074 rtx new_reg; 3075 3076 lra_assert (ad->base == ad->base_term); 3077 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code, 3078 get_index_code (ad)); 3079 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX, 3080 cl, "base + disp"); 3081 lra_emit_add (new_reg, *ad->base_term, disp); 3082 return new_reg; 3083 } 3084 3085 /* Make reload of index part of address AD. Return the new 3086 pseudo. */ 3087 static rtx 3088 index_part_to_reg (struct address_info *ad) 3089 { 3090 rtx new_reg; 3091 3092 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX, 3093 INDEX_REG_CLASS, "index term"); 3094 expand_mult (GET_MODE (*ad->index), *ad->index_term, 3095 GEN_INT (get_index_scale (ad)), new_reg, 1); 3096 return new_reg; 3097 } 3098 3099 /* Return true if we can add a displacement to address AD, even if that 3100 makes the address invalid. The fix-up code requires any new address 3101 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */ 3102 static bool 3103 can_add_disp_p (struct address_info *ad) 3104 { 3105 return (!ad->autoinc_p 3106 && ad->segment == NULL 3107 && ad->base == ad->base_term 3108 && ad->disp == ad->disp_term); 3109 } 3110 3111 /* Make equiv substitution in address AD. Return true if a substitution 3112 was made. */ 3113 static bool 3114 equiv_address_substitution (struct address_info *ad) 3115 { 3116 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term; 3117 poly_int64 disp; 3118 HOST_WIDE_INT scale; 3119 bool change_p; 3120 3121 base_term = strip_subreg (ad->base_term); 3122 if (base_term == NULL) 3123 base_reg = new_base_reg = NULL_RTX; 3124 else 3125 { 3126 base_reg = *base_term; 3127 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn); 3128 } 3129 index_term = strip_subreg (ad->index_term); 3130 if (index_term == NULL) 3131 index_reg = new_index_reg = NULL_RTX; 3132 else 3133 { 3134 index_reg = *index_term; 3135 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn); 3136 } 3137 if (base_reg == new_base_reg && index_reg == new_index_reg) 3138 return false; 3139 disp = 0; 3140 change_p = false; 3141 if (lra_dump_file != NULL) 3142 { 3143 fprintf (lra_dump_file, "Changing address in insn %d ", 3144 INSN_UID (curr_insn)); 3145 dump_value_slim (lra_dump_file, *ad->outer, 1); 3146 } 3147 if (base_reg != new_base_reg) 3148 { 3149 poly_int64 offset; 3150 if (REG_P (new_base_reg)) 3151 { 3152 *base_term = new_base_reg; 3153 change_p = true; 3154 } 3155 else if (GET_CODE (new_base_reg) == PLUS 3156 && REG_P (XEXP (new_base_reg, 0)) 3157 && poly_int_rtx_p (XEXP (new_base_reg, 1), &offset) 3158 && can_add_disp_p (ad)) 3159 { 3160 disp += offset; 3161 *base_term = XEXP (new_base_reg, 0); 3162 change_p = true; 3163 } 3164 if (ad->base_term2 != NULL) 3165 *ad->base_term2 = *ad->base_term; 3166 } 3167 if (index_reg != new_index_reg) 3168 { 3169 poly_int64 offset; 3170 if (REG_P (new_index_reg)) 3171 { 3172 *index_term = new_index_reg; 3173 change_p = true; 3174 } 3175 else if (GET_CODE (new_index_reg) == PLUS 3176 && REG_P (XEXP (new_index_reg, 0)) 3177 && poly_int_rtx_p (XEXP (new_index_reg, 1), &offset) 3178 && can_add_disp_p (ad) 3179 && (scale = get_index_scale (ad))) 3180 { 3181 disp += offset * scale; 3182 *index_term = XEXP (new_index_reg, 0); 3183 change_p = true; 3184 } 3185 } 3186 if (maybe_ne (disp, 0)) 3187 { 3188 if (ad->disp != NULL) 3189 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp); 3190 else 3191 { 3192 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp); 3193 update_address (ad); 3194 } 3195 change_p = true; 3196 } 3197 if (lra_dump_file != NULL) 3198 { 3199 if (! change_p) 3200 fprintf (lra_dump_file, " -- no change\n"); 3201 else 3202 { 3203 fprintf (lra_dump_file, " on equiv "); 3204 dump_value_slim (lra_dump_file, *ad->outer, 1); 3205 fprintf (lra_dump_file, "\n"); 3206 } 3207 } 3208 return change_p; 3209 } 3210 3211 /* Major function to make reloads for an address in operand NOP or 3212 check its correctness (If CHECK_ONLY_P is true). The supported 3213 cases are: 3214 3215 1) an address that existed before LRA started, at which point it 3216 must have been valid. These addresses are subject to elimination 3217 and may have become invalid due to the elimination offset being out 3218 of range. 3219 3220 2) an address created by forcing a constant to memory 3221 (force_const_to_mem). The initial form of these addresses might 3222 not be valid, and it is this function's job to make them valid. 3223 3224 3) a frame address formed from a register and a (possibly zero) 3225 constant offset. As above, these addresses might not be valid and 3226 this function must make them so. 3227 3228 Add reloads to the lists *BEFORE and *AFTER. We might need to add 3229 reloads to *AFTER because of inc/dec, {pre, post} modify in the 3230 address. Return true for any RTL change. 3231 3232 The function is a helper function which does not produce all 3233 transformations (when CHECK_ONLY_P is false) which can be 3234 necessary. It does just basic steps. To do all necessary 3235 transformations use function process_address. */ 3236 static bool 3237 process_address_1 (int nop, bool check_only_p, 3238 rtx_insn **before, rtx_insn **after) 3239 { 3240 struct address_info ad; 3241 rtx new_reg; 3242 HOST_WIDE_INT scale; 3243 rtx op = *curr_id->operand_loc[nop]; 3244 const char *constraint = curr_static_id->operand[nop].constraint; 3245 enum constraint_num cn = lookup_constraint (constraint); 3246 bool change_p = false; 3247 3248 if (MEM_P (op) 3249 && GET_MODE (op) == BLKmode 3250 && GET_CODE (XEXP (op, 0)) == SCRATCH) 3251 return false; 3252 3253 if (insn_extra_address_constraint (cn) 3254 /* When we find an asm operand with an address constraint that 3255 doesn't satisfy address_operand to begin with, we clear 3256 is_address, so that we don't try to make a non-address fit. 3257 If the asm statement got this far, it's because other 3258 constraints are available, and we'll use them, disregarding 3259 the unsatisfiable address ones. */ 3260 && curr_static_id->operand[nop].is_address) 3261 decompose_lea_address (&ad, curr_id->operand_loc[nop]); 3262 /* Do not attempt to decompose arbitrary addresses generated by combine 3263 for asm operands with loose constraints, e.g 'X'. */ 3264 else if (MEM_P (op) 3265 && !(INSN_CODE (curr_insn) < 0 3266 && get_constraint_type (cn) == CT_FIXED_FORM 3267 && constraint_satisfied_p (op, cn))) 3268 decompose_mem_address (&ad, op); 3269 else if (GET_CODE (op) == SUBREG 3270 && MEM_P (SUBREG_REG (op))) 3271 decompose_mem_address (&ad, SUBREG_REG (op)); 3272 else 3273 return false; 3274 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to 3275 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both 3276 when INDEX_REG_CLASS is a single register class. */ 3277 if (ad.base_term != NULL 3278 && ad.index_term != NULL 3279 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1 3280 && REG_P (*ad.base_term) 3281 && REG_P (*ad.index_term) 3282 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL) 3283 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL)) 3284 { 3285 std::swap (ad.base, ad.index); 3286 std::swap (ad.base_term, ad.index_term); 3287 } 3288 if (! check_only_p) 3289 change_p = equiv_address_substitution (&ad); 3290 if (ad.base_term != NULL 3291 && (process_addr_reg 3292 (ad.base_term, check_only_p, before, 3293 (ad.autoinc_p 3294 && !(REG_P (*ad.base_term) 3295 && find_regno_note (curr_insn, REG_DEAD, 3296 REGNO (*ad.base_term)) != NULL_RTX) 3297 ? after : NULL), 3298 base_reg_class (ad.mode, ad.as, ad.base_outer_code, 3299 get_index_code (&ad))))) 3300 { 3301 change_p = true; 3302 if (ad.base_term2 != NULL) 3303 *ad.base_term2 = *ad.base_term; 3304 } 3305 if (ad.index_term != NULL 3306 && process_addr_reg (ad.index_term, check_only_p, 3307 before, NULL, INDEX_REG_CLASS)) 3308 change_p = true; 3309 3310 /* Target hooks sometimes don't treat extra-constraint addresses as 3311 legitimate address_operands, so handle them specially. */ 3312 if (insn_extra_address_constraint (cn) 3313 && satisfies_address_constraint_p (&ad, cn)) 3314 return change_p; 3315 3316 if (check_only_p) 3317 return change_p; 3318 3319 /* There are three cases where the shape of *AD.INNER may now be invalid: 3320 3321 1) the original address was valid, but either elimination or 3322 equiv_address_substitution was applied and that made 3323 the address invalid. 3324 3325 2) the address is an invalid symbolic address created by 3326 force_const_to_mem. 3327 3328 3) the address is a frame address with an invalid offset. 3329 3330 4) the address is a frame address with an invalid base. 3331 3332 All these cases involve a non-autoinc address, so there is no 3333 point revalidating other types. */ 3334 if (ad.autoinc_p || valid_address_p (&ad)) 3335 return change_p; 3336 3337 /* Any index existed before LRA started, so we can assume that the 3338 presence and shape of the index is valid. */ 3339 push_to_sequence (*before); 3340 lra_assert (ad.disp == ad.disp_term); 3341 if (ad.base == NULL) 3342 { 3343 if (ad.index == NULL) 3344 { 3345 rtx_insn *insn; 3346 rtx_insn *last = get_last_insn (); 3347 int code = -1; 3348 enum reg_class cl = base_reg_class (ad.mode, ad.as, 3349 SCRATCH, SCRATCH); 3350 rtx addr = *ad.inner; 3351 3352 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr"); 3353 if (HAVE_lo_sum) 3354 { 3355 /* addr => lo_sum (new_base, addr), case (2) above. */ 3356 insn = emit_insn (gen_rtx_SET 3357 (new_reg, 3358 gen_rtx_HIGH (Pmode, copy_rtx (addr)))); 3359 code = recog_memoized (insn); 3360 if (code >= 0) 3361 { 3362 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr); 3363 if (! valid_address_p (ad.mode, *ad.outer, ad.as)) 3364 { 3365 /* Try to put lo_sum into register. */ 3366 insn = emit_insn (gen_rtx_SET 3367 (new_reg, 3368 gen_rtx_LO_SUM (Pmode, new_reg, addr))); 3369 code = recog_memoized (insn); 3370 if (code >= 0) 3371 { 3372 *ad.inner = new_reg; 3373 if (! valid_address_p (ad.mode, *ad.outer, ad.as)) 3374 { 3375 *ad.inner = addr; 3376 code = -1; 3377 } 3378 } 3379 3380 } 3381 } 3382 if (code < 0) 3383 delete_insns_since (last); 3384 } 3385 3386 if (code < 0) 3387 { 3388 /* addr => new_base, case (2) above. */ 3389 lra_emit_move (new_reg, addr); 3390 3391 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last); 3392 insn != NULL_RTX; 3393 insn = NEXT_INSN (insn)) 3394 if (recog_memoized (insn) < 0) 3395 break; 3396 if (insn != NULL_RTX) 3397 { 3398 /* Do nothing if we cannot generate right insns. 3399 This is analogous to reload pass behavior. */ 3400 delete_insns_since (last); 3401 end_sequence (); 3402 return false; 3403 } 3404 *ad.inner = new_reg; 3405 } 3406 } 3407 else 3408 { 3409 /* index * scale + disp => new base + index * scale, 3410 case (1) above. */ 3411 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS, 3412 GET_CODE (*ad.index)); 3413 3414 lra_assert (INDEX_REG_CLASS != NO_REGS); 3415 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp"); 3416 lra_emit_move (new_reg, *ad.disp); 3417 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), 3418 new_reg, *ad.index); 3419 } 3420 } 3421 else if (ad.index == NULL) 3422 { 3423 int regno; 3424 enum reg_class cl; 3425 rtx set; 3426 rtx_insn *insns, *last_insn; 3427 /* Try to reload base into register only if the base is invalid 3428 for the address but with valid offset, case (4) above. */ 3429 start_sequence (); 3430 new_reg = base_to_reg (&ad); 3431 3432 /* base + disp => new base, cases (1) and (3) above. */ 3433 /* Another option would be to reload the displacement into an 3434 index register. However, postreload has code to optimize 3435 address reloads that have the same base and different 3436 displacements, so reloading into an index register would 3437 not necessarily be a win. */ 3438 if (new_reg == NULL_RTX) 3439 { 3440 /* See if the target can split the displacement into a 3441 legitimate new displacement from a local anchor. */ 3442 gcc_assert (ad.disp == ad.disp_term); 3443 poly_int64 orig_offset; 3444 rtx offset1, offset2; 3445 if (poly_int_rtx_p (*ad.disp, &orig_offset) 3446 && targetm.legitimize_address_displacement (&offset1, &offset2, 3447 orig_offset, 3448 ad.mode)) 3449 { 3450 new_reg = base_plus_disp_to_reg (&ad, offset1); 3451 new_reg = gen_rtx_PLUS (GET_MODE (new_reg), new_reg, offset2); 3452 } 3453 else 3454 new_reg = base_plus_disp_to_reg (&ad, *ad.disp); 3455 } 3456 insns = get_insns (); 3457 last_insn = get_last_insn (); 3458 /* If we generated at least two insns, try last insn source as 3459 an address. If we succeed, we generate one less insn. */ 3460 if (REG_P (new_reg) 3461 && last_insn != insns 3462 && (set = single_set (last_insn)) != NULL_RTX 3463 && GET_CODE (SET_SRC (set)) == PLUS 3464 && REG_P (XEXP (SET_SRC (set), 0)) 3465 && CONSTANT_P (XEXP (SET_SRC (set), 1))) 3466 { 3467 *ad.inner = SET_SRC (set); 3468 if (valid_address_p (ad.mode, *ad.outer, ad.as)) 3469 { 3470 *ad.base_term = XEXP (SET_SRC (set), 0); 3471 *ad.disp_term = XEXP (SET_SRC (set), 1); 3472 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code, 3473 get_index_code (&ad)); 3474 regno = REGNO (*ad.base_term); 3475 if (regno >= FIRST_PSEUDO_REGISTER 3476 && cl != lra_get_allocno_class (regno)) 3477 lra_change_class (regno, cl, " Change to", true); 3478 new_reg = SET_SRC (set); 3479 delete_insns_since (PREV_INSN (last_insn)); 3480 } 3481 } 3482 end_sequence (); 3483 emit_insn (insns); 3484 *ad.inner = new_reg; 3485 } 3486 else if (ad.disp_term != NULL) 3487 { 3488 /* base + scale * index + disp => new base + scale * index, 3489 case (1) above. */ 3490 gcc_assert (ad.disp == ad.disp_term); 3491 new_reg = base_plus_disp_to_reg (&ad, *ad.disp); 3492 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), 3493 new_reg, *ad.index); 3494 } 3495 else if ((scale = get_index_scale (&ad)) == 1) 3496 { 3497 /* The last transformation to one reg will be made in 3498 curr_insn_transform function. */ 3499 end_sequence (); 3500 return false; 3501 } 3502 else if (scale != 0) 3503 { 3504 /* base + scale * index => base + new_reg, 3505 case (1) above. 3506 Index part of address may become invalid. For example, we 3507 changed pseudo on the equivalent memory and a subreg of the 3508 pseudo onto the memory of different mode for which the scale is 3509 prohibitted. */ 3510 new_reg = index_part_to_reg (&ad); 3511 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), 3512 *ad.base_term, new_reg); 3513 } 3514 else 3515 { 3516 enum reg_class cl = base_reg_class (ad.mode, ad.as, 3517 SCRATCH, SCRATCH); 3518 rtx addr = *ad.inner; 3519 3520 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr"); 3521 /* addr => new_base. */ 3522 lra_emit_move (new_reg, addr); 3523 *ad.inner = new_reg; 3524 } 3525 *before = get_insns (); 3526 end_sequence (); 3527 return true; 3528 } 3529 3530 /* If CHECK_ONLY_P is false, do address reloads until it is necessary. 3531 Use process_address_1 as a helper function. Return true for any 3532 RTL changes. 3533 3534 If CHECK_ONLY_P is true, just check address correctness. Return 3535 false if the address correct. */ 3536 static bool 3537 process_address (int nop, bool check_only_p, 3538 rtx_insn **before, rtx_insn **after) 3539 { 3540 bool res = false; 3541 3542 while (process_address_1 (nop, check_only_p, before, after)) 3543 { 3544 if (check_only_p) 3545 return true; 3546 res = true; 3547 } 3548 return res; 3549 } 3550 3551 /* Emit insns to reload VALUE into a new register. VALUE is an 3552 auto-increment or auto-decrement RTX whose operand is a register or 3553 memory location; so reloading involves incrementing that location. 3554 IN is either identical to VALUE, or some cheaper place to reload 3555 value being incremented/decremented from. 3556 3557 INC_AMOUNT is the number to increment or decrement by (always 3558 positive and ignored for POST_MODIFY/PRE_MODIFY). 3559 3560 Return pseudo containing the result. */ 3561 static rtx 3562 emit_inc (enum reg_class new_rclass, rtx in, rtx value, poly_int64 inc_amount) 3563 { 3564 /* REG or MEM to be copied and incremented. */ 3565 rtx incloc = XEXP (value, 0); 3566 /* Nonzero if increment after copying. */ 3567 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC 3568 || GET_CODE (value) == POST_MODIFY); 3569 rtx_insn *last; 3570 rtx inc; 3571 rtx_insn *add_insn; 3572 int code; 3573 rtx real_in = in == value ? incloc : in; 3574 rtx result; 3575 bool plus_p = true; 3576 3577 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY) 3578 { 3579 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS 3580 || GET_CODE (XEXP (value, 1)) == MINUS); 3581 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0))); 3582 plus_p = GET_CODE (XEXP (value, 1)) == PLUS; 3583 inc = XEXP (XEXP (value, 1), 1); 3584 } 3585 else 3586 { 3587 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC) 3588 inc_amount = -inc_amount; 3589 3590 inc = gen_int_mode (inc_amount, GET_MODE (value)); 3591 } 3592 3593 if (! post && REG_P (incloc)) 3594 result = incloc; 3595 else 3596 result = lra_create_new_reg (GET_MODE (value), value, new_rclass, 3597 "INC/DEC result"); 3598 3599 if (real_in != result) 3600 { 3601 /* First copy the location to the result register. */ 3602 lra_assert (REG_P (result)); 3603 emit_insn (gen_move_insn (result, real_in)); 3604 } 3605 3606 /* We suppose that there are insns to add/sub with the constant 3607 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the 3608 old reload worked with this assumption. If the assumption 3609 becomes wrong, we should use approach in function 3610 base_plus_disp_to_reg. */ 3611 if (in == value) 3612 { 3613 /* See if we can directly increment INCLOC. */ 3614 last = get_last_insn (); 3615 add_insn = emit_insn (plus_p 3616 ? gen_add2_insn (incloc, inc) 3617 : gen_sub2_insn (incloc, inc)); 3618 3619 code = recog_memoized (add_insn); 3620 if (code >= 0) 3621 { 3622 if (! post && result != incloc) 3623 emit_insn (gen_move_insn (result, incloc)); 3624 return result; 3625 } 3626 delete_insns_since (last); 3627 } 3628 3629 /* If couldn't do the increment directly, must increment in RESULT. 3630 The way we do this depends on whether this is pre- or 3631 post-increment. For pre-increment, copy INCLOC to the reload 3632 register, increment it there, then save back. */ 3633 if (! post) 3634 { 3635 if (real_in != result) 3636 emit_insn (gen_move_insn (result, real_in)); 3637 if (plus_p) 3638 emit_insn (gen_add2_insn (result, inc)); 3639 else 3640 emit_insn (gen_sub2_insn (result, inc)); 3641 if (result != incloc) 3642 emit_insn (gen_move_insn (incloc, result)); 3643 } 3644 else 3645 { 3646 /* Post-increment. 3647 3648 Because this might be a jump insn or a compare, and because 3649 RESULT may not be available after the insn in an input 3650 reload, we must do the incrementing before the insn being 3651 reloaded for. 3652 3653 We have already copied IN to RESULT. Increment the copy in 3654 RESULT, save that back, then decrement RESULT so it has 3655 the original value. */ 3656 if (plus_p) 3657 emit_insn (gen_add2_insn (result, inc)); 3658 else 3659 emit_insn (gen_sub2_insn (result, inc)); 3660 emit_insn (gen_move_insn (incloc, result)); 3661 /* Restore non-modified value for the result. We prefer this 3662 way because it does not require an additional hard 3663 register. */ 3664 if (plus_p) 3665 { 3666 poly_int64 offset; 3667 if (poly_int_rtx_p (inc, &offset)) 3668 emit_insn (gen_add2_insn (result, 3669 gen_int_mode (-offset, 3670 GET_MODE (result)))); 3671 else 3672 emit_insn (gen_sub2_insn (result, inc)); 3673 } 3674 else 3675 emit_insn (gen_add2_insn (result, inc)); 3676 } 3677 return result; 3678 } 3679 3680 /* Return true if the current move insn does not need processing as we 3681 already know that it satisfies its constraints. */ 3682 static bool 3683 simple_move_p (void) 3684 { 3685 rtx dest, src; 3686 enum reg_class dclass, sclass; 3687 3688 lra_assert (curr_insn_set != NULL_RTX); 3689 dest = SET_DEST (curr_insn_set); 3690 src = SET_SRC (curr_insn_set); 3691 3692 /* If the instruction has multiple sets we need to process it even if it 3693 is single_set. This can happen if one or more of the SETs are dead. 3694 See PR73650. */ 3695 if (multiple_sets (curr_insn)) 3696 return false; 3697 3698 return ((dclass = get_op_class (dest)) != NO_REGS 3699 && (sclass = get_op_class (src)) != NO_REGS 3700 /* The backend guarantees that register moves of cost 2 3701 never need reloads. */ 3702 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2); 3703 } 3704 3705 /* Swap operands NOP and NOP + 1. */ 3706 static inline void 3707 swap_operands (int nop) 3708 { 3709 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]); 3710 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]); 3711 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]); 3712 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]); 3713 /* Swap the duplicates too. */ 3714 lra_update_dup (curr_id, nop); 3715 lra_update_dup (curr_id, nop + 1); 3716 } 3717 3718 /* Main entry point of the constraint code: search the body of the 3719 current insn to choose the best alternative. It is mimicking insn 3720 alternative cost calculation model of former reload pass. That is 3721 because machine descriptions were written to use this model. This 3722 model can be changed in future. Make commutative operand exchange 3723 if it is chosen. 3724 3725 if CHECK_ONLY_P is false, do RTL changes to satisfy the 3726 constraints. Return true if any change happened during function 3727 call. 3728 3729 If CHECK_ONLY_P is true then don't do any transformation. Just 3730 check that the insn satisfies all constraints. If the insn does 3731 not satisfy any constraint, return true. */ 3732 static bool 3733 curr_insn_transform (bool check_only_p) 3734 { 3735 int i, j, k; 3736 int n_operands; 3737 int n_alternatives; 3738 int n_outputs; 3739 int commutative; 3740 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS]; 3741 signed char match_inputs[MAX_RECOG_OPERANDS + 1]; 3742 signed char outputs[MAX_RECOG_OPERANDS + 1]; 3743 rtx_insn *before, *after; 3744 bool alt_p = false; 3745 /* Flag that the insn has been changed through a transformation. */ 3746 bool change_p; 3747 bool sec_mem_p; 3748 bool use_sec_mem_p; 3749 int max_regno_before; 3750 int reused_alternative_num; 3751 3752 curr_insn_set = single_set (curr_insn); 3753 if (curr_insn_set != NULL_RTX && simple_move_p ()) 3754 { 3755 /* We assume that the corresponding insn alternative has no 3756 earlier clobbers. If it is not the case, don't define move 3757 cost equal to 2 for the corresponding register classes. */ 3758 lra_set_used_insn_alternative (curr_insn, LRA_NON_CLOBBERED_ALT); 3759 return false; 3760 } 3761 3762 no_input_reloads_p = no_output_reloads_p = false; 3763 goal_alt_number = -1; 3764 change_p = sec_mem_p = false; 3765 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output 3766 reloads; neither are insns that SET cc0. Insns that use CC0 are 3767 not allowed to have any input reloads. */ 3768 if (JUMP_P (curr_insn) || CALL_P (curr_insn)) 3769 no_output_reloads_p = true; 3770 3771 if (HAVE_cc0 && reg_referenced_p (cc0_rtx, PATTERN (curr_insn))) 3772 no_input_reloads_p = true; 3773 if (HAVE_cc0 && reg_set_p (cc0_rtx, PATTERN (curr_insn))) 3774 no_output_reloads_p = true; 3775 3776 n_operands = curr_static_id->n_operands; 3777 n_alternatives = curr_static_id->n_alternatives; 3778 3779 /* Just return "no reloads" if insn has no operands with 3780 constraints. */ 3781 if (n_operands == 0 || n_alternatives == 0) 3782 return false; 3783 3784 max_regno_before = max_reg_num (); 3785 3786 for (i = 0; i < n_operands; i++) 3787 { 3788 goal_alt_matched[i][0] = -1; 3789 goal_alt_matches[i] = -1; 3790 } 3791 3792 commutative = curr_static_id->commutative; 3793 3794 /* Now see what we need for pseudos that didn't get hard regs or got 3795 the wrong kind of hard reg. For this, we must consider all the 3796 operands together against the register constraints. */ 3797 3798 best_losers = best_overall = INT_MAX; 3799 best_reload_sum = 0; 3800 3801 curr_swapped = false; 3802 goal_alt_swapped = false; 3803 3804 if (! check_only_p) 3805 /* Make equivalence substitution and memory subreg elimination 3806 before address processing because an address legitimacy can 3807 depend on memory mode. */ 3808 for (i = 0; i < n_operands; i++) 3809 { 3810 rtx op, subst, old; 3811 bool op_change_p = false; 3812 3813 if (curr_static_id->operand[i].is_operator) 3814 continue; 3815 3816 old = op = *curr_id->operand_loc[i]; 3817 if (GET_CODE (old) == SUBREG) 3818 old = SUBREG_REG (old); 3819 subst = get_equiv_with_elimination (old, curr_insn); 3820 original_subreg_reg_mode[i] = VOIDmode; 3821 equiv_substition_p[i] = false; 3822 if (subst != old) 3823 { 3824 equiv_substition_p[i] = true; 3825 subst = copy_rtx (subst); 3826 lra_assert (REG_P (old)); 3827 if (GET_CODE (op) != SUBREG) 3828 *curr_id->operand_loc[i] = subst; 3829 else 3830 { 3831 SUBREG_REG (op) = subst; 3832 if (GET_MODE (subst) == VOIDmode) 3833 original_subreg_reg_mode[i] = GET_MODE (old); 3834 } 3835 if (lra_dump_file != NULL) 3836 { 3837 fprintf (lra_dump_file, 3838 "Changing pseudo %d in operand %i of insn %u on equiv ", 3839 REGNO (old), i, INSN_UID (curr_insn)); 3840 dump_value_slim (lra_dump_file, subst, 1); 3841 fprintf (lra_dump_file, "\n"); 3842 } 3843 op_change_p = change_p = true; 3844 } 3845 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p) 3846 { 3847 change_p = true; 3848 lra_update_dup (curr_id, i); 3849 } 3850 } 3851 3852 /* Reload address registers and displacements. We do it before 3853 finding an alternative because of memory constraints. */ 3854 before = after = NULL; 3855 for (i = 0; i < n_operands; i++) 3856 if (! curr_static_id->operand[i].is_operator 3857 && process_address (i, check_only_p, &before, &after)) 3858 { 3859 if (check_only_p) 3860 return true; 3861 change_p = true; 3862 lra_update_dup (curr_id, i); 3863 } 3864 3865 if (change_p) 3866 /* If we've changed the instruction then any alternative that 3867 we chose previously may no longer be valid. */ 3868 lra_set_used_insn_alternative (curr_insn, LRA_UNKNOWN_ALT); 3869 3870 if (! check_only_p && curr_insn_set != NULL_RTX 3871 && check_and_process_move (&change_p, &sec_mem_p)) 3872 return change_p; 3873 3874 try_swapped: 3875 3876 reused_alternative_num = check_only_p ? LRA_UNKNOWN_ALT : curr_id->used_insn_alternative; 3877 if (lra_dump_file != NULL && reused_alternative_num >= 0) 3878 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n", 3879 reused_alternative_num, INSN_UID (curr_insn)); 3880 3881 if (process_alt_operands (reused_alternative_num)) 3882 alt_p = true; 3883 3884 if (check_only_p) 3885 return ! alt_p || best_losers != 0; 3886 3887 /* If insn is commutative (it's safe to exchange a certain pair of 3888 operands) then we need to try each alternative twice, the second 3889 time matching those two operands as if we had exchanged them. To 3890 do this, really exchange them in operands. 3891 3892 If we have just tried the alternatives the second time, return 3893 operands to normal and drop through. */ 3894 3895 if (reused_alternative_num < 0 && commutative >= 0) 3896 { 3897 curr_swapped = !curr_swapped; 3898 if (curr_swapped) 3899 { 3900 swap_operands (commutative); 3901 goto try_swapped; 3902 } 3903 else 3904 swap_operands (commutative); 3905 } 3906 3907 if (! alt_p && ! sec_mem_p) 3908 { 3909 /* No alternative works with reloads?? */ 3910 if (INSN_CODE (curr_insn) >= 0) 3911 fatal_insn ("unable to generate reloads for:", curr_insn); 3912 error_for_asm (curr_insn, 3913 "inconsistent operand constraints in an %<asm%>"); 3914 /* Avoid further trouble with this insn. Don't generate use 3915 pattern here as we could use the insn SP offset. */ 3916 lra_set_insn_deleted (curr_insn); 3917 return true; 3918 } 3919 3920 /* If the best alternative is with operands 1 and 2 swapped, swap 3921 them. Update the operand numbers of any reloads already 3922 pushed. */ 3923 3924 if (goal_alt_swapped) 3925 { 3926 if (lra_dump_file != NULL) 3927 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n", 3928 INSN_UID (curr_insn)); 3929 3930 /* Swap the duplicates too. */ 3931 swap_operands (commutative); 3932 change_p = true; 3933 } 3934 3935 /* Some targets' TARGET_SECONDARY_MEMORY_NEEDED (e.g. x86) are defined 3936 too conservatively. So we use the secondary memory only if there 3937 is no any alternative without reloads. */ 3938 use_sec_mem_p = false; 3939 if (! alt_p) 3940 use_sec_mem_p = true; 3941 else if (sec_mem_p) 3942 { 3943 for (i = 0; i < n_operands; i++) 3944 if (! goal_alt_win[i] && ! goal_alt_match_win[i]) 3945 break; 3946 use_sec_mem_p = i < n_operands; 3947 } 3948 3949 if (use_sec_mem_p) 3950 { 3951 int in = -1, out = -1; 3952 rtx new_reg, src, dest, rld; 3953 machine_mode sec_mode, rld_mode; 3954 3955 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p); 3956 dest = SET_DEST (curr_insn_set); 3957 src = SET_SRC (curr_insn_set); 3958 for (i = 0; i < n_operands; i++) 3959 if (*curr_id->operand_loc[i] == dest) 3960 out = i; 3961 else if (*curr_id->operand_loc[i] == src) 3962 in = i; 3963 for (i = 0; i < curr_static_id->n_dups; i++) 3964 if (out < 0 && *curr_id->dup_loc[i] == dest) 3965 out = curr_static_id->dup_num[i]; 3966 else if (in < 0 && *curr_id->dup_loc[i] == src) 3967 in = curr_static_id->dup_num[i]; 3968 lra_assert (out >= 0 && in >= 0 3969 && curr_static_id->operand[out].type == OP_OUT 3970 && curr_static_id->operand[in].type == OP_IN); 3971 rld = partial_subreg_p (GET_MODE (src), GET_MODE (dest)) ? src : dest; 3972 rld_mode = GET_MODE (rld); 3973 sec_mode = targetm.secondary_memory_needed_mode (rld_mode); 3974 new_reg = lra_create_new_reg (sec_mode, NULL_RTX, 3975 NO_REGS, "secondary"); 3976 /* If the mode is changed, it should be wider. */ 3977 lra_assert (!partial_subreg_p (sec_mode, rld_mode)); 3978 if (sec_mode != rld_mode) 3979 { 3980 /* If the target says specifically to use another mode for 3981 secondary memory moves we can not reuse the original 3982 insn. */ 3983 after = emit_spill_move (false, new_reg, dest); 3984 lra_process_new_insns (curr_insn, NULL, after, 3985 "Inserting the sec. move"); 3986 /* We may have non null BEFORE here (e.g. after address 3987 processing. */ 3988 push_to_sequence (before); 3989 before = emit_spill_move (true, new_reg, src); 3990 emit_insn (before); 3991 before = get_insns (); 3992 end_sequence (); 3993 lra_process_new_insns (curr_insn, before, NULL, "Changing on"); 3994 lra_set_insn_deleted (curr_insn); 3995 } 3996 else if (dest == rld) 3997 { 3998 *curr_id->operand_loc[out] = new_reg; 3999 lra_update_dup (curr_id, out); 4000 after = emit_spill_move (false, new_reg, dest); 4001 lra_process_new_insns (curr_insn, NULL, after, 4002 "Inserting the sec. move"); 4003 } 4004 else 4005 { 4006 *curr_id->operand_loc[in] = new_reg; 4007 lra_update_dup (curr_id, in); 4008 /* See comments above. */ 4009 push_to_sequence (before); 4010 before = emit_spill_move (true, new_reg, src); 4011 emit_insn (before); 4012 before = get_insns (); 4013 end_sequence (); 4014 lra_process_new_insns (curr_insn, before, NULL, 4015 "Inserting the sec. move"); 4016 } 4017 lra_update_insn_regno_info (curr_insn); 4018 return true; 4019 } 4020 4021 lra_assert (goal_alt_number >= 0); 4022 lra_set_used_insn_alternative (curr_insn, goal_alt_number); 4023 4024 if (lra_dump_file != NULL) 4025 { 4026 const char *p; 4027 4028 fprintf (lra_dump_file, " Choosing alt %d in insn %u:", 4029 goal_alt_number, INSN_UID (curr_insn)); 4030 for (i = 0; i < n_operands; i++) 4031 { 4032 p = (curr_static_id->operand_alternative 4033 [goal_alt_number * n_operands + i].constraint); 4034 if (*p == '\0') 4035 continue; 4036 fprintf (lra_dump_file, " (%d) ", i); 4037 for (; *p != '\0' && *p != ',' && *p != '#'; p++) 4038 fputc (*p, lra_dump_file); 4039 } 4040 if (INSN_CODE (curr_insn) >= 0 4041 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL) 4042 fprintf (lra_dump_file, " {%s}", p); 4043 if (maybe_ne (curr_id->sp_offset, 0)) 4044 { 4045 fprintf (lra_dump_file, " (sp_off="); 4046 print_dec (curr_id->sp_offset, lra_dump_file); 4047 fprintf (lra_dump_file, ")"); 4048 } 4049 fprintf (lra_dump_file, "\n"); 4050 } 4051 4052 /* Right now, for any pair of operands I and J that are required to 4053 match, with J < I, goal_alt_matches[I] is J. Add I to 4054 goal_alt_matched[J]. */ 4055 4056 for (i = 0; i < n_operands; i++) 4057 if ((j = goal_alt_matches[i]) >= 0) 4058 { 4059 for (k = 0; goal_alt_matched[j][k] >= 0; k++) 4060 ; 4061 /* We allow matching one output operand and several input 4062 operands. */ 4063 lra_assert (k == 0 4064 || (curr_static_id->operand[j].type == OP_OUT 4065 && curr_static_id->operand[i].type == OP_IN 4066 && (curr_static_id->operand 4067 [goal_alt_matched[j][0]].type == OP_IN))); 4068 goal_alt_matched[j][k] = i; 4069 goal_alt_matched[j][k + 1] = -1; 4070 } 4071 4072 for (i = 0; i < n_operands; i++) 4073 goal_alt_win[i] |= goal_alt_match_win[i]; 4074 4075 /* Any constants that aren't allowed and can't be reloaded into 4076 registers are here changed into memory references. */ 4077 for (i = 0; i < n_operands; i++) 4078 if (goal_alt_win[i]) 4079 { 4080 int regno; 4081 enum reg_class new_class; 4082 rtx reg = *curr_id->operand_loc[i]; 4083 4084 if (GET_CODE (reg) == SUBREG) 4085 reg = SUBREG_REG (reg); 4086 4087 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER) 4088 { 4089 bool ok_p = in_class_p (reg, goal_alt[i], &new_class); 4090 4091 if (new_class != NO_REGS && get_reg_class (regno) != new_class) 4092 { 4093 lra_assert (ok_p); 4094 lra_change_class (regno, new_class, " Change to", true); 4095 } 4096 } 4097 } 4098 else 4099 { 4100 const char *constraint; 4101 char c; 4102 rtx op = *curr_id->operand_loc[i]; 4103 rtx subreg = NULL_RTX; 4104 machine_mode mode = curr_operand_mode[i]; 4105 4106 if (GET_CODE (op) == SUBREG) 4107 { 4108 subreg = op; 4109 op = SUBREG_REG (op); 4110 mode = GET_MODE (op); 4111 } 4112 4113 if (CONST_POOL_OK_P (mode, op) 4114 && ((targetm.preferred_reload_class 4115 (op, (enum reg_class) goal_alt[i]) == NO_REGS) 4116 || no_input_reloads_p)) 4117 { 4118 rtx tem = force_const_mem (mode, op); 4119 4120 change_p = true; 4121 if (subreg != NULL_RTX) 4122 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg)); 4123 4124 *curr_id->operand_loc[i] = tem; 4125 lra_update_dup (curr_id, i); 4126 process_address (i, false, &before, &after); 4127 4128 /* If the alternative accepts constant pool refs directly 4129 there will be no reload needed at all. */ 4130 if (subreg != NULL_RTX) 4131 continue; 4132 /* Skip alternatives before the one requested. */ 4133 constraint = (curr_static_id->operand_alternative 4134 [goal_alt_number * n_operands + i].constraint); 4135 for (; 4136 (c = *constraint) && c != ',' && c != '#'; 4137 constraint += CONSTRAINT_LEN (c, constraint)) 4138 { 4139 enum constraint_num cn = lookup_constraint (constraint); 4140 if ((insn_extra_memory_constraint (cn) 4141 || insn_extra_special_memory_constraint (cn)) 4142 && satisfies_memory_constraint_p (tem, cn)) 4143 break; 4144 } 4145 if (c == '\0' || c == ',' || c == '#') 4146 continue; 4147 4148 goal_alt_win[i] = true; 4149 } 4150 } 4151 4152 n_outputs = 0; 4153 outputs[0] = -1; 4154 for (i = 0; i < n_operands; i++) 4155 { 4156 int regno; 4157 bool optional_p = false; 4158 rtx old, new_reg; 4159 rtx op = *curr_id->operand_loc[i]; 4160 4161 if (goal_alt_win[i]) 4162 { 4163 if (goal_alt[i] == NO_REGS 4164 && REG_P (op) 4165 /* When we assign NO_REGS it means that we will not 4166 assign a hard register to the scratch pseudo by 4167 assigment pass and the scratch pseudo will be 4168 spilled. Spilled scratch pseudos are transformed 4169 back to scratches at the LRA end. */ 4170 && lra_former_scratch_operand_p (curr_insn, i) 4171 && lra_former_scratch_p (REGNO (op))) 4172 { 4173 int regno = REGNO (op); 4174 lra_change_class (regno, NO_REGS, " Change to", true); 4175 if (lra_get_regno_hard_regno (regno) >= 0) 4176 /* We don't have to mark all insn affected by the 4177 spilled pseudo as there is only one such insn, the 4178 current one. */ 4179 reg_renumber[regno] = -1; 4180 lra_assert (bitmap_single_bit_set_p 4181 (&lra_reg_info[REGNO (op)].insn_bitmap)); 4182 } 4183 /* We can do an optional reload. If the pseudo got a hard 4184 reg, we might improve the code through inheritance. If 4185 it does not get a hard register we coalesce memory/memory 4186 moves later. Ignore move insns to avoid cycling. */ 4187 if (! lra_simple_p 4188 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES 4189 && goal_alt[i] != NO_REGS && REG_P (op) 4190 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER 4191 && regno < new_regno_start 4192 && ! lra_former_scratch_p (regno) 4193 && reg_renumber[regno] < 0 4194 /* Check that the optional reload pseudo will be able to 4195 hold given mode value. */ 4196 && ! (prohibited_class_reg_set_mode_p 4197 (goal_alt[i], reg_class_contents[goal_alt[i]], 4198 PSEUDO_REGNO_MODE (regno))) 4199 && (curr_insn_set == NULL_RTX 4200 || !((REG_P (SET_SRC (curr_insn_set)) 4201 || MEM_P (SET_SRC (curr_insn_set)) 4202 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG) 4203 && (REG_P (SET_DEST (curr_insn_set)) 4204 || MEM_P (SET_DEST (curr_insn_set)) 4205 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG)))) 4206 optional_p = true; 4207 else 4208 continue; 4209 } 4210 4211 /* Operands that match previous ones have already been handled. */ 4212 if (goal_alt_matches[i] >= 0) 4213 continue; 4214 4215 /* We should not have an operand with a non-offsettable address 4216 appearing where an offsettable address will do. It also may 4217 be a case when the address should be special in other words 4218 not a general one (e.g. it needs no index reg). */ 4219 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op)) 4220 { 4221 enum reg_class rclass; 4222 rtx *loc = &XEXP (op, 0); 4223 enum rtx_code code = GET_CODE (*loc); 4224 4225 push_to_sequence (before); 4226 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op), 4227 MEM, SCRATCH); 4228 if (GET_RTX_CLASS (code) == RTX_AUTOINC) 4229 new_reg = emit_inc (rclass, *loc, *loc, 4230 /* This value does not matter for MODIFY. */ 4231 GET_MODE_SIZE (GET_MODE (op))); 4232 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE, 4233 "offsetable address", &new_reg)) 4234 { 4235 rtx addr = *loc; 4236 enum rtx_code code = GET_CODE (addr); 4237 4238 if (code == AND && CONST_INT_P (XEXP (addr, 1))) 4239 /* (and ... (const_int -X)) is used to align to X bytes. */ 4240 addr = XEXP (*loc, 0); 4241 lra_emit_move (new_reg, addr); 4242 if (addr != *loc) 4243 emit_move_insn (new_reg, gen_rtx_AND (GET_MODE (new_reg), new_reg, XEXP (*loc, 1))); 4244 } 4245 before = get_insns (); 4246 end_sequence (); 4247 *loc = new_reg; 4248 lra_update_dup (curr_id, i); 4249 } 4250 else if (goal_alt_matched[i][0] == -1) 4251 { 4252 machine_mode mode; 4253 rtx reg, *loc; 4254 int hard_regno; 4255 enum op_type type = curr_static_id->operand[i].type; 4256 4257 loc = curr_id->operand_loc[i]; 4258 mode = curr_operand_mode[i]; 4259 if (GET_CODE (*loc) == SUBREG) 4260 { 4261 reg = SUBREG_REG (*loc); 4262 poly_int64 byte = SUBREG_BYTE (*loc); 4263 if (REG_P (reg) 4264 /* Strict_low_part requires reloading the register and not 4265 just the subreg. Likewise for a strict subreg no wider 4266 than a word for WORD_REGISTER_OPERATIONS targets. */ 4267 && (curr_static_id->operand[i].strict_low 4268 || (!paradoxical_subreg_p (mode, GET_MODE (reg)) 4269 && (hard_regno 4270 = get_try_hard_regno (REGNO (reg))) >= 0 4271 && (simplify_subreg_regno 4272 (hard_regno, 4273 GET_MODE (reg), byte, mode) < 0) 4274 && (goal_alt[i] == NO_REGS 4275 || (simplify_subreg_regno 4276 (ira_class_hard_regs[goal_alt[i]][0], 4277 GET_MODE (reg), byte, mode) >= 0))) 4278 || (partial_subreg_p (mode, GET_MODE (reg)) 4279 && known_le (GET_MODE_SIZE (GET_MODE (reg)), 4280 UNITS_PER_WORD) 4281 && WORD_REGISTER_OPERATIONS))) 4282 { 4283 /* An OP_INOUT is required when reloading a subreg of a 4284 mode wider than a word to ensure that data beyond the 4285 word being reloaded is preserved. Also automatically 4286 ensure that strict_low_part reloads are made into 4287 OP_INOUT which should already be true from the backend 4288 constraints. */ 4289 if (type == OP_OUT 4290 && (curr_static_id->operand[i].strict_low 4291 || read_modify_subreg_p (*loc))) 4292 type = OP_INOUT; 4293 loc = &SUBREG_REG (*loc); 4294 mode = GET_MODE (*loc); 4295 } 4296 } 4297 old = *loc; 4298 if (get_reload_reg (type, mode, old, goal_alt[i], 4299 loc != curr_id->operand_loc[i], "", &new_reg) 4300 && type != OP_OUT) 4301 { 4302 push_to_sequence (before); 4303 lra_emit_move (new_reg, old); 4304 before = get_insns (); 4305 end_sequence (); 4306 } 4307 *loc = new_reg; 4308 if (type != OP_IN 4309 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX) 4310 { 4311 start_sequence (); 4312 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg); 4313 emit_insn (after); 4314 after = get_insns (); 4315 end_sequence (); 4316 *loc = new_reg; 4317 } 4318 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++) 4319 if (goal_alt_dont_inherit_ops[j] == i) 4320 { 4321 lra_set_regno_unique_value (REGNO (new_reg)); 4322 break; 4323 } 4324 lra_update_dup (curr_id, i); 4325 } 4326 else if (curr_static_id->operand[i].type == OP_IN 4327 && (curr_static_id->operand[goal_alt_matched[i][0]].type 4328 == OP_OUT 4329 || (curr_static_id->operand[goal_alt_matched[i][0]].type 4330 == OP_INOUT 4331 && (operands_match_p 4332 (*curr_id->operand_loc[i], 4333 *curr_id->operand_loc[goal_alt_matched[i][0]], 4334 -1))))) 4335 { 4336 /* generate reloads for input and matched outputs. */ 4337 match_inputs[0] = i; 4338 match_inputs[1] = -1; 4339 match_reload (goal_alt_matched[i][0], match_inputs, outputs, 4340 goal_alt[i], &before, &after, 4341 curr_static_id->operand_alternative 4342 [goal_alt_number * n_operands + goal_alt_matched[i][0]] 4343 .earlyclobber); 4344 } 4345 else if ((curr_static_id->operand[i].type == OP_OUT 4346 || (curr_static_id->operand[i].type == OP_INOUT 4347 && (operands_match_p 4348 (*curr_id->operand_loc[i], 4349 *curr_id->operand_loc[goal_alt_matched[i][0]], 4350 -1)))) 4351 && (curr_static_id->operand[goal_alt_matched[i][0]].type 4352 == OP_IN)) 4353 /* Generate reloads for output and matched inputs. */ 4354 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before, 4355 &after, curr_static_id->operand_alternative 4356 [goal_alt_number * n_operands + i].earlyclobber); 4357 else if (curr_static_id->operand[i].type == OP_IN 4358 && (curr_static_id->operand[goal_alt_matched[i][0]].type 4359 == OP_IN)) 4360 { 4361 /* Generate reloads for matched inputs. */ 4362 match_inputs[0] = i; 4363 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++) 4364 match_inputs[j + 1] = k; 4365 match_inputs[j + 1] = -1; 4366 match_reload (-1, match_inputs, outputs, goal_alt[i], &before, 4367 &after, false); 4368 } 4369 else 4370 /* We must generate code in any case when function 4371 process_alt_operands decides that it is possible. */ 4372 gcc_unreachable (); 4373 4374 /* Memorise processed outputs so that output remaining to be processed 4375 can avoid using the same register value (see match_reload). */ 4376 if (curr_static_id->operand[i].type == OP_OUT) 4377 { 4378 outputs[n_outputs++] = i; 4379 outputs[n_outputs] = -1; 4380 } 4381 4382 if (optional_p) 4383 { 4384 rtx reg = op; 4385 4386 lra_assert (REG_P (reg)); 4387 regno = REGNO (reg); 4388 op = *curr_id->operand_loc[i]; /* Substitution. */ 4389 if (GET_CODE (op) == SUBREG) 4390 op = SUBREG_REG (op); 4391 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start); 4392 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op)); 4393 lra_reg_info[REGNO (op)].restore_rtx = reg; 4394 if (lra_dump_file != NULL) 4395 fprintf (lra_dump_file, 4396 " Making reload reg %d for reg %d optional\n", 4397 REGNO (op), regno); 4398 } 4399 } 4400 if (before != NULL_RTX || after != NULL_RTX 4401 || max_regno_before != max_reg_num ()) 4402 change_p = true; 4403 if (change_p) 4404 { 4405 lra_update_operator_dups (curr_id); 4406 /* Something changes -- process the insn. */ 4407 lra_update_insn_regno_info (curr_insn); 4408 } 4409 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload"); 4410 return change_p; 4411 } 4412 4413 /* Return true if INSN satisfies all constraints. In other words, no 4414 reload insns are needed. */ 4415 bool 4416 lra_constrain_insn (rtx_insn *insn) 4417 { 4418 int saved_new_regno_start = new_regno_start; 4419 int saved_new_insn_uid_start = new_insn_uid_start; 4420 bool change_p; 4421 4422 curr_insn = insn; 4423 curr_id = lra_get_insn_recog_data (curr_insn); 4424 curr_static_id = curr_id->insn_static_data; 4425 new_insn_uid_start = get_max_uid (); 4426 new_regno_start = max_reg_num (); 4427 change_p = curr_insn_transform (true); 4428 new_regno_start = saved_new_regno_start; 4429 new_insn_uid_start = saved_new_insn_uid_start; 4430 return ! change_p; 4431 } 4432 4433 /* Return true if X is in LIST. */ 4434 static bool 4435 in_list_p (rtx x, rtx list) 4436 { 4437 for (; list != NULL_RTX; list = XEXP (list, 1)) 4438 if (XEXP (list, 0) == x) 4439 return true; 4440 return false; 4441 } 4442 4443 /* Return true if X contains an allocatable hard register (if 4444 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */ 4445 static bool 4446 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p) 4447 { 4448 int i, j; 4449 const char *fmt; 4450 enum rtx_code code; 4451 4452 code = GET_CODE (x); 4453 if (REG_P (x)) 4454 { 4455 int regno = REGNO (x); 4456 HARD_REG_SET alloc_regs; 4457 4458 if (hard_reg_p) 4459 { 4460 if (regno >= FIRST_PSEUDO_REGISTER) 4461 regno = lra_get_regno_hard_regno (regno); 4462 if (regno < 0) 4463 return false; 4464 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs); 4465 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno); 4466 } 4467 else 4468 { 4469 if (regno < FIRST_PSEUDO_REGISTER) 4470 return false; 4471 if (! spilled_p) 4472 return true; 4473 return lra_get_regno_hard_regno (regno) < 0; 4474 } 4475 } 4476 fmt = GET_RTX_FORMAT (code); 4477 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) 4478 { 4479 if (fmt[i] == 'e') 4480 { 4481 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p)) 4482 return true; 4483 } 4484 else if (fmt[i] == 'E') 4485 { 4486 for (j = XVECLEN (x, i) - 1; j >= 0; j--) 4487 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p)) 4488 return true; 4489 } 4490 } 4491 return false; 4492 } 4493 4494 /* Process all regs in location *LOC and change them on equivalent 4495 substitution. Return true if any change was done. */ 4496 static bool 4497 loc_equivalence_change_p (rtx *loc) 4498 { 4499 rtx subst, reg, x = *loc; 4500 bool result = false; 4501 enum rtx_code code = GET_CODE (x); 4502 const char *fmt; 4503 int i, j; 4504 4505 if (code == SUBREG) 4506 { 4507 reg = SUBREG_REG (x); 4508 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg 4509 && GET_MODE (subst) == VOIDmode) 4510 { 4511 /* We cannot reload debug location. Simplify subreg here 4512 while we know the inner mode. */ 4513 *loc = simplify_gen_subreg (GET_MODE (x), subst, 4514 GET_MODE (reg), SUBREG_BYTE (x)); 4515 return true; 4516 } 4517 } 4518 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x) 4519 { 4520 *loc = subst; 4521 return true; 4522 } 4523 4524 /* Scan all the operand sub-expressions. */ 4525 fmt = GET_RTX_FORMAT (code); 4526 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) 4527 { 4528 if (fmt[i] == 'e') 4529 result = loc_equivalence_change_p (&XEXP (x, i)) || result; 4530 else if (fmt[i] == 'E') 4531 for (j = XVECLEN (x, i) - 1; j >= 0; j--) 4532 result 4533 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result; 4534 } 4535 return result; 4536 } 4537 4538 /* Similar to loc_equivalence_change_p, but for use as 4539 simplify_replace_fn_rtx callback. DATA is insn for which the 4540 elimination is done. If it null we don't do the elimination. */ 4541 static rtx 4542 loc_equivalence_callback (rtx loc, const_rtx, void *data) 4543 { 4544 if (!REG_P (loc)) 4545 return NULL_RTX; 4546 4547 rtx subst = (data == NULL 4548 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data)); 4549 if (subst != loc) 4550 return subst; 4551 4552 return NULL_RTX; 4553 } 4554 4555 /* Maximum number of generated reload insns per an insn. It is for 4556 preventing this pass cycling in a bug case. */ 4557 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS 4558 4559 /* The current iteration number of this LRA pass. */ 4560 int lra_constraint_iter; 4561 4562 /* True if we substituted equiv which needs checking register 4563 allocation correctness because the equivalent value contains 4564 allocatable hard registers or when we restore multi-register 4565 pseudo. */ 4566 bool lra_risky_transformations_p; 4567 4568 /* Return true if REGNO is referenced in more than one block. */ 4569 static bool 4570 multi_block_pseudo_p (int regno) 4571 { 4572 basic_block bb = NULL; 4573 unsigned int uid; 4574 bitmap_iterator bi; 4575 4576 if (regno < FIRST_PSEUDO_REGISTER) 4577 return false; 4578 4579 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi) 4580 if (bb == NULL) 4581 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn); 4582 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb) 4583 return true; 4584 return false; 4585 } 4586 4587 /* Return true if LIST contains a deleted insn. */ 4588 static bool 4589 contains_deleted_insn_p (rtx_insn_list *list) 4590 { 4591 for (; list != NULL_RTX; list = list->next ()) 4592 if (NOTE_P (list->insn ()) 4593 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED) 4594 return true; 4595 return false; 4596 } 4597 4598 /* Return true if X contains a pseudo dying in INSN. */ 4599 static bool 4600 dead_pseudo_p (rtx x, rtx_insn *insn) 4601 { 4602 int i, j; 4603 const char *fmt; 4604 enum rtx_code code; 4605 4606 if (REG_P (x)) 4607 return (insn != NULL_RTX 4608 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX); 4609 code = GET_CODE (x); 4610 fmt = GET_RTX_FORMAT (code); 4611 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) 4612 { 4613 if (fmt[i] == 'e') 4614 { 4615 if (dead_pseudo_p (XEXP (x, i), insn)) 4616 return true; 4617 } 4618 else if (fmt[i] == 'E') 4619 { 4620 for (j = XVECLEN (x, i) - 1; j >= 0; j--) 4621 if (dead_pseudo_p (XVECEXP (x, i, j), insn)) 4622 return true; 4623 } 4624 } 4625 return false; 4626 } 4627 4628 /* Return true if INSN contains a dying pseudo in INSN right hand 4629 side. */ 4630 static bool 4631 insn_rhs_dead_pseudo_p (rtx_insn *insn) 4632 { 4633 rtx set = single_set (insn); 4634 4635 gcc_assert (set != NULL); 4636 return dead_pseudo_p (SET_SRC (set), insn); 4637 } 4638 4639 /* Return true if any init insn of REGNO contains a dying pseudo in 4640 insn right hand side. */ 4641 static bool 4642 init_insn_rhs_dead_pseudo_p (int regno) 4643 { 4644 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns; 4645 4646 if (insns == NULL) 4647 return false; 4648 for (; insns != NULL_RTX; insns = insns->next ()) 4649 if (insn_rhs_dead_pseudo_p (insns->insn ())) 4650 return true; 4651 return false; 4652 } 4653 4654 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is 4655 reverse only if we have one init insn with given REGNO as a 4656 source. */ 4657 static bool 4658 reverse_equiv_p (int regno) 4659 { 4660 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns; 4661 rtx set; 4662 4663 if (insns == NULL) 4664 return false; 4665 if (! INSN_P (insns->insn ()) 4666 || insns->next () != NULL) 4667 return false; 4668 if ((set = single_set (insns->insn ())) == NULL_RTX) 4669 return false; 4670 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno; 4671 } 4672 4673 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We 4674 call this function only for non-reverse equivalence. */ 4675 static bool 4676 contains_reloaded_insn_p (int regno) 4677 { 4678 rtx set; 4679 rtx_insn_list *list = ira_reg_equiv[regno].init_insns; 4680 4681 for (; list != NULL; list = list->next ()) 4682 if ((set = single_set (list->insn ())) == NULL_RTX 4683 || ! REG_P (SET_DEST (set)) 4684 || (int) REGNO (SET_DEST (set)) != regno) 4685 return true; 4686 return false; 4687 } 4688 4689 /* Entry function of LRA constraint pass. Return true if the 4690 constraint pass did change the code. */ 4691 bool 4692 lra_constraints (bool first_p) 4693 { 4694 bool changed_p; 4695 int i, hard_regno, new_insns_num; 4696 unsigned int min_len, new_min_len, uid; 4697 rtx set, x, reg, dest_reg; 4698 basic_block last_bb; 4699 bitmap_iterator bi; 4700 4701 lra_constraint_iter++; 4702 if (lra_dump_file != NULL) 4703 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n", 4704 lra_constraint_iter); 4705 changed_p = false; 4706 if (pic_offset_table_rtx 4707 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER) 4708 lra_risky_transformations_p = true; 4709 else 4710 /* On the first iteration we should check IRA assignment 4711 correctness. In rare cases, the assignments can be wrong as 4712 early clobbers operands are ignored in IRA. */ 4713 lra_risky_transformations_p = first_p; 4714 new_insn_uid_start = get_max_uid (); 4715 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num (); 4716 /* Mark used hard regs for target stack size calulations. */ 4717 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++) 4718 if (lra_reg_info[i].nrefs != 0 4719 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0) 4720 { 4721 int j, nregs; 4722 4723 nregs = hard_regno_nregs (hard_regno, lra_reg_info[i].biggest_mode); 4724 for (j = 0; j < nregs; j++) 4725 df_set_regs_ever_live (hard_regno + j, true); 4726 } 4727 /* Do elimination before the equivalence processing as we can spill 4728 some pseudos during elimination. */ 4729 lra_eliminate (false, first_p); 4730 auto_bitmap equiv_insn_bitmap (®_obstack); 4731 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++) 4732 if (lra_reg_info[i].nrefs != 0) 4733 { 4734 ira_reg_equiv[i].profitable_p = true; 4735 reg = regno_reg_rtx[i]; 4736 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg) 4737 { 4738 bool pseudo_p = contains_reg_p (x, false, false); 4739 4740 /* After RTL transformation, we can not guarantee that 4741 pseudo in the substitution was not reloaded which might 4742 make equivalence invalid. For example, in reverse 4743 equiv of p0 4744 4745 p0 <- ... 4746 ... 4747 equiv_mem <- p0 4748 4749 the memory address register was reloaded before the 2nd 4750 insn. */ 4751 if ((! first_p && pseudo_p) 4752 /* We don't use DF for compilation speed sake. So it 4753 is problematic to update live info when we use an 4754 equivalence containing pseudos in more than one 4755 BB. */ 4756 || (pseudo_p && multi_block_pseudo_p (i)) 4757 /* If an init insn was deleted for some reason, cancel 4758 the equiv. We could update the equiv insns after 4759 transformations including an equiv insn deletion 4760 but it is not worthy as such cases are extremely 4761 rare. */ 4762 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns) 4763 /* If it is not a reverse equivalence, we check that a 4764 pseudo in rhs of the init insn is not dying in the 4765 insn. Otherwise, the live info at the beginning of 4766 the corresponding BB might be wrong after we 4767 removed the insn. When the equiv can be a 4768 constant, the right hand side of the init insn can 4769 be a pseudo. */ 4770 || (! reverse_equiv_p (i) 4771 && (init_insn_rhs_dead_pseudo_p (i) 4772 /* If we reloaded the pseudo in an equivalence 4773 init insn, we can not remove the equiv init 4774 insns and the init insns might write into 4775 const memory in this case. */ 4776 || contains_reloaded_insn_p (i))) 4777 /* Prevent access beyond equivalent memory for 4778 paradoxical subregs. */ 4779 || (MEM_P (x) 4780 && maybe_gt (GET_MODE_SIZE (lra_reg_info[i].biggest_mode), 4781 GET_MODE_SIZE (GET_MODE (x)))) 4782 || (pic_offset_table_rtx 4783 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x) 4784 && (targetm.preferred_reload_class 4785 (x, lra_get_allocno_class (i)) == NO_REGS)) 4786 || contains_symbol_ref_p (x)))) 4787 ira_reg_equiv[i].defined_p = false; 4788 if (contains_reg_p (x, false, true)) 4789 ira_reg_equiv[i].profitable_p = false; 4790 if (get_equiv (reg) != reg) 4791 bitmap_ior_into (equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap); 4792 } 4793 } 4794 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++) 4795 update_equiv (i); 4796 /* We should add all insns containing pseudos which should be 4797 substituted by their equivalences. */ 4798 EXECUTE_IF_SET_IN_BITMAP (equiv_insn_bitmap, 0, uid, bi) 4799 lra_push_insn_by_uid (uid); 4800 min_len = lra_insn_stack_length (); 4801 new_insns_num = 0; 4802 last_bb = NULL; 4803 changed_p = false; 4804 while ((new_min_len = lra_insn_stack_length ()) != 0) 4805 { 4806 curr_insn = lra_pop_insn (); 4807 --new_min_len; 4808 curr_bb = BLOCK_FOR_INSN (curr_insn); 4809 if (curr_bb != last_bb) 4810 { 4811 last_bb = curr_bb; 4812 bb_reload_num = lra_curr_reload_num; 4813 } 4814 if (min_len > new_min_len) 4815 { 4816 min_len = new_min_len; 4817 new_insns_num = 0; 4818 } 4819 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER) 4820 internal_error 4821 ("Max. number of generated reload insns per insn is achieved (%d)\n", 4822 MAX_RELOAD_INSNS_NUMBER); 4823 new_insns_num++; 4824 if (DEBUG_INSN_P (curr_insn)) 4825 { 4826 /* We need to check equivalence in debug insn and change 4827 pseudo to the equivalent value if necessary. */ 4828 curr_id = lra_get_insn_recog_data (curr_insn); 4829 if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn))) 4830 { 4831 rtx old = *curr_id->operand_loc[0]; 4832 *curr_id->operand_loc[0] 4833 = simplify_replace_fn_rtx (old, NULL_RTX, 4834 loc_equivalence_callback, curr_insn); 4835 if (old != *curr_id->operand_loc[0]) 4836 { 4837 lra_update_insn_regno_info (curr_insn); 4838 changed_p = true; 4839 } 4840 } 4841 } 4842 else if (INSN_P (curr_insn)) 4843 { 4844 if ((set = single_set (curr_insn)) != NULL_RTX) 4845 { 4846 dest_reg = SET_DEST (set); 4847 /* The equivalence pseudo could be set up as SUBREG in a 4848 case when it is a call restore insn in a mode 4849 different from the pseudo mode. */ 4850 if (GET_CODE (dest_reg) == SUBREG) 4851 dest_reg = SUBREG_REG (dest_reg); 4852 if ((REG_P (dest_reg) 4853 && (x = get_equiv (dest_reg)) != dest_reg 4854 /* Remove insns which set up a pseudo whose value 4855 can not be changed. Such insns might be not in 4856 init_insns because we don't update equiv data 4857 during insn transformations. 4858 4859 As an example, let suppose that a pseudo got 4860 hard register and on the 1st pass was not 4861 changed to equivalent constant. We generate an 4862 additional insn setting up the pseudo because of 4863 secondary memory movement. Then the pseudo is 4864 spilled and we use the equiv constant. In this 4865 case we should remove the additional insn and 4866 this insn is not init_insns list. */ 4867 && (! MEM_P (x) || MEM_READONLY_P (x) 4868 /* Check that this is actually an insn setting 4869 up the equivalence. */ 4870 || in_list_p (curr_insn, 4871 ira_reg_equiv 4872 [REGNO (dest_reg)].init_insns))) 4873 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set)) 4874 && in_list_p (curr_insn, 4875 ira_reg_equiv 4876 [REGNO (SET_SRC (set))].init_insns))) 4877 { 4878 /* This is equiv init insn of pseudo which did not get a 4879 hard register -- remove the insn. */ 4880 if (lra_dump_file != NULL) 4881 { 4882 fprintf (lra_dump_file, 4883 " Removing equiv init insn %i (freq=%d)\n", 4884 INSN_UID (curr_insn), 4885 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn))); 4886 dump_insn_slim (lra_dump_file, curr_insn); 4887 } 4888 if (contains_reg_p (x, true, false)) 4889 lra_risky_transformations_p = true; 4890 lra_set_insn_deleted (curr_insn); 4891 continue; 4892 } 4893 } 4894 curr_id = lra_get_insn_recog_data (curr_insn); 4895 curr_static_id = curr_id->insn_static_data; 4896 init_curr_insn_input_reloads (); 4897 init_curr_operand_mode (); 4898 if (curr_insn_transform (false)) 4899 changed_p = true; 4900 /* Check non-transformed insns too for equiv change as USE 4901 or CLOBBER don't need reloads but can contain pseudos 4902 being changed on their equivalences. */ 4903 else if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn)) 4904 && loc_equivalence_change_p (&PATTERN (curr_insn))) 4905 { 4906 lra_update_insn_regno_info (curr_insn); 4907 changed_p = true; 4908 } 4909 } 4910 } 4911 4912 /* If we used a new hard regno, changed_p should be true because the 4913 hard reg is assigned to a new pseudo. */ 4914 if (flag_checking && !changed_p) 4915 { 4916 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++) 4917 if (lra_reg_info[i].nrefs != 0 4918 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0) 4919 { 4920 int j, nregs = hard_regno_nregs (hard_regno, 4921 PSEUDO_REGNO_MODE (i)); 4922 4923 for (j = 0; j < nregs; j++) 4924 lra_assert (df_regs_ever_live_p (hard_regno + j)); 4925 } 4926 } 4927 return changed_p; 4928 } 4929 4930 static void initiate_invariants (void); 4931 static void finish_invariants (void); 4932 4933 /* Initiate the LRA constraint pass. It is done once per 4934 function. */ 4935 void 4936 lra_constraints_init (void) 4937 { 4938 initiate_invariants (); 4939 } 4940 4941 /* Finalize the LRA constraint pass. It is done once per 4942 function. */ 4943 void 4944 lra_constraints_finish (void) 4945 { 4946 finish_invariants (); 4947 } 4948 4949 4950 4951 /* Structure describes invariants for ineheritance. */ 4952 struct lra_invariant 4953 { 4954 /* The order number of the invariant. */ 4955 int num; 4956 /* The invariant RTX. */ 4957 rtx invariant_rtx; 4958 /* The origin insn of the invariant. */ 4959 rtx_insn *insn; 4960 }; 4961 4962 typedef lra_invariant invariant_t; 4963 typedef invariant_t *invariant_ptr_t; 4964 typedef const invariant_t *const_invariant_ptr_t; 4965 4966 /* Pointer to the inheritance invariants. */ 4967 static vec<invariant_ptr_t> invariants; 4968 4969 /* Allocation pool for the invariants. */ 4970 static object_allocator<lra_invariant> *invariants_pool; 4971 4972 /* Hash table for the invariants. */ 4973 static htab_t invariant_table; 4974 4975 /* Hash function for INVARIANT. */ 4976 static hashval_t 4977 invariant_hash (const void *invariant) 4978 { 4979 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx; 4980 return lra_rtx_hash (inv); 4981 } 4982 4983 /* Equal function for invariants INVARIANT1 and INVARIANT2. */ 4984 static int 4985 invariant_eq_p (const void *invariant1, const void *invariant2) 4986 { 4987 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx; 4988 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx; 4989 4990 return rtx_equal_p (inv1, inv2); 4991 } 4992 4993 /* Insert INVARIANT_RTX into the table if it is not there yet. Return 4994 invariant which is in the table. */ 4995 static invariant_ptr_t 4996 insert_invariant (rtx invariant_rtx) 4997 { 4998 void **entry_ptr; 4999 invariant_t invariant; 5000 invariant_ptr_t invariant_ptr; 5001 5002 invariant.invariant_rtx = invariant_rtx; 5003 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT); 5004 if (*entry_ptr == NULL) 5005 { 5006 invariant_ptr = invariants_pool->allocate (); 5007 invariant_ptr->invariant_rtx = invariant_rtx; 5008 invariant_ptr->insn = NULL; 5009 invariants.safe_push (invariant_ptr); 5010 *entry_ptr = (void *) invariant_ptr; 5011 } 5012 return (invariant_ptr_t) *entry_ptr; 5013 } 5014 5015 /* Initiate the invariant table. */ 5016 static void 5017 initiate_invariants (void) 5018 { 5019 invariants.create (100); 5020 invariants_pool 5021 = new object_allocator<lra_invariant> ("Inheritance invariants"); 5022 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL); 5023 } 5024 5025 /* Finish the invariant table. */ 5026 static void 5027 finish_invariants (void) 5028 { 5029 htab_delete (invariant_table); 5030 delete invariants_pool; 5031 invariants.release (); 5032 } 5033 5034 /* Make the invariant table empty. */ 5035 static void 5036 clear_invariants (void) 5037 { 5038 htab_empty (invariant_table); 5039 invariants_pool->release (); 5040 invariants.truncate (0); 5041 } 5042 5043 5044 5045 /* This page contains code to do inheritance/split 5046 transformations. */ 5047 5048 /* Number of reloads passed so far in current EBB. */ 5049 static int reloads_num; 5050 5051 /* Number of calls passed so far in current EBB. */ 5052 static int calls_num; 5053 5054 /* Current reload pseudo check for validity of elements in 5055 USAGE_INSNS. */ 5056 static int curr_usage_insns_check; 5057 5058 /* Info about last usage of registers in EBB to do inheritance/split 5059 transformation. Inheritance transformation is done from a spilled 5060 pseudo and split transformations from a hard register or a pseudo 5061 assigned to a hard register. */ 5062 struct usage_insns 5063 { 5064 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member 5065 value INSNS is valid. The insns is chain of optional debug insns 5066 and a finishing non-debug insn using the corresponding reg. The 5067 value is also used to mark the registers which are set up in the 5068 current insn. The negated insn uid is used for this. */ 5069 int check; 5070 /* Value of global reloads_num at the last insn in INSNS. */ 5071 int reloads_num; 5072 /* Value of global reloads_nums at the last insn in INSNS. */ 5073 int calls_num; 5074 /* It can be true only for splitting. And it means that the restore 5075 insn should be put after insn given by the following member. */ 5076 bool after_p; 5077 /* Next insns in the current EBB which use the original reg and the 5078 original reg value is not changed between the current insn and 5079 the next insns. In order words, e.g. for inheritance, if we need 5080 to use the original reg value again in the next insns we can try 5081 to use the value in a hard register from a reload insn of the 5082 current insn. */ 5083 rtx insns; 5084 }; 5085 5086 /* Map: regno -> corresponding pseudo usage insns. */ 5087 static struct usage_insns *usage_insns; 5088 5089 static void 5090 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p) 5091 { 5092 usage_insns[regno].check = curr_usage_insns_check; 5093 usage_insns[regno].insns = insn; 5094 usage_insns[regno].reloads_num = reloads_num; 5095 usage_insns[regno].calls_num = calls_num; 5096 usage_insns[regno].after_p = after_p; 5097 } 5098 5099 /* The function is used to form list REGNO usages which consists of 5100 optional debug insns finished by a non-debug insn using REGNO. 5101 RELOADS_NUM is current number of reload insns processed so far. */ 5102 static void 5103 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num) 5104 { 5105 rtx next_usage_insns; 5106 5107 if (usage_insns[regno].check == curr_usage_insns_check 5108 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX 5109 && DEBUG_INSN_P (insn)) 5110 { 5111 /* Check that we did not add the debug insn yet. */ 5112 if (next_usage_insns != insn 5113 && (GET_CODE (next_usage_insns) != INSN_LIST 5114 || XEXP (next_usage_insns, 0) != insn)) 5115 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn, 5116 next_usage_insns); 5117 } 5118 else if (NONDEBUG_INSN_P (insn)) 5119 setup_next_usage_insn (regno, insn, reloads_num, false); 5120 else 5121 usage_insns[regno].check = 0; 5122 } 5123 5124 /* Return first non-debug insn in list USAGE_INSNS. */ 5125 static rtx_insn * 5126 skip_usage_debug_insns (rtx usage_insns) 5127 { 5128 rtx insn; 5129 5130 /* Skip debug insns. */ 5131 for (insn = usage_insns; 5132 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST; 5133 insn = XEXP (insn, 1)) 5134 ; 5135 return safe_as_a <rtx_insn *> (insn); 5136 } 5137 5138 /* Return true if we need secondary memory moves for insn in 5139 USAGE_INSNS after inserting inherited pseudo of class INHER_CL 5140 into the insn. */ 5141 static bool 5142 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED, 5143 rtx usage_insns ATTRIBUTE_UNUSED) 5144 { 5145 rtx_insn *insn; 5146 rtx set, dest; 5147 enum reg_class cl; 5148 5149 if (inher_cl == ALL_REGS 5150 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX) 5151 return false; 5152 lra_assert (INSN_P (insn)); 5153 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set))) 5154 return false; 5155 dest = SET_DEST (set); 5156 if (! REG_P (dest)) 5157 return false; 5158 lra_assert (inher_cl != NO_REGS); 5159 cl = get_reg_class (REGNO (dest)); 5160 return (cl != NO_REGS && cl != ALL_REGS 5161 && targetm.secondary_memory_needed (GET_MODE (dest), inher_cl, cl)); 5162 } 5163 5164 /* Registers involved in inheritance/split in the current EBB 5165 (inheritance/split pseudos and original registers). */ 5166 static bitmap_head check_only_regs; 5167 5168 /* Reload pseudos can not be involded in invariant inheritance in the 5169 current EBB. */ 5170 static bitmap_head invalid_invariant_regs; 5171 5172 /* Do inheritance transformations for insn INSN, which defines (if 5173 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which 5174 instruction in the EBB next uses ORIGINAL_REGNO; it has the same 5175 form as the "insns" field of usage_insns. Return true if we 5176 succeed in such transformation. 5177 5178 The transformations look like: 5179 5180 p <- ... i <- ... 5181 ... p <- i (new insn) 5182 ... => 5183 <- ... p ... <- ... i ... 5184 or 5185 ... i <- p (new insn) 5186 <- ... p ... <- ... i ... 5187 ... => 5188 <- ... p ... <- ... i ... 5189 where p is a spilled original pseudo and i is a new inheritance pseudo. 5190 5191 5192 The inheritance pseudo has the smallest class of two classes CL and 5193 class of ORIGINAL REGNO. */ 5194 static bool 5195 inherit_reload_reg (bool def_p, int original_regno, 5196 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns) 5197 { 5198 if (optimize_function_for_size_p (cfun)) 5199 return false; 5200 5201 enum reg_class rclass = lra_get_allocno_class (original_regno); 5202 rtx original_reg = regno_reg_rtx[original_regno]; 5203 rtx new_reg, usage_insn; 5204 rtx_insn *new_insns; 5205 5206 lra_assert (! usage_insns[original_regno].after_p); 5207 if (lra_dump_file != NULL) 5208 fprintf (lra_dump_file, 5209 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n"); 5210 if (! ira_reg_classes_intersect_p[cl][rclass]) 5211 { 5212 if (lra_dump_file != NULL) 5213 { 5214 fprintf (lra_dump_file, 5215 " Rejecting inheritance for %d " 5216 "because of disjoint classes %s and %s\n", 5217 original_regno, reg_class_names[cl], 5218 reg_class_names[rclass]); 5219 fprintf (lra_dump_file, 5220 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n"); 5221 } 5222 return false; 5223 } 5224 if ((ira_class_subset_p[cl][rclass] && cl != rclass) 5225 /* We don't use a subset of two classes because it can be 5226 NO_REGS. This transformation is still profitable in most 5227 cases even if the classes are not intersected as register 5228 move is probably cheaper than a memory load. */ 5229 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass]) 5230 { 5231 if (lra_dump_file != NULL) 5232 fprintf (lra_dump_file, " Use smallest class of %s and %s\n", 5233 reg_class_names[cl], reg_class_names[rclass]); 5234 5235 rclass = cl; 5236 } 5237 if (check_secondary_memory_needed_p (rclass, next_usage_insns)) 5238 { 5239 /* Reject inheritance resulting in secondary memory moves. 5240 Otherwise, there is a danger in LRA cycling. Also such 5241 transformation will be unprofitable. */ 5242 if (lra_dump_file != NULL) 5243 { 5244 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns); 5245 rtx set = single_set (insn); 5246 5247 lra_assert (set != NULL_RTX); 5248 5249 rtx dest = SET_DEST (set); 5250 5251 lra_assert (REG_P (dest)); 5252 fprintf (lra_dump_file, 5253 " Rejecting inheritance for insn %d(%s)<-%d(%s) " 5254 "as secondary mem is needed\n", 5255 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))], 5256 original_regno, reg_class_names[rclass]); 5257 fprintf (lra_dump_file, 5258 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n"); 5259 } 5260 return false; 5261 } 5262 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg, 5263 rclass, "inheritance"); 5264 start_sequence (); 5265 if (def_p) 5266 lra_emit_move (original_reg, new_reg); 5267 else 5268 lra_emit_move (new_reg, original_reg); 5269 new_insns = get_insns (); 5270 end_sequence (); 5271 if (NEXT_INSN (new_insns) != NULL_RTX) 5272 { 5273 if (lra_dump_file != NULL) 5274 { 5275 fprintf (lra_dump_file, 5276 " Rejecting inheritance %d->%d " 5277 "as it results in 2 or more insns:\n", 5278 original_regno, REGNO (new_reg)); 5279 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0); 5280 fprintf (lra_dump_file, 5281 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n"); 5282 } 5283 return false; 5284 } 5285 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false); 5286 lra_update_insn_regno_info (insn); 5287 if (! def_p) 5288 /* We now have a new usage insn for original regno. */ 5289 setup_next_usage_insn (original_regno, new_insns, reloads_num, false); 5290 if (lra_dump_file != NULL) 5291 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n", 5292 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index); 5293 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno]; 5294 bitmap_set_bit (&check_only_regs, REGNO (new_reg)); 5295 bitmap_set_bit (&check_only_regs, original_regno); 5296 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg)); 5297 if (def_p) 5298 lra_process_new_insns (insn, NULL, new_insns, 5299 "Add original<-inheritance"); 5300 else 5301 lra_process_new_insns (insn, new_insns, NULL, 5302 "Add inheritance<-original"); 5303 while (next_usage_insns != NULL_RTX) 5304 { 5305 if (GET_CODE (next_usage_insns) != INSN_LIST) 5306 { 5307 usage_insn = next_usage_insns; 5308 lra_assert (NONDEBUG_INSN_P (usage_insn)); 5309 next_usage_insns = NULL; 5310 } 5311 else 5312 { 5313 usage_insn = XEXP (next_usage_insns, 0); 5314 lra_assert (DEBUG_INSN_P (usage_insn)); 5315 next_usage_insns = XEXP (next_usage_insns, 1); 5316 } 5317 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false, 5318 DEBUG_INSN_P (usage_insn)); 5319 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn)); 5320 if (lra_dump_file != NULL) 5321 { 5322 basic_block bb = BLOCK_FOR_INSN (usage_insn); 5323 fprintf (lra_dump_file, 5324 " Inheritance reuse change %d->%d (bb%d):\n", 5325 original_regno, REGNO (new_reg), 5326 bb ? bb->index : -1); 5327 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn)); 5328 } 5329 } 5330 if (lra_dump_file != NULL) 5331 fprintf (lra_dump_file, 5332 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n"); 5333 return true; 5334 } 5335 5336 /* Return true if we need a caller save/restore for pseudo REGNO which 5337 was assigned to a hard register. */ 5338 static inline bool 5339 need_for_call_save_p (int regno) 5340 { 5341 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0); 5342 return (usage_insns[regno].calls_num < calls_num 5343 && (overlaps_hard_reg_set_p 5344 ((flag_ipa_ra && 5345 ! hard_reg_set_empty_p (lra_reg_info[regno].actual_call_used_reg_set)) 5346 ? lra_reg_info[regno].actual_call_used_reg_set 5347 : call_used_reg_set, 5348 PSEUDO_REGNO_MODE (regno), reg_renumber[regno]) 5349 || (targetm.hard_regno_call_part_clobbered 5350 (reg_renumber[regno], PSEUDO_REGNO_MODE (regno))))); 5351 } 5352 5353 /* Global registers occurring in the current EBB. */ 5354 static bitmap_head ebb_global_regs; 5355 5356 /* Return true if we need a split for hard register REGNO or pseudo 5357 REGNO which was assigned to a hard register. 5358 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be 5359 used for reloads since the EBB end. It is an approximation of the 5360 used hard registers in the split range. The exact value would 5361 require expensive calculations. If we were aggressive with 5362 splitting because of the approximation, the split pseudo will save 5363 the same hard register assignment and will be removed in the undo 5364 pass. We still need the approximation because too aggressive 5365 splitting would result in too inaccurate cost calculation in the 5366 assignment pass because of too many generated moves which will be 5367 probably removed in the undo pass. */ 5368 static inline bool 5369 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno) 5370 { 5371 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno]; 5372 5373 lra_assert (hard_regno >= 0); 5374 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno) 5375 /* Don't split eliminable hard registers, otherwise we can 5376 split hard registers like hard frame pointer, which 5377 lives on BB start/end according to DF-infrastructure, 5378 when there is a pseudo assigned to the register and 5379 living in the same BB. */ 5380 && (regno >= FIRST_PSEUDO_REGISTER 5381 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno)) 5382 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno) 5383 /* Don't split call clobbered hard regs living through 5384 calls, otherwise we might have a check problem in the 5385 assign sub-pass as in the most cases (exception is a 5386 situation when lra_risky_transformations_p value is 5387 true) the assign pass assumes that all pseudos living 5388 through calls are assigned to call saved hard regs. */ 5389 && (regno >= FIRST_PSEUDO_REGISTER 5390 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno) 5391 || usage_insns[regno].calls_num == calls_num) 5392 /* We need at least 2 reloads to make pseudo splitting 5393 profitable. We should provide hard regno splitting in 5394 any case to solve 1st insn scheduling problem when 5395 moving hard register definition up might result in 5396 impossibility to find hard register for reload pseudo of 5397 small register class. */ 5398 && (usage_insns[regno].reloads_num 5399 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num) 5400 && (regno < FIRST_PSEUDO_REGISTER 5401 /* For short living pseudos, spilling + inheritance can 5402 be considered a substitution for splitting. 5403 Therefore we do not splitting for local pseudos. It 5404 decreases also aggressiveness of splitting. The 5405 minimal number of references is chosen taking into 5406 account that for 2 references splitting has no sense 5407 as we can just spill the pseudo. */ 5408 || (regno >= FIRST_PSEUDO_REGISTER 5409 && lra_reg_info[regno].nrefs > 3 5410 && bitmap_bit_p (&ebb_global_regs, regno)))) 5411 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno))); 5412 } 5413 5414 /* Return class for the split pseudo created from original pseudo with 5415 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We 5416 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and 5417 results in no secondary memory movements. */ 5418 static enum reg_class 5419 choose_split_class (enum reg_class allocno_class, 5420 int hard_regno ATTRIBUTE_UNUSED, 5421 machine_mode mode ATTRIBUTE_UNUSED) 5422 { 5423 int i; 5424 enum reg_class cl, best_cl = NO_REGS; 5425 enum reg_class hard_reg_class ATTRIBUTE_UNUSED 5426 = REGNO_REG_CLASS (hard_regno); 5427 5428 if (! targetm.secondary_memory_needed (mode, allocno_class, allocno_class) 5429 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno)) 5430 return allocno_class; 5431 for (i = 0; 5432 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES; 5433 i++) 5434 if (! targetm.secondary_memory_needed (mode, cl, hard_reg_class) 5435 && ! targetm.secondary_memory_needed (mode, hard_reg_class, cl) 5436 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno) 5437 && (best_cl == NO_REGS 5438 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl])) 5439 best_cl = cl; 5440 return best_cl; 5441 } 5442 5443 /* Copy any equivalence information from ORIGINAL_REGNO to NEW_REGNO. 5444 It only makes sense to call this function if NEW_REGNO is always 5445 equal to ORIGINAL_REGNO. */ 5446 5447 static void 5448 lra_copy_reg_equiv (unsigned int new_regno, unsigned int original_regno) 5449 { 5450 if (!ira_reg_equiv[original_regno].defined_p) 5451 return; 5452 5453 ira_expand_reg_equiv (); 5454 ira_reg_equiv[new_regno].defined_p = true; 5455 if (ira_reg_equiv[original_regno].memory) 5456 ira_reg_equiv[new_regno].memory 5457 = copy_rtx (ira_reg_equiv[original_regno].memory); 5458 if (ira_reg_equiv[original_regno].constant) 5459 ira_reg_equiv[new_regno].constant 5460 = copy_rtx (ira_reg_equiv[original_regno].constant); 5461 if (ira_reg_equiv[original_regno].invariant) 5462 ira_reg_equiv[new_regno].invariant 5463 = copy_rtx (ira_reg_equiv[original_regno].invariant); 5464 } 5465 5466 /* Do split transformations for insn INSN, which defines or uses 5467 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in 5468 the EBB next uses ORIGINAL_REGNO; it has the same form as the 5469 "insns" field of usage_insns. If TO is not NULL, we don't use 5470 usage_insns, we put restore insns after TO insn. 5471 5472 The transformations look like: 5473 5474 p <- ... p <- ... 5475 ... s <- p (new insn -- save) 5476 ... => 5477 ... p <- s (new insn -- restore) 5478 <- ... p ... <- ... p ... 5479 or 5480 <- ... p ... <- ... p ... 5481 ... s <- p (new insn -- save) 5482 ... => 5483 ... p <- s (new insn -- restore) 5484 <- ... p ... <- ... p ... 5485 5486 where p is an original pseudo got a hard register or a hard 5487 register and s is a new split pseudo. The save is put before INSN 5488 if BEFORE_P is true. Return true if we succeed in such 5489 transformation. */ 5490 static bool 5491 split_reg (bool before_p, int original_regno, rtx_insn *insn, 5492 rtx next_usage_insns, rtx_insn *to) 5493 { 5494 enum reg_class rclass; 5495 rtx original_reg; 5496 int hard_regno, nregs; 5497 rtx new_reg, usage_insn; 5498 rtx_insn *restore, *save; 5499 bool after_p; 5500 bool call_save_p; 5501 machine_mode mode; 5502 5503 if (original_regno < FIRST_PSEUDO_REGISTER) 5504 { 5505 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)]; 5506 hard_regno = original_regno; 5507 call_save_p = false; 5508 nregs = 1; 5509 mode = lra_reg_info[hard_regno].biggest_mode; 5510 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]); 5511 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen 5512 as part of a multi-word register. In that case, or if the biggest 5513 mode was larger than a register, just use the reg_rtx. Otherwise, 5514 limit the size to that of the biggest access in the function. */ 5515 if (mode == VOIDmode 5516 || paradoxical_subreg_p (mode, reg_rtx_mode)) 5517 { 5518 original_reg = regno_reg_rtx[hard_regno]; 5519 mode = reg_rtx_mode; 5520 } 5521 else 5522 original_reg = gen_rtx_REG (mode, hard_regno); 5523 } 5524 else 5525 { 5526 mode = PSEUDO_REGNO_MODE (original_regno); 5527 hard_regno = reg_renumber[original_regno]; 5528 nregs = hard_regno_nregs (hard_regno, mode); 5529 rclass = lra_get_allocno_class (original_regno); 5530 original_reg = regno_reg_rtx[original_regno]; 5531 call_save_p = need_for_call_save_p (original_regno); 5532 } 5533 lra_assert (hard_regno >= 0); 5534 if (lra_dump_file != NULL) 5535 fprintf (lra_dump_file, 5536 " ((((((((((((((((((((((((((((((((((((((((((((((((\n"); 5537 5538 if (call_save_p) 5539 { 5540 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno, 5541 hard_regno_nregs (hard_regno, mode), 5542 mode); 5543 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save"); 5544 } 5545 else 5546 { 5547 rclass = choose_split_class (rclass, hard_regno, mode); 5548 if (rclass == NO_REGS) 5549 { 5550 if (lra_dump_file != NULL) 5551 { 5552 fprintf (lra_dump_file, 5553 " Rejecting split of %d(%s): " 5554 "no good reg class for %d(%s)\n", 5555 original_regno, 5556 reg_class_names[lra_get_allocno_class (original_regno)], 5557 hard_regno, 5558 reg_class_names[REGNO_REG_CLASS (hard_regno)]); 5559 fprintf 5560 (lra_dump_file, 5561 " ))))))))))))))))))))))))))))))))))))))))))))))))\n"); 5562 } 5563 return false; 5564 } 5565 /* Split_if_necessary can split hard registers used as part of a 5566 multi-register mode but splits each register individually. The 5567 mode used for each independent register may not be supported 5568 so reject the split. Splitting the wider mode should theoretically 5569 be possible but is not implemented. */ 5570 if (!targetm.hard_regno_mode_ok (hard_regno, mode)) 5571 { 5572 if (lra_dump_file != NULL) 5573 { 5574 fprintf (lra_dump_file, 5575 " Rejecting split of %d(%s): unsuitable mode %s\n", 5576 original_regno, 5577 reg_class_names[lra_get_allocno_class (original_regno)], 5578 GET_MODE_NAME (mode)); 5579 fprintf 5580 (lra_dump_file, 5581 " ))))))))))))))))))))))))))))))))))))))))))))))))\n"); 5582 } 5583 return false; 5584 } 5585 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split"); 5586 reg_renumber[REGNO (new_reg)] = hard_regno; 5587 } 5588 int new_regno = REGNO (new_reg); 5589 save = emit_spill_move (true, new_reg, original_reg); 5590 if (NEXT_INSN (save) != NULL_RTX && !call_save_p) 5591 { 5592 if (lra_dump_file != NULL) 5593 { 5594 fprintf 5595 (lra_dump_file, 5596 " Rejecting split %d->%d resulting in > 2 save insns:\n", 5597 original_regno, new_regno); 5598 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0); 5599 fprintf (lra_dump_file, 5600 " ))))))))))))))))))))))))))))))))))))))))))))))))\n"); 5601 } 5602 return false; 5603 } 5604 restore = emit_spill_move (false, new_reg, original_reg); 5605 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p) 5606 { 5607 if (lra_dump_file != NULL) 5608 { 5609 fprintf (lra_dump_file, 5610 " Rejecting split %d->%d " 5611 "resulting in > 2 restore insns:\n", 5612 original_regno, new_regno); 5613 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0); 5614 fprintf (lra_dump_file, 5615 " ))))))))))))))))))))))))))))))))))))))))))))))))\n"); 5616 } 5617 return false; 5618 } 5619 /* Transfer equivalence information to the spill register, so that 5620 if we fail to allocate the spill register, we have the option of 5621 rematerializing the original value instead of spilling to the stack. */ 5622 if (!HARD_REGISTER_NUM_P (original_regno) 5623 && mode == PSEUDO_REGNO_MODE (original_regno)) 5624 lra_copy_reg_equiv (new_regno, original_regno); 5625 lra_reg_info[new_regno].restore_rtx = regno_reg_rtx[original_regno]; 5626 bitmap_set_bit (&check_only_regs, new_regno); 5627 bitmap_set_bit (&check_only_regs, original_regno); 5628 bitmap_set_bit (&lra_split_regs, new_regno); 5629 if (to != NULL) 5630 { 5631 usage_insn = to; 5632 after_p = TRUE; 5633 } 5634 else 5635 { 5636 after_p = usage_insns[original_regno].after_p; 5637 for (;;) 5638 { 5639 if (GET_CODE (next_usage_insns) != INSN_LIST) 5640 { 5641 usage_insn = next_usage_insns; 5642 break; 5643 } 5644 usage_insn = XEXP (next_usage_insns, 0); 5645 lra_assert (DEBUG_INSN_P (usage_insn)); 5646 next_usage_insns = XEXP (next_usage_insns, 1); 5647 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false, 5648 true); 5649 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn)); 5650 if (lra_dump_file != NULL) 5651 { 5652 fprintf (lra_dump_file, " Split reuse change %d->%d:\n", 5653 original_regno, new_regno); 5654 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn)); 5655 } 5656 } 5657 } 5658 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn)); 5659 lra_assert (usage_insn != insn || (after_p && before_p)); 5660 lra_process_new_insns (as_a <rtx_insn *> (usage_insn), 5661 after_p ? NULL : restore, 5662 after_p ? restore : NULL, 5663 call_save_p 5664 ? "Add reg<-save" : "Add reg<-split"); 5665 lra_process_new_insns (insn, before_p ? save : NULL, 5666 before_p ? NULL : save, 5667 call_save_p 5668 ? "Add save<-reg" : "Add split<-reg"); 5669 if (nregs > 1) 5670 /* If we are trying to split multi-register. We should check 5671 conflicts on the next assignment sub-pass. IRA can allocate on 5672 sub-register levels, LRA do this on pseudos level right now and 5673 this discrepancy may create allocation conflicts after 5674 splitting. */ 5675 lra_risky_transformations_p = true; 5676 if (lra_dump_file != NULL) 5677 fprintf (lra_dump_file, 5678 " ))))))))))))))))))))))))))))))))))))))))))))))))\n"); 5679 return true; 5680 } 5681 5682 /* Split a hard reg for reload pseudo REGNO having RCLASS and living 5683 in the range [FROM, TO]. Return true if did a split. Otherwise, 5684 return false. */ 5685 bool 5686 spill_hard_reg_in_range (int regno, enum reg_class rclass, rtx_insn *from, rtx_insn *to) 5687 { 5688 int i, hard_regno; 5689 int rclass_size; 5690 rtx_insn *insn; 5691 unsigned int uid; 5692 bitmap_iterator bi; 5693 HARD_REG_SET ignore; 5694 5695 lra_assert (from != NULL && to != NULL); 5696 CLEAR_HARD_REG_SET (ignore); 5697 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi) 5698 { 5699 lra_insn_recog_data_t id = lra_insn_recog_data[uid]; 5700 struct lra_static_insn_data *static_id = id->insn_static_data; 5701 struct lra_insn_reg *reg; 5702 5703 for (reg = id->regs; reg != NULL; reg = reg->next) 5704 if (reg->regno <= FIRST_PSEUDO_REGISTER) 5705 SET_HARD_REG_BIT (ignore, reg->regno); 5706 for (reg = static_id->hard_regs; reg != NULL; reg = reg->next) 5707 SET_HARD_REG_BIT (ignore, reg->regno); 5708 } 5709 rclass_size = ira_class_hard_regs_num[rclass]; 5710 for (i = 0; i < rclass_size; i++) 5711 { 5712 hard_regno = ira_class_hard_regs[rclass][i]; 5713 if (! TEST_HARD_REG_BIT (lra_reg_info[regno].conflict_hard_regs, hard_regno) 5714 || TEST_HARD_REG_BIT (ignore, hard_regno)) 5715 continue; 5716 for (insn = from; insn != NEXT_INSN (to); insn = NEXT_INSN (insn)) 5717 if (bitmap_bit_p (&lra_reg_info[hard_regno].insn_bitmap, 5718 INSN_UID (insn))) 5719 break; 5720 if (insn != NEXT_INSN (to)) 5721 continue; 5722 if (split_reg (TRUE, hard_regno, from, NULL, to)) 5723 return true; 5724 } 5725 return false; 5726 } 5727 5728 /* Recognize that we need a split transformation for insn INSN, which 5729 defines or uses REGNO in its insn biggest MODE (we use it only if 5730 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains 5731 hard registers which might be used for reloads since the EBB end. 5732 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla 5733 uid before starting INSN processing. Return true if we succeed in 5734 such transformation. */ 5735 static bool 5736 split_if_necessary (int regno, machine_mode mode, 5737 HARD_REG_SET potential_reload_hard_regs, 5738 bool before_p, rtx_insn *insn, int max_uid) 5739 { 5740 bool res = false; 5741 int i, nregs = 1; 5742 rtx next_usage_insns; 5743 5744 if (regno < FIRST_PSEUDO_REGISTER) 5745 nregs = hard_regno_nregs (regno, mode); 5746 for (i = 0; i < nregs; i++) 5747 if (usage_insns[regno + i].check == curr_usage_insns_check 5748 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX 5749 /* To avoid processing the register twice or more. */ 5750 && ((GET_CODE (next_usage_insns) != INSN_LIST 5751 && INSN_UID (next_usage_insns) < max_uid) 5752 || (GET_CODE (next_usage_insns) == INSN_LIST 5753 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid))) 5754 && need_for_split_p (potential_reload_hard_regs, regno + i) 5755 && split_reg (before_p, regno + i, insn, next_usage_insns, NULL)) 5756 res = true; 5757 return res; 5758 } 5759 5760 /* Return TRUE if rtx X is considered as an invariant for 5761 inheritance. */ 5762 static bool 5763 invariant_p (const_rtx x) 5764 { 5765 machine_mode mode; 5766 const char *fmt; 5767 enum rtx_code code; 5768 int i, j; 5769 5770 code = GET_CODE (x); 5771 mode = GET_MODE (x); 5772 if (code == SUBREG) 5773 { 5774 x = SUBREG_REG (x); 5775 code = GET_CODE (x); 5776 mode = wider_subreg_mode (mode, GET_MODE (x)); 5777 } 5778 5779 if (MEM_P (x)) 5780 return false; 5781 5782 if (REG_P (x)) 5783 { 5784 int i, nregs, regno = REGNO (x); 5785 5786 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM 5787 || TEST_HARD_REG_BIT (eliminable_regset, regno) 5788 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC) 5789 return false; 5790 nregs = hard_regno_nregs (regno, mode); 5791 for (i = 0; i < nregs; i++) 5792 if (! fixed_regs[regno + i] 5793 /* A hard register may be clobbered in the current insn 5794 but we can ignore this case because if the hard 5795 register is used it should be set somewhere after the 5796 clobber. */ 5797 || bitmap_bit_p (&invalid_invariant_regs, regno + i)) 5798 return false; 5799 } 5800 fmt = GET_RTX_FORMAT (code); 5801 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) 5802 { 5803 if (fmt[i] == 'e') 5804 { 5805 if (! invariant_p (XEXP (x, i))) 5806 return false; 5807 } 5808 else if (fmt[i] == 'E') 5809 { 5810 for (j = XVECLEN (x, i) - 1; j >= 0; j--) 5811 if (! invariant_p (XVECEXP (x, i, j))) 5812 return false; 5813 } 5814 } 5815 return true; 5816 } 5817 5818 /* We have 'dest_reg <- invariant'. Let us try to make an invariant 5819 inheritance transformation (using dest_reg instead invariant in a 5820 subsequent insn). */ 5821 static bool 5822 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx) 5823 { 5824 invariant_ptr_t invariant_ptr; 5825 rtx_insn *insn, *new_insns; 5826 rtx insn_set, insn_reg, new_reg; 5827 int insn_regno; 5828 bool succ_p = false; 5829 int dst_regno = REGNO (dst_reg); 5830 machine_mode dst_mode = GET_MODE (dst_reg); 5831 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl; 5832 5833 invariant_ptr = insert_invariant (invariant_rtx); 5834 if ((insn = invariant_ptr->insn) != NULL_RTX) 5835 { 5836 /* We have a subsequent insn using the invariant. */ 5837 insn_set = single_set (insn); 5838 lra_assert (insn_set != NULL); 5839 insn_reg = SET_DEST (insn_set); 5840 lra_assert (REG_P (insn_reg)); 5841 insn_regno = REGNO (insn_reg); 5842 insn_reg_cl = lra_get_allocno_class (insn_regno); 5843 5844 if (dst_mode == GET_MODE (insn_reg) 5845 /* We should consider only result move reg insns which are 5846 cheap. */ 5847 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2 5848 && targetm.register_move_cost (dst_mode, cl, cl) == 2) 5849 { 5850 if (lra_dump_file != NULL) 5851 fprintf (lra_dump_file, 5852 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n"); 5853 new_reg = lra_create_new_reg (dst_mode, dst_reg, 5854 cl, "invariant inheritance"); 5855 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg)); 5856 bitmap_set_bit (&check_only_regs, REGNO (new_reg)); 5857 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn); 5858 start_sequence (); 5859 lra_emit_move (new_reg, dst_reg); 5860 new_insns = get_insns (); 5861 end_sequence (); 5862 lra_process_new_insns (curr_insn, NULL, new_insns, 5863 "Add invariant inheritance<-original"); 5864 start_sequence (); 5865 lra_emit_move (SET_DEST (insn_set), new_reg); 5866 new_insns = get_insns (); 5867 end_sequence (); 5868 lra_process_new_insns (insn, NULL, new_insns, 5869 "Changing reload<-inheritance"); 5870 lra_set_insn_deleted (insn); 5871 succ_p = true; 5872 if (lra_dump_file != NULL) 5873 { 5874 fprintf (lra_dump_file, 5875 " Invariant inheritance reuse change %d (bb%d):\n", 5876 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index); 5877 dump_insn_slim (lra_dump_file, insn); 5878 fprintf (lra_dump_file, 5879 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n"); 5880 } 5881 } 5882 } 5883 invariant_ptr->insn = curr_insn; 5884 return succ_p; 5885 } 5886 5887 /* Check only registers living at the current program point in the 5888 current EBB. */ 5889 static bitmap_head live_regs; 5890 5891 /* Update live info in EBB given by its HEAD and TAIL insns after 5892 inheritance/split transformation. The function removes dead moves 5893 too. */ 5894 static void 5895 update_ebb_live_info (rtx_insn *head, rtx_insn *tail) 5896 { 5897 unsigned int j; 5898 int i, regno; 5899 bool live_p; 5900 rtx_insn *prev_insn; 5901 rtx set; 5902 bool remove_p; 5903 basic_block last_bb, prev_bb, curr_bb; 5904 bitmap_iterator bi; 5905 struct lra_insn_reg *reg; 5906 edge e; 5907 edge_iterator ei; 5908 5909 last_bb = BLOCK_FOR_INSN (tail); 5910 prev_bb = NULL; 5911 for (curr_insn = tail; 5912 curr_insn != PREV_INSN (head); 5913 curr_insn = prev_insn) 5914 { 5915 prev_insn = PREV_INSN (curr_insn); 5916 /* We need to process empty blocks too. They contain 5917 NOTE_INSN_BASIC_BLOCK referring for the basic block. */ 5918 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK) 5919 continue; 5920 curr_bb = BLOCK_FOR_INSN (curr_insn); 5921 if (curr_bb != prev_bb) 5922 { 5923 if (prev_bb != NULL) 5924 { 5925 /* Update df_get_live_in (prev_bb): */ 5926 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi) 5927 if (bitmap_bit_p (&live_regs, j)) 5928 bitmap_set_bit (df_get_live_in (prev_bb), j); 5929 else 5930 bitmap_clear_bit (df_get_live_in (prev_bb), j); 5931 } 5932 if (curr_bb != last_bb) 5933 { 5934 /* Update df_get_live_out (curr_bb): */ 5935 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi) 5936 { 5937 live_p = bitmap_bit_p (&live_regs, j); 5938 if (! live_p) 5939 FOR_EACH_EDGE (e, ei, curr_bb->succs) 5940 if (bitmap_bit_p (df_get_live_in (e->dest), j)) 5941 { 5942 live_p = true; 5943 break; 5944 } 5945 if (live_p) 5946 bitmap_set_bit (df_get_live_out (curr_bb), j); 5947 else 5948 bitmap_clear_bit (df_get_live_out (curr_bb), j); 5949 } 5950 } 5951 prev_bb = curr_bb; 5952 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb)); 5953 } 5954 if (! NONDEBUG_INSN_P (curr_insn)) 5955 continue; 5956 curr_id = lra_get_insn_recog_data (curr_insn); 5957 curr_static_id = curr_id->insn_static_data; 5958 remove_p = false; 5959 if ((set = single_set (curr_insn)) != NULL_RTX 5960 && REG_P (SET_DEST (set)) 5961 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER 5962 && SET_DEST (set) != pic_offset_table_rtx 5963 && bitmap_bit_p (&check_only_regs, regno) 5964 && ! bitmap_bit_p (&live_regs, regno)) 5965 remove_p = true; 5966 /* See which defined values die here. */ 5967 for (reg = curr_id->regs; reg != NULL; reg = reg->next) 5968 if (reg->type == OP_OUT && ! reg->subreg_p) 5969 bitmap_clear_bit (&live_regs, reg->regno); 5970 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next) 5971 if (reg->type == OP_OUT && ! reg->subreg_p) 5972 bitmap_clear_bit (&live_regs, reg->regno); 5973 if (curr_id->arg_hard_regs != NULL) 5974 /* Make clobbered argument hard registers die. */ 5975 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++) 5976 if (regno >= FIRST_PSEUDO_REGISTER) 5977 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER); 5978 /* Mark each used value as live. */ 5979 for (reg = curr_id->regs; reg != NULL; reg = reg->next) 5980 if (reg->type != OP_OUT 5981 && bitmap_bit_p (&check_only_regs, reg->regno)) 5982 bitmap_set_bit (&live_regs, reg->regno); 5983 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next) 5984 if (reg->type != OP_OUT 5985 && bitmap_bit_p (&check_only_regs, reg->regno)) 5986 bitmap_set_bit (&live_regs, reg->regno); 5987 if (curr_id->arg_hard_regs != NULL) 5988 /* Make used argument hard registers live. */ 5989 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++) 5990 if (regno < FIRST_PSEUDO_REGISTER 5991 && bitmap_bit_p (&check_only_regs, regno)) 5992 bitmap_set_bit (&live_regs, regno); 5993 /* It is quite important to remove dead move insns because it 5994 means removing dead store. We don't need to process them for 5995 constraints. */ 5996 if (remove_p) 5997 { 5998 if (lra_dump_file != NULL) 5999 { 6000 fprintf (lra_dump_file, " Removing dead insn:\n "); 6001 dump_insn_slim (lra_dump_file, curr_insn); 6002 } 6003 lra_set_insn_deleted (curr_insn); 6004 } 6005 } 6006 } 6007 6008 /* The structure describes info to do an inheritance for the current 6009 insn. We need to collect such info first before doing the 6010 transformations because the transformations change the insn 6011 internal representation. */ 6012 struct to_inherit 6013 { 6014 /* Original regno. */ 6015 int regno; 6016 /* Subsequent insns which can inherit original reg value. */ 6017 rtx insns; 6018 }; 6019 6020 /* Array containing all info for doing inheritance from the current 6021 insn. */ 6022 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS]; 6023 6024 /* Number elements in the previous array. */ 6025 static int to_inherit_num; 6026 6027 /* Add inheritance info REGNO and INSNS. Their meaning is described in 6028 structure to_inherit. */ 6029 static void 6030 add_to_inherit (int regno, rtx insns) 6031 { 6032 int i; 6033 6034 for (i = 0; i < to_inherit_num; i++) 6035 if (to_inherit[i].regno == regno) 6036 return; 6037 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS); 6038 to_inherit[to_inherit_num].regno = regno; 6039 to_inherit[to_inherit_num++].insns = insns; 6040 } 6041 6042 /* Return the last non-debug insn in basic block BB, or the block begin 6043 note if none. */ 6044 static rtx_insn * 6045 get_last_insertion_point (basic_block bb) 6046 { 6047 rtx_insn *insn; 6048 6049 FOR_BB_INSNS_REVERSE (bb, insn) 6050 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn)) 6051 return insn; 6052 gcc_unreachable (); 6053 } 6054 6055 /* Set up RES by registers living on edges FROM except the edge (FROM, 6056 TO) or by registers set up in a jump insn in BB FROM. */ 6057 static void 6058 get_live_on_other_edges (basic_block from, basic_block to, bitmap res) 6059 { 6060 rtx_insn *last; 6061 struct lra_insn_reg *reg; 6062 edge e; 6063 edge_iterator ei; 6064 6065 lra_assert (to != NULL); 6066 bitmap_clear (res); 6067 FOR_EACH_EDGE (e, ei, from->succs) 6068 if (e->dest != to) 6069 bitmap_ior_into (res, df_get_live_in (e->dest)); 6070 last = get_last_insertion_point (from); 6071 if (! JUMP_P (last)) 6072 return; 6073 curr_id = lra_get_insn_recog_data (last); 6074 for (reg = curr_id->regs; reg != NULL; reg = reg->next) 6075 if (reg->type != OP_IN) 6076 bitmap_set_bit (res, reg->regno); 6077 } 6078 6079 /* Used as a temporary results of some bitmap calculations. */ 6080 static bitmap_head temp_bitmap; 6081 6082 /* We split for reloads of small class of hard regs. The following 6083 defines how many hard regs the class should have to be qualified as 6084 small. The code is mostly oriented to x86/x86-64 architecture 6085 where some insns need to use only specific register or pair of 6086 registers and these register can live in RTL explicitly, e.g. for 6087 parameter passing. */ 6088 static const int max_small_class_regs_num = 2; 6089 6090 /* Do inheritance/split transformations in EBB starting with HEAD and 6091 finishing on TAIL. We process EBB insns in the reverse order. 6092 Return true if we did any inheritance/split transformation in the 6093 EBB. 6094 6095 We should avoid excessive splitting which results in worse code 6096 because of inaccurate cost calculations for spilling new split 6097 pseudos in such case. To achieve this we do splitting only if 6098 register pressure is high in given basic block and there are reload 6099 pseudos requiring hard registers. We could do more register 6100 pressure calculations at any given program point to avoid necessary 6101 splitting even more but it is to expensive and the current approach 6102 works well enough. */ 6103 static bool 6104 inherit_in_ebb (rtx_insn *head, rtx_insn *tail) 6105 { 6106 int i, src_regno, dst_regno, nregs; 6107 bool change_p, succ_p, update_reloads_num_p; 6108 rtx_insn *prev_insn, *last_insn; 6109 rtx next_usage_insns, curr_set; 6110 enum reg_class cl; 6111 struct lra_insn_reg *reg; 6112 basic_block last_processed_bb, curr_bb = NULL; 6113 HARD_REG_SET potential_reload_hard_regs, live_hard_regs; 6114 bitmap to_process; 6115 unsigned int j; 6116 bitmap_iterator bi; 6117 bool head_p, after_p; 6118 6119 change_p = false; 6120 curr_usage_insns_check++; 6121 clear_invariants (); 6122 reloads_num = calls_num = 0; 6123 bitmap_clear (&check_only_regs); 6124 bitmap_clear (&invalid_invariant_regs); 6125 last_processed_bb = NULL; 6126 CLEAR_HARD_REG_SET (potential_reload_hard_regs); 6127 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset); 6128 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs); 6129 /* We don't process new insns generated in the loop. */ 6130 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn) 6131 { 6132 prev_insn = PREV_INSN (curr_insn); 6133 if (BLOCK_FOR_INSN (curr_insn) != NULL) 6134 curr_bb = BLOCK_FOR_INSN (curr_insn); 6135 if (last_processed_bb != curr_bb) 6136 { 6137 /* We are at the end of BB. Add qualified living 6138 pseudos for potential splitting. */ 6139 to_process = df_get_live_out (curr_bb); 6140 if (last_processed_bb != NULL) 6141 { 6142 /* We are somewhere in the middle of EBB. */ 6143 get_live_on_other_edges (curr_bb, last_processed_bb, 6144 &temp_bitmap); 6145 to_process = &temp_bitmap; 6146 } 6147 last_processed_bb = curr_bb; 6148 last_insn = get_last_insertion_point (curr_bb); 6149 after_p = (! JUMP_P (last_insn) 6150 && (! CALL_P (last_insn) 6151 || (find_reg_note (last_insn, 6152 REG_NORETURN, NULL_RTX) == NULL_RTX 6153 && ! SIBLING_CALL_P (last_insn)))); 6154 CLEAR_HARD_REG_SET (potential_reload_hard_regs); 6155 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi) 6156 { 6157 if ((int) j >= lra_constraint_new_regno_start) 6158 break; 6159 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0) 6160 { 6161 if (j < FIRST_PSEUDO_REGISTER) 6162 SET_HARD_REG_BIT (live_hard_regs, j); 6163 else 6164 add_to_hard_reg_set (&live_hard_regs, 6165 PSEUDO_REGNO_MODE (j), 6166 reg_renumber[j]); 6167 setup_next_usage_insn (j, last_insn, reloads_num, after_p); 6168 } 6169 } 6170 } 6171 src_regno = dst_regno = -1; 6172 curr_set = single_set (curr_insn); 6173 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set))) 6174 dst_regno = REGNO (SET_DEST (curr_set)); 6175 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set))) 6176 src_regno = REGNO (SET_SRC (curr_set)); 6177 update_reloads_num_p = true; 6178 if (src_regno < lra_constraint_new_regno_start 6179 && src_regno >= FIRST_PSEUDO_REGISTER 6180 && reg_renumber[src_regno] < 0 6181 && dst_regno >= lra_constraint_new_regno_start 6182 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS) 6183 { 6184 /* 'reload_pseudo <- original_pseudo'. */ 6185 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num) 6186 reloads_num++; 6187 update_reloads_num_p = false; 6188 succ_p = false; 6189 if (usage_insns[src_regno].check == curr_usage_insns_check 6190 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX) 6191 succ_p = inherit_reload_reg (false, src_regno, cl, 6192 curr_insn, next_usage_insns); 6193 if (succ_p) 6194 change_p = true; 6195 else 6196 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false); 6197 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs)) 6198 IOR_HARD_REG_SET (potential_reload_hard_regs, 6199 reg_class_contents[cl]); 6200 } 6201 else if (src_regno < 0 6202 && dst_regno >= lra_constraint_new_regno_start 6203 && invariant_p (SET_SRC (curr_set)) 6204 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS 6205 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno) 6206 && ! bitmap_bit_p (&invalid_invariant_regs, 6207 ORIGINAL_REGNO(regno_reg_rtx[dst_regno]))) 6208 { 6209 /* 'reload_pseudo <- invariant'. */ 6210 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num) 6211 reloads_num++; 6212 update_reloads_num_p = false; 6213 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set))) 6214 change_p = true; 6215 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs)) 6216 IOR_HARD_REG_SET (potential_reload_hard_regs, 6217 reg_class_contents[cl]); 6218 } 6219 else if (src_regno >= lra_constraint_new_regno_start 6220 && dst_regno < lra_constraint_new_regno_start 6221 && dst_regno >= FIRST_PSEUDO_REGISTER 6222 && reg_renumber[dst_regno] < 0 6223 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS 6224 && usage_insns[dst_regno].check == curr_usage_insns_check 6225 && (next_usage_insns 6226 = usage_insns[dst_regno].insns) != NULL_RTX) 6227 { 6228 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num) 6229 reloads_num++; 6230 update_reloads_num_p = false; 6231 /* 'original_pseudo <- reload_pseudo'. */ 6232 if (! JUMP_P (curr_insn) 6233 && inherit_reload_reg (true, dst_regno, cl, 6234 curr_insn, next_usage_insns)) 6235 change_p = true; 6236 /* Invalidate. */ 6237 usage_insns[dst_regno].check = 0; 6238 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs)) 6239 IOR_HARD_REG_SET (potential_reload_hard_regs, 6240 reg_class_contents[cl]); 6241 } 6242 else if (INSN_P (curr_insn)) 6243 { 6244 int iter; 6245 int max_uid = get_max_uid (); 6246 6247 curr_id = lra_get_insn_recog_data (curr_insn); 6248 curr_static_id = curr_id->insn_static_data; 6249 to_inherit_num = 0; 6250 /* Process insn definitions. */ 6251 for (iter = 0; iter < 2; iter++) 6252 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs; 6253 reg != NULL; 6254 reg = reg->next) 6255 if (reg->type != OP_IN 6256 && (dst_regno = reg->regno) < lra_constraint_new_regno_start) 6257 { 6258 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT 6259 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p 6260 && usage_insns[dst_regno].check == curr_usage_insns_check 6261 && (next_usage_insns 6262 = usage_insns[dst_regno].insns) != NULL_RTX) 6263 { 6264 struct lra_insn_reg *r; 6265 6266 for (r = curr_id->regs; r != NULL; r = r->next) 6267 if (r->type != OP_OUT && r->regno == dst_regno) 6268 break; 6269 /* Don't do inheritance if the pseudo is also 6270 used in the insn. */ 6271 if (r == NULL) 6272 /* We can not do inheritance right now 6273 because the current insn reg info (chain 6274 regs) can change after that. */ 6275 add_to_inherit (dst_regno, next_usage_insns); 6276 } 6277 /* We can not process one reg twice here because of 6278 usage_insns invalidation. */ 6279 if ((dst_regno < FIRST_PSEUDO_REGISTER 6280 || reg_renumber[dst_regno] >= 0) 6281 && ! reg->subreg_p && reg->type != OP_IN) 6282 { 6283 HARD_REG_SET s; 6284 6285 if (split_if_necessary (dst_regno, reg->biggest_mode, 6286 potential_reload_hard_regs, 6287 false, curr_insn, max_uid)) 6288 change_p = true; 6289 CLEAR_HARD_REG_SET (s); 6290 if (dst_regno < FIRST_PSEUDO_REGISTER) 6291 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno); 6292 else 6293 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno), 6294 reg_renumber[dst_regno]); 6295 AND_COMPL_HARD_REG_SET (live_hard_regs, s); 6296 } 6297 /* We should invalidate potential inheritance or 6298 splitting for the current insn usages to the next 6299 usage insns (see code below) as the output pseudo 6300 prevents this. */ 6301 if ((dst_regno >= FIRST_PSEUDO_REGISTER 6302 && reg_renumber[dst_regno] < 0) 6303 || (reg->type == OP_OUT && ! reg->subreg_p 6304 && (dst_regno < FIRST_PSEUDO_REGISTER 6305 || reg_renumber[dst_regno] >= 0))) 6306 { 6307 /* Invalidate and mark definitions. */ 6308 if (dst_regno >= FIRST_PSEUDO_REGISTER) 6309 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn); 6310 else 6311 { 6312 nregs = hard_regno_nregs (dst_regno, 6313 reg->biggest_mode); 6314 for (i = 0; i < nregs; i++) 6315 usage_insns[dst_regno + i].check 6316 = -(int) INSN_UID (curr_insn); 6317 } 6318 } 6319 } 6320 /* Process clobbered call regs. */ 6321 if (curr_id->arg_hard_regs != NULL) 6322 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++) 6323 if (dst_regno >= FIRST_PSEUDO_REGISTER) 6324 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check 6325 = -(int) INSN_UID (curr_insn); 6326 if (! JUMP_P (curr_insn)) 6327 for (i = 0; i < to_inherit_num; i++) 6328 if (inherit_reload_reg (true, to_inherit[i].regno, 6329 ALL_REGS, curr_insn, 6330 to_inherit[i].insns)) 6331 change_p = true; 6332 if (CALL_P (curr_insn)) 6333 { 6334 rtx cheap, pat, dest; 6335 rtx_insn *restore; 6336 int regno, hard_regno; 6337 6338 calls_num++; 6339 if ((cheap = find_reg_note (curr_insn, 6340 REG_RETURNED, NULL_RTX)) != NULL_RTX 6341 && ((cheap = XEXP (cheap, 0)), true) 6342 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER 6343 && (hard_regno = reg_renumber[regno]) >= 0 6344 && usage_insns[regno].check == curr_usage_insns_check 6345 /* If there are pending saves/restores, the 6346 optimization is not worth. */ 6347 && usage_insns[regno].calls_num == calls_num - 1 6348 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno)) 6349 { 6350 /* Restore the pseudo from the call result as 6351 REG_RETURNED note says that the pseudo value is 6352 in the call result and the pseudo is an argument 6353 of the call. */ 6354 pat = PATTERN (curr_insn); 6355 if (GET_CODE (pat) == PARALLEL) 6356 pat = XVECEXP (pat, 0, 0); 6357 dest = SET_DEST (pat); 6358 /* For multiple return values dest is PARALLEL. 6359 Currently we handle only single return value case. */ 6360 if (REG_P (dest)) 6361 { 6362 start_sequence (); 6363 emit_move_insn (cheap, copy_rtx (dest)); 6364 restore = get_insns (); 6365 end_sequence (); 6366 lra_process_new_insns (curr_insn, NULL, restore, 6367 "Inserting call parameter restore"); 6368 /* We don't need to save/restore of the pseudo from 6369 this call. */ 6370 usage_insns[regno].calls_num = calls_num; 6371 bitmap_set_bit (&check_only_regs, regno); 6372 } 6373 } 6374 } 6375 to_inherit_num = 0; 6376 /* Process insn usages. */ 6377 for (iter = 0; iter < 2; iter++) 6378 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs; 6379 reg != NULL; 6380 reg = reg->next) 6381 if ((reg->type != OP_OUT 6382 || (reg->type == OP_OUT && reg->subreg_p)) 6383 && (src_regno = reg->regno) < lra_constraint_new_regno_start) 6384 { 6385 if (src_regno >= FIRST_PSEUDO_REGISTER 6386 && reg_renumber[src_regno] < 0 && reg->type == OP_IN) 6387 { 6388 if (usage_insns[src_regno].check == curr_usage_insns_check 6389 && (next_usage_insns 6390 = usage_insns[src_regno].insns) != NULL_RTX 6391 && NONDEBUG_INSN_P (curr_insn)) 6392 add_to_inherit (src_regno, next_usage_insns); 6393 else if (usage_insns[src_regno].check 6394 != -(int) INSN_UID (curr_insn)) 6395 /* Add usages but only if the reg is not set up 6396 in the same insn. */ 6397 add_next_usage_insn (src_regno, curr_insn, reloads_num); 6398 } 6399 else if (src_regno < FIRST_PSEUDO_REGISTER 6400 || reg_renumber[src_regno] >= 0) 6401 { 6402 bool before_p; 6403 rtx_insn *use_insn = curr_insn; 6404 6405 before_p = (JUMP_P (curr_insn) 6406 || (CALL_P (curr_insn) && reg->type == OP_IN)); 6407 if (NONDEBUG_INSN_P (curr_insn) 6408 && (! JUMP_P (curr_insn) || reg->type == OP_IN) 6409 && split_if_necessary (src_regno, reg->biggest_mode, 6410 potential_reload_hard_regs, 6411 before_p, curr_insn, max_uid)) 6412 { 6413 if (reg->subreg_p) 6414 lra_risky_transformations_p = true; 6415 change_p = true; 6416 /* Invalidate. */ 6417 usage_insns[src_regno].check = 0; 6418 if (before_p) 6419 use_insn = PREV_INSN (curr_insn); 6420 } 6421 if (NONDEBUG_INSN_P (curr_insn)) 6422 { 6423 if (src_regno < FIRST_PSEUDO_REGISTER) 6424 add_to_hard_reg_set (&live_hard_regs, 6425 reg->biggest_mode, src_regno); 6426 else 6427 add_to_hard_reg_set (&live_hard_regs, 6428 PSEUDO_REGNO_MODE (src_regno), 6429 reg_renumber[src_regno]); 6430 } 6431 if (src_regno >= FIRST_PSEUDO_REGISTER) 6432 add_next_usage_insn (src_regno, use_insn, reloads_num); 6433 else 6434 { 6435 for (i = 0; i < hard_regno_nregs (src_regno, reg->biggest_mode); i++) 6436 add_next_usage_insn (src_regno + i, use_insn, reloads_num); 6437 } 6438 } 6439 } 6440 /* Process used call regs. */ 6441 if (curr_id->arg_hard_regs != NULL) 6442 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++) 6443 if (src_regno < FIRST_PSEUDO_REGISTER) 6444 { 6445 SET_HARD_REG_BIT (live_hard_regs, src_regno); 6446 add_next_usage_insn (src_regno, curr_insn, reloads_num); 6447 } 6448 for (i = 0; i < to_inherit_num; i++) 6449 { 6450 src_regno = to_inherit[i].regno; 6451 if (inherit_reload_reg (false, src_regno, ALL_REGS, 6452 curr_insn, to_inherit[i].insns)) 6453 change_p = true; 6454 else 6455 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false); 6456 } 6457 } 6458 if (update_reloads_num_p 6459 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX) 6460 { 6461 int regno = -1; 6462 if ((REG_P (SET_DEST (curr_set)) 6463 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start 6464 && reg_renumber[regno] < 0 6465 && (cl = lra_get_allocno_class (regno)) != NO_REGS) 6466 || (REG_P (SET_SRC (curr_set)) 6467 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start 6468 && reg_renumber[regno] < 0 6469 && (cl = lra_get_allocno_class (regno)) != NO_REGS)) 6470 { 6471 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num) 6472 reloads_num++; 6473 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs)) 6474 IOR_HARD_REG_SET (potential_reload_hard_regs, 6475 reg_class_contents[cl]); 6476 } 6477 } 6478 if (NONDEBUG_INSN_P (curr_insn)) 6479 { 6480 int regno; 6481 6482 /* Invalidate invariants with changed regs. */ 6483 curr_id = lra_get_insn_recog_data (curr_insn); 6484 for (reg = curr_id->regs; reg != NULL; reg = reg->next) 6485 if (reg->type != OP_IN) 6486 { 6487 bitmap_set_bit (&invalid_invariant_regs, reg->regno); 6488 bitmap_set_bit (&invalid_invariant_regs, 6489 ORIGINAL_REGNO (regno_reg_rtx[reg->regno])); 6490 } 6491 curr_static_id = curr_id->insn_static_data; 6492 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next) 6493 if (reg->type != OP_IN) 6494 bitmap_set_bit (&invalid_invariant_regs, reg->regno); 6495 if (curr_id->arg_hard_regs != NULL) 6496 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++) 6497 if (regno >= FIRST_PSEUDO_REGISTER) 6498 bitmap_set_bit (&invalid_invariant_regs, 6499 regno - FIRST_PSEUDO_REGISTER); 6500 } 6501 /* We reached the start of the current basic block. */ 6502 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head) 6503 || BLOCK_FOR_INSN (prev_insn) != curr_bb) 6504 { 6505 /* We reached the beginning of the current block -- do 6506 rest of spliting in the current BB. */ 6507 to_process = df_get_live_in (curr_bb); 6508 if (BLOCK_FOR_INSN (head) != curr_bb) 6509 { 6510 /* We are somewhere in the middle of EBB. */ 6511 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src, 6512 curr_bb, &temp_bitmap); 6513 to_process = &temp_bitmap; 6514 } 6515 head_p = true; 6516 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi) 6517 { 6518 if ((int) j >= lra_constraint_new_regno_start) 6519 break; 6520 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0) 6521 && usage_insns[j].check == curr_usage_insns_check 6522 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX) 6523 { 6524 if (need_for_split_p (potential_reload_hard_regs, j)) 6525 { 6526 if (lra_dump_file != NULL && head_p) 6527 { 6528 fprintf (lra_dump_file, 6529 " ----------------------------------\n"); 6530 head_p = false; 6531 } 6532 if (split_reg (false, j, bb_note (curr_bb), 6533 next_usage_insns, NULL)) 6534 change_p = true; 6535 } 6536 usage_insns[j].check = 0; 6537 } 6538 } 6539 } 6540 } 6541 return change_p; 6542 } 6543 6544 /* This value affects EBB forming. If probability of edge from EBB to 6545 a BB is not greater than the following value, we don't add the BB 6546 to EBB. */ 6547 #define EBB_PROBABILITY_CUTOFF \ 6548 ((REG_BR_PROB_BASE * LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF) / 100) 6549 6550 /* Current number of inheritance/split iteration. */ 6551 int lra_inheritance_iter; 6552 6553 /* Entry function for inheritance/split pass. */ 6554 void 6555 lra_inheritance (void) 6556 { 6557 int i; 6558 basic_block bb, start_bb; 6559 edge e; 6560 6561 lra_inheritance_iter++; 6562 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES) 6563 return; 6564 timevar_push (TV_LRA_INHERITANCE); 6565 if (lra_dump_file != NULL) 6566 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n", 6567 lra_inheritance_iter); 6568 curr_usage_insns_check = 0; 6569 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start); 6570 for (i = 0; i < lra_constraint_new_regno_start; i++) 6571 usage_insns[i].check = 0; 6572 bitmap_initialize (&check_only_regs, ®_obstack); 6573 bitmap_initialize (&invalid_invariant_regs, ®_obstack); 6574 bitmap_initialize (&live_regs, ®_obstack); 6575 bitmap_initialize (&temp_bitmap, ®_obstack); 6576 bitmap_initialize (&ebb_global_regs, ®_obstack); 6577 FOR_EACH_BB_FN (bb, cfun) 6578 { 6579 start_bb = bb; 6580 if (lra_dump_file != NULL) 6581 fprintf (lra_dump_file, "EBB"); 6582 /* Form a EBB starting with BB. */ 6583 bitmap_clear (&ebb_global_regs); 6584 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb)); 6585 for (;;) 6586 { 6587 if (lra_dump_file != NULL) 6588 fprintf (lra_dump_file, " %d", bb->index); 6589 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun) 6590 || LABEL_P (BB_HEAD (bb->next_bb))) 6591 break; 6592 e = find_fallthru_edge (bb->succs); 6593 if (! e) 6594 break; 6595 if (e->probability.initialized_p () 6596 && e->probability.to_reg_br_prob_base () < EBB_PROBABILITY_CUTOFF) 6597 break; 6598 bb = bb->next_bb; 6599 } 6600 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb)); 6601 if (lra_dump_file != NULL) 6602 fprintf (lra_dump_file, "\n"); 6603 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb))) 6604 /* Remember that the EBB head and tail can change in 6605 inherit_in_ebb. */ 6606 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb)); 6607 } 6608 bitmap_clear (&ebb_global_regs); 6609 bitmap_clear (&temp_bitmap); 6610 bitmap_clear (&live_regs); 6611 bitmap_clear (&invalid_invariant_regs); 6612 bitmap_clear (&check_only_regs); 6613 free (usage_insns); 6614 6615 timevar_pop (TV_LRA_INHERITANCE); 6616 } 6617 6618 6619 6620 /* This page contains code to undo failed inheritance/split 6621 transformations. */ 6622 6623 /* Current number of iteration undoing inheritance/split. */ 6624 int lra_undo_inheritance_iter; 6625 6626 /* Fix BB live info LIVE after removing pseudos created on pass doing 6627 inheritance/split which are REMOVED_PSEUDOS. */ 6628 static void 6629 fix_bb_live_info (bitmap live, bitmap removed_pseudos) 6630 { 6631 unsigned int regno; 6632 bitmap_iterator bi; 6633 6634 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi) 6635 if (bitmap_clear_bit (live, regno) 6636 && REG_P (lra_reg_info[regno].restore_rtx)) 6637 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx)); 6638 } 6639 6640 /* Return regno of the (subreg of) REG. Otherwise, return a negative 6641 number. */ 6642 static int 6643 get_regno (rtx reg) 6644 { 6645 if (GET_CODE (reg) == SUBREG) 6646 reg = SUBREG_REG (reg); 6647 if (REG_P (reg)) 6648 return REGNO (reg); 6649 return -1; 6650 } 6651 6652 /* Delete a move INSN with destination reg DREGNO and a previous 6653 clobber insn with the same regno. The inheritance/split code can 6654 generate moves with preceding clobber and when we delete such moves 6655 we should delete the clobber insn too to keep the correct life 6656 info. */ 6657 static void 6658 delete_move_and_clobber (rtx_insn *insn, int dregno) 6659 { 6660 rtx_insn *prev_insn = PREV_INSN (insn); 6661 6662 lra_set_insn_deleted (insn); 6663 lra_assert (dregno >= 0); 6664 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn) 6665 && GET_CODE (PATTERN (prev_insn)) == CLOBBER 6666 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0))) 6667 lra_set_insn_deleted (prev_insn); 6668 } 6669 6670 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and 6671 return true if we did any change. The undo transformations for 6672 inheritance looks like 6673 i <- i2 6674 p <- i => p <- i2 6675 or removing 6676 p <- i, i <- p, and i <- i3 6677 where p is original pseudo from which inheritance pseudo i was 6678 created, i and i3 are removed inheritance pseudos, i2 is another 6679 not removed inheritance pseudo. All split pseudos or other 6680 occurrences of removed inheritance pseudos are changed on the 6681 corresponding original pseudos. 6682 6683 The function also schedules insns changed and created during 6684 inheritance/split pass for processing by the subsequent constraint 6685 pass. */ 6686 static bool 6687 remove_inheritance_pseudos (bitmap remove_pseudos) 6688 { 6689 basic_block bb; 6690 int regno, sregno, prev_sregno, dregno; 6691 rtx restore_rtx; 6692 rtx set, prev_set; 6693 rtx_insn *prev_insn; 6694 bool change_p, done_p; 6695 6696 change_p = ! bitmap_empty_p (remove_pseudos); 6697 /* We can not finish the function right away if CHANGE_P is true 6698 because we need to marks insns affected by previous 6699 inheritance/split pass for processing by the subsequent 6700 constraint pass. */ 6701 FOR_EACH_BB_FN (bb, cfun) 6702 { 6703 fix_bb_live_info (df_get_live_in (bb), remove_pseudos); 6704 fix_bb_live_info (df_get_live_out (bb), remove_pseudos); 6705 FOR_BB_INSNS_REVERSE (bb, curr_insn) 6706 { 6707 if (! INSN_P (curr_insn)) 6708 continue; 6709 done_p = false; 6710 sregno = dregno = -1; 6711 if (change_p && NONDEBUG_INSN_P (curr_insn) 6712 && (set = single_set (curr_insn)) != NULL_RTX) 6713 { 6714 dregno = get_regno (SET_DEST (set)); 6715 sregno = get_regno (SET_SRC (set)); 6716 } 6717 6718 if (sregno >= 0 && dregno >= 0) 6719 { 6720 if (bitmap_bit_p (remove_pseudos, dregno) 6721 && ! REG_P (lra_reg_info[dregno].restore_rtx)) 6722 { 6723 /* invariant inheritance pseudo <- original pseudo */ 6724 if (lra_dump_file != NULL) 6725 { 6726 fprintf (lra_dump_file, " Removing invariant inheritance:\n"); 6727 dump_insn_slim (lra_dump_file, curr_insn); 6728 fprintf (lra_dump_file, "\n"); 6729 } 6730 delete_move_and_clobber (curr_insn, dregno); 6731 done_p = true; 6732 } 6733 else if (bitmap_bit_p (remove_pseudos, sregno) 6734 && ! REG_P (lra_reg_info[sregno].restore_rtx)) 6735 { 6736 /* reload pseudo <- invariant inheritance pseudo */ 6737 start_sequence (); 6738 /* We can not just change the source. It might be 6739 an insn different from the move. */ 6740 emit_insn (lra_reg_info[sregno].restore_rtx); 6741 rtx_insn *new_insns = get_insns (); 6742 end_sequence (); 6743 lra_assert (single_set (new_insns) != NULL 6744 && SET_DEST (set) == SET_DEST (single_set (new_insns))); 6745 lra_process_new_insns (curr_insn, NULL, new_insns, 6746 "Changing reload<-invariant inheritance"); 6747 delete_move_and_clobber (curr_insn, dregno); 6748 done_p = true; 6749 } 6750 else if ((bitmap_bit_p (remove_pseudos, sregno) 6751 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno 6752 || (bitmap_bit_p (remove_pseudos, dregno) 6753 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0 6754 && (get_regno (lra_reg_info[sregno].restore_rtx) 6755 == get_regno (lra_reg_info[dregno].restore_rtx))))) 6756 || (bitmap_bit_p (remove_pseudos, dregno) 6757 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno)) 6758 /* One of the following cases: 6759 original <- removed inheritance pseudo 6760 removed inherit pseudo <- another removed inherit pseudo 6761 removed inherit pseudo <- original pseudo 6762 Or 6763 removed_split_pseudo <- original_reg 6764 original_reg <- removed_split_pseudo */ 6765 { 6766 if (lra_dump_file != NULL) 6767 { 6768 fprintf (lra_dump_file, " Removing %s:\n", 6769 bitmap_bit_p (&lra_split_regs, sregno) 6770 || bitmap_bit_p (&lra_split_regs, dregno) 6771 ? "split" : "inheritance"); 6772 dump_insn_slim (lra_dump_file, curr_insn); 6773 } 6774 delete_move_and_clobber (curr_insn, dregno); 6775 done_p = true; 6776 } 6777 else if (bitmap_bit_p (remove_pseudos, sregno) 6778 && bitmap_bit_p (&lra_inheritance_pseudos, sregno)) 6779 { 6780 /* Search the following pattern: 6781 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2 6782 original_pseudo <- inherit_or_split_pseudo1 6783 where the 2nd insn is the current insn and 6784 inherit_or_split_pseudo2 is not removed. If it is found, 6785 change the current insn onto: 6786 original_pseudo <- inherit_or_split_pseudo2. */ 6787 for (prev_insn = PREV_INSN (curr_insn); 6788 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn); 6789 prev_insn = PREV_INSN (prev_insn)) 6790 ; 6791 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb 6792 && (prev_set = single_set (prev_insn)) != NULL_RTX 6793 /* There should be no subregs in insn we are 6794 searching because only the original reg might 6795 be in subreg when we changed the mode of 6796 load/store for splitting. */ 6797 && REG_P (SET_DEST (prev_set)) 6798 && REG_P (SET_SRC (prev_set)) 6799 && (int) REGNO (SET_DEST (prev_set)) == sregno 6800 && ((prev_sregno = REGNO (SET_SRC (prev_set))) 6801 >= FIRST_PSEUDO_REGISTER) 6802 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX 6803 || 6804 /* As we consider chain of inheritance or 6805 splitting described in above comment we should 6806 check that sregno and prev_sregno were 6807 inheritance/split pseudos created from the 6808 same original regno. */ 6809 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0 6810 && (get_regno (lra_reg_info[sregno].restore_rtx) 6811 == get_regno (lra_reg_info[prev_sregno].restore_rtx)))) 6812 && ! bitmap_bit_p (remove_pseudos, prev_sregno)) 6813 { 6814 lra_assert (GET_MODE (SET_SRC (prev_set)) 6815 == GET_MODE (regno_reg_rtx[sregno])); 6816 /* Although we have a single set, the insn can 6817 contain more one sregno register occurrence 6818 as a source. Change all occurrences. */ 6819 lra_substitute_pseudo_within_insn (curr_insn, sregno, 6820 SET_SRC (prev_set), 6821 false); 6822 /* As we are finishing with processing the insn 6823 here, check the destination too as it might 6824 inheritance pseudo for another pseudo. */ 6825 if (bitmap_bit_p (remove_pseudos, dregno) 6826 && bitmap_bit_p (&lra_inheritance_pseudos, dregno) 6827 && (restore_rtx 6828 = lra_reg_info[dregno].restore_rtx) != NULL_RTX) 6829 { 6830 if (GET_CODE (SET_DEST (set)) == SUBREG) 6831 SUBREG_REG (SET_DEST (set)) = restore_rtx; 6832 else 6833 SET_DEST (set) = restore_rtx; 6834 } 6835 lra_push_insn_and_update_insn_regno_info (curr_insn); 6836 lra_set_used_insn_alternative_by_uid 6837 (INSN_UID (curr_insn), LRA_UNKNOWN_ALT); 6838 done_p = true; 6839 if (lra_dump_file != NULL) 6840 { 6841 fprintf (lra_dump_file, " Change reload insn:\n"); 6842 dump_insn_slim (lra_dump_file, curr_insn); 6843 } 6844 } 6845 } 6846 } 6847 if (! done_p) 6848 { 6849 struct lra_insn_reg *reg; 6850 bool restored_regs_p = false; 6851 bool kept_regs_p = false; 6852 6853 curr_id = lra_get_insn_recog_data (curr_insn); 6854 for (reg = curr_id->regs; reg != NULL; reg = reg->next) 6855 { 6856 regno = reg->regno; 6857 restore_rtx = lra_reg_info[regno].restore_rtx; 6858 if (restore_rtx != NULL_RTX) 6859 { 6860 if (change_p && bitmap_bit_p (remove_pseudos, regno)) 6861 { 6862 lra_substitute_pseudo_within_insn 6863 (curr_insn, regno, restore_rtx, false); 6864 restored_regs_p = true; 6865 } 6866 else 6867 kept_regs_p = true; 6868 } 6869 } 6870 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p) 6871 { 6872 /* The instruction has changed since the previous 6873 constraints pass. */ 6874 lra_push_insn_and_update_insn_regno_info (curr_insn); 6875 lra_set_used_insn_alternative_by_uid 6876 (INSN_UID (curr_insn), LRA_UNKNOWN_ALT); 6877 } 6878 else if (restored_regs_p) 6879 /* The instruction has been restored to the form that 6880 it had during the previous constraints pass. */ 6881 lra_update_insn_regno_info (curr_insn); 6882 if (restored_regs_p && lra_dump_file != NULL) 6883 { 6884 fprintf (lra_dump_file, " Insn after restoring regs:\n"); 6885 dump_insn_slim (lra_dump_file, curr_insn); 6886 } 6887 } 6888 } 6889 } 6890 return change_p; 6891 } 6892 6893 /* If optional reload pseudos failed to get a hard register or was not 6894 inherited, it is better to remove optional reloads. We do this 6895 transformation after undoing inheritance to figure out necessity to 6896 remove optional reloads easier. Return true if we do any 6897 change. */ 6898 static bool 6899 undo_optional_reloads (void) 6900 { 6901 bool change_p, keep_p; 6902 unsigned int regno, uid; 6903 bitmap_iterator bi, bi2; 6904 rtx_insn *insn; 6905 rtx set, src, dest; 6906 auto_bitmap removed_optional_reload_pseudos (®_obstack); 6907 6908 bitmap_copy (removed_optional_reload_pseudos, &lra_optional_reload_pseudos); 6909 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi) 6910 { 6911 keep_p = false; 6912 /* Keep optional reloads from previous subpasses. */ 6913 if (lra_reg_info[regno].restore_rtx == NULL_RTX 6914 /* If the original pseudo changed its allocation, just 6915 removing the optional pseudo is dangerous as the original 6916 pseudo will have longer live range. */ 6917 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0) 6918 keep_p = true; 6919 else if (reg_renumber[regno] >= 0) 6920 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2) 6921 { 6922 insn = lra_insn_recog_data[uid]->insn; 6923 if ((set = single_set (insn)) == NULL_RTX) 6924 continue; 6925 src = SET_SRC (set); 6926 dest = SET_DEST (set); 6927 if (! REG_P (src) || ! REG_P (dest)) 6928 continue; 6929 if (REGNO (dest) == regno 6930 /* Ignore insn for optional reloads itself. */ 6931 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src) 6932 /* Check only inheritance on last inheritance pass. */ 6933 && (int) REGNO (src) >= new_regno_start 6934 /* Check that the optional reload was inherited. */ 6935 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src))) 6936 { 6937 keep_p = true; 6938 break; 6939 } 6940 } 6941 if (keep_p) 6942 { 6943 bitmap_clear_bit (removed_optional_reload_pseudos, regno); 6944 if (lra_dump_file != NULL) 6945 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno); 6946 } 6947 } 6948 change_p = ! bitmap_empty_p (removed_optional_reload_pseudos); 6949 auto_bitmap insn_bitmap (®_obstack); 6950 EXECUTE_IF_SET_IN_BITMAP (removed_optional_reload_pseudos, 0, regno, bi) 6951 { 6952 if (lra_dump_file != NULL) 6953 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno); 6954 bitmap_copy (insn_bitmap, &lra_reg_info[regno].insn_bitmap); 6955 EXECUTE_IF_SET_IN_BITMAP (insn_bitmap, 0, uid, bi2) 6956 { 6957 insn = lra_insn_recog_data[uid]->insn; 6958 if ((set = single_set (insn)) != NULL_RTX) 6959 { 6960 src = SET_SRC (set); 6961 dest = SET_DEST (set); 6962 if (REG_P (src) && REG_P (dest) 6963 && ((REGNO (src) == regno 6964 && (REGNO (lra_reg_info[regno].restore_rtx) 6965 == REGNO (dest))) 6966 || (REGNO (dest) == regno 6967 && (REGNO (lra_reg_info[regno].restore_rtx) 6968 == REGNO (src))))) 6969 { 6970 if (lra_dump_file != NULL) 6971 { 6972 fprintf (lra_dump_file, " Deleting move %u\n", 6973 INSN_UID (insn)); 6974 dump_insn_slim (lra_dump_file, insn); 6975 } 6976 delete_move_and_clobber (insn, REGNO (dest)); 6977 continue; 6978 } 6979 /* We should not worry about generation memory-memory 6980 moves here as if the corresponding inheritance did 6981 not work (inheritance pseudo did not get a hard reg), 6982 we remove the inheritance pseudo and the optional 6983 reload. */ 6984 } 6985 lra_substitute_pseudo_within_insn 6986 (insn, regno, lra_reg_info[regno].restore_rtx, false); 6987 lra_update_insn_regno_info (insn); 6988 if (lra_dump_file != NULL) 6989 { 6990 fprintf (lra_dump_file, 6991 " Restoring original insn:\n"); 6992 dump_insn_slim (lra_dump_file, insn); 6993 } 6994 } 6995 } 6996 /* Clear restore_regnos. */ 6997 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi) 6998 lra_reg_info[regno].restore_rtx = NULL_RTX; 6999 return change_p; 7000 } 7001 7002 /* Entry function for undoing inheritance/split transformation. Return true 7003 if we did any RTL change in this pass. */ 7004 bool 7005 lra_undo_inheritance (void) 7006 { 7007 unsigned int regno; 7008 int hard_regno; 7009 int n_all_inherit, n_inherit, n_all_split, n_split; 7010 rtx restore_rtx; 7011 bitmap_iterator bi; 7012 bool change_p; 7013 7014 lra_undo_inheritance_iter++; 7015 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES) 7016 return false; 7017 if (lra_dump_file != NULL) 7018 fprintf (lra_dump_file, 7019 "\n********** Undoing inheritance #%d: **********\n\n", 7020 lra_undo_inheritance_iter); 7021 auto_bitmap remove_pseudos (®_obstack); 7022 n_inherit = n_all_inherit = 0; 7023 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi) 7024 if (lra_reg_info[regno].restore_rtx != NULL_RTX) 7025 { 7026 n_all_inherit++; 7027 if (reg_renumber[regno] < 0 7028 /* If the original pseudo changed its allocation, just 7029 removing inheritance is dangerous as for changing 7030 allocation we used shorter live-ranges. */ 7031 && (! REG_P (lra_reg_info[regno].restore_rtx) 7032 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0)) 7033 bitmap_set_bit (remove_pseudos, regno); 7034 else 7035 n_inherit++; 7036 } 7037 if (lra_dump_file != NULL && n_all_inherit != 0) 7038 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n", 7039 n_inherit, n_all_inherit, 7040 (double) n_inherit / n_all_inherit * 100); 7041 n_split = n_all_split = 0; 7042 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi) 7043 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX) 7044 { 7045 int restore_regno = REGNO (restore_rtx); 7046 7047 n_all_split++; 7048 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER 7049 ? reg_renumber[restore_regno] : restore_regno); 7050 if (hard_regno < 0 || reg_renumber[regno] == hard_regno) 7051 bitmap_set_bit (remove_pseudos, regno); 7052 else 7053 { 7054 n_split++; 7055 if (lra_dump_file != NULL) 7056 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n", 7057 regno, restore_regno); 7058 } 7059 } 7060 if (lra_dump_file != NULL && n_all_split != 0) 7061 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n", 7062 n_split, n_all_split, 7063 (double) n_split / n_all_split * 100); 7064 change_p = remove_inheritance_pseudos (remove_pseudos); 7065 /* Clear restore_regnos. */ 7066 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi) 7067 lra_reg_info[regno].restore_rtx = NULL_RTX; 7068 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi) 7069 lra_reg_info[regno].restore_rtx = NULL_RTX; 7070 change_p = undo_optional_reloads () || change_p; 7071 return change_p; 7072 } 7073