1 /* Instruction scheduling pass. Selective scheduler and pipeliner. 2 Copyright (C) 2006-2019 Free Software Foundation, Inc. 3 4 This file is part of GCC. 5 6 GCC is free software; you can redistribute it and/or modify it under 7 the terms of the GNU General Public License as published by the Free 8 Software Foundation; either version 3, or (at your option) any later 9 version. 10 11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY 12 WARRANTY; without even the implied warranty of MERCHANTABILITY or 13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 14 for more details. 15 16 You should have received a copy of the GNU General Public License 17 along with GCC; see the file COPYING3. If not see 18 <http://www.gnu.org/licenses/>. */ 19 20 #include "config.h" 21 #include "system.h" 22 #include "coretypes.h" 23 #include "backend.h" 24 #include "cfghooks.h" 25 #include "tree.h" 26 #include "rtl.h" 27 #include "df.h" 28 #include "memmodel.h" 29 #include "tm_p.h" 30 #include "cfgrtl.h" 31 #include "cfganal.h" 32 #include "cfgbuild.h" 33 #include "insn-config.h" 34 #include "insn-attr.h" 35 #include "recog.h" 36 #include "params.h" 37 #include "target.h" 38 #include "sched-int.h" 39 #include "emit-rtl.h" /* FIXME: Can go away once crtl is moved to rtl.h. */ 40 41 #ifdef INSN_SCHEDULING 42 #include "regset.h" 43 #include "cfgloop.h" 44 #include "sel-sched-ir.h" 45 /* We don't have to use it except for sel_print_insn. */ 46 #include "sel-sched-dump.h" 47 48 /* A vector holding bb info for whole scheduling pass. */ 49 vec<sel_global_bb_info_def> sel_global_bb_info; 50 51 /* A vector holding bb info. */ 52 vec<sel_region_bb_info_def> sel_region_bb_info; 53 54 /* A pool for allocating all lists. */ 55 object_allocator<_list_node> sched_lists_pool ("sel-sched-lists"); 56 57 /* This contains information about successors for compute_av_set. */ 58 struct succs_info current_succs; 59 60 /* Data structure to describe interaction with the generic scheduler utils. */ 61 static struct common_sched_info_def sel_common_sched_info; 62 63 /* The loop nest being pipelined. */ 64 struct loop *current_loop_nest; 65 66 /* LOOP_NESTS is a vector containing the corresponding loop nest for 67 each region. */ 68 static vec<loop_p> loop_nests; 69 70 /* Saves blocks already in loop regions, indexed by bb->index. */ 71 static sbitmap bbs_in_loop_rgns = NULL; 72 73 /* CFG hooks that are saved before changing create_basic_block hook. */ 74 static struct cfg_hooks orig_cfg_hooks; 75 76 77 /* Array containing reverse topological index of function basic blocks, 78 indexed by BB->INDEX. */ 79 static int *rev_top_order_index = NULL; 80 81 /* Length of the above array. */ 82 static int rev_top_order_index_len = -1; 83 84 /* A regset pool structure. */ 85 static struct 86 { 87 /* The stack to which regsets are returned. */ 88 regset *v; 89 90 /* Its pointer. */ 91 int n; 92 93 /* Its size. */ 94 int s; 95 96 /* In VV we save all generated regsets so that, when destructing the 97 pool, we can compare it with V and check that every regset was returned 98 back to pool. */ 99 regset *vv; 100 101 /* The pointer of VV stack. */ 102 int nn; 103 104 /* Its size. */ 105 int ss; 106 107 /* The difference between allocated and returned regsets. */ 108 int diff; 109 } regset_pool = { NULL, 0, 0, NULL, 0, 0, 0 }; 110 111 /* This represents the nop pool. */ 112 static struct 113 { 114 /* The vector which holds previously emitted nops. */ 115 insn_t *v; 116 117 /* Its pointer. */ 118 int n; 119 120 /* Its size. */ 121 int s; 122 } nop_pool = { NULL, 0, 0 }; 123 124 /* The pool for basic block notes. */ 125 static vec<rtx_note *> bb_note_pool; 126 127 /* A NOP pattern used to emit placeholder insns. */ 128 rtx nop_pattern = NULL_RTX; 129 /* A special instruction that resides in EXIT_BLOCK. 130 EXIT_INSN is successor of the insns that lead to EXIT_BLOCK. */ 131 rtx_insn *exit_insn = NULL; 132 133 /* TRUE if while scheduling current region, which is loop, its preheader 134 was removed. */ 135 bool preheader_removed = false; 136 137 138 /* Forward static declarations. */ 139 static void fence_clear (fence_t); 140 141 static void deps_init_id (idata_t, insn_t, bool); 142 static void init_id_from_df (idata_t, insn_t, bool); 143 static expr_t set_insn_init (expr_t, vinsn_t, int); 144 145 static void cfg_preds (basic_block, insn_t **, int *); 146 static void prepare_insn_expr (insn_t, int); 147 static void free_history_vect (vec<expr_history_def> &); 148 149 static void move_bb_info (basic_block, basic_block); 150 static void remove_empty_bb (basic_block, bool); 151 static void sel_merge_blocks (basic_block, basic_block); 152 static void sel_remove_loop_preheader (void); 153 static bool bb_has_removable_jump_to_p (basic_block, basic_block); 154 155 static bool insn_is_the_only_one_in_bb_p (insn_t); 156 static void create_initial_data_sets (basic_block); 157 158 static void free_av_set (basic_block); 159 static void invalidate_av_set (basic_block); 160 static void extend_insn_data (void); 161 static void sel_init_new_insn (insn_t, int, int = -1); 162 static void finish_insns (void); 163 164 /* Various list functions. */ 165 166 /* Copy an instruction list L. */ 167 ilist_t 168 ilist_copy (ilist_t l) 169 { 170 ilist_t head = NULL, *tailp = &head; 171 172 while (l) 173 { 174 ilist_add (tailp, ILIST_INSN (l)); 175 tailp = &ILIST_NEXT (*tailp); 176 l = ILIST_NEXT (l); 177 } 178 179 return head; 180 } 181 182 /* Invert an instruction list L. */ 183 ilist_t 184 ilist_invert (ilist_t l) 185 { 186 ilist_t res = NULL; 187 188 while (l) 189 { 190 ilist_add (&res, ILIST_INSN (l)); 191 l = ILIST_NEXT (l); 192 } 193 194 return res; 195 } 196 197 /* Add a new boundary to the LP list with parameters TO, PTR, and DC. */ 198 void 199 blist_add (blist_t *lp, insn_t to, ilist_t ptr, deps_t dc) 200 { 201 bnd_t bnd; 202 203 _list_add (lp); 204 bnd = BLIST_BND (*lp); 205 206 BND_TO (bnd) = to; 207 BND_PTR (bnd) = ptr; 208 BND_AV (bnd) = NULL; 209 BND_AV1 (bnd) = NULL; 210 BND_DC (bnd) = dc; 211 } 212 213 /* Remove the list note pointed to by LP. */ 214 void 215 blist_remove (blist_t *lp) 216 { 217 bnd_t b = BLIST_BND (*lp); 218 219 av_set_clear (&BND_AV (b)); 220 av_set_clear (&BND_AV1 (b)); 221 ilist_clear (&BND_PTR (b)); 222 223 _list_remove (lp); 224 } 225 226 /* Init a fence tail L. */ 227 void 228 flist_tail_init (flist_tail_t l) 229 { 230 FLIST_TAIL_HEAD (l) = NULL; 231 FLIST_TAIL_TAILP (l) = &FLIST_TAIL_HEAD (l); 232 } 233 234 /* Try to find fence corresponding to INSN in L. */ 235 fence_t 236 flist_lookup (flist_t l, insn_t insn) 237 { 238 while (l) 239 { 240 if (FENCE_INSN (FLIST_FENCE (l)) == insn) 241 return FLIST_FENCE (l); 242 243 l = FLIST_NEXT (l); 244 } 245 246 return NULL; 247 } 248 249 /* Init the fields of F before running fill_insns. */ 250 static void 251 init_fence_for_scheduling (fence_t f) 252 { 253 FENCE_BNDS (f) = NULL; 254 FENCE_PROCESSED_P (f) = false; 255 FENCE_SCHEDULED_P (f) = false; 256 } 257 258 /* Add new fence consisting of INSN and STATE to the list pointed to by LP. */ 259 static void 260 flist_add (flist_t *lp, insn_t insn, state_t state, deps_t dc, void *tc, 261 insn_t last_scheduled_insn, vec<rtx_insn *, va_gc> *executing_insns, 262 int *ready_ticks, int ready_ticks_size, insn_t sched_next, 263 int cycle, int cycle_issued_insns, int issue_more, 264 bool starts_cycle_p, bool after_stall_p) 265 { 266 fence_t f; 267 268 _list_add (lp); 269 f = FLIST_FENCE (*lp); 270 271 FENCE_INSN (f) = insn; 272 273 gcc_assert (state != NULL); 274 FENCE_STATE (f) = state; 275 276 FENCE_CYCLE (f) = cycle; 277 FENCE_ISSUED_INSNS (f) = cycle_issued_insns; 278 FENCE_STARTS_CYCLE_P (f) = starts_cycle_p; 279 FENCE_AFTER_STALL_P (f) = after_stall_p; 280 281 gcc_assert (dc != NULL); 282 FENCE_DC (f) = dc; 283 284 gcc_assert (tc != NULL || targetm.sched.alloc_sched_context == NULL); 285 FENCE_TC (f) = tc; 286 287 FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn; 288 FENCE_ISSUE_MORE (f) = issue_more; 289 FENCE_EXECUTING_INSNS (f) = executing_insns; 290 FENCE_READY_TICKS (f) = ready_ticks; 291 FENCE_READY_TICKS_SIZE (f) = ready_ticks_size; 292 FENCE_SCHED_NEXT (f) = sched_next; 293 294 init_fence_for_scheduling (f); 295 } 296 297 /* Remove the head node of the list pointed to by LP. */ 298 static void 299 flist_remove (flist_t *lp) 300 { 301 if (FENCE_INSN (FLIST_FENCE (*lp))) 302 fence_clear (FLIST_FENCE (*lp)); 303 _list_remove (lp); 304 } 305 306 /* Clear the fence list pointed to by LP. */ 307 void 308 flist_clear (flist_t *lp) 309 { 310 while (*lp) 311 flist_remove (lp); 312 } 313 314 /* Add ORIGINAL_INSN the def list DL honoring CROSSES_CALL. */ 315 void 316 def_list_add (def_list_t *dl, insn_t original_insn, bool crosses_call) 317 { 318 def_t d; 319 320 _list_add (dl); 321 d = DEF_LIST_DEF (*dl); 322 323 d->orig_insn = original_insn; 324 d->crosses_call = crosses_call; 325 } 326 327 328 /* Functions to work with target contexts. */ 329 330 /* Bulk target context. It is convenient for debugging purposes to ensure 331 that there are no uninitialized (null) target contexts. */ 332 static tc_t bulk_tc = (tc_t) 1; 333 334 /* Target hooks wrappers. In the future we can provide some default 335 implementations for them. */ 336 337 /* Allocate a store for the target context. */ 338 static tc_t 339 alloc_target_context (void) 340 { 341 return (targetm.sched.alloc_sched_context 342 ? targetm.sched.alloc_sched_context () : bulk_tc); 343 } 344 345 /* Init target context TC. 346 If CLEAN_P is true, then make TC as it is beginning of the scheduler. 347 Overwise, copy current backend context to TC. */ 348 static void 349 init_target_context (tc_t tc, bool clean_p) 350 { 351 if (targetm.sched.init_sched_context) 352 targetm.sched.init_sched_context (tc, clean_p); 353 } 354 355 /* Allocate and initialize a target context. Meaning of CLEAN_P is the same as 356 int init_target_context (). */ 357 tc_t 358 create_target_context (bool clean_p) 359 { 360 tc_t tc = alloc_target_context (); 361 362 init_target_context (tc, clean_p); 363 return tc; 364 } 365 366 /* Copy TC to the current backend context. */ 367 void 368 set_target_context (tc_t tc) 369 { 370 if (targetm.sched.set_sched_context) 371 targetm.sched.set_sched_context (tc); 372 } 373 374 /* TC is about to be destroyed. Free any internal data. */ 375 static void 376 clear_target_context (tc_t tc) 377 { 378 if (targetm.sched.clear_sched_context) 379 targetm.sched.clear_sched_context (tc); 380 } 381 382 /* Clear and free it. */ 383 static void 384 delete_target_context (tc_t tc) 385 { 386 clear_target_context (tc); 387 388 if (targetm.sched.free_sched_context) 389 targetm.sched.free_sched_context (tc); 390 } 391 392 /* Make a copy of FROM in TO. 393 NB: May be this should be a hook. */ 394 static void 395 copy_target_context (tc_t to, tc_t from) 396 { 397 tc_t tmp = create_target_context (false); 398 399 set_target_context (from); 400 init_target_context (to, false); 401 402 set_target_context (tmp); 403 delete_target_context (tmp); 404 } 405 406 /* Create a copy of TC. */ 407 static tc_t 408 create_copy_of_target_context (tc_t tc) 409 { 410 tc_t copy = alloc_target_context (); 411 412 copy_target_context (copy, tc); 413 414 return copy; 415 } 416 417 /* Clear TC and initialize it according to CLEAN_P. The meaning of CLEAN_P 418 is the same as in init_target_context (). */ 419 void 420 reset_target_context (tc_t tc, bool clean_p) 421 { 422 clear_target_context (tc); 423 init_target_context (tc, clean_p); 424 } 425 426 /* Functions to work with dependence contexts. 427 Dc (aka deps context, aka deps_t, aka struct deps_desc *) is short for dependence 428 context. It accumulates information about processed insns to decide if 429 current insn is dependent on the processed ones. */ 430 431 /* Make a copy of FROM in TO. */ 432 static void 433 copy_deps_context (deps_t to, deps_t from) 434 { 435 init_deps (to, false); 436 deps_join (to, from); 437 } 438 439 /* Allocate store for dep context. */ 440 static deps_t 441 alloc_deps_context (void) 442 { 443 return XNEW (struct deps_desc); 444 } 445 446 /* Allocate and initialize dep context. */ 447 static deps_t 448 create_deps_context (void) 449 { 450 deps_t dc = alloc_deps_context (); 451 452 init_deps (dc, false); 453 return dc; 454 } 455 456 /* Create a copy of FROM. */ 457 static deps_t 458 create_copy_of_deps_context (deps_t from) 459 { 460 deps_t to = alloc_deps_context (); 461 462 copy_deps_context (to, from); 463 return to; 464 } 465 466 /* Clean up internal data of DC. */ 467 static void 468 clear_deps_context (deps_t dc) 469 { 470 free_deps (dc); 471 } 472 473 /* Clear and free DC. */ 474 static void 475 delete_deps_context (deps_t dc) 476 { 477 clear_deps_context (dc); 478 free (dc); 479 } 480 481 /* Clear and init DC. */ 482 static void 483 reset_deps_context (deps_t dc) 484 { 485 clear_deps_context (dc); 486 init_deps (dc, false); 487 } 488 489 /* This structure describes the dependence analysis hooks for advancing 490 dependence context. */ 491 static struct sched_deps_info_def advance_deps_context_sched_deps_info = 492 { 493 NULL, 494 495 NULL, /* start_insn */ 496 NULL, /* finish_insn */ 497 NULL, /* start_lhs */ 498 NULL, /* finish_lhs */ 499 NULL, /* start_rhs */ 500 NULL, /* finish_rhs */ 501 haifa_note_reg_set, 502 haifa_note_reg_clobber, 503 haifa_note_reg_use, 504 NULL, /* note_mem_dep */ 505 NULL, /* note_dep */ 506 507 0, 0, 0 508 }; 509 510 /* Process INSN and add its impact on DC. */ 511 void 512 advance_deps_context (deps_t dc, insn_t insn) 513 { 514 sched_deps_info = &advance_deps_context_sched_deps_info; 515 deps_analyze_insn (dc, insn); 516 } 517 518 519 /* Functions to work with DFA states. */ 520 521 /* Allocate store for a DFA state. */ 522 static state_t 523 state_alloc (void) 524 { 525 return xmalloc (dfa_state_size); 526 } 527 528 /* Allocate and initialize DFA state. */ 529 static state_t 530 state_create (void) 531 { 532 state_t state = state_alloc (); 533 534 state_reset (state); 535 advance_state (state); 536 return state; 537 } 538 539 /* Free DFA state. */ 540 static void 541 state_free (state_t state) 542 { 543 free (state); 544 } 545 546 /* Make a copy of FROM in TO. */ 547 static void 548 state_copy (state_t to, state_t from) 549 { 550 memcpy (to, from, dfa_state_size); 551 } 552 553 /* Create a copy of FROM. */ 554 static state_t 555 state_create_copy (state_t from) 556 { 557 state_t to = state_alloc (); 558 559 state_copy (to, from); 560 return to; 561 } 562 563 564 /* Functions to work with fences. */ 565 566 /* Clear the fence. */ 567 static void 568 fence_clear (fence_t f) 569 { 570 state_t s = FENCE_STATE (f); 571 deps_t dc = FENCE_DC (f); 572 void *tc = FENCE_TC (f); 573 574 ilist_clear (&FENCE_BNDS (f)); 575 576 gcc_assert ((s != NULL && dc != NULL && tc != NULL) 577 || (s == NULL && dc == NULL && tc == NULL)); 578 579 free (s); 580 581 if (dc != NULL) 582 delete_deps_context (dc); 583 584 if (tc != NULL) 585 delete_target_context (tc); 586 vec_free (FENCE_EXECUTING_INSNS (f)); 587 free (FENCE_READY_TICKS (f)); 588 FENCE_READY_TICKS (f) = NULL; 589 } 590 591 /* Init a list of fences with successors of OLD_FENCE. */ 592 void 593 init_fences (insn_t old_fence) 594 { 595 insn_t succ; 596 succ_iterator si; 597 bool first = true; 598 int ready_ticks_size = get_max_uid () + 1; 599 600 FOR_EACH_SUCC_1 (succ, si, old_fence, 601 SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) 602 { 603 604 if (first) 605 first = false; 606 else 607 gcc_assert (flag_sel_sched_pipelining_outer_loops); 608 609 flist_add (&fences, succ, 610 state_create (), 611 create_deps_context () /* dc */, 612 create_target_context (true) /* tc */, 613 NULL /* last_scheduled_insn */, 614 NULL, /* executing_insns */ 615 XCNEWVEC (int, ready_ticks_size), /* ready_ticks */ 616 ready_ticks_size, 617 NULL /* sched_next */, 618 1 /* cycle */, 0 /* cycle_issued_insns */, 619 issue_rate, /* issue_more */ 620 1 /* starts_cycle_p */, 0 /* after_stall_p */); 621 } 622 } 623 624 /* Merges two fences (filling fields of fence F with resulting values) by 625 following rules: 1) state, target context and last scheduled insn are 626 propagated from fallthrough edge if it is available; 627 2) deps context and cycle is propagated from more probable edge; 628 3) all other fields are set to corresponding constant values. 629 630 INSN, STATE, DC, TC, LAST_SCHEDULED_INSN, EXECUTING_INSNS, 631 READY_TICKS, READY_TICKS_SIZE, SCHED_NEXT, CYCLE, ISSUE_MORE 632 and AFTER_STALL_P are the corresponding fields of the second fence. */ 633 static void 634 merge_fences (fence_t f, insn_t insn, 635 state_t state, deps_t dc, void *tc, 636 rtx_insn *last_scheduled_insn, 637 vec<rtx_insn *, va_gc> *executing_insns, 638 int *ready_ticks, int ready_ticks_size, 639 rtx sched_next, int cycle, int issue_more, bool after_stall_p) 640 { 641 insn_t last_scheduled_insn_old = FENCE_LAST_SCHEDULED_INSN (f); 642 643 gcc_assert (sel_bb_head_p (FENCE_INSN (f)) 644 && !sched_next && !FENCE_SCHED_NEXT (f)); 645 646 /* Check if we can decide which path fences came. 647 If we can't (or don't want to) - reset all. */ 648 if (last_scheduled_insn == NULL 649 || last_scheduled_insn_old == NULL 650 /* This is a case when INSN is reachable on several paths from 651 one insn (this can happen when pipelining of outer loops is on and 652 there are two edges: one going around of inner loop and the other - 653 right through it; in such case just reset everything). */ 654 || last_scheduled_insn == last_scheduled_insn_old) 655 { 656 state_reset (FENCE_STATE (f)); 657 state_free (state); 658 659 reset_deps_context (FENCE_DC (f)); 660 delete_deps_context (dc); 661 662 reset_target_context (FENCE_TC (f), true); 663 delete_target_context (tc); 664 665 if (cycle > FENCE_CYCLE (f)) 666 FENCE_CYCLE (f) = cycle; 667 668 FENCE_LAST_SCHEDULED_INSN (f) = NULL; 669 FENCE_ISSUE_MORE (f) = issue_rate; 670 vec_free (executing_insns); 671 free (ready_ticks); 672 if (FENCE_EXECUTING_INSNS (f)) 673 FENCE_EXECUTING_INSNS (f)->block_remove (0, 674 FENCE_EXECUTING_INSNS (f)->length ()); 675 if (FENCE_READY_TICKS (f)) 676 memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f)); 677 } 678 else 679 { 680 edge edge_old = NULL, edge_new = NULL; 681 edge candidate; 682 succ_iterator si; 683 insn_t succ; 684 685 /* Find fallthrough edge. */ 686 gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb); 687 candidate = find_fallthru_edge_from (BLOCK_FOR_INSN (insn)->prev_bb); 688 689 if (!candidate 690 || (candidate->src != BLOCK_FOR_INSN (last_scheduled_insn) 691 && candidate->src != BLOCK_FOR_INSN (last_scheduled_insn_old))) 692 { 693 /* No fallthrough edge leading to basic block of INSN. */ 694 state_reset (FENCE_STATE (f)); 695 state_free (state); 696 697 reset_target_context (FENCE_TC (f), true); 698 delete_target_context (tc); 699 700 FENCE_LAST_SCHEDULED_INSN (f) = NULL; 701 FENCE_ISSUE_MORE (f) = issue_rate; 702 } 703 else 704 if (candidate->src == BLOCK_FOR_INSN (last_scheduled_insn)) 705 { 706 state_free (FENCE_STATE (f)); 707 FENCE_STATE (f) = state; 708 709 delete_target_context (FENCE_TC (f)); 710 FENCE_TC (f) = tc; 711 712 FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn; 713 FENCE_ISSUE_MORE (f) = issue_more; 714 } 715 else 716 { 717 /* Leave STATE, TC and LAST_SCHEDULED_INSN fields untouched. */ 718 state_free (state); 719 delete_target_context (tc); 720 721 gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb 722 != BLOCK_FOR_INSN (last_scheduled_insn)); 723 } 724 725 /* Find edge of first predecessor (last_scheduled_insn_old->insn). */ 726 FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn_old, 727 SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) 728 { 729 if (succ == insn) 730 { 731 /* No same successor allowed from several edges. */ 732 gcc_assert (!edge_old); 733 edge_old = si.e1; 734 } 735 } 736 /* Find edge of second predecessor (last_scheduled_insn->insn). */ 737 FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn, 738 SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) 739 { 740 if (succ == insn) 741 { 742 /* No same successor allowed from several edges. */ 743 gcc_assert (!edge_new); 744 edge_new = si.e1; 745 } 746 } 747 748 /* Check if we can choose most probable predecessor. */ 749 if (edge_old == NULL || edge_new == NULL) 750 { 751 reset_deps_context (FENCE_DC (f)); 752 delete_deps_context (dc); 753 vec_free (executing_insns); 754 free (ready_ticks); 755 756 FENCE_CYCLE (f) = MAX (FENCE_CYCLE (f), cycle); 757 if (FENCE_EXECUTING_INSNS (f)) 758 FENCE_EXECUTING_INSNS (f)->block_remove (0, 759 FENCE_EXECUTING_INSNS (f)->length ()); 760 if (FENCE_READY_TICKS (f)) 761 memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f)); 762 } 763 else 764 if (edge_new->probability > edge_old->probability) 765 { 766 delete_deps_context (FENCE_DC (f)); 767 FENCE_DC (f) = dc; 768 vec_free (FENCE_EXECUTING_INSNS (f)); 769 FENCE_EXECUTING_INSNS (f) = executing_insns; 770 free (FENCE_READY_TICKS (f)); 771 FENCE_READY_TICKS (f) = ready_ticks; 772 FENCE_READY_TICKS_SIZE (f) = ready_ticks_size; 773 FENCE_CYCLE (f) = cycle; 774 } 775 else 776 { 777 /* Leave DC and CYCLE untouched. */ 778 delete_deps_context (dc); 779 vec_free (executing_insns); 780 free (ready_ticks); 781 } 782 } 783 784 /* Fill remaining invariant fields. */ 785 if (after_stall_p) 786 FENCE_AFTER_STALL_P (f) = 1; 787 788 FENCE_ISSUED_INSNS (f) = 0; 789 FENCE_STARTS_CYCLE_P (f) = 1; 790 FENCE_SCHED_NEXT (f) = NULL; 791 } 792 793 /* Add a new fence to NEW_FENCES list, initializing it from all 794 other parameters. */ 795 static void 796 add_to_fences (flist_tail_t new_fences, insn_t insn, 797 state_t state, deps_t dc, void *tc, 798 rtx_insn *last_scheduled_insn, 799 vec<rtx_insn *, va_gc> *executing_insns, int *ready_ticks, 800 int ready_ticks_size, rtx_insn *sched_next, int cycle, 801 int cycle_issued_insns, int issue_rate, 802 bool starts_cycle_p, bool after_stall_p) 803 { 804 fence_t f = flist_lookup (FLIST_TAIL_HEAD (new_fences), insn); 805 806 if (! f) 807 { 808 flist_add (FLIST_TAIL_TAILP (new_fences), insn, state, dc, tc, 809 last_scheduled_insn, executing_insns, ready_ticks, 810 ready_ticks_size, sched_next, cycle, cycle_issued_insns, 811 issue_rate, starts_cycle_p, after_stall_p); 812 813 FLIST_TAIL_TAILP (new_fences) 814 = &FLIST_NEXT (*FLIST_TAIL_TAILP (new_fences)); 815 } 816 else 817 { 818 merge_fences (f, insn, state, dc, tc, last_scheduled_insn, 819 executing_insns, ready_ticks, ready_ticks_size, 820 sched_next, cycle, issue_rate, after_stall_p); 821 } 822 } 823 824 /* Move the first fence in the OLD_FENCES list to NEW_FENCES. */ 825 void 826 move_fence_to_fences (flist_t old_fences, flist_tail_t new_fences) 827 { 828 fence_t f, old; 829 flist_t *tailp = FLIST_TAIL_TAILP (new_fences); 830 831 old = FLIST_FENCE (old_fences); 832 f = flist_lookup (FLIST_TAIL_HEAD (new_fences), 833 FENCE_INSN (FLIST_FENCE (old_fences))); 834 if (f) 835 { 836 merge_fences (f, old->insn, old->state, old->dc, old->tc, 837 old->last_scheduled_insn, old->executing_insns, 838 old->ready_ticks, old->ready_ticks_size, 839 old->sched_next, old->cycle, old->issue_more, 840 old->after_stall_p); 841 } 842 else 843 { 844 _list_add (tailp); 845 FLIST_TAIL_TAILP (new_fences) = &FLIST_NEXT (*tailp); 846 *FLIST_FENCE (*tailp) = *old; 847 init_fence_for_scheduling (FLIST_FENCE (*tailp)); 848 } 849 FENCE_INSN (old) = NULL; 850 } 851 852 /* Add a new fence to NEW_FENCES list and initialize most of its data 853 as a clean one. */ 854 void 855 add_clean_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence) 856 { 857 int ready_ticks_size = get_max_uid () + 1; 858 859 add_to_fences (new_fences, 860 succ, state_create (), create_deps_context (), 861 create_target_context (true), 862 NULL, NULL, 863 XCNEWVEC (int, ready_ticks_size), ready_ticks_size, 864 NULL, FENCE_CYCLE (fence) + 1, 865 0, issue_rate, 1, FENCE_AFTER_STALL_P (fence)); 866 } 867 868 /* Add a new fence to NEW_FENCES list and initialize all of its data 869 from FENCE and SUCC. */ 870 void 871 add_dirty_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence) 872 { 873 int * new_ready_ticks 874 = XNEWVEC (int, FENCE_READY_TICKS_SIZE (fence)); 875 876 memcpy (new_ready_ticks, FENCE_READY_TICKS (fence), 877 FENCE_READY_TICKS_SIZE (fence) * sizeof (int)); 878 add_to_fences (new_fences, 879 succ, state_create_copy (FENCE_STATE (fence)), 880 create_copy_of_deps_context (FENCE_DC (fence)), 881 create_copy_of_target_context (FENCE_TC (fence)), 882 FENCE_LAST_SCHEDULED_INSN (fence), 883 vec_safe_copy (FENCE_EXECUTING_INSNS (fence)), 884 new_ready_ticks, 885 FENCE_READY_TICKS_SIZE (fence), 886 FENCE_SCHED_NEXT (fence), 887 FENCE_CYCLE (fence), 888 FENCE_ISSUED_INSNS (fence), 889 FENCE_ISSUE_MORE (fence), 890 FENCE_STARTS_CYCLE_P (fence), 891 FENCE_AFTER_STALL_P (fence)); 892 } 893 894 895 /* Functions to work with regset and nop pools. */ 896 897 /* Returns the new regset from pool. It might have some of the bits set 898 from the previous usage. */ 899 regset 900 get_regset_from_pool (void) 901 { 902 regset rs; 903 904 if (regset_pool.n != 0) 905 rs = regset_pool.v[--regset_pool.n]; 906 else 907 /* We need to create the regset. */ 908 { 909 rs = ALLOC_REG_SET (®_obstack); 910 911 if (regset_pool.nn == regset_pool.ss) 912 regset_pool.vv = XRESIZEVEC (regset, regset_pool.vv, 913 (regset_pool.ss = 2 * regset_pool.ss + 1)); 914 regset_pool.vv[regset_pool.nn++] = rs; 915 } 916 917 regset_pool.diff++; 918 919 return rs; 920 } 921 922 /* Same as above, but returns the empty regset. */ 923 regset 924 get_clear_regset_from_pool (void) 925 { 926 regset rs = get_regset_from_pool (); 927 928 CLEAR_REG_SET (rs); 929 return rs; 930 } 931 932 /* Return regset RS to the pool for future use. */ 933 void 934 return_regset_to_pool (regset rs) 935 { 936 gcc_assert (rs); 937 regset_pool.diff--; 938 939 if (regset_pool.n == regset_pool.s) 940 regset_pool.v = XRESIZEVEC (regset, regset_pool.v, 941 (regset_pool.s = 2 * regset_pool.s + 1)); 942 regset_pool.v[regset_pool.n++] = rs; 943 } 944 945 /* This is used as a qsort callback for sorting regset pool stacks. 946 X and XX are addresses of two regsets. They are never equal. */ 947 static int 948 cmp_v_in_regset_pool (const void *x, const void *xx) 949 { 950 uintptr_t r1 = (uintptr_t) *((const regset *) x); 951 uintptr_t r2 = (uintptr_t) *((const regset *) xx); 952 if (r1 > r2) 953 return 1; 954 else if (r1 < r2) 955 return -1; 956 gcc_unreachable (); 957 } 958 959 /* Free the regset pool possibly checking for memory leaks. */ 960 void 961 free_regset_pool (void) 962 { 963 if (flag_checking) 964 { 965 regset *v = regset_pool.v; 966 int i = 0; 967 int n = regset_pool.n; 968 969 regset *vv = regset_pool.vv; 970 int ii = 0; 971 int nn = regset_pool.nn; 972 973 int diff = 0; 974 975 gcc_assert (n <= nn); 976 977 /* Sort both vectors so it will be possible to compare them. */ 978 qsort (v, n, sizeof (*v), cmp_v_in_regset_pool); 979 qsort (vv, nn, sizeof (*vv), cmp_v_in_regset_pool); 980 981 while (ii < nn) 982 { 983 if (v[i] == vv[ii]) 984 i++; 985 else 986 /* VV[II] was lost. */ 987 diff++; 988 989 ii++; 990 } 991 992 gcc_assert (diff == regset_pool.diff); 993 } 994 995 /* If not true - we have a memory leak. */ 996 gcc_assert (regset_pool.diff == 0); 997 998 while (regset_pool.n) 999 { 1000 --regset_pool.n; 1001 FREE_REG_SET (regset_pool.v[regset_pool.n]); 1002 } 1003 1004 free (regset_pool.v); 1005 regset_pool.v = NULL; 1006 regset_pool.s = 0; 1007 1008 free (regset_pool.vv); 1009 regset_pool.vv = NULL; 1010 regset_pool.nn = 0; 1011 regset_pool.ss = 0; 1012 1013 regset_pool.diff = 0; 1014 } 1015 1016 1017 /* Functions to work with nop pools. NOP insns are used as temporary 1018 placeholders of the insns being scheduled to allow correct update of 1019 the data sets. When update is finished, NOPs are deleted. */ 1020 1021 /* A vinsn that is used to represent a nop. This vinsn is shared among all 1022 nops sel-sched generates. */ 1023 static vinsn_t nop_vinsn = NULL; 1024 1025 /* Emit a nop before INSN, taking it from pool. */ 1026 insn_t 1027 get_nop_from_pool (insn_t insn) 1028 { 1029 rtx nop_pat; 1030 insn_t nop; 1031 bool old_p = nop_pool.n != 0; 1032 int flags; 1033 1034 if (old_p) 1035 nop_pat = nop_pool.v[--nop_pool.n]; 1036 else 1037 nop_pat = nop_pattern; 1038 1039 nop = emit_insn_before (nop_pat, insn); 1040 1041 if (old_p) 1042 flags = INSN_INIT_TODO_SSID; 1043 else 1044 flags = INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID; 1045 1046 set_insn_init (INSN_EXPR (insn), nop_vinsn, INSN_SEQNO (insn)); 1047 sel_init_new_insn (nop, flags); 1048 1049 return nop; 1050 } 1051 1052 /* Remove NOP from the instruction stream and return it to the pool. */ 1053 void 1054 return_nop_to_pool (insn_t nop, bool full_tidying) 1055 { 1056 gcc_assert (INSN_IN_STREAM_P (nop)); 1057 sel_remove_insn (nop, false, full_tidying); 1058 1059 /* We'll recycle this nop. */ 1060 nop->set_undeleted (); 1061 1062 if (nop_pool.n == nop_pool.s) 1063 nop_pool.v = XRESIZEVEC (rtx_insn *, nop_pool.v, 1064 (nop_pool.s = 2 * nop_pool.s + 1)); 1065 nop_pool.v[nop_pool.n++] = nop; 1066 } 1067 1068 /* Free the nop pool. */ 1069 void 1070 free_nop_pool (void) 1071 { 1072 nop_pool.n = 0; 1073 nop_pool.s = 0; 1074 free (nop_pool.v); 1075 nop_pool.v = NULL; 1076 } 1077 1078 1079 /* Skip unspec to support ia64 speculation. Called from rtx_equal_p_cb. 1080 The callback is given two rtxes XX and YY and writes the new rtxes 1081 to NX and NY in case some needs to be skipped. */ 1082 static int 1083 skip_unspecs_callback (const_rtx *xx, const_rtx *yy, rtx *nx, rtx* ny) 1084 { 1085 const_rtx x = *xx; 1086 const_rtx y = *yy; 1087 1088 if (GET_CODE (x) == UNSPEC 1089 && (targetm.sched.skip_rtx_p == NULL 1090 || targetm.sched.skip_rtx_p (x))) 1091 { 1092 *nx = XVECEXP (x, 0, 0); 1093 *ny = CONST_CAST_RTX (y); 1094 return 1; 1095 } 1096 1097 if (GET_CODE (y) == UNSPEC 1098 && (targetm.sched.skip_rtx_p == NULL 1099 || targetm.sched.skip_rtx_p (y))) 1100 { 1101 *nx = CONST_CAST_RTX (x); 1102 *ny = XVECEXP (y, 0, 0); 1103 return 1; 1104 } 1105 1106 return 0; 1107 } 1108 1109 /* Callback, called from hash_rtx_cb. Helps to hash UNSPEC rtx X in a correct way 1110 to support ia64 speculation. When changes are needed, new rtx X and new mode 1111 NMODE are written, and the callback returns true. */ 1112 static int 1113 hash_with_unspec_callback (const_rtx x, machine_mode mode ATTRIBUTE_UNUSED, 1114 rtx *nx, machine_mode* nmode) 1115 { 1116 if (GET_CODE (x) == UNSPEC 1117 && targetm.sched.skip_rtx_p 1118 && targetm.sched.skip_rtx_p (x)) 1119 { 1120 *nx = XVECEXP (x, 0 ,0); 1121 *nmode = VOIDmode; 1122 return 1; 1123 } 1124 1125 return 0; 1126 } 1127 1128 /* Returns LHS and RHS are ok to be scheduled separately. */ 1129 static bool 1130 lhs_and_rhs_separable_p (rtx lhs, rtx rhs) 1131 { 1132 if (lhs == NULL || rhs == NULL) 1133 return false; 1134 1135 /* Do not schedule constants as rhs: no point to use reg, if const 1136 can be used. Moreover, scheduling const as rhs may lead to mode 1137 mismatch cause consts don't have modes but they could be merged 1138 from branches where the same const used in different modes. */ 1139 if (CONSTANT_P (rhs)) 1140 return false; 1141 1142 /* ??? Do not rename predicate registers to avoid ICEs in bundling. */ 1143 if (COMPARISON_P (rhs)) 1144 return false; 1145 1146 /* Do not allow single REG to be an rhs. */ 1147 if (REG_P (rhs)) 1148 return false; 1149 1150 /* See comment at find_used_regs_1 (*1) for explanation of this 1151 restriction. */ 1152 /* FIXME: remove this later. */ 1153 if (MEM_P (lhs)) 1154 return false; 1155 1156 /* This will filter all tricky things like ZERO_EXTRACT etc. 1157 For now we don't handle it. */ 1158 if (!REG_P (lhs) && !MEM_P (lhs)) 1159 return false; 1160 1161 return true; 1162 } 1163 1164 /* Initialize vinsn VI for INSN. Only for use from vinsn_create (). When 1165 FORCE_UNIQUE_P is true, the resulting vinsn will not be clonable. This is 1166 used e.g. for insns from recovery blocks. */ 1167 static void 1168 vinsn_init (vinsn_t vi, insn_t insn, bool force_unique_p) 1169 { 1170 hash_rtx_callback_function hrcf; 1171 int insn_class; 1172 1173 VINSN_INSN_RTX (vi) = insn; 1174 VINSN_COUNT (vi) = 0; 1175 vi->cost = -1; 1176 1177 if (INSN_NOP_P (insn)) 1178 return; 1179 1180 if (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL) 1181 init_id_from_df (VINSN_ID (vi), insn, force_unique_p); 1182 else 1183 deps_init_id (VINSN_ID (vi), insn, force_unique_p); 1184 1185 /* Hash vinsn depending on whether it is separable or not. */ 1186 hrcf = targetm.sched.skip_rtx_p ? hash_with_unspec_callback : NULL; 1187 if (VINSN_SEPARABLE_P (vi)) 1188 { 1189 rtx rhs = VINSN_RHS (vi); 1190 1191 VINSN_HASH (vi) = hash_rtx_cb (rhs, GET_MODE (rhs), 1192 NULL, NULL, false, hrcf); 1193 VINSN_HASH_RTX (vi) = hash_rtx_cb (VINSN_PATTERN (vi), 1194 VOIDmode, NULL, NULL, 1195 false, hrcf); 1196 } 1197 else 1198 { 1199 VINSN_HASH (vi) = hash_rtx_cb (VINSN_PATTERN (vi), VOIDmode, 1200 NULL, NULL, false, hrcf); 1201 VINSN_HASH_RTX (vi) = VINSN_HASH (vi); 1202 } 1203 1204 insn_class = haifa_classify_insn (insn); 1205 if (insn_class >= 2 1206 && (!targetm.sched.get_insn_spec_ds 1207 || ((targetm.sched.get_insn_spec_ds (insn) & BEGIN_CONTROL) 1208 == 0))) 1209 VINSN_MAY_TRAP_P (vi) = true; 1210 else 1211 VINSN_MAY_TRAP_P (vi) = false; 1212 } 1213 1214 /* Indicate that VI has become the part of an rtx object. */ 1215 void 1216 vinsn_attach (vinsn_t vi) 1217 { 1218 /* Assert that VI is not pending for deletion. */ 1219 gcc_assert (VINSN_INSN_RTX (vi)); 1220 1221 VINSN_COUNT (vi)++; 1222 } 1223 1224 /* Create and init VI from the INSN. Use UNIQUE_P for determining the correct 1225 VINSN_TYPE (VI). */ 1226 static vinsn_t 1227 vinsn_create (insn_t insn, bool force_unique_p) 1228 { 1229 vinsn_t vi = XCNEW (struct vinsn_def); 1230 1231 vinsn_init (vi, insn, force_unique_p); 1232 return vi; 1233 } 1234 1235 /* Return a copy of VI. When REATTACH_P is true, detach VI and attach 1236 the copy. */ 1237 vinsn_t 1238 vinsn_copy (vinsn_t vi, bool reattach_p) 1239 { 1240 rtx_insn *copy; 1241 bool unique = VINSN_UNIQUE_P (vi); 1242 vinsn_t new_vi; 1243 1244 copy = create_copy_of_insn_rtx (VINSN_INSN_RTX (vi)); 1245 new_vi = create_vinsn_from_insn_rtx (copy, unique); 1246 if (reattach_p) 1247 { 1248 vinsn_detach (vi); 1249 vinsn_attach (new_vi); 1250 } 1251 1252 return new_vi; 1253 } 1254 1255 /* Delete the VI vinsn and free its data. */ 1256 static void 1257 vinsn_delete (vinsn_t vi) 1258 { 1259 gcc_assert (VINSN_COUNT (vi) == 0); 1260 1261 if (!INSN_NOP_P (VINSN_INSN_RTX (vi))) 1262 { 1263 return_regset_to_pool (VINSN_REG_SETS (vi)); 1264 return_regset_to_pool (VINSN_REG_USES (vi)); 1265 return_regset_to_pool (VINSN_REG_CLOBBERS (vi)); 1266 } 1267 1268 free (vi); 1269 } 1270 1271 /* Indicate that VI is no longer a part of some rtx object. 1272 Remove VI if it is no longer needed. */ 1273 void 1274 vinsn_detach (vinsn_t vi) 1275 { 1276 gcc_assert (VINSN_COUNT (vi) > 0); 1277 1278 if (--VINSN_COUNT (vi) == 0) 1279 vinsn_delete (vi); 1280 } 1281 1282 /* Returns TRUE if VI is a branch. */ 1283 bool 1284 vinsn_cond_branch_p (vinsn_t vi) 1285 { 1286 insn_t insn; 1287 1288 if (!VINSN_UNIQUE_P (vi)) 1289 return false; 1290 1291 insn = VINSN_INSN_RTX (vi); 1292 if (BB_END (BLOCK_FOR_INSN (insn)) != insn) 1293 return false; 1294 1295 return control_flow_insn_p (insn); 1296 } 1297 1298 /* Return latency of INSN. */ 1299 static int 1300 sel_insn_rtx_cost (rtx_insn *insn) 1301 { 1302 int cost; 1303 1304 /* A USE insn, or something else we don't need to 1305 understand. We can't pass these directly to 1306 result_ready_cost or insn_default_latency because it will 1307 trigger a fatal error for unrecognizable insns. */ 1308 if (recog_memoized (insn) < 0) 1309 cost = 0; 1310 else 1311 { 1312 cost = insn_default_latency (insn); 1313 1314 if (cost < 0) 1315 cost = 0; 1316 } 1317 1318 return cost; 1319 } 1320 1321 /* Return the cost of the VI. 1322 !!! FIXME: Unify with haifa-sched.c: insn_sched_cost (). */ 1323 int 1324 sel_vinsn_cost (vinsn_t vi) 1325 { 1326 int cost = vi->cost; 1327 1328 if (cost < 0) 1329 { 1330 cost = sel_insn_rtx_cost (VINSN_INSN_RTX (vi)); 1331 vi->cost = cost; 1332 } 1333 1334 return cost; 1335 } 1336 1337 1338 /* Functions for insn emitting. */ 1339 1340 /* Emit new insn after AFTER based on PATTERN and initialize its data from 1341 EXPR and SEQNO. */ 1342 insn_t 1343 sel_gen_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, insn_t after) 1344 { 1345 insn_t new_insn; 1346 1347 gcc_assert (EXPR_TARGET_AVAILABLE (expr) == true); 1348 1349 new_insn = emit_insn_after (pattern, after); 1350 set_insn_init (expr, NULL, seqno); 1351 sel_init_new_insn (new_insn, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID); 1352 1353 return new_insn; 1354 } 1355 1356 /* Force newly generated vinsns to be unique. */ 1357 static bool init_insn_force_unique_p = false; 1358 1359 /* Emit new speculation recovery insn after AFTER based on PATTERN and 1360 initialize its data from EXPR and SEQNO. */ 1361 insn_t 1362 sel_gen_recovery_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, 1363 insn_t after) 1364 { 1365 insn_t insn; 1366 1367 gcc_assert (!init_insn_force_unique_p); 1368 1369 init_insn_force_unique_p = true; 1370 insn = sel_gen_insn_from_rtx_after (pattern, expr, seqno, after); 1371 CANT_MOVE (insn) = 1; 1372 init_insn_force_unique_p = false; 1373 1374 return insn; 1375 } 1376 1377 /* Emit new insn after AFTER based on EXPR and SEQNO. If VINSN is not NULL, 1378 take it as a new vinsn instead of EXPR's vinsn. 1379 We simplify insns later, after scheduling region in 1380 simplify_changed_insns. */ 1381 insn_t 1382 sel_gen_insn_from_expr_after (expr_t expr, vinsn_t vinsn, int seqno, 1383 insn_t after) 1384 { 1385 expr_t emit_expr; 1386 insn_t insn; 1387 int flags; 1388 1389 emit_expr = set_insn_init (expr, vinsn ? vinsn : EXPR_VINSN (expr), 1390 seqno); 1391 insn = EXPR_INSN_RTX (emit_expr); 1392 1393 /* The insn may come from the transformation cache, which may hold already 1394 deleted insns, so mark it as not deleted. */ 1395 insn->set_undeleted (); 1396 1397 add_insn_after (insn, after, BLOCK_FOR_INSN (insn)); 1398 1399 flags = INSN_INIT_TODO_SSID; 1400 if (INSN_LUID (insn) == 0) 1401 flags |= INSN_INIT_TODO_LUID; 1402 sel_init_new_insn (insn, flags); 1403 1404 return insn; 1405 } 1406 1407 /* Move insn from EXPR after AFTER. */ 1408 insn_t 1409 sel_move_insn (expr_t expr, int seqno, insn_t after) 1410 { 1411 insn_t insn = EXPR_INSN_RTX (expr); 1412 basic_block bb = BLOCK_FOR_INSN (after); 1413 insn_t next = NEXT_INSN (after); 1414 1415 /* Assert that in move_op we disconnected this insn properly. */ 1416 gcc_assert (EXPR_VINSN (INSN_EXPR (insn)) != NULL); 1417 SET_PREV_INSN (insn) = after; 1418 SET_NEXT_INSN (insn) = next; 1419 1420 SET_NEXT_INSN (after) = insn; 1421 SET_PREV_INSN (next) = insn; 1422 1423 /* Update links from insn to bb and vice versa. */ 1424 df_insn_change_bb (insn, bb); 1425 if (BB_END (bb) == after) 1426 BB_END (bb) = insn; 1427 1428 prepare_insn_expr (insn, seqno); 1429 return insn; 1430 } 1431 1432 1433 /* Functions to work with right-hand sides. */ 1434 1435 /* Search for a hash value determined by UID/NEW_VINSN in a sorted vector 1436 VECT and return true when found. Use NEW_VINSN for comparison only when 1437 COMPARE_VINSNS is true. Write to INDP the index on which 1438 the search has stopped, such that inserting the new element at INDP will 1439 retain VECT's sort order. */ 1440 static bool 1441 find_in_history_vect_1 (vec<expr_history_def> vect, 1442 unsigned uid, vinsn_t new_vinsn, 1443 bool compare_vinsns, int *indp) 1444 { 1445 expr_history_def *arr; 1446 int i, j, len = vect.length (); 1447 1448 if (len == 0) 1449 { 1450 *indp = 0; 1451 return false; 1452 } 1453 1454 arr = vect.address (); 1455 i = 0, j = len - 1; 1456 1457 while (i <= j) 1458 { 1459 unsigned auid = arr[i].uid; 1460 vinsn_t avinsn = arr[i].new_expr_vinsn; 1461 1462 if (auid == uid 1463 /* When undoing transformation on a bookkeeping copy, the new vinsn 1464 may not be exactly equal to the one that is saved in the vector. 1465 This is because the insn whose copy we're checking was possibly 1466 substituted itself. */ 1467 && (! compare_vinsns 1468 || vinsn_equal_p (avinsn, new_vinsn))) 1469 { 1470 *indp = i; 1471 return true; 1472 } 1473 else if (auid > uid) 1474 break; 1475 i++; 1476 } 1477 1478 *indp = i; 1479 return false; 1480 } 1481 1482 /* Search for a uid of INSN and NEW_VINSN in a sorted vector VECT. Return 1483 the position found or -1, if no such value is in vector. 1484 Search also for UIDs of insn's originators, if ORIGINATORS_P is true. */ 1485 int 1486 find_in_history_vect (vec<expr_history_def> vect, rtx insn, 1487 vinsn_t new_vinsn, bool originators_p) 1488 { 1489 int ind; 1490 1491 if (find_in_history_vect_1 (vect, INSN_UID (insn), new_vinsn, 1492 false, &ind)) 1493 return ind; 1494 1495 if (INSN_ORIGINATORS (insn) && originators_p) 1496 { 1497 unsigned uid; 1498 bitmap_iterator bi; 1499 1500 EXECUTE_IF_SET_IN_BITMAP (INSN_ORIGINATORS (insn), 0, uid, bi) 1501 if (find_in_history_vect_1 (vect, uid, new_vinsn, false, &ind)) 1502 return ind; 1503 } 1504 1505 return -1; 1506 } 1507 1508 /* Insert new element in a sorted history vector pointed to by PVECT, 1509 if it is not there already. The element is searched using 1510 UID/NEW_EXPR_VINSN pair. TYPE, OLD_EXPR_VINSN and SPEC_DS save 1511 the history of a transformation. */ 1512 void 1513 insert_in_history_vect (vec<expr_history_def> *pvect, 1514 unsigned uid, enum local_trans_type type, 1515 vinsn_t old_expr_vinsn, vinsn_t new_expr_vinsn, 1516 ds_t spec_ds) 1517 { 1518 vec<expr_history_def> vect = *pvect; 1519 expr_history_def temp; 1520 bool res; 1521 int ind; 1522 1523 res = find_in_history_vect_1 (vect, uid, new_expr_vinsn, true, &ind); 1524 1525 if (res) 1526 { 1527 expr_history_def *phist = &vect[ind]; 1528 1529 /* It is possible that speculation types of expressions that were 1530 propagated through different paths will be different here. In this 1531 case, merge the status to get the correct check later. */ 1532 if (phist->spec_ds != spec_ds) 1533 phist->spec_ds = ds_max_merge (phist->spec_ds, spec_ds); 1534 return; 1535 } 1536 1537 temp.uid = uid; 1538 temp.old_expr_vinsn = old_expr_vinsn; 1539 temp.new_expr_vinsn = new_expr_vinsn; 1540 temp.spec_ds = spec_ds; 1541 temp.type = type; 1542 1543 vinsn_attach (old_expr_vinsn); 1544 vinsn_attach (new_expr_vinsn); 1545 vect.safe_insert (ind, temp); 1546 *pvect = vect; 1547 } 1548 1549 /* Free history vector PVECT. */ 1550 static void 1551 free_history_vect (vec<expr_history_def> &pvect) 1552 { 1553 unsigned i; 1554 expr_history_def *phist; 1555 1556 if (! pvect.exists ()) 1557 return; 1558 1559 for (i = 0; pvect.iterate (i, &phist); i++) 1560 { 1561 vinsn_detach (phist->old_expr_vinsn); 1562 vinsn_detach (phist->new_expr_vinsn); 1563 } 1564 1565 pvect.release (); 1566 } 1567 1568 /* Merge vector FROM to PVECT. */ 1569 static void 1570 merge_history_vect (vec<expr_history_def> *pvect, 1571 vec<expr_history_def> from) 1572 { 1573 expr_history_def *phist; 1574 int i; 1575 1576 /* We keep this vector sorted. */ 1577 for (i = 0; from.iterate (i, &phist); i++) 1578 insert_in_history_vect (pvect, phist->uid, phist->type, 1579 phist->old_expr_vinsn, phist->new_expr_vinsn, 1580 phist->spec_ds); 1581 } 1582 1583 /* Compare two vinsns as rhses if possible and as vinsns otherwise. */ 1584 bool 1585 vinsn_equal_p (vinsn_t x, vinsn_t y) 1586 { 1587 rtx_equal_p_callback_function repcf; 1588 1589 if (x == y) 1590 return true; 1591 1592 if (VINSN_TYPE (x) != VINSN_TYPE (y)) 1593 return false; 1594 1595 if (VINSN_HASH (x) != VINSN_HASH (y)) 1596 return false; 1597 1598 repcf = targetm.sched.skip_rtx_p ? skip_unspecs_callback : NULL; 1599 if (VINSN_SEPARABLE_P (x)) 1600 { 1601 /* Compare RHSes of VINSNs. */ 1602 gcc_assert (VINSN_RHS (x)); 1603 gcc_assert (VINSN_RHS (y)); 1604 1605 return rtx_equal_p_cb (VINSN_RHS (x), VINSN_RHS (y), repcf); 1606 } 1607 1608 return rtx_equal_p_cb (VINSN_PATTERN (x), VINSN_PATTERN (y), repcf); 1609 } 1610 1611 1612 /* Functions for working with expressions. */ 1613 1614 /* Initialize EXPR. */ 1615 static void 1616 init_expr (expr_t expr, vinsn_t vi, int spec, int use, int priority, 1617 int sched_times, int orig_bb_index, ds_t spec_done_ds, 1618 ds_t spec_to_check_ds, int orig_sched_cycle, 1619 vec<expr_history_def> history, 1620 signed char target_available, 1621 bool was_substituted, bool was_renamed, bool needs_spec_check_p, 1622 bool cant_move) 1623 { 1624 vinsn_attach (vi); 1625 1626 EXPR_VINSN (expr) = vi; 1627 EXPR_SPEC (expr) = spec; 1628 EXPR_USEFULNESS (expr) = use; 1629 EXPR_PRIORITY (expr) = priority; 1630 EXPR_PRIORITY_ADJ (expr) = 0; 1631 EXPR_SCHED_TIMES (expr) = sched_times; 1632 EXPR_ORIG_BB_INDEX (expr) = orig_bb_index; 1633 EXPR_ORIG_SCHED_CYCLE (expr) = orig_sched_cycle; 1634 EXPR_SPEC_DONE_DS (expr) = spec_done_ds; 1635 EXPR_SPEC_TO_CHECK_DS (expr) = spec_to_check_ds; 1636 1637 if (history.exists ()) 1638 EXPR_HISTORY_OF_CHANGES (expr) = history; 1639 else 1640 EXPR_HISTORY_OF_CHANGES (expr).create (0); 1641 1642 EXPR_TARGET_AVAILABLE (expr) = target_available; 1643 EXPR_WAS_SUBSTITUTED (expr) = was_substituted; 1644 EXPR_WAS_RENAMED (expr) = was_renamed; 1645 EXPR_NEEDS_SPEC_CHECK_P (expr) = needs_spec_check_p; 1646 EXPR_CANT_MOVE (expr) = cant_move; 1647 } 1648 1649 /* Make a copy of the expr FROM into the expr TO. */ 1650 void 1651 copy_expr (expr_t to, expr_t from) 1652 { 1653 vec<expr_history_def> temp = vNULL; 1654 1655 if (EXPR_HISTORY_OF_CHANGES (from).exists ()) 1656 { 1657 unsigned i; 1658 expr_history_def *phist; 1659 1660 temp = EXPR_HISTORY_OF_CHANGES (from).copy (); 1661 for (i = 0; 1662 temp.iterate (i, &phist); 1663 i++) 1664 { 1665 vinsn_attach (phist->old_expr_vinsn); 1666 vinsn_attach (phist->new_expr_vinsn); 1667 } 1668 } 1669 1670 init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from), 1671 EXPR_USEFULNESS (from), EXPR_PRIORITY (from), 1672 EXPR_SCHED_TIMES (from), EXPR_ORIG_BB_INDEX (from), 1673 EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), 1674 EXPR_ORIG_SCHED_CYCLE (from), temp, 1675 EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from), 1676 EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from), 1677 EXPR_CANT_MOVE (from)); 1678 } 1679 1680 /* Same, but the final expr will not ever be in av sets, so don't copy 1681 "uninteresting" data such as bitmap cache. */ 1682 void 1683 copy_expr_onside (expr_t to, expr_t from) 1684 { 1685 init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from), EXPR_USEFULNESS (from), 1686 EXPR_PRIORITY (from), EXPR_SCHED_TIMES (from), 0, 1687 EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), 0, 1688 vNULL, 1689 EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from), 1690 EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from), 1691 EXPR_CANT_MOVE (from)); 1692 } 1693 1694 /* Prepare the expr of INSN for scheduling. Used when moving insn and when 1695 initializing new insns. */ 1696 static void 1697 prepare_insn_expr (insn_t insn, int seqno) 1698 { 1699 expr_t expr = INSN_EXPR (insn); 1700 ds_t ds; 1701 1702 INSN_SEQNO (insn) = seqno; 1703 EXPR_ORIG_BB_INDEX (expr) = BLOCK_NUM (insn); 1704 EXPR_SPEC (expr) = 0; 1705 EXPR_ORIG_SCHED_CYCLE (expr) = 0; 1706 EXPR_WAS_SUBSTITUTED (expr) = 0; 1707 EXPR_WAS_RENAMED (expr) = 0; 1708 EXPR_TARGET_AVAILABLE (expr) = 1; 1709 INSN_LIVE_VALID_P (insn) = false; 1710 1711 /* ??? If this expression is speculative, make its dependence 1712 as weak as possible. We can filter this expression later 1713 in process_spec_exprs, because we do not distinguish 1714 between the status we got during compute_av_set and the 1715 existing status. To be fixed. */ 1716 ds = EXPR_SPEC_DONE_DS (expr); 1717 if (ds) 1718 EXPR_SPEC_DONE_DS (expr) = ds_get_max_dep_weak (ds); 1719 1720 free_history_vect (EXPR_HISTORY_OF_CHANGES (expr)); 1721 } 1722 1723 /* Update target_available bits when merging exprs TO and FROM. SPLIT_POINT 1724 is non-null when expressions are merged from different successors at 1725 a split point. */ 1726 static void 1727 update_target_availability (expr_t to, expr_t from, insn_t split_point) 1728 { 1729 if (EXPR_TARGET_AVAILABLE (to) < 0 1730 || EXPR_TARGET_AVAILABLE (from) < 0) 1731 EXPR_TARGET_AVAILABLE (to) = -1; 1732 else 1733 { 1734 /* We try to detect the case when one of the expressions 1735 can only be reached through another one. In this case, 1736 we can do better. */ 1737 if (split_point == NULL) 1738 { 1739 int toind, fromind; 1740 1741 toind = EXPR_ORIG_BB_INDEX (to); 1742 fromind = EXPR_ORIG_BB_INDEX (from); 1743 1744 if (toind && toind == fromind) 1745 /* Do nothing -- everything is done in 1746 merge_with_other_exprs. */ 1747 ; 1748 else 1749 EXPR_TARGET_AVAILABLE (to) = -1; 1750 } 1751 else if (EXPR_TARGET_AVAILABLE (from) == 0 1752 && EXPR_LHS (from) 1753 && REG_P (EXPR_LHS (from)) 1754 && REGNO (EXPR_LHS (to)) != REGNO (EXPR_LHS (from))) 1755 EXPR_TARGET_AVAILABLE (to) = -1; 1756 else 1757 EXPR_TARGET_AVAILABLE (to) &= EXPR_TARGET_AVAILABLE (from); 1758 } 1759 } 1760 1761 /* Update speculation bits when merging exprs TO and FROM. SPLIT_POINT 1762 is non-null when expressions are merged from different successors at 1763 a split point. */ 1764 static void 1765 update_speculative_bits (expr_t to, expr_t from, insn_t split_point) 1766 { 1767 ds_t old_to_ds, old_from_ds; 1768 1769 old_to_ds = EXPR_SPEC_DONE_DS (to); 1770 old_from_ds = EXPR_SPEC_DONE_DS (from); 1771 1772 EXPR_SPEC_DONE_DS (to) = ds_max_merge (old_to_ds, old_from_ds); 1773 EXPR_SPEC_TO_CHECK_DS (to) |= EXPR_SPEC_TO_CHECK_DS (from); 1774 EXPR_NEEDS_SPEC_CHECK_P (to) |= EXPR_NEEDS_SPEC_CHECK_P (from); 1775 1776 /* When merging e.g. control & data speculative exprs, or a control 1777 speculative with a control&data speculative one, we really have 1778 to change vinsn too. Also, when speculative status is changed, 1779 we also need to record this as a transformation in expr's history. */ 1780 if ((old_to_ds & SPECULATIVE) || (old_from_ds & SPECULATIVE)) 1781 { 1782 old_to_ds = ds_get_speculation_types (old_to_ds); 1783 old_from_ds = ds_get_speculation_types (old_from_ds); 1784 1785 if (old_to_ds != old_from_ds) 1786 { 1787 ds_t record_ds; 1788 1789 /* When both expressions are speculative, we need to change 1790 the vinsn first. */ 1791 if ((old_to_ds & SPECULATIVE) && (old_from_ds & SPECULATIVE)) 1792 { 1793 int res; 1794 1795 res = speculate_expr (to, EXPR_SPEC_DONE_DS (to)); 1796 gcc_assert (res >= 0); 1797 } 1798 1799 if (split_point != NULL) 1800 { 1801 /* Record the change with proper status. */ 1802 record_ds = EXPR_SPEC_DONE_DS (to) & SPECULATIVE; 1803 record_ds &= ~(old_to_ds & SPECULATIVE); 1804 record_ds &= ~(old_from_ds & SPECULATIVE); 1805 1806 insert_in_history_vect (&EXPR_HISTORY_OF_CHANGES (to), 1807 INSN_UID (split_point), TRANS_SPECULATION, 1808 EXPR_VINSN (from), EXPR_VINSN (to), 1809 record_ds); 1810 } 1811 } 1812 } 1813 } 1814 1815 1816 /* Merge bits of FROM expr to TO expr. When SPLIT_POINT is not NULL, 1817 this is done along different paths. */ 1818 void 1819 merge_expr_data (expr_t to, expr_t from, insn_t split_point) 1820 { 1821 /* Choose the maximum of the specs of merged exprs. This is required 1822 for correctness of bookkeeping. */ 1823 if (EXPR_SPEC (to) < EXPR_SPEC (from)) 1824 EXPR_SPEC (to) = EXPR_SPEC (from); 1825 1826 if (split_point) 1827 EXPR_USEFULNESS (to) += EXPR_USEFULNESS (from); 1828 else 1829 EXPR_USEFULNESS (to) = MAX (EXPR_USEFULNESS (to), 1830 EXPR_USEFULNESS (from)); 1831 1832 if (EXPR_PRIORITY (to) < EXPR_PRIORITY (from)) 1833 EXPR_PRIORITY (to) = EXPR_PRIORITY (from); 1834 1835 /* We merge sched-times half-way to the larger value to avoid the endless 1836 pipelining of unneeded insns. The average seems to be good compromise 1837 between pipelining opportunities and avoiding extra work. */ 1838 if (EXPR_SCHED_TIMES (to) != EXPR_SCHED_TIMES (from)) 1839 EXPR_SCHED_TIMES (to) = ((EXPR_SCHED_TIMES (from) + EXPR_SCHED_TIMES (to) 1840 + 1) / 2); 1841 1842 if (EXPR_ORIG_BB_INDEX (to) != EXPR_ORIG_BB_INDEX (from)) 1843 EXPR_ORIG_BB_INDEX (to) = 0; 1844 1845 EXPR_ORIG_SCHED_CYCLE (to) = MIN (EXPR_ORIG_SCHED_CYCLE (to), 1846 EXPR_ORIG_SCHED_CYCLE (from)); 1847 1848 EXPR_WAS_SUBSTITUTED (to) |= EXPR_WAS_SUBSTITUTED (from); 1849 EXPR_WAS_RENAMED (to) |= EXPR_WAS_RENAMED (from); 1850 EXPR_CANT_MOVE (to) |= EXPR_CANT_MOVE (from); 1851 1852 merge_history_vect (&EXPR_HISTORY_OF_CHANGES (to), 1853 EXPR_HISTORY_OF_CHANGES (from)); 1854 update_target_availability (to, from, split_point); 1855 update_speculative_bits (to, from, split_point); 1856 } 1857 1858 /* Merge bits of FROM expr to TO expr. Vinsns in the exprs should be equal 1859 in terms of vinsn_equal_p. SPLIT_POINT is non-null when expressions 1860 are merged from different successors at a split point. */ 1861 void 1862 merge_expr (expr_t to, expr_t from, insn_t split_point) 1863 { 1864 vinsn_t to_vi = EXPR_VINSN (to); 1865 vinsn_t from_vi = EXPR_VINSN (from); 1866 1867 gcc_assert (vinsn_equal_p (to_vi, from_vi)); 1868 1869 /* Make sure that speculative pattern is propagated into exprs that 1870 have non-speculative one. This will provide us with consistent 1871 speculative bits and speculative patterns inside expr. */ 1872 if (EXPR_SPEC_DONE_DS (to) == 0 1873 && (EXPR_SPEC_DONE_DS (from) != 0 1874 /* Do likewise for volatile insns, so that we always retain 1875 the may_trap_p bit on the resulting expression. However, 1876 avoid propagating the trapping bit into the instructions 1877 already speculated. This would result in replacing the 1878 speculative pattern with the non-speculative one and breaking 1879 the speculation support. */ 1880 || (!VINSN_MAY_TRAP_P (EXPR_VINSN (to)) 1881 && VINSN_MAY_TRAP_P (EXPR_VINSN (from))))) 1882 change_vinsn_in_expr (to, EXPR_VINSN (from)); 1883 1884 merge_expr_data (to, from, split_point); 1885 gcc_assert (EXPR_USEFULNESS (to) <= REG_BR_PROB_BASE); 1886 } 1887 1888 /* Clear the information of this EXPR. */ 1889 void 1890 clear_expr (expr_t expr) 1891 { 1892 1893 vinsn_detach (EXPR_VINSN (expr)); 1894 EXPR_VINSN (expr) = NULL; 1895 1896 free_history_vect (EXPR_HISTORY_OF_CHANGES (expr)); 1897 } 1898 1899 /* For a given LV_SET, mark EXPR having unavailable target register. */ 1900 static void 1901 set_unavailable_target_for_expr (expr_t expr, regset lv_set) 1902 { 1903 if (EXPR_SEPARABLE_P (expr)) 1904 { 1905 if (REG_P (EXPR_LHS (expr)) 1906 && register_unavailable_p (lv_set, EXPR_LHS (expr))) 1907 { 1908 /* If it's an insn like r1 = use (r1, ...), and it exists in 1909 different forms in each of the av_sets being merged, we can't say 1910 whether original destination register is available or not. 1911 However, this still works if destination register is not used 1912 in the original expression: if the branch at which LV_SET we're 1913 looking here is not actually 'other branch' in sense that same 1914 expression is available through it (but it can't be determined 1915 at computation stage because of transformations on one of the 1916 branches), it still won't affect the availability. 1917 Liveness of a register somewhere on a code motion path means 1918 it's either read somewhere on a codemotion path, live on 1919 'other' branch, live at the point immediately following 1920 the original operation, or is read by the original operation. 1921 The latter case is filtered out in the condition below. 1922 It still doesn't cover the case when register is defined and used 1923 somewhere within the code motion path, and in this case we could 1924 miss a unifying code motion along both branches using a renamed 1925 register, but it won't affect a code correctness since upon 1926 an actual code motion a bookkeeping code would be generated. */ 1927 if (register_unavailable_p (VINSN_REG_USES (EXPR_VINSN (expr)), 1928 EXPR_LHS (expr))) 1929 EXPR_TARGET_AVAILABLE (expr) = -1; 1930 else 1931 EXPR_TARGET_AVAILABLE (expr) = false; 1932 } 1933 } 1934 else 1935 { 1936 unsigned regno; 1937 reg_set_iterator rsi; 1938 1939 EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_SETS (EXPR_VINSN (expr)), 1940 0, regno, rsi) 1941 if (bitmap_bit_p (lv_set, regno)) 1942 { 1943 EXPR_TARGET_AVAILABLE (expr) = false; 1944 break; 1945 } 1946 1947 EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_CLOBBERS (EXPR_VINSN (expr)), 1948 0, regno, rsi) 1949 if (bitmap_bit_p (lv_set, regno)) 1950 { 1951 EXPR_TARGET_AVAILABLE (expr) = false; 1952 break; 1953 } 1954 } 1955 } 1956 1957 /* Try to make EXPR speculative. Return 1 when EXPR's pattern 1958 or dependence status have changed, 2 when also the target register 1959 became unavailable, 0 if nothing had to be changed. */ 1960 int 1961 speculate_expr (expr_t expr, ds_t ds) 1962 { 1963 int res; 1964 rtx_insn *orig_insn_rtx; 1965 rtx spec_pat; 1966 ds_t target_ds, current_ds; 1967 1968 /* Obtain the status we need to put on EXPR. */ 1969 target_ds = (ds & SPECULATIVE); 1970 current_ds = EXPR_SPEC_DONE_DS (expr); 1971 ds = ds_full_merge (current_ds, target_ds, NULL_RTX, NULL_RTX); 1972 1973 orig_insn_rtx = EXPR_INSN_RTX (expr); 1974 1975 res = sched_speculate_insn (orig_insn_rtx, ds, &spec_pat); 1976 1977 switch (res) 1978 { 1979 case 0: 1980 EXPR_SPEC_DONE_DS (expr) = ds; 1981 return current_ds != ds ? 1 : 0; 1982 1983 case 1: 1984 { 1985 rtx_insn *spec_insn_rtx = 1986 create_insn_rtx_from_pattern (spec_pat, NULL_RTX); 1987 vinsn_t spec_vinsn = create_vinsn_from_insn_rtx (spec_insn_rtx, false); 1988 1989 change_vinsn_in_expr (expr, spec_vinsn); 1990 EXPR_SPEC_DONE_DS (expr) = ds; 1991 EXPR_NEEDS_SPEC_CHECK_P (expr) = true; 1992 1993 /* Do not allow clobbering the address register of speculative 1994 insns. */ 1995 if (register_unavailable_p (VINSN_REG_USES (EXPR_VINSN (expr)), 1996 expr_dest_reg (expr))) 1997 { 1998 EXPR_TARGET_AVAILABLE (expr) = false; 1999 return 2; 2000 } 2001 2002 return 1; 2003 } 2004 2005 case -1: 2006 return -1; 2007 2008 default: 2009 gcc_unreachable (); 2010 return -1; 2011 } 2012 } 2013 2014 /* Return a destination register, if any, of EXPR. */ 2015 rtx 2016 expr_dest_reg (expr_t expr) 2017 { 2018 rtx dest = VINSN_LHS (EXPR_VINSN (expr)); 2019 2020 if (dest != NULL_RTX && REG_P (dest)) 2021 return dest; 2022 2023 return NULL_RTX; 2024 } 2025 2026 /* Returns the REGNO of the R's destination. */ 2027 unsigned 2028 expr_dest_regno (expr_t expr) 2029 { 2030 rtx dest = expr_dest_reg (expr); 2031 2032 gcc_assert (dest != NULL_RTX); 2033 return REGNO (dest); 2034 } 2035 2036 /* For a given LV_SET, mark all expressions in JOIN_SET, but not present in 2037 AV_SET having unavailable target register. */ 2038 void 2039 mark_unavailable_targets (av_set_t join_set, av_set_t av_set, regset lv_set) 2040 { 2041 expr_t expr; 2042 av_set_iterator avi; 2043 2044 FOR_EACH_EXPR (expr, avi, join_set) 2045 if (av_set_lookup (av_set, EXPR_VINSN (expr)) == NULL) 2046 set_unavailable_target_for_expr (expr, lv_set); 2047 } 2048 2049 2050 /* Returns true if REG (at least partially) is present in REGS. */ 2051 bool 2052 register_unavailable_p (regset regs, rtx reg) 2053 { 2054 unsigned regno, end_regno; 2055 2056 regno = REGNO (reg); 2057 if (bitmap_bit_p (regs, regno)) 2058 return true; 2059 2060 end_regno = END_REGNO (reg); 2061 2062 while (++regno < end_regno) 2063 if (bitmap_bit_p (regs, regno)) 2064 return true; 2065 2066 return false; 2067 } 2068 2069 /* Av set functions. */ 2070 2071 /* Add a new element to av set SETP. 2072 Return the element added. */ 2073 static av_set_t 2074 av_set_add_element (av_set_t *setp) 2075 { 2076 /* Insert at the beginning of the list. */ 2077 _list_add (setp); 2078 return *setp; 2079 } 2080 2081 /* Add EXPR to SETP. */ 2082 void 2083 av_set_add (av_set_t *setp, expr_t expr) 2084 { 2085 av_set_t elem; 2086 2087 gcc_assert (!INSN_NOP_P (EXPR_INSN_RTX (expr))); 2088 elem = av_set_add_element (setp); 2089 copy_expr (_AV_SET_EXPR (elem), expr); 2090 } 2091 2092 /* Same, but do not copy EXPR. */ 2093 static void 2094 av_set_add_nocopy (av_set_t *setp, expr_t expr) 2095 { 2096 av_set_t elem; 2097 2098 elem = av_set_add_element (setp); 2099 *_AV_SET_EXPR (elem) = *expr; 2100 } 2101 2102 /* Remove expr pointed to by IP from the av_set. */ 2103 void 2104 av_set_iter_remove (av_set_iterator *ip) 2105 { 2106 clear_expr (_AV_SET_EXPR (*ip->lp)); 2107 _list_iter_remove (ip); 2108 } 2109 2110 /* Search for an expr in SET, such that it's equivalent to SOUGHT_VINSN in the 2111 sense of vinsn_equal_p function. Return NULL if no such expr is 2112 in SET was found. */ 2113 expr_t 2114 av_set_lookup (av_set_t set, vinsn_t sought_vinsn) 2115 { 2116 expr_t expr; 2117 av_set_iterator i; 2118 2119 FOR_EACH_EXPR (expr, i, set) 2120 if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn)) 2121 return expr; 2122 return NULL; 2123 } 2124 2125 /* Same, but also remove the EXPR found. */ 2126 static expr_t 2127 av_set_lookup_and_remove (av_set_t *setp, vinsn_t sought_vinsn) 2128 { 2129 expr_t expr; 2130 av_set_iterator i; 2131 2132 FOR_EACH_EXPR_1 (expr, i, setp) 2133 if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn)) 2134 { 2135 _list_iter_remove_nofree (&i); 2136 return expr; 2137 } 2138 return NULL; 2139 } 2140 2141 /* Search for an expr in SET, such that it's equivalent to EXPR in the 2142 sense of vinsn_equal_p function of their vinsns, but not EXPR itself. 2143 Returns NULL if no such expr is in SET was found. */ 2144 static expr_t 2145 av_set_lookup_other_equiv_expr (av_set_t set, expr_t expr) 2146 { 2147 expr_t cur_expr; 2148 av_set_iterator i; 2149 2150 FOR_EACH_EXPR (cur_expr, i, set) 2151 { 2152 if (cur_expr == expr) 2153 continue; 2154 if (vinsn_equal_p (EXPR_VINSN (cur_expr), EXPR_VINSN (expr))) 2155 return cur_expr; 2156 } 2157 2158 return NULL; 2159 } 2160 2161 /* If other expression is already in AVP, remove one of them. */ 2162 expr_t 2163 merge_with_other_exprs (av_set_t *avp, av_set_iterator *ip, expr_t expr) 2164 { 2165 expr_t expr2; 2166 2167 expr2 = av_set_lookup_other_equiv_expr (*avp, expr); 2168 if (expr2 != NULL) 2169 { 2170 /* Reset target availability on merge, since taking it only from one 2171 of the exprs would be controversial for different code. */ 2172 EXPR_TARGET_AVAILABLE (expr2) = -1; 2173 EXPR_USEFULNESS (expr2) = 0; 2174 2175 merge_expr (expr2, expr, NULL); 2176 2177 /* Fix usefulness as it should be now REG_BR_PROB_BASE. */ 2178 EXPR_USEFULNESS (expr2) = REG_BR_PROB_BASE; 2179 2180 av_set_iter_remove (ip); 2181 return expr2; 2182 } 2183 2184 return expr; 2185 } 2186 2187 /* Return true if there is an expr that correlates to VI in SET. */ 2188 bool 2189 av_set_is_in_p (av_set_t set, vinsn_t vi) 2190 { 2191 return av_set_lookup (set, vi) != NULL; 2192 } 2193 2194 /* Return a copy of SET. */ 2195 av_set_t 2196 av_set_copy (av_set_t set) 2197 { 2198 expr_t expr; 2199 av_set_iterator i; 2200 av_set_t res = NULL; 2201 2202 FOR_EACH_EXPR (expr, i, set) 2203 av_set_add (&res, expr); 2204 2205 return res; 2206 } 2207 2208 /* Join two av sets that do not have common elements by attaching second set 2209 (pointed to by FROMP) to the end of first set (TO_TAILP must point to 2210 _AV_SET_NEXT of first set's last element). */ 2211 static void 2212 join_distinct_sets (av_set_t *to_tailp, av_set_t *fromp) 2213 { 2214 gcc_assert (*to_tailp == NULL); 2215 *to_tailp = *fromp; 2216 *fromp = NULL; 2217 } 2218 2219 /* Makes set pointed to by TO to be the union of TO and FROM. Clear av_set 2220 pointed to by FROMP afterwards. */ 2221 void 2222 av_set_union_and_clear (av_set_t *top, av_set_t *fromp, insn_t insn) 2223 { 2224 expr_t expr1; 2225 av_set_iterator i; 2226 2227 /* Delete from TOP all exprs, that present in FROMP. */ 2228 FOR_EACH_EXPR_1 (expr1, i, top) 2229 { 2230 expr_t expr2 = av_set_lookup (*fromp, EXPR_VINSN (expr1)); 2231 2232 if (expr2) 2233 { 2234 merge_expr (expr2, expr1, insn); 2235 av_set_iter_remove (&i); 2236 } 2237 } 2238 2239 join_distinct_sets (i.lp, fromp); 2240 } 2241 2242 /* Same as above, but also update availability of target register in 2243 TOP judging by TO_LV_SET and FROM_LV_SET. */ 2244 void 2245 av_set_union_and_live (av_set_t *top, av_set_t *fromp, regset to_lv_set, 2246 regset from_lv_set, insn_t insn) 2247 { 2248 expr_t expr1; 2249 av_set_iterator i; 2250 av_set_t *to_tailp, in_both_set = NULL; 2251 2252 /* Delete from TOP all expres, that present in FROMP. */ 2253 FOR_EACH_EXPR_1 (expr1, i, top) 2254 { 2255 expr_t expr2 = av_set_lookup_and_remove (fromp, EXPR_VINSN (expr1)); 2256 2257 if (expr2) 2258 { 2259 /* It may be that the expressions have different destination 2260 registers, in which case we need to check liveness here. */ 2261 if (EXPR_SEPARABLE_P (expr1)) 2262 { 2263 int regno1 = (REG_P (EXPR_LHS (expr1)) 2264 ? (int) expr_dest_regno (expr1) : -1); 2265 int regno2 = (REG_P (EXPR_LHS (expr2)) 2266 ? (int) expr_dest_regno (expr2) : -1); 2267 2268 /* ??? We don't have a way to check restrictions for 2269 *other* register on the current path, we did it only 2270 for the current target register. Give up. */ 2271 if (regno1 != regno2) 2272 EXPR_TARGET_AVAILABLE (expr2) = -1; 2273 } 2274 else if (EXPR_INSN_RTX (expr1) != EXPR_INSN_RTX (expr2)) 2275 EXPR_TARGET_AVAILABLE (expr2) = -1; 2276 2277 merge_expr (expr2, expr1, insn); 2278 av_set_add_nocopy (&in_both_set, expr2); 2279 av_set_iter_remove (&i); 2280 } 2281 else 2282 /* EXPR1 is present in TOP, but not in FROMP. Check it on 2283 FROM_LV_SET. */ 2284 set_unavailable_target_for_expr (expr1, from_lv_set); 2285 } 2286 to_tailp = i.lp; 2287 2288 /* These expressions are not present in TOP. Check liveness 2289 restrictions on TO_LV_SET. */ 2290 FOR_EACH_EXPR (expr1, i, *fromp) 2291 set_unavailable_target_for_expr (expr1, to_lv_set); 2292 2293 join_distinct_sets (i.lp, &in_both_set); 2294 join_distinct_sets (to_tailp, fromp); 2295 } 2296 2297 /* Clear av_set pointed to by SETP. */ 2298 void 2299 av_set_clear (av_set_t *setp) 2300 { 2301 expr_t expr; 2302 av_set_iterator i; 2303 2304 FOR_EACH_EXPR_1 (expr, i, setp) 2305 av_set_iter_remove (&i); 2306 2307 gcc_assert (*setp == NULL); 2308 } 2309 2310 /* Leave only one non-speculative element in the SETP. */ 2311 void 2312 av_set_leave_one_nonspec (av_set_t *setp) 2313 { 2314 expr_t expr; 2315 av_set_iterator i; 2316 bool has_one_nonspec = false; 2317 2318 /* Keep all speculative exprs, and leave one non-speculative 2319 (the first one). */ 2320 FOR_EACH_EXPR_1 (expr, i, setp) 2321 { 2322 if (!EXPR_SPEC_DONE_DS (expr)) 2323 { 2324 if (has_one_nonspec) 2325 av_set_iter_remove (&i); 2326 else 2327 has_one_nonspec = true; 2328 } 2329 } 2330 } 2331 2332 /* Return the N'th element of the SET. */ 2333 expr_t 2334 av_set_element (av_set_t set, int n) 2335 { 2336 expr_t expr; 2337 av_set_iterator i; 2338 2339 FOR_EACH_EXPR (expr, i, set) 2340 if (n-- == 0) 2341 return expr; 2342 2343 gcc_unreachable (); 2344 return NULL; 2345 } 2346 2347 /* Deletes all expressions from AVP that are conditional branches (IFs). */ 2348 void 2349 av_set_substract_cond_branches (av_set_t *avp) 2350 { 2351 av_set_iterator i; 2352 expr_t expr; 2353 2354 FOR_EACH_EXPR_1 (expr, i, avp) 2355 if (vinsn_cond_branch_p (EXPR_VINSN (expr))) 2356 av_set_iter_remove (&i); 2357 } 2358 2359 /* Multiplies usefulness attribute of each member of av-set *AVP by 2360 value PROB / ALL_PROB. */ 2361 void 2362 av_set_split_usefulness (av_set_t av, int prob, int all_prob) 2363 { 2364 av_set_iterator i; 2365 expr_t expr; 2366 2367 FOR_EACH_EXPR (expr, i, av) 2368 EXPR_USEFULNESS (expr) = (all_prob 2369 ? (EXPR_USEFULNESS (expr) * prob) / all_prob 2370 : 0); 2371 } 2372 2373 /* Leave in AVP only those expressions, which are present in AV, 2374 and return it, merging history expressions. */ 2375 void 2376 av_set_code_motion_filter (av_set_t *avp, av_set_t av) 2377 { 2378 av_set_iterator i; 2379 expr_t expr, expr2; 2380 2381 FOR_EACH_EXPR_1 (expr, i, avp) 2382 if ((expr2 = av_set_lookup (av, EXPR_VINSN (expr))) == NULL) 2383 av_set_iter_remove (&i); 2384 else 2385 /* When updating av sets in bookkeeping blocks, we can add more insns 2386 there which will be transformed but the upper av sets will not 2387 reflect those transformations. We then fail to undo those 2388 when searching for such insns. So merge the history saved 2389 in the av set of the block we are processing. */ 2390 merge_history_vect (&EXPR_HISTORY_OF_CHANGES (expr), 2391 EXPR_HISTORY_OF_CHANGES (expr2)); 2392 } 2393 2394 2395 2396 /* Dependence hooks to initialize insn data. */ 2397 2398 /* This is used in hooks callable from dependence analysis when initializing 2399 instruction's data. */ 2400 static struct 2401 { 2402 /* Where the dependence was found (lhs/rhs). */ 2403 deps_where_t where; 2404 2405 /* The actual data object to initialize. */ 2406 idata_t id; 2407 2408 /* True when the insn should not be made clonable. */ 2409 bool force_unique_p; 2410 2411 /* True when insn should be treated as of type USE, i.e. never renamed. */ 2412 bool force_use_p; 2413 } deps_init_id_data; 2414 2415 2416 /* Setup ID for INSN. FORCE_UNIQUE_P is true when INSN should not be 2417 clonable. */ 2418 static void 2419 setup_id_for_insn (idata_t id, insn_t insn, bool force_unique_p) 2420 { 2421 int type; 2422 2423 /* Determine whether INSN could be cloned and return appropriate vinsn type. 2424 That clonable insns which can be separated into lhs and rhs have type SET. 2425 Other clonable insns have type USE. */ 2426 type = GET_CODE (insn); 2427 2428 /* Only regular insns could be cloned. */ 2429 if (type == INSN && !force_unique_p) 2430 type = SET; 2431 else if (type == JUMP_INSN && simplejump_p (insn)) 2432 type = PC; 2433 else if (type == DEBUG_INSN) 2434 type = !force_unique_p ? USE : INSN; 2435 2436 IDATA_TYPE (id) = type; 2437 IDATA_REG_SETS (id) = get_clear_regset_from_pool (); 2438 IDATA_REG_USES (id) = get_clear_regset_from_pool (); 2439 IDATA_REG_CLOBBERS (id) = get_clear_regset_from_pool (); 2440 } 2441 2442 /* Start initializing insn data. */ 2443 static void 2444 deps_init_id_start_insn (insn_t insn) 2445 { 2446 gcc_assert (deps_init_id_data.where == DEPS_IN_NOWHERE); 2447 2448 setup_id_for_insn (deps_init_id_data.id, insn, 2449 deps_init_id_data.force_unique_p); 2450 deps_init_id_data.where = DEPS_IN_INSN; 2451 } 2452 2453 /* Start initializing lhs data. */ 2454 static void 2455 deps_init_id_start_lhs (rtx lhs) 2456 { 2457 gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); 2458 gcc_assert (IDATA_LHS (deps_init_id_data.id) == NULL); 2459 2460 if (IDATA_TYPE (deps_init_id_data.id) == SET) 2461 { 2462 IDATA_LHS (deps_init_id_data.id) = lhs; 2463 deps_init_id_data.where = DEPS_IN_LHS; 2464 } 2465 } 2466 2467 /* Finish initializing lhs data. */ 2468 static void 2469 deps_init_id_finish_lhs (void) 2470 { 2471 deps_init_id_data.where = DEPS_IN_INSN; 2472 } 2473 2474 /* Note a set of REGNO. */ 2475 static void 2476 deps_init_id_note_reg_set (int regno) 2477 { 2478 haifa_note_reg_set (regno); 2479 2480 if (deps_init_id_data.where == DEPS_IN_RHS) 2481 deps_init_id_data.force_use_p = true; 2482 2483 if (IDATA_TYPE (deps_init_id_data.id) != PC) 2484 SET_REGNO_REG_SET (IDATA_REG_SETS (deps_init_id_data.id), regno); 2485 2486 #ifdef STACK_REGS 2487 /* Make instructions that set stack registers to be ineligible for 2488 renaming to avoid issues with find_used_regs. */ 2489 if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) 2490 deps_init_id_data.force_use_p = true; 2491 #endif 2492 } 2493 2494 /* Note a clobber of REGNO. */ 2495 static void 2496 deps_init_id_note_reg_clobber (int regno) 2497 { 2498 haifa_note_reg_clobber (regno); 2499 2500 if (deps_init_id_data.where == DEPS_IN_RHS) 2501 deps_init_id_data.force_use_p = true; 2502 2503 if (IDATA_TYPE (deps_init_id_data.id) != PC) 2504 SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (deps_init_id_data.id), regno); 2505 } 2506 2507 /* Note a use of REGNO. */ 2508 static void 2509 deps_init_id_note_reg_use (int regno) 2510 { 2511 haifa_note_reg_use (regno); 2512 2513 if (IDATA_TYPE (deps_init_id_data.id) != PC) 2514 SET_REGNO_REG_SET (IDATA_REG_USES (deps_init_id_data.id), regno); 2515 } 2516 2517 /* Start initializing rhs data. */ 2518 static void 2519 deps_init_id_start_rhs (rtx rhs) 2520 { 2521 gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); 2522 2523 /* And there was no sel_deps_reset_to_insn (). */ 2524 if (IDATA_LHS (deps_init_id_data.id) != NULL) 2525 { 2526 IDATA_RHS (deps_init_id_data.id) = rhs; 2527 deps_init_id_data.where = DEPS_IN_RHS; 2528 } 2529 } 2530 2531 /* Finish initializing rhs data. */ 2532 static void 2533 deps_init_id_finish_rhs (void) 2534 { 2535 gcc_assert (deps_init_id_data.where == DEPS_IN_RHS 2536 || deps_init_id_data.where == DEPS_IN_INSN); 2537 deps_init_id_data.where = DEPS_IN_INSN; 2538 } 2539 2540 /* Finish initializing insn data. */ 2541 static void 2542 deps_init_id_finish_insn (void) 2543 { 2544 gcc_assert (deps_init_id_data.where == DEPS_IN_INSN); 2545 2546 if (IDATA_TYPE (deps_init_id_data.id) == SET) 2547 { 2548 rtx lhs = IDATA_LHS (deps_init_id_data.id); 2549 rtx rhs = IDATA_RHS (deps_init_id_data.id); 2550 2551 if (lhs == NULL || rhs == NULL || !lhs_and_rhs_separable_p (lhs, rhs) 2552 || deps_init_id_data.force_use_p) 2553 { 2554 /* This should be a USE, as we don't want to schedule its RHS 2555 separately. However, we still want to have them recorded 2556 for the purposes of substitution. That's why we don't 2557 simply call downgrade_to_use () here. */ 2558 gcc_assert (IDATA_TYPE (deps_init_id_data.id) == SET); 2559 gcc_assert (!lhs == !rhs); 2560 2561 IDATA_TYPE (deps_init_id_data.id) = USE; 2562 } 2563 } 2564 2565 deps_init_id_data.where = DEPS_IN_NOWHERE; 2566 } 2567 2568 /* This is dependence info used for initializing insn's data. */ 2569 static struct sched_deps_info_def deps_init_id_sched_deps_info; 2570 2571 /* This initializes most of the static part of the above structure. */ 2572 static const struct sched_deps_info_def const_deps_init_id_sched_deps_info = 2573 { 2574 NULL, 2575 2576 deps_init_id_start_insn, 2577 deps_init_id_finish_insn, 2578 deps_init_id_start_lhs, 2579 deps_init_id_finish_lhs, 2580 deps_init_id_start_rhs, 2581 deps_init_id_finish_rhs, 2582 deps_init_id_note_reg_set, 2583 deps_init_id_note_reg_clobber, 2584 deps_init_id_note_reg_use, 2585 NULL, /* note_mem_dep */ 2586 NULL, /* note_dep */ 2587 2588 0, /* use_cselib */ 2589 0, /* use_deps_list */ 2590 0 /* generate_spec_deps */ 2591 }; 2592 2593 /* Initialize INSN's lhs and rhs in ID. When FORCE_UNIQUE_P is true, 2594 we don't actually need information about lhs and rhs. */ 2595 static void 2596 setup_id_lhs_rhs (idata_t id, insn_t insn, bool force_unique_p) 2597 { 2598 rtx pat = PATTERN (insn); 2599 2600 if (NONJUMP_INSN_P (insn) 2601 && GET_CODE (pat) == SET 2602 && !force_unique_p) 2603 { 2604 IDATA_RHS (id) = SET_SRC (pat); 2605 IDATA_LHS (id) = SET_DEST (pat); 2606 } 2607 else 2608 IDATA_LHS (id) = IDATA_RHS (id) = NULL; 2609 } 2610 2611 /* Possibly downgrade INSN to USE. */ 2612 static void 2613 maybe_downgrade_id_to_use (idata_t id, insn_t insn) 2614 { 2615 bool must_be_use = false; 2616 df_ref def; 2617 rtx lhs = IDATA_LHS (id); 2618 rtx rhs = IDATA_RHS (id); 2619 2620 /* We downgrade only SETs. */ 2621 if (IDATA_TYPE (id) != SET) 2622 return; 2623 2624 if (!lhs || !lhs_and_rhs_separable_p (lhs, rhs)) 2625 { 2626 IDATA_TYPE (id) = USE; 2627 return; 2628 } 2629 2630 FOR_EACH_INSN_DEF (def, insn) 2631 { 2632 if (DF_REF_INSN (def) 2633 && DF_REF_FLAGS_IS_SET (def, DF_REF_PRE_POST_MODIFY) 2634 && loc_mentioned_in_p (DF_REF_LOC (def), IDATA_RHS (id))) 2635 { 2636 must_be_use = true; 2637 break; 2638 } 2639 2640 #ifdef STACK_REGS 2641 /* Make instructions that set stack registers to be ineligible for 2642 renaming to avoid issues with find_used_regs. */ 2643 if (IN_RANGE (DF_REF_REGNO (def), FIRST_STACK_REG, LAST_STACK_REG)) 2644 { 2645 must_be_use = true; 2646 break; 2647 } 2648 #endif 2649 } 2650 2651 if (must_be_use) 2652 IDATA_TYPE (id) = USE; 2653 } 2654 2655 /* Setup implicit register clobbers calculated by sched-deps for INSN 2656 before reload and save them in ID. */ 2657 static void 2658 setup_id_implicit_regs (idata_t id, insn_t insn) 2659 { 2660 if (reload_completed) 2661 return; 2662 2663 HARD_REG_SET temp; 2664 unsigned regno; 2665 hard_reg_set_iterator hrsi; 2666 2667 get_implicit_reg_pending_clobbers (&temp, insn); 2668 EXECUTE_IF_SET_IN_HARD_REG_SET (temp, 0, regno, hrsi) 2669 SET_REGNO_REG_SET (IDATA_REG_SETS (id), regno); 2670 } 2671 2672 /* Setup register sets describing INSN in ID. */ 2673 static void 2674 setup_id_reg_sets (idata_t id, insn_t insn) 2675 { 2676 struct df_insn_info *insn_info = DF_INSN_INFO_GET (insn); 2677 df_ref def, use; 2678 regset tmp = get_clear_regset_from_pool (); 2679 2680 FOR_EACH_INSN_INFO_DEF (def, insn_info) 2681 { 2682 unsigned int regno = DF_REF_REGNO (def); 2683 2684 /* Post modifies are treated like clobbers by sched-deps.c. */ 2685 if (DF_REF_FLAGS_IS_SET (def, (DF_REF_MUST_CLOBBER 2686 | DF_REF_PRE_POST_MODIFY))) 2687 SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (id), regno); 2688 else if (! DF_REF_FLAGS_IS_SET (def, DF_REF_MAY_CLOBBER)) 2689 { 2690 SET_REGNO_REG_SET (IDATA_REG_SETS (id), regno); 2691 2692 #ifdef STACK_REGS 2693 /* For stack registers, treat writes to them as writes 2694 to the first one to be consistent with sched-deps.c. */ 2695 if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) 2696 SET_REGNO_REG_SET (IDATA_REG_SETS (id), FIRST_STACK_REG); 2697 #endif 2698 } 2699 /* Mark special refs that generate read/write def pair. */ 2700 if (DF_REF_FLAGS_IS_SET (def, DF_REF_CONDITIONAL) 2701 || regno == STACK_POINTER_REGNUM) 2702 bitmap_set_bit (tmp, regno); 2703 } 2704 2705 FOR_EACH_INSN_INFO_USE (use, insn_info) 2706 { 2707 unsigned int regno = DF_REF_REGNO (use); 2708 2709 /* When these refs are met for the first time, skip them, as 2710 these uses are just counterparts of some defs. */ 2711 if (bitmap_bit_p (tmp, regno)) 2712 bitmap_clear_bit (tmp, regno); 2713 else if (! DF_REF_FLAGS_IS_SET (use, DF_REF_CALL_STACK_USAGE)) 2714 { 2715 SET_REGNO_REG_SET (IDATA_REG_USES (id), regno); 2716 2717 #ifdef STACK_REGS 2718 /* For stack registers, treat reads from them as reads from 2719 the first one to be consistent with sched-deps.c. */ 2720 if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)) 2721 SET_REGNO_REG_SET (IDATA_REG_USES (id), FIRST_STACK_REG); 2722 #endif 2723 } 2724 } 2725 2726 /* Also get implicit reg clobbers from sched-deps. */ 2727 setup_id_implicit_regs (id, insn); 2728 2729 return_regset_to_pool (tmp); 2730 } 2731 2732 /* Initialize instruction data for INSN in ID using DF's data. */ 2733 static void 2734 init_id_from_df (idata_t id, insn_t insn, bool force_unique_p) 2735 { 2736 gcc_assert (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL); 2737 2738 setup_id_for_insn (id, insn, force_unique_p); 2739 setup_id_lhs_rhs (id, insn, force_unique_p); 2740 2741 if (INSN_NOP_P (insn)) 2742 return; 2743 2744 maybe_downgrade_id_to_use (id, insn); 2745 setup_id_reg_sets (id, insn); 2746 } 2747 2748 /* Initialize instruction data for INSN in ID. */ 2749 static void 2750 deps_init_id (idata_t id, insn_t insn, bool force_unique_p) 2751 { 2752 struct deps_desc _dc, *dc = &_dc; 2753 2754 deps_init_id_data.where = DEPS_IN_NOWHERE; 2755 deps_init_id_data.id = id; 2756 deps_init_id_data.force_unique_p = force_unique_p; 2757 deps_init_id_data.force_use_p = false; 2758 2759 init_deps (dc, false); 2760 memcpy (&deps_init_id_sched_deps_info, 2761 &const_deps_init_id_sched_deps_info, 2762 sizeof (deps_init_id_sched_deps_info)); 2763 if (spec_info != NULL) 2764 deps_init_id_sched_deps_info.generate_spec_deps = 1; 2765 sched_deps_info = &deps_init_id_sched_deps_info; 2766 2767 deps_analyze_insn (dc, insn); 2768 /* Implicit reg clobbers received from sched-deps separately. */ 2769 setup_id_implicit_regs (id, insn); 2770 2771 free_deps (dc); 2772 deps_init_id_data.id = NULL; 2773 } 2774 2775 2776 struct sched_scan_info_def 2777 { 2778 /* This hook notifies scheduler frontend to extend its internal per basic 2779 block data structures. This hook should be called once before a series of 2780 calls to bb_init (). */ 2781 void (*extend_bb) (void); 2782 2783 /* This hook makes scheduler frontend to initialize its internal data 2784 structures for the passed basic block. */ 2785 void (*init_bb) (basic_block); 2786 2787 /* This hook notifies scheduler frontend to extend its internal per insn data 2788 structures. This hook should be called once before a series of calls to 2789 insn_init (). */ 2790 void (*extend_insn) (void); 2791 2792 /* This hook makes scheduler frontend to initialize its internal data 2793 structures for the passed insn. */ 2794 void (*init_insn) (insn_t); 2795 }; 2796 2797 /* A driver function to add a set of basic blocks (BBS) to the 2798 scheduling region. */ 2799 static void 2800 sched_scan (const struct sched_scan_info_def *ssi, bb_vec_t bbs) 2801 { 2802 unsigned i; 2803 basic_block bb; 2804 2805 if (ssi->extend_bb) 2806 ssi->extend_bb (); 2807 2808 if (ssi->init_bb) 2809 FOR_EACH_VEC_ELT (bbs, i, bb) 2810 ssi->init_bb (bb); 2811 2812 if (ssi->extend_insn) 2813 ssi->extend_insn (); 2814 2815 if (ssi->init_insn) 2816 FOR_EACH_VEC_ELT (bbs, i, bb) 2817 { 2818 rtx_insn *insn; 2819 2820 FOR_BB_INSNS (bb, insn) 2821 ssi->init_insn (insn); 2822 } 2823 } 2824 2825 /* Implement hooks for collecting fundamental insn properties like if insn is 2826 an ASM or is within a SCHED_GROUP. */ 2827 2828 /* True when a "one-time init" data for INSN was already inited. */ 2829 static bool 2830 first_time_insn_init (insn_t insn) 2831 { 2832 return INSN_LIVE (insn) == NULL; 2833 } 2834 2835 /* Hash an entry in a transformed_insns hashtable. */ 2836 static hashval_t 2837 hash_transformed_insns (const void *p) 2838 { 2839 return VINSN_HASH_RTX (((const struct transformed_insns *) p)->vinsn_old); 2840 } 2841 2842 /* Compare the entries in a transformed_insns hashtable. */ 2843 static int 2844 eq_transformed_insns (const void *p, const void *q) 2845 { 2846 rtx_insn *i1 = 2847 VINSN_INSN_RTX (((const struct transformed_insns *) p)->vinsn_old); 2848 rtx_insn *i2 = 2849 VINSN_INSN_RTX (((const struct transformed_insns *) q)->vinsn_old); 2850 2851 if (INSN_UID (i1) == INSN_UID (i2)) 2852 return 1; 2853 return rtx_equal_p (PATTERN (i1), PATTERN (i2)); 2854 } 2855 2856 /* Free an entry in a transformed_insns hashtable. */ 2857 static void 2858 free_transformed_insns (void *p) 2859 { 2860 struct transformed_insns *pti = (struct transformed_insns *) p; 2861 2862 vinsn_detach (pti->vinsn_old); 2863 vinsn_detach (pti->vinsn_new); 2864 free (pti); 2865 } 2866 2867 /* Init the s_i_d data for INSN which should be inited just once, when 2868 we first see the insn. */ 2869 static void 2870 init_first_time_insn_data (insn_t insn) 2871 { 2872 /* This should not be set if this is the first time we init data for 2873 insn. */ 2874 gcc_assert (first_time_insn_init (insn)); 2875 2876 /* These are needed for nops too. */ 2877 INSN_LIVE (insn) = get_regset_from_pool (); 2878 INSN_LIVE_VALID_P (insn) = false; 2879 2880 if (!INSN_NOP_P (insn)) 2881 { 2882 INSN_ANALYZED_DEPS (insn) = BITMAP_ALLOC (NULL); 2883 INSN_FOUND_DEPS (insn) = BITMAP_ALLOC (NULL); 2884 INSN_TRANSFORMED_INSNS (insn) 2885 = htab_create (16, hash_transformed_insns, 2886 eq_transformed_insns, free_transformed_insns); 2887 init_deps (&INSN_DEPS_CONTEXT (insn), true); 2888 } 2889 } 2890 2891 /* Free almost all above data for INSN that is scheduled already. 2892 Used for extra-large basic blocks. */ 2893 void 2894 free_data_for_scheduled_insn (insn_t insn) 2895 { 2896 gcc_assert (! first_time_insn_init (insn)); 2897 2898 if (! INSN_ANALYZED_DEPS (insn)) 2899 return; 2900 2901 BITMAP_FREE (INSN_ANALYZED_DEPS (insn)); 2902 BITMAP_FREE (INSN_FOUND_DEPS (insn)); 2903 htab_delete (INSN_TRANSFORMED_INSNS (insn)); 2904 2905 /* This is allocated only for bookkeeping insns. */ 2906 if (INSN_ORIGINATORS (insn)) 2907 BITMAP_FREE (INSN_ORIGINATORS (insn)); 2908 free_deps (&INSN_DEPS_CONTEXT (insn)); 2909 2910 INSN_ANALYZED_DEPS (insn) = NULL; 2911 2912 /* Clear the readonly flag so we would ICE when trying to recalculate 2913 the deps context (as we believe that it should not happen). */ 2914 (&INSN_DEPS_CONTEXT (insn))->readonly = 0; 2915 } 2916 2917 /* Free the same data as above for INSN. */ 2918 static void 2919 free_first_time_insn_data (insn_t insn) 2920 { 2921 gcc_assert (! first_time_insn_init (insn)); 2922 2923 free_data_for_scheduled_insn (insn); 2924 return_regset_to_pool (INSN_LIVE (insn)); 2925 INSN_LIVE (insn) = NULL; 2926 INSN_LIVE_VALID_P (insn) = false; 2927 } 2928 2929 /* Initialize region-scope data structures for basic blocks. */ 2930 static void 2931 init_global_and_expr_for_bb (basic_block bb) 2932 { 2933 if (sel_bb_empty_p (bb)) 2934 return; 2935 2936 invalidate_av_set (bb); 2937 } 2938 2939 /* Data for global dependency analysis (to initialize CANT_MOVE and 2940 SCHED_GROUP_P). */ 2941 static struct 2942 { 2943 /* Previous insn. */ 2944 insn_t prev_insn; 2945 } init_global_data; 2946 2947 /* Determine if INSN is in the sched_group, is an asm or should not be 2948 cloned. After that initialize its expr. */ 2949 static void 2950 init_global_and_expr_for_insn (insn_t insn) 2951 { 2952 if (LABEL_P (insn)) 2953 return; 2954 2955 if (NOTE_INSN_BASIC_BLOCK_P (insn)) 2956 { 2957 init_global_data.prev_insn = NULL; 2958 return; 2959 } 2960 2961 gcc_assert (INSN_P (insn)); 2962 2963 if (SCHED_GROUP_P (insn)) 2964 /* Setup a sched_group. */ 2965 { 2966 insn_t prev_insn = init_global_data.prev_insn; 2967 2968 if (prev_insn) 2969 INSN_SCHED_NEXT (prev_insn) = insn; 2970 2971 init_global_data.prev_insn = insn; 2972 } 2973 else 2974 init_global_data.prev_insn = NULL; 2975 2976 if (GET_CODE (PATTERN (insn)) == ASM_INPUT 2977 || asm_noperands (PATTERN (insn)) >= 0) 2978 /* Mark INSN as an asm. */ 2979 INSN_ASM_P (insn) = true; 2980 2981 { 2982 bool force_unique_p; 2983 ds_t spec_done_ds; 2984 2985 /* Certain instructions cannot be cloned, and frame related insns and 2986 the insn adjacent to NOTE_INSN_EPILOGUE_BEG cannot be moved out of 2987 their block. */ 2988 if (prologue_epilogue_contains (insn)) 2989 { 2990 if (RTX_FRAME_RELATED_P (insn)) 2991 CANT_MOVE (insn) = 1; 2992 else 2993 { 2994 rtx note; 2995 for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) 2996 if (REG_NOTE_KIND (note) == REG_SAVE_NOTE 2997 && ((enum insn_note) INTVAL (XEXP (note, 0)) 2998 == NOTE_INSN_EPILOGUE_BEG)) 2999 { 3000 CANT_MOVE (insn) = 1; 3001 break; 3002 } 3003 } 3004 force_unique_p = true; 3005 } 3006 else 3007 if (CANT_MOVE (insn) 3008 || INSN_ASM_P (insn) 3009 || SCHED_GROUP_P (insn) 3010 || CALL_P (insn) 3011 /* Exception handling insns are always unique. */ 3012 || (cfun->can_throw_non_call_exceptions && can_throw_internal (insn)) 3013 /* TRAP_IF though have an INSN code is control_flow_insn_p (). */ 3014 || control_flow_insn_p (insn) 3015 || volatile_insn_p (PATTERN (insn)) 3016 || (targetm.cannot_copy_insn_p 3017 && targetm.cannot_copy_insn_p (insn))) 3018 force_unique_p = true; 3019 else 3020 force_unique_p = false; 3021 3022 if (targetm.sched.get_insn_spec_ds) 3023 { 3024 spec_done_ds = targetm.sched.get_insn_spec_ds (insn); 3025 spec_done_ds = ds_get_max_dep_weak (spec_done_ds); 3026 } 3027 else 3028 spec_done_ds = 0; 3029 3030 /* Initialize INSN's expr. */ 3031 init_expr (INSN_EXPR (insn), vinsn_create (insn, force_unique_p), 0, 3032 REG_BR_PROB_BASE, INSN_PRIORITY (insn), 0, BLOCK_NUM (insn), 3033 spec_done_ds, 0, 0, vNULL, true, 3034 false, false, false, CANT_MOVE (insn)); 3035 } 3036 3037 init_first_time_insn_data (insn); 3038 } 3039 3040 /* Scan the region and initialize instruction data for basic blocks BBS. */ 3041 void 3042 sel_init_global_and_expr (bb_vec_t bbs) 3043 { 3044 /* ??? It would be nice to implement push / pop scheme for sched_infos. */ 3045 const struct sched_scan_info_def ssi = 3046 { 3047 NULL, /* extend_bb */ 3048 init_global_and_expr_for_bb, /* init_bb */ 3049 extend_insn_data, /* extend_insn */ 3050 init_global_and_expr_for_insn /* init_insn */ 3051 }; 3052 3053 sched_scan (&ssi, bbs); 3054 } 3055 3056 /* Finalize region-scope data structures for basic blocks. */ 3057 static void 3058 finish_global_and_expr_for_bb (basic_block bb) 3059 { 3060 av_set_clear (&BB_AV_SET (bb)); 3061 BB_AV_LEVEL (bb) = 0; 3062 } 3063 3064 /* Finalize INSN's data. */ 3065 static void 3066 finish_global_and_expr_insn (insn_t insn) 3067 { 3068 if (LABEL_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn)) 3069 return; 3070 3071 gcc_assert (INSN_P (insn)); 3072 3073 if (INSN_LUID (insn) > 0) 3074 { 3075 free_first_time_insn_data (insn); 3076 INSN_WS_LEVEL (insn) = 0; 3077 CANT_MOVE (insn) = 0; 3078 3079 /* We can no longer assert this, as vinsns of this insn could be 3080 easily live in other insn's caches. This should be changed to 3081 a counter-like approach among all vinsns. */ 3082 gcc_assert (true || VINSN_COUNT (INSN_VINSN (insn)) == 1); 3083 clear_expr (INSN_EXPR (insn)); 3084 } 3085 } 3086 3087 /* Finalize per instruction data for the whole region. */ 3088 void 3089 sel_finish_global_and_expr (void) 3090 { 3091 { 3092 bb_vec_t bbs; 3093 int i; 3094 3095 bbs.create (current_nr_blocks); 3096 3097 for (i = 0; i < current_nr_blocks; i++) 3098 bbs.quick_push (BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i))); 3099 3100 /* Clear AV_SETs and INSN_EXPRs. */ 3101 { 3102 const struct sched_scan_info_def ssi = 3103 { 3104 NULL, /* extend_bb */ 3105 finish_global_and_expr_for_bb, /* init_bb */ 3106 NULL, /* extend_insn */ 3107 finish_global_and_expr_insn /* init_insn */ 3108 }; 3109 3110 sched_scan (&ssi, bbs); 3111 } 3112 3113 bbs.release (); 3114 } 3115 3116 finish_insns (); 3117 } 3118 3119 3120 /* In the below hooks, we merely calculate whether or not a dependence 3121 exists, and in what part of insn. However, we will need more data 3122 when we'll start caching dependence requests. */ 3123 3124 /* Container to hold information for dependency analysis. */ 3125 static struct 3126 { 3127 deps_t dc; 3128 3129 /* A variable to track which part of rtx we are scanning in 3130 sched-deps.c: sched_analyze_insn (). */ 3131 deps_where_t where; 3132 3133 /* Current producer. */ 3134 insn_t pro; 3135 3136 /* Current consumer. */ 3137 vinsn_t con; 3138 3139 /* Is SEL_DEPS_HAS_DEP_P[DEPS_IN_X] is true, then X has a dependence. 3140 X is from { INSN, LHS, RHS }. */ 3141 ds_t has_dep_p[DEPS_IN_NOWHERE]; 3142 } has_dependence_data; 3143 3144 /* Start analyzing dependencies of INSN. */ 3145 static void 3146 has_dependence_start_insn (insn_t insn ATTRIBUTE_UNUSED) 3147 { 3148 gcc_assert (has_dependence_data.where == DEPS_IN_NOWHERE); 3149 3150 has_dependence_data.where = DEPS_IN_INSN; 3151 } 3152 3153 /* Finish analyzing dependencies of an insn. */ 3154 static void 3155 has_dependence_finish_insn (void) 3156 { 3157 gcc_assert (has_dependence_data.where == DEPS_IN_INSN); 3158 3159 has_dependence_data.where = DEPS_IN_NOWHERE; 3160 } 3161 3162 /* Start analyzing dependencies of LHS. */ 3163 static void 3164 has_dependence_start_lhs (rtx lhs ATTRIBUTE_UNUSED) 3165 { 3166 gcc_assert (has_dependence_data.where == DEPS_IN_INSN); 3167 3168 if (VINSN_LHS (has_dependence_data.con) != NULL) 3169 has_dependence_data.where = DEPS_IN_LHS; 3170 } 3171 3172 /* Finish analyzing dependencies of an lhs. */ 3173 static void 3174 has_dependence_finish_lhs (void) 3175 { 3176 has_dependence_data.where = DEPS_IN_INSN; 3177 } 3178 3179 /* Start analyzing dependencies of RHS. */ 3180 static void 3181 has_dependence_start_rhs (rtx rhs ATTRIBUTE_UNUSED) 3182 { 3183 gcc_assert (has_dependence_data.where == DEPS_IN_INSN); 3184 3185 if (VINSN_RHS (has_dependence_data.con) != NULL) 3186 has_dependence_data.where = DEPS_IN_RHS; 3187 } 3188 3189 /* Start analyzing dependencies of an rhs. */ 3190 static void 3191 has_dependence_finish_rhs (void) 3192 { 3193 gcc_assert (has_dependence_data.where == DEPS_IN_RHS 3194 || has_dependence_data.where == DEPS_IN_INSN); 3195 3196 has_dependence_data.where = DEPS_IN_INSN; 3197 } 3198 3199 /* Note a set of REGNO. */ 3200 static void 3201 has_dependence_note_reg_set (int regno) 3202 { 3203 struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; 3204 3205 if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, 3206 VINSN_INSN_RTX 3207 (has_dependence_data.con))) 3208 { 3209 ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; 3210 3211 if (reg_last->sets != NULL 3212 || reg_last->clobbers != NULL) 3213 *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT; 3214 3215 if (reg_last->uses || reg_last->implicit_sets) 3216 *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; 3217 } 3218 } 3219 3220 /* Note a clobber of REGNO. */ 3221 static void 3222 has_dependence_note_reg_clobber (int regno) 3223 { 3224 struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; 3225 3226 if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, 3227 VINSN_INSN_RTX 3228 (has_dependence_data.con))) 3229 { 3230 ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; 3231 3232 if (reg_last->sets) 3233 *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT; 3234 3235 if (reg_last->uses || reg_last->implicit_sets) 3236 *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; 3237 } 3238 } 3239 3240 /* Note a use of REGNO. */ 3241 static void 3242 has_dependence_note_reg_use (int regno) 3243 { 3244 struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno]; 3245 3246 if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, 3247 VINSN_INSN_RTX 3248 (has_dependence_data.con))) 3249 { 3250 ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; 3251 3252 if (reg_last->sets) 3253 *dsp = (*dsp & ~SPECULATIVE) | DEP_TRUE; 3254 3255 if (reg_last->clobbers || reg_last->implicit_sets) 3256 *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI; 3257 3258 /* Merge BE_IN_SPEC bits into *DSP when the dependency producer 3259 is actually a check insn. We need to do this for any register 3260 read-read dependency with the check unless we track properly 3261 all registers written by BE_IN_SPEC-speculated insns, as 3262 we don't have explicit dependence lists. See PR 53975. */ 3263 if (reg_last->uses) 3264 { 3265 ds_t pro_spec_checked_ds; 3266 3267 pro_spec_checked_ds = INSN_SPEC_CHECKED_DS (has_dependence_data.pro); 3268 pro_spec_checked_ds = ds_get_max_dep_weak (pro_spec_checked_ds); 3269 3270 if (pro_spec_checked_ds != 0) 3271 *dsp = ds_full_merge (*dsp, pro_spec_checked_ds, 3272 NULL_RTX, NULL_RTX); 3273 } 3274 } 3275 } 3276 3277 /* Note a memory dependence. */ 3278 static void 3279 has_dependence_note_mem_dep (rtx mem ATTRIBUTE_UNUSED, 3280 rtx pending_mem ATTRIBUTE_UNUSED, 3281 insn_t pending_insn ATTRIBUTE_UNUSED, 3282 ds_t ds ATTRIBUTE_UNUSED) 3283 { 3284 if (!sched_insns_conditions_mutex_p (has_dependence_data.pro, 3285 VINSN_INSN_RTX (has_dependence_data.con))) 3286 { 3287 ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; 3288 3289 *dsp = ds_full_merge (ds, *dsp, pending_mem, mem); 3290 } 3291 } 3292 3293 /* Note a dependence. */ 3294 static void 3295 has_dependence_note_dep (insn_t pro, ds_t ds ATTRIBUTE_UNUSED) 3296 { 3297 insn_t real_pro = has_dependence_data.pro; 3298 insn_t real_con = VINSN_INSN_RTX (has_dependence_data.con); 3299 3300 /* We do not allow for debug insns to move through others unless they 3301 are at the start of bb. This movement may create bookkeeping copies 3302 that later would not be able to move up, violating the invariant 3303 that a bookkeeping copy should be movable as the original insn. 3304 Detect that here and allow that movement if we allowed it before 3305 in the first place. */ 3306 if (DEBUG_INSN_P (real_con) && !DEBUG_INSN_P (real_pro) 3307 && INSN_UID (NEXT_INSN (pro)) == INSN_UID (real_con)) 3308 return; 3309 3310 if (!sched_insns_conditions_mutex_p (real_pro, real_con)) 3311 { 3312 ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where]; 3313 3314 *dsp = ds_full_merge (ds, *dsp, NULL_RTX, NULL_RTX); 3315 } 3316 } 3317 3318 /* Mark the insn as having a hard dependence that prevents speculation. */ 3319 void 3320 sel_mark_hard_insn (rtx insn) 3321 { 3322 int i; 3323 3324 /* Only work when we're in has_dependence_p mode. 3325 ??? This is a hack, this should actually be a hook. */ 3326 if (!has_dependence_data.dc || !has_dependence_data.pro) 3327 return; 3328 3329 gcc_assert (insn == VINSN_INSN_RTX (has_dependence_data.con)); 3330 gcc_assert (has_dependence_data.where == DEPS_IN_INSN); 3331 3332 for (i = 0; i < DEPS_IN_NOWHERE; i++) 3333 has_dependence_data.has_dep_p[i] &= ~SPECULATIVE; 3334 } 3335 3336 /* This structure holds the hooks for the dependency analysis used when 3337 actually processing dependencies in the scheduler. */ 3338 static struct sched_deps_info_def has_dependence_sched_deps_info; 3339 3340 /* This initializes most of the fields of the above structure. */ 3341 static const struct sched_deps_info_def const_has_dependence_sched_deps_info = 3342 { 3343 NULL, 3344 3345 has_dependence_start_insn, 3346 has_dependence_finish_insn, 3347 has_dependence_start_lhs, 3348 has_dependence_finish_lhs, 3349 has_dependence_start_rhs, 3350 has_dependence_finish_rhs, 3351 has_dependence_note_reg_set, 3352 has_dependence_note_reg_clobber, 3353 has_dependence_note_reg_use, 3354 has_dependence_note_mem_dep, 3355 has_dependence_note_dep, 3356 3357 0, /* use_cselib */ 3358 0, /* use_deps_list */ 3359 0 /* generate_spec_deps */ 3360 }; 3361 3362 /* Initialize has_dependence_sched_deps_info with extra spec field. */ 3363 static void 3364 setup_has_dependence_sched_deps_info (void) 3365 { 3366 memcpy (&has_dependence_sched_deps_info, 3367 &const_has_dependence_sched_deps_info, 3368 sizeof (has_dependence_sched_deps_info)); 3369 3370 if (spec_info != NULL) 3371 has_dependence_sched_deps_info.generate_spec_deps = 1; 3372 3373 sched_deps_info = &has_dependence_sched_deps_info; 3374 } 3375 3376 /* Remove all dependences found and recorded in has_dependence_data array. */ 3377 void 3378 sel_clear_has_dependence (void) 3379 { 3380 int i; 3381 3382 for (i = 0; i < DEPS_IN_NOWHERE; i++) 3383 has_dependence_data.has_dep_p[i] = 0; 3384 } 3385 3386 /* Return nonzero if EXPR has is dependent upon PRED. Return the pointer 3387 to the dependence information array in HAS_DEP_PP. */ 3388 ds_t 3389 has_dependence_p (expr_t expr, insn_t pred, ds_t **has_dep_pp) 3390 { 3391 int i; 3392 ds_t ds; 3393 struct deps_desc *dc; 3394 3395 if (INSN_SIMPLEJUMP_P (pred)) 3396 /* Unconditional jump is just a transfer of control flow. 3397 Ignore it. */ 3398 return false; 3399 3400 dc = &INSN_DEPS_CONTEXT (pred); 3401 3402 /* We init this field lazily. */ 3403 if (dc->reg_last == NULL) 3404 init_deps_reg_last (dc); 3405 3406 if (!dc->readonly) 3407 { 3408 has_dependence_data.pro = NULL; 3409 /* Initialize empty dep context with information about PRED. */ 3410 advance_deps_context (dc, pred); 3411 dc->readonly = 1; 3412 } 3413 3414 has_dependence_data.where = DEPS_IN_NOWHERE; 3415 has_dependence_data.pro = pred; 3416 has_dependence_data.con = EXPR_VINSN (expr); 3417 has_dependence_data.dc = dc; 3418 3419 sel_clear_has_dependence (); 3420 3421 /* Now catch all dependencies that would be generated between PRED and 3422 INSN. */ 3423 setup_has_dependence_sched_deps_info (); 3424 deps_analyze_insn (dc, EXPR_INSN_RTX (expr)); 3425 has_dependence_data.dc = NULL; 3426 3427 /* When a barrier was found, set DEPS_IN_INSN bits. */ 3428 if (dc->last_reg_pending_barrier == TRUE_BARRIER) 3429 has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_TRUE; 3430 else if (dc->last_reg_pending_barrier == MOVE_BARRIER) 3431 has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI; 3432 3433 /* Do not allow stores to memory to move through checks. Currently 3434 we don't move this to sched-deps.c as the check doesn't have 3435 obvious places to which this dependence can be attached. 3436 FIMXE: this should go to a hook. */ 3437 if (EXPR_LHS (expr) 3438 && MEM_P (EXPR_LHS (expr)) 3439 && sel_insn_is_speculation_check (pred)) 3440 has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI; 3441 3442 *has_dep_pp = has_dependence_data.has_dep_p; 3443 ds = 0; 3444 for (i = 0; i < DEPS_IN_NOWHERE; i++) 3445 ds = ds_full_merge (ds, has_dependence_data.has_dep_p[i], 3446 NULL_RTX, NULL_RTX); 3447 3448 return ds; 3449 } 3450 3451 3452 /* Dependence hooks implementation that checks dependence latency constraints 3453 on the insns being scheduled. The entry point for these routines is 3454 tick_check_p predicate. */ 3455 3456 static struct 3457 { 3458 /* An expr we are currently checking. */ 3459 expr_t expr; 3460 3461 /* A minimal cycle for its scheduling. */ 3462 int cycle; 3463 3464 /* Whether we have seen a true dependence while checking. */ 3465 bool seen_true_dep_p; 3466 } tick_check_data; 3467 3468 /* Update minimal scheduling cycle for tick_check_insn given that it depends 3469 on PRO with status DS and weight DW. */ 3470 static void 3471 tick_check_dep_with_dw (insn_t pro_insn, ds_t ds, dw_t dw) 3472 { 3473 expr_t con_expr = tick_check_data.expr; 3474 insn_t con_insn = EXPR_INSN_RTX (con_expr); 3475 3476 if (con_insn != pro_insn) 3477 { 3478 enum reg_note dt; 3479 int tick; 3480 3481 if (/* PROducer was removed from above due to pipelining. */ 3482 !INSN_IN_STREAM_P (pro_insn) 3483 /* Or PROducer was originally on the next iteration regarding the 3484 CONsumer. */ 3485 || (INSN_SCHED_TIMES (pro_insn) 3486 - EXPR_SCHED_TIMES (con_expr)) > 1) 3487 /* Don't count this dependence. */ 3488 return; 3489 3490 dt = ds_to_dt (ds); 3491 if (dt == REG_DEP_TRUE) 3492 tick_check_data.seen_true_dep_p = true; 3493 3494 gcc_assert (INSN_SCHED_CYCLE (pro_insn) > 0); 3495 3496 { 3497 dep_def _dep, *dep = &_dep; 3498 3499 init_dep (dep, pro_insn, con_insn, dt); 3500 3501 tick = INSN_SCHED_CYCLE (pro_insn) + dep_cost_1 (dep, dw); 3502 } 3503 3504 /* When there are several kinds of dependencies between pro and con, 3505 only REG_DEP_TRUE should be taken into account. */ 3506 if (tick > tick_check_data.cycle 3507 && (dt == REG_DEP_TRUE || !tick_check_data.seen_true_dep_p)) 3508 tick_check_data.cycle = tick; 3509 } 3510 } 3511 3512 /* An implementation of note_dep hook. */ 3513 static void 3514 tick_check_note_dep (insn_t pro, ds_t ds) 3515 { 3516 tick_check_dep_with_dw (pro, ds, 0); 3517 } 3518 3519 /* An implementation of note_mem_dep hook. */ 3520 static void 3521 tick_check_note_mem_dep (rtx mem1, rtx mem2, insn_t pro, ds_t ds) 3522 { 3523 dw_t dw; 3524 3525 dw = (ds_to_dt (ds) == REG_DEP_TRUE 3526 ? estimate_dep_weak (mem1, mem2) 3527 : 0); 3528 3529 tick_check_dep_with_dw (pro, ds, dw); 3530 } 3531 3532 /* This structure contains hooks for dependence analysis used when determining 3533 whether an insn is ready for scheduling. */ 3534 static struct sched_deps_info_def tick_check_sched_deps_info = 3535 { 3536 NULL, 3537 3538 NULL, 3539 NULL, 3540 NULL, 3541 NULL, 3542 NULL, 3543 NULL, 3544 haifa_note_reg_set, 3545 haifa_note_reg_clobber, 3546 haifa_note_reg_use, 3547 tick_check_note_mem_dep, 3548 tick_check_note_dep, 3549 3550 0, 0, 0 3551 }; 3552 3553 /* Estimate number of cycles from the current cycle of FENCE until EXPR can be 3554 scheduled. Return 0 if all data from producers in DC is ready. */ 3555 int 3556 tick_check_p (expr_t expr, deps_t dc, fence_t fence) 3557 { 3558 int cycles_left; 3559 /* Initialize variables. */ 3560 tick_check_data.expr = expr; 3561 tick_check_data.cycle = 0; 3562 tick_check_data.seen_true_dep_p = false; 3563 sched_deps_info = &tick_check_sched_deps_info; 3564 3565 gcc_assert (!dc->readonly); 3566 dc->readonly = 1; 3567 deps_analyze_insn (dc, EXPR_INSN_RTX (expr)); 3568 dc->readonly = 0; 3569 3570 cycles_left = tick_check_data.cycle - FENCE_CYCLE (fence); 3571 3572 return cycles_left >= 0 ? cycles_left : 0; 3573 } 3574 3575 3576 /* Functions to work with insns. */ 3577 3578 /* Returns true if LHS of INSN is the same as DEST of an insn 3579 being moved. */ 3580 bool 3581 lhs_of_insn_equals_to_dest_p (insn_t insn, rtx dest) 3582 { 3583 rtx lhs = INSN_LHS (insn); 3584 3585 if (lhs == NULL || dest == NULL) 3586 return false; 3587 3588 return rtx_equal_p (lhs, dest); 3589 } 3590 3591 /* Return s_i_d entry of INSN. Callable from debugger. */ 3592 sel_insn_data_def 3593 insn_sid (insn_t insn) 3594 { 3595 return *SID (insn); 3596 } 3597 3598 /* True when INSN is a speculative check. We can tell this by looking 3599 at the data structures of the selective scheduler, not by examining 3600 the pattern. */ 3601 bool 3602 sel_insn_is_speculation_check (rtx insn) 3603 { 3604 return s_i_d.exists () && !! INSN_SPEC_CHECKED_DS (insn); 3605 } 3606 3607 /* Extracts machine mode MODE and destination location DST_LOC 3608 for given INSN. */ 3609 void 3610 get_dest_and_mode (rtx insn, rtx *dst_loc, machine_mode *mode) 3611 { 3612 rtx pat = PATTERN (insn); 3613 3614 gcc_assert (dst_loc); 3615 gcc_assert (GET_CODE (pat) == SET); 3616 3617 *dst_loc = SET_DEST (pat); 3618 3619 gcc_assert (*dst_loc); 3620 gcc_assert (MEM_P (*dst_loc) || REG_P (*dst_loc)); 3621 3622 if (mode) 3623 *mode = GET_MODE (*dst_loc); 3624 } 3625 3626 /* Returns true when moving through JUMP will result in bookkeeping 3627 creation. */ 3628 bool 3629 bookkeeping_can_be_created_if_moved_through_p (insn_t jump) 3630 { 3631 insn_t succ; 3632 succ_iterator si; 3633 3634 FOR_EACH_SUCC (succ, si, jump) 3635 if (sel_num_cfg_preds_gt_1 (succ)) 3636 return true; 3637 3638 return false; 3639 } 3640 3641 /* Return 'true' if INSN is the only one in its basic block. */ 3642 static bool 3643 insn_is_the_only_one_in_bb_p (insn_t insn) 3644 { 3645 return sel_bb_head_p (insn) && sel_bb_end_p (insn); 3646 } 3647 3648 /* Check that the region we're scheduling still has at most one 3649 backedge. */ 3650 static void 3651 verify_backedges (void) 3652 { 3653 if (pipelining_p) 3654 { 3655 int i, n = 0; 3656 edge e; 3657 edge_iterator ei; 3658 3659 for (i = 0; i < current_nr_blocks; i++) 3660 FOR_EACH_EDGE (e, ei, BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i))->succs) 3661 if (in_current_region_p (e->dest) 3662 && BLOCK_TO_BB (e->dest->index) < i) 3663 n++; 3664 3665 gcc_assert (n <= 1); 3666 } 3667 } 3668 3669 3670 /* Functions to work with control flow. */ 3671 3672 /* Recompute BLOCK_TO_BB and BB_FOR_BLOCK for current region so that blocks 3673 are sorted in topological order (it might have been invalidated by 3674 redirecting an edge). */ 3675 static void 3676 sel_recompute_toporder (void) 3677 { 3678 int i, n, rgn; 3679 int *postorder, n_blocks; 3680 3681 postorder = XALLOCAVEC (int, n_basic_blocks_for_fn (cfun)); 3682 n_blocks = post_order_compute (postorder, false, false); 3683 3684 rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); 3685 for (n = 0, i = n_blocks - 1; i >= 0; i--) 3686 if (CONTAINING_RGN (postorder[i]) == rgn) 3687 { 3688 BLOCK_TO_BB (postorder[i]) = n; 3689 BB_TO_BLOCK (n) = postorder[i]; 3690 n++; 3691 } 3692 3693 /* Assert that we updated info for all blocks. We may miss some blocks if 3694 this function is called when redirecting an edge made a block 3695 unreachable, but that block is not deleted yet. */ 3696 gcc_assert (n == RGN_NR_BLOCKS (rgn)); 3697 } 3698 3699 /* Tidy the possibly empty block BB. */ 3700 static bool 3701 maybe_tidy_empty_bb (basic_block bb) 3702 { 3703 basic_block succ_bb, pred_bb, note_bb; 3704 vec<basic_block> dom_bbs; 3705 edge e; 3706 edge_iterator ei; 3707 bool rescan_p; 3708 3709 /* Keep empty bb only if this block immediately precedes EXIT and 3710 has incoming non-fallthrough edge, or it has no predecessors or 3711 successors. Otherwise remove it. */ 3712 if (!sel_bb_empty_p (bb) 3713 || (single_succ_p (bb) 3714 && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun) 3715 && (!single_pred_p (bb) 3716 || !(single_pred_edge (bb)->flags & EDGE_FALLTHRU))) 3717 || EDGE_COUNT (bb->preds) == 0 3718 || EDGE_COUNT (bb->succs) == 0) 3719 return false; 3720 3721 /* Do not attempt to redirect complex edges. */ 3722 FOR_EACH_EDGE (e, ei, bb->preds) 3723 if (e->flags & EDGE_COMPLEX) 3724 return false; 3725 else if (e->flags & EDGE_FALLTHRU) 3726 { 3727 rtx note; 3728 /* If prev bb ends with asm goto, see if any of the 3729 ASM_OPERANDS_LABELs don't point to the fallthru 3730 label. Do not attempt to redirect it in that case. */ 3731 if (JUMP_P (BB_END (e->src)) 3732 && (note = extract_asm_operands (PATTERN (BB_END (e->src))))) 3733 { 3734 int i, n = ASM_OPERANDS_LABEL_LENGTH (note); 3735 3736 for (i = 0; i < n; ++i) 3737 if (XEXP (ASM_OPERANDS_LABEL (note, i), 0) == BB_HEAD (bb)) 3738 return false; 3739 } 3740 } 3741 3742 free_data_sets (bb); 3743 3744 /* Do not delete BB if it has more than one successor. 3745 That can occur when we moving a jump. */ 3746 if (!single_succ_p (bb)) 3747 { 3748 gcc_assert (can_merge_blocks_p (bb->prev_bb, bb)); 3749 sel_merge_blocks (bb->prev_bb, bb); 3750 return true; 3751 } 3752 3753 succ_bb = single_succ (bb); 3754 rescan_p = true; 3755 pred_bb = NULL; 3756 dom_bbs.create (0); 3757 3758 /* Save a pred/succ from the current region to attach the notes to. */ 3759 note_bb = NULL; 3760 FOR_EACH_EDGE (e, ei, bb->preds) 3761 if (in_current_region_p (e->src)) 3762 { 3763 note_bb = e->src; 3764 break; 3765 } 3766 if (note_bb == NULL) 3767 note_bb = succ_bb; 3768 3769 /* Redirect all non-fallthru edges to the next bb. */ 3770 while (rescan_p) 3771 { 3772 rescan_p = false; 3773 3774 FOR_EACH_EDGE (e, ei, bb->preds) 3775 { 3776 pred_bb = e->src; 3777 3778 if (!(e->flags & EDGE_FALLTHRU)) 3779 { 3780 /* We cannot invalidate computed topological order by moving 3781 the edge destination block (E->SUCC) along a fallthru edge. 3782 3783 We will update dominators here only when we'll get 3784 an unreachable block when redirecting, otherwise 3785 sel_redirect_edge_and_branch will take care of it. */ 3786 if (e->dest != bb 3787 && single_pred_p (e->dest)) 3788 dom_bbs.safe_push (e->dest); 3789 sel_redirect_edge_and_branch (e, succ_bb); 3790 rescan_p = true; 3791 break; 3792 } 3793 /* If the edge is fallthru, but PRED_BB ends in a conditional jump 3794 to BB (so there is no non-fallthru edge from PRED_BB to BB), we 3795 still have to adjust it. */ 3796 else if (single_succ_p (pred_bb) && any_condjump_p (BB_END (pred_bb))) 3797 { 3798 /* If possible, try to remove the unneeded conditional jump. */ 3799 if (INSN_SCHED_TIMES (BB_END (pred_bb)) == 0 3800 && !IN_CURRENT_FENCE_P (BB_END (pred_bb))) 3801 { 3802 if (!sel_remove_insn (BB_END (pred_bb), false, false)) 3803 tidy_fallthru_edge (e); 3804 } 3805 else 3806 sel_redirect_edge_and_branch (e, succ_bb); 3807 rescan_p = true; 3808 break; 3809 } 3810 } 3811 } 3812 3813 if (can_merge_blocks_p (bb->prev_bb, bb)) 3814 sel_merge_blocks (bb->prev_bb, bb); 3815 else 3816 { 3817 /* This is a block without fallthru predecessor. Just delete it. */ 3818 gcc_assert (note_bb); 3819 move_bb_info (note_bb, bb); 3820 remove_empty_bb (bb, true); 3821 } 3822 3823 if (!dom_bbs.is_empty ()) 3824 { 3825 dom_bbs.safe_push (succ_bb); 3826 iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, false); 3827 dom_bbs.release (); 3828 } 3829 3830 return true; 3831 } 3832 3833 /* Tidy the control flow after we have removed original insn from 3834 XBB. Return true if we have removed some blocks. When FULL_TIDYING 3835 is true, also try to optimize control flow on non-empty blocks. */ 3836 bool 3837 tidy_control_flow (basic_block xbb, bool full_tidying) 3838 { 3839 bool changed = true; 3840 insn_t first, last; 3841 3842 /* First check whether XBB is empty. */ 3843 changed = maybe_tidy_empty_bb (xbb); 3844 if (changed || !full_tidying) 3845 return changed; 3846 3847 /* Check if there is a unnecessary jump after insn left. */ 3848 if (bb_has_removable_jump_to_p (xbb, xbb->next_bb) 3849 && INSN_SCHED_TIMES (BB_END (xbb)) == 0 3850 && !IN_CURRENT_FENCE_P (BB_END (xbb))) 3851 { 3852 /* We used to call sel_remove_insn here that can trigger tidy_control_flow 3853 before we fix up the fallthru edge. Correct that ordering by 3854 explicitly doing the latter before the former. */ 3855 clear_expr (INSN_EXPR (BB_END (xbb))); 3856 tidy_fallthru_edge (EDGE_SUCC (xbb, 0)); 3857 if (tidy_control_flow (xbb, false)) 3858 return true; 3859 } 3860 3861 first = sel_bb_head (xbb); 3862 last = sel_bb_end (xbb); 3863 if (MAY_HAVE_DEBUG_INSNS) 3864 { 3865 if (first != last && DEBUG_INSN_P (first)) 3866 do 3867 first = NEXT_INSN (first); 3868 while (first != last && (DEBUG_INSN_P (first) || NOTE_P (first))); 3869 3870 if (first != last && DEBUG_INSN_P (last)) 3871 do 3872 last = PREV_INSN (last); 3873 while (first != last && (DEBUG_INSN_P (last) || NOTE_P (last))); 3874 } 3875 /* Check if there is an unnecessary jump in previous basic block leading 3876 to next basic block left after removing INSN from stream. 3877 If it is so, remove that jump and redirect edge to current 3878 basic block (where there was INSN before deletion). This way 3879 when NOP will be deleted several instructions later with its 3880 basic block we will not get a jump to next instruction, which 3881 can be harmful. */ 3882 if (first == last 3883 && !sel_bb_empty_p (xbb) 3884 && INSN_NOP_P (last) 3885 /* Flow goes fallthru from current block to the next. */ 3886 && EDGE_COUNT (xbb->succs) == 1 3887 && (EDGE_SUCC (xbb, 0)->flags & EDGE_FALLTHRU) 3888 /* When successor is an EXIT block, it may not be the next block. */ 3889 && single_succ (xbb) != EXIT_BLOCK_PTR_FOR_FN (cfun) 3890 /* And unconditional jump in previous basic block leads to 3891 next basic block of XBB and this jump can be safely removed. */ 3892 && in_current_region_p (xbb->prev_bb) 3893 && bb_has_removable_jump_to_p (xbb->prev_bb, xbb->next_bb) 3894 && INSN_SCHED_TIMES (BB_END (xbb->prev_bb)) == 0 3895 /* Also this jump is not at the scheduling boundary. */ 3896 && !IN_CURRENT_FENCE_P (BB_END (xbb->prev_bb))) 3897 { 3898 bool recompute_toporder_p; 3899 /* Clear data structures of jump - jump itself will be removed 3900 by sel_redirect_edge_and_branch. */ 3901 clear_expr (INSN_EXPR (BB_END (xbb->prev_bb))); 3902 recompute_toporder_p 3903 = sel_redirect_edge_and_branch (EDGE_SUCC (xbb->prev_bb, 0), xbb); 3904 3905 gcc_assert (EDGE_SUCC (xbb->prev_bb, 0)->flags & EDGE_FALLTHRU); 3906 3907 /* We could have skipped some debug insns which did not get removed with the block, 3908 and the seqnos could become incorrect. Fix them up here. */ 3909 if (MAY_HAVE_DEBUG_INSNS && (sel_bb_head (xbb) != first || sel_bb_end (xbb) != last)) 3910 { 3911 if (!sel_bb_empty_p (xbb->prev_bb)) 3912 { 3913 int prev_seqno = INSN_SEQNO (sel_bb_end (xbb->prev_bb)); 3914 if (prev_seqno > INSN_SEQNO (sel_bb_head (xbb))) 3915 for (insn_t insn = sel_bb_head (xbb); insn != first; insn = NEXT_INSN (insn)) 3916 INSN_SEQNO (insn) = prev_seqno + 1; 3917 } 3918 } 3919 3920 /* It can turn out that after removing unused jump, basic block 3921 that contained that jump, becomes empty too. In such case 3922 remove it too. */ 3923 if (sel_bb_empty_p (xbb->prev_bb)) 3924 changed = maybe_tidy_empty_bb (xbb->prev_bb); 3925 if (recompute_toporder_p) 3926 sel_recompute_toporder (); 3927 } 3928 3929 /* TODO: use separate flag for CFG checking. */ 3930 if (flag_checking) 3931 { 3932 verify_backedges (); 3933 verify_dominators (CDI_DOMINATORS); 3934 } 3935 3936 return changed; 3937 } 3938 3939 /* Purge meaningless empty blocks in the middle of a region. */ 3940 void 3941 purge_empty_blocks (void) 3942 { 3943 int i; 3944 3945 /* Do not attempt to delete the first basic block in the region. */ 3946 for (i = 1; i < current_nr_blocks; ) 3947 { 3948 basic_block b = BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i)); 3949 3950 if (maybe_tidy_empty_bb (b)) 3951 continue; 3952 3953 i++; 3954 } 3955 } 3956 3957 /* Rip-off INSN from the insn stream. When ONLY_DISCONNECT is true, 3958 do not delete insn's data, because it will be later re-emitted. 3959 Return true if we have removed some blocks afterwards. */ 3960 bool 3961 sel_remove_insn (insn_t insn, bool only_disconnect, bool full_tidying) 3962 { 3963 basic_block bb = BLOCK_FOR_INSN (insn); 3964 3965 gcc_assert (INSN_IN_STREAM_P (insn)); 3966 3967 if (DEBUG_INSN_P (insn) && BB_AV_SET_VALID_P (bb)) 3968 { 3969 expr_t expr; 3970 av_set_iterator i; 3971 3972 /* When we remove a debug insn that is head of a BB, it remains 3973 in the AV_SET of the block, but it shouldn't. */ 3974 FOR_EACH_EXPR_1 (expr, i, &BB_AV_SET (bb)) 3975 if (EXPR_INSN_RTX (expr) == insn) 3976 { 3977 av_set_iter_remove (&i); 3978 break; 3979 } 3980 } 3981 3982 if (only_disconnect) 3983 remove_insn (insn); 3984 else 3985 { 3986 delete_insn (insn); 3987 clear_expr (INSN_EXPR (insn)); 3988 } 3989 3990 /* It is necessary to NULL these fields in case we are going to re-insert 3991 INSN into the insns stream, as will usually happen in the ONLY_DISCONNECT 3992 case, but also for NOPs that we will return to the nop pool. */ 3993 SET_PREV_INSN (insn) = NULL_RTX; 3994 SET_NEXT_INSN (insn) = NULL_RTX; 3995 set_block_for_insn (insn, NULL); 3996 3997 return tidy_control_flow (bb, full_tidying); 3998 } 3999 4000 /* Estimate number of the insns in BB. */ 4001 static int 4002 sel_estimate_number_of_insns (basic_block bb) 4003 { 4004 int res = 0; 4005 insn_t insn = NEXT_INSN (BB_HEAD (bb)), next_tail = NEXT_INSN (BB_END (bb)); 4006 4007 for (; insn != next_tail; insn = NEXT_INSN (insn)) 4008 if (NONDEBUG_INSN_P (insn)) 4009 res++; 4010 4011 return res; 4012 } 4013 4014 /* We don't need separate luids for notes or labels. */ 4015 static int 4016 sel_luid_for_non_insn (rtx x) 4017 { 4018 gcc_assert (NOTE_P (x) || LABEL_P (x)); 4019 4020 return -1; 4021 } 4022 4023 /* Find the proper seqno for inserting at INSN by successors. 4024 Return -1 if no successors with positive seqno exist. */ 4025 static int 4026 get_seqno_by_succs (rtx_insn *insn) 4027 { 4028 basic_block bb = BLOCK_FOR_INSN (insn); 4029 rtx_insn *tmp = insn, *end = BB_END (bb); 4030 int seqno; 4031 insn_t succ = NULL; 4032 succ_iterator si; 4033 4034 while (tmp != end) 4035 { 4036 tmp = NEXT_INSN (tmp); 4037 if (INSN_P (tmp)) 4038 return INSN_SEQNO (tmp); 4039 } 4040 4041 seqno = INT_MAX; 4042 4043 FOR_EACH_SUCC_1 (succ, si, end, SUCCS_NORMAL) 4044 if (INSN_SEQNO (succ) > 0) 4045 seqno = MIN (seqno, INSN_SEQNO (succ)); 4046 4047 if (seqno == INT_MAX) 4048 return -1; 4049 4050 return seqno; 4051 } 4052 4053 /* Compute seqno for INSN by its preds or succs. Use OLD_SEQNO to compute 4054 seqno in corner cases. */ 4055 static int 4056 get_seqno_for_a_jump (insn_t insn, int old_seqno) 4057 { 4058 int seqno; 4059 4060 gcc_assert (INSN_SIMPLEJUMP_P (insn)); 4061 4062 if (!sel_bb_head_p (insn)) 4063 seqno = INSN_SEQNO (PREV_INSN (insn)); 4064 else 4065 { 4066 basic_block bb = BLOCK_FOR_INSN (insn); 4067 4068 if (single_pred_p (bb) 4069 && !in_current_region_p (single_pred (bb))) 4070 { 4071 /* We can have preds outside a region when splitting edges 4072 for pipelining of an outer loop. Use succ instead. 4073 There should be only one of them. */ 4074 insn_t succ = NULL; 4075 succ_iterator si; 4076 bool first = true; 4077 4078 gcc_assert (flag_sel_sched_pipelining_outer_loops 4079 && current_loop_nest); 4080 FOR_EACH_SUCC_1 (succ, si, insn, 4081 SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS) 4082 { 4083 gcc_assert (first); 4084 first = false; 4085 } 4086 4087 gcc_assert (succ != NULL); 4088 seqno = INSN_SEQNO (succ); 4089 } 4090 else 4091 { 4092 insn_t *preds; 4093 int n; 4094 4095 cfg_preds (BLOCK_FOR_INSN (insn), &preds, &n); 4096 4097 gcc_assert (n > 0); 4098 /* For one predecessor, use simple method. */ 4099 if (n == 1) 4100 seqno = INSN_SEQNO (preds[0]); 4101 else 4102 seqno = get_seqno_by_preds (insn); 4103 4104 free (preds); 4105 } 4106 } 4107 4108 /* We were unable to find a good seqno among preds. */ 4109 if (seqno < 0) 4110 seqno = get_seqno_by_succs (insn); 4111 4112 if (seqno < 0) 4113 { 4114 /* The only case where this could be here legally is that the only 4115 unscheduled insn was a conditional jump that got removed and turned 4116 into this unconditional one. Initialize from the old seqno 4117 of that jump passed down to here. */ 4118 seqno = old_seqno; 4119 } 4120 4121 gcc_assert (seqno >= 0); 4122 return seqno; 4123 } 4124 4125 /* Find the proper seqno for inserting at INSN. Returns -1 if no predecessors 4126 with positive seqno exist. */ 4127 int 4128 get_seqno_by_preds (rtx_insn *insn) 4129 { 4130 basic_block bb = BLOCK_FOR_INSN (insn); 4131 rtx_insn *tmp = insn, *head = BB_HEAD (bb); 4132 insn_t *preds; 4133 int n, i, seqno; 4134 4135 /* Loop backwards from INSN to HEAD including both. */ 4136 while (1) 4137 { 4138 if (INSN_P (tmp)) 4139 return INSN_SEQNO (tmp); 4140 if (tmp == head) 4141 break; 4142 tmp = PREV_INSN (tmp); 4143 } 4144 4145 cfg_preds (bb, &preds, &n); 4146 for (i = 0, seqno = -1; i < n; i++) 4147 seqno = MAX (seqno, INSN_SEQNO (preds[i])); 4148 4149 return seqno; 4150 } 4151 4152 4153 4154 /* Extend pass-scope data structures for basic blocks. */ 4155 void 4156 sel_extend_global_bb_info (void) 4157 { 4158 sel_global_bb_info.safe_grow_cleared (last_basic_block_for_fn (cfun)); 4159 } 4160 4161 /* Extend region-scope data structures for basic blocks. */ 4162 static void 4163 extend_region_bb_info (void) 4164 { 4165 sel_region_bb_info.safe_grow_cleared (last_basic_block_for_fn (cfun)); 4166 } 4167 4168 /* Extend all data structures to fit for all basic blocks. */ 4169 static void 4170 extend_bb_info (void) 4171 { 4172 sel_extend_global_bb_info (); 4173 extend_region_bb_info (); 4174 } 4175 4176 /* Finalize pass-scope data structures for basic blocks. */ 4177 void 4178 sel_finish_global_bb_info (void) 4179 { 4180 sel_global_bb_info.release (); 4181 } 4182 4183 /* Finalize region-scope data structures for basic blocks. */ 4184 static void 4185 finish_region_bb_info (void) 4186 { 4187 sel_region_bb_info.release (); 4188 } 4189 4190 4191 /* Data for each insn in current region. */ 4192 vec<sel_insn_data_def> s_i_d; 4193 4194 /* Extend data structures for insns from current region. */ 4195 static void 4196 extend_insn_data (void) 4197 { 4198 int reserve; 4199 4200 sched_extend_target (); 4201 sched_deps_init (false); 4202 4203 /* Extend data structures for insns from current region. */ 4204 reserve = (sched_max_luid + 1 - s_i_d.length ()); 4205 if (reserve > 0 && ! s_i_d.space (reserve)) 4206 { 4207 int size; 4208 4209 if (sched_max_luid / 2 > 1024) 4210 size = sched_max_luid + 1024; 4211 else 4212 size = 3 * sched_max_luid / 2; 4213 4214 4215 s_i_d.safe_grow_cleared (size); 4216 } 4217 } 4218 4219 /* Finalize data structures for insns from current region. */ 4220 static void 4221 finish_insns (void) 4222 { 4223 unsigned i; 4224 4225 /* Clear here all dependence contexts that may have left from insns that were 4226 removed during the scheduling. */ 4227 for (i = 0; i < s_i_d.length (); i++) 4228 { 4229 sel_insn_data_def *sid_entry = &s_i_d[i]; 4230 4231 if (sid_entry->live) 4232 return_regset_to_pool (sid_entry->live); 4233 if (sid_entry->analyzed_deps) 4234 { 4235 BITMAP_FREE (sid_entry->analyzed_deps); 4236 BITMAP_FREE (sid_entry->found_deps); 4237 htab_delete (sid_entry->transformed_insns); 4238 free_deps (&sid_entry->deps_context); 4239 } 4240 if (EXPR_VINSN (&sid_entry->expr)) 4241 { 4242 clear_expr (&sid_entry->expr); 4243 4244 /* Also, clear CANT_MOVE bit here, because we really don't want it 4245 to be passed to the next region. */ 4246 CANT_MOVE_BY_LUID (i) = 0; 4247 } 4248 } 4249 4250 s_i_d.release (); 4251 } 4252 4253 /* A proxy to pass initialization data to init_insn (). */ 4254 static sel_insn_data_def _insn_init_ssid; 4255 static sel_insn_data_t insn_init_ssid = &_insn_init_ssid; 4256 4257 /* If true create a new vinsn. Otherwise use the one from EXPR. */ 4258 static bool insn_init_create_new_vinsn_p; 4259 4260 /* Set all necessary data for initialization of the new insn[s]. */ 4261 static expr_t 4262 set_insn_init (expr_t expr, vinsn_t vi, int seqno) 4263 { 4264 expr_t x = &insn_init_ssid->expr; 4265 4266 copy_expr_onside (x, expr); 4267 if (vi != NULL) 4268 { 4269 insn_init_create_new_vinsn_p = false; 4270 change_vinsn_in_expr (x, vi); 4271 } 4272 else 4273 insn_init_create_new_vinsn_p = true; 4274 4275 insn_init_ssid->seqno = seqno; 4276 return x; 4277 } 4278 4279 /* Init data for INSN. */ 4280 static void 4281 init_insn_data (insn_t insn) 4282 { 4283 expr_t expr; 4284 sel_insn_data_t ssid = insn_init_ssid; 4285 4286 /* The fields mentioned below are special and hence are not being 4287 propagated to the new insns. */ 4288 gcc_assert (!ssid->asm_p && ssid->sched_next == NULL 4289 && !ssid->after_stall_p && ssid->sched_cycle == 0); 4290 gcc_assert (INSN_P (insn) && INSN_LUID (insn) > 0); 4291 4292 expr = INSN_EXPR (insn); 4293 copy_expr (expr, &ssid->expr); 4294 prepare_insn_expr (insn, ssid->seqno); 4295 4296 if (insn_init_create_new_vinsn_p) 4297 change_vinsn_in_expr (expr, vinsn_create (insn, init_insn_force_unique_p)); 4298 4299 if (first_time_insn_init (insn)) 4300 init_first_time_insn_data (insn); 4301 } 4302 4303 /* This is used to initialize spurious jumps generated by 4304 sel_redirect_edge (). OLD_SEQNO is used for initializing seqnos 4305 in corner cases within get_seqno_for_a_jump. */ 4306 static void 4307 init_simplejump_data (insn_t insn, int old_seqno) 4308 { 4309 init_expr (INSN_EXPR (insn), vinsn_create (insn, false), 0, 4310 REG_BR_PROB_BASE, 0, 0, 0, 0, 0, 0, 4311 vNULL, true, false, false, 4312 false, true); 4313 INSN_SEQNO (insn) = get_seqno_for_a_jump (insn, old_seqno); 4314 init_first_time_insn_data (insn); 4315 } 4316 4317 /* Perform deferred initialization of insns. This is used to process 4318 a new jump that may be created by redirect_edge. OLD_SEQNO is used 4319 for initializing simplejumps in init_simplejump_data. */ 4320 static void 4321 sel_init_new_insn (insn_t insn, int flags, int old_seqno) 4322 { 4323 /* We create data structures for bb when the first insn is emitted in it. */ 4324 if (INSN_P (insn) 4325 && INSN_IN_STREAM_P (insn) 4326 && insn_is_the_only_one_in_bb_p (insn)) 4327 { 4328 extend_bb_info (); 4329 create_initial_data_sets (BLOCK_FOR_INSN (insn)); 4330 } 4331 4332 if (flags & INSN_INIT_TODO_LUID) 4333 { 4334 sched_extend_luids (); 4335 sched_init_insn_luid (insn); 4336 } 4337 4338 if (flags & INSN_INIT_TODO_SSID) 4339 { 4340 extend_insn_data (); 4341 init_insn_data (insn); 4342 clear_expr (&insn_init_ssid->expr); 4343 } 4344 4345 if (flags & INSN_INIT_TODO_SIMPLEJUMP) 4346 { 4347 extend_insn_data (); 4348 init_simplejump_data (insn, old_seqno); 4349 } 4350 4351 gcc_assert (CONTAINING_RGN (BLOCK_NUM (insn)) 4352 == CONTAINING_RGN (BB_TO_BLOCK (0))); 4353 } 4354 4355 4356 /* Functions to init/finish work with lv sets. */ 4357 4358 /* Init BB_LV_SET of BB from DF_LR_IN set of BB. */ 4359 static void 4360 init_lv_set (basic_block bb) 4361 { 4362 gcc_assert (!BB_LV_SET_VALID_P (bb)); 4363 4364 BB_LV_SET (bb) = get_regset_from_pool (); 4365 COPY_REG_SET (BB_LV_SET (bb), DF_LR_IN (bb)); 4366 BB_LV_SET_VALID_P (bb) = true; 4367 } 4368 4369 /* Copy liveness information to BB from FROM_BB. */ 4370 static void 4371 copy_lv_set_from (basic_block bb, basic_block from_bb) 4372 { 4373 gcc_assert (!BB_LV_SET_VALID_P (bb)); 4374 4375 COPY_REG_SET (BB_LV_SET (bb), BB_LV_SET (from_bb)); 4376 BB_LV_SET_VALID_P (bb) = true; 4377 } 4378 4379 /* Initialize lv set of all bb headers. */ 4380 void 4381 init_lv_sets (void) 4382 { 4383 basic_block bb; 4384 4385 /* Initialize of LV sets. */ 4386 FOR_EACH_BB_FN (bb, cfun) 4387 init_lv_set (bb); 4388 4389 /* Don't forget EXIT_BLOCK. */ 4390 init_lv_set (EXIT_BLOCK_PTR_FOR_FN (cfun)); 4391 } 4392 4393 /* Release lv set of HEAD. */ 4394 static void 4395 free_lv_set (basic_block bb) 4396 { 4397 gcc_assert (BB_LV_SET (bb) != NULL); 4398 4399 return_regset_to_pool (BB_LV_SET (bb)); 4400 BB_LV_SET (bb) = NULL; 4401 BB_LV_SET_VALID_P (bb) = false; 4402 } 4403 4404 /* Finalize lv sets of all bb headers. */ 4405 void 4406 free_lv_sets (void) 4407 { 4408 basic_block bb; 4409 4410 /* Don't forget EXIT_BLOCK. */ 4411 free_lv_set (EXIT_BLOCK_PTR_FOR_FN (cfun)); 4412 4413 /* Free LV sets. */ 4414 FOR_EACH_BB_FN (bb, cfun) 4415 if (BB_LV_SET (bb)) 4416 free_lv_set (bb); 4417 } 4418 4419 /* Mark AV_SET for BB as invalid, so this set will be updated the next time 4420 compute_av() processes BB. This function is called when creating new basic 4421 blocks, as well as for blocks (either new or existing) where new jumps are 4422 created when the control flow is being updated. */ 4423 static void 4424 invalidate_av_set (basic_block bb) 4425 { 4426 BB_AV_LEVEL (bb) = -1; 4427 } 4428 4429 /* Create initial data sets for BB (they will be invalid). */ 4430 static void 4431 create_initial_data_sets (basic_block bb) 4432 { 4433 if (BB_LV_SET (bb)) 4434 BB_LV_SET_VALID_P (bb) = false; 4435 else 4436 BB_LV_SET (bb) = get_regset_from_pool (); 4437 invalidate_av_set (bb); 4438 } 4439 4440 /* Free av set of BB. */ 4441 static void 4442 free_av_set (basic_block bb) 4443 { 4444 av_set_clear (&BB_AV_SET (bb)); 4445 BB_AV_LEVEL (bb) = 0; 4446 } 4447 4448 /* Free data sets of BB. */ 4449 void 4450 free_data_sets (basic_block bb) 4451 { 4452 free_lv_set (bb); 4453 free_av_set (bb); 4454 } 4455 4456 /* Exchange data sets of TO and FROM. */ 4457 void 4458 exchange_data_sets (basic_block to, basic_block from) 4459 { 4460 /* Exchange lv sets of TO and FROM. */ 4461 std::swap (BB_LV_SET (from), BB_LV_SET (to)); 4462 std::swap (BB_LV_SET_VALID_P (from), BB_LV_SET_VALID_P (to)); 4463 4464 /* Exchange av sets of TO and FROM. */ 4465 std::swap (BB_AV_SET (from), BB_AV_SET (to)); 4466 std::swap (BB_AV_LEVEL (from), BB_AV_LEVEL (to)); 4467 } 4468 4469 /* Copy data sets of FROM to TO. */ 4470 void 4471 copy_data_sets (basic_block to, basic_block from) 4472 { 4473 gcc_assert (!BB_LV_SET_VALID_P (to) && !BB_AV_SET_VALID_P (to)); 4474 gcc_assert (BB_AV_SET (to) == NULL); 4475 4476 BB_AV_LEVEL (to) = BB_AV_LEVEL (from); 4477 BB_LV_SET_VALID_P (to) = BB_LV_SET_VALID_P (from); 4478 4479 if (BB_AV_SET_VALID_P (from)) 4480 { 4481 BB_AV_SET (to) = av_set_copy (BB_AV_SET (from)); 4482 } 4483 if (BB_LV_SET_VALID_P (from)) 4484 { 4485 gcc_assert (BB_LV_SET (to) != NULL); 4486 COPY_REG_SET (BB_LV_SET (to), BB_LV_SET (from)); 4487 } 4488 } 4489 4490 /* Return an av set for INSN, if any. */ 4491 av_set_t 4492 get_av_set (insn_t insn) 4493 { 4494 av_set_t av_set; 4495 4496 gcc_assert (AV_SET_VALID_P (insn)); 4497 4498 if (sel_bb_head_p (insn)) 4499 av_set = BB_AV_SET (BLOCK_FOR_INSN (insn)); 4500 else 4501 av_set = NULL; 4502 4503 return av_set; 4504 } 4505 4506 /* Implementation of AV_LEVEL () macro. Return AV_LEVEL () of INSN. */ 4507 int 4508 get_av_level (insn_t insn) 4509 { 4510 int av_level; 4511 4512 gcc_assert (INSN_P (insn)); 4513 4514 if (sel_bb_head_p (insn)) 4515 av_level = BB_AV_LEVEL (BLOCK_FOR_INSN (insn)); 4516 else 4517 av_level = INSN_WS_LEVEL (insn); 4518 4519 return av_level; 4520 } 4521 4522 4523 4524 /* Variables to work with control-flow graph. */ 4525 4526 /* The basic block that already has been processed by the sched_data_update (), 4527 but hasn't been in sel_add_bb () yet. */ 4528 static vec<basic_block> last_added_blocks; 4529 4530 /* A pool for allocating successor infos. */ 4531 static struct 4532 { 4533 /* A stack for saving succs_info structures. */ 4534 struct succs_info *stack; 4535 4536 /* Its size. */ 4537 int size; 4538 4539 /* Top of the stack. */ 4540 int top; 4541 4542 /* Maximal value of the top. */ 4543 int max_top; 4544 } succs_info_pool; 4545 4546 /* Functions to work with control-flow graph. */ 4547 4548 /* Return basic block note of BB. */ 4549 rtx_insn * 4550 sel_bb_head (basic_block bb) 4551 { 4552 rtx_insn *head; 4553 4554 if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun)) 4555 { 4556 gcc_assert (exit_insn != NULL_RTX); 4557 head = exit_insn; 4558 } 4559 else 4560 { 4561 rtx_note *note = bb_note (bb); 4562 head = next_nonnote_insn (note); 4563 4564 if (head && (BARRIER_P (head) || BLOCK_FOR_INSN (head) != bb)) 4565 head = NULL; 4566 } 4567 4568 return head; 4569 } 4570 4571 /* Return true if INSN is a basic block header. */ 4572 bool 4573 sel_bb_head_p (insn_t insn) 4574 { 4575 return sel_bb_head (BLOCK_FOR_INSN (insn)) == insn; 4576 } 4577 4578 /* Return last insn of BB. */ 4579 rtx_insn * 4580 sel_bb_end (basic_block bb) 4581 { 4582 if (sel_bb_empty_p (bb)) 4583 return NULL; 4584 4585 gcc_assert (bb != EXIT_BLOCK_PTR_FOR_FN (cfun)); 4586 4587 return BB_END (bb); 4588 } 4589 4590 /* Return true if INSN is the last insn in its basic block. */ 4591 bool 4592 sel_bb_end_p (insn_t insn) 4593 { 4594 return insn == sel_bb_end (BLOCK_FOR_INSN (insn)); 4595 } 4596 4597 /* Return true if BB consist of single NOTE_INSN_BASIC_BLOCK. */ 4598 bool 4599 sel_bb_empty_p (basic_block bb) 4600 { 4601 return sel_bb_head (bb) == NULL; 4602 } 4603 4604 /* True when BB belongs to the current scheduling region. */ 4605 bool 4606 in_current_region_p (basic_block bb) 4607 { 4608 if (bb->index < NUM_FIXED_BLOCKS) 4609 return false; 4610 4611 return CONTAINING_RGN (bb->index) == CONTAINING_RGN (BB_TO_BLOCK (0)); 4612 } 4613 4614 /* Return the block which is a fallthru bb of a conditional jump JUMP. */ 4615 basic_block 4616 fallthru_bb_of_jump (const rtx_insn *jump) 4617 { 4618 if (!JUMP_P (jump)) 4619 return NULL; 4620 4621 if (!any_condjump_p (jump)) 4622 return NULL; 4623 4624 /* A basic block that ends with a conditional jump may still have one successor 4625 (and be followed by a barrier), we are not interested. */ 4626 if (single_succ_p (BLOCK_FOR_INSN (jump))) 4627 return NULL; 4628 4629 return FALLTHRU_EDGE (BLOCK_FOR_INSN (jump))->dest; 4630 } 4631 4632 /* Remove all notes from BB. */ 4633 static void 4634 init_bb (basic_block bb) 4635 { 4636 remove_notes (bb_note (bb), BB_END (bb)); 4637 BB_NOTE_LIST (bb) = note_list; 4638 } 4639 4640 void 4641 sel_init_bbs (bb_vec_t bbs) 4642 { 4643 const struct sched_scan_info_def ssi = 4644 { 4645 extend_bb_info, /* extend_bb */ 4646 init_bb, /* init_bb */ 4647 NULL, /* extend_insn */ 4648 NULL /* init_insn */ 4649 }; 4650 4651 sched_scan (&ssi, bbs); 4652 } 4653 4654 /* Restore notes for the whole region. */ 4655 static void 4656 sel_restore_notes (void) 4657 { 4658 int bb; 4659 insn_t insn; 4660 4661 for (bb = 0; bb < current_nr_blocks; bb++) 4662 { 4663 basic_block first, last; 4664 4665 first = EBB_FIRST_BB (bb); 4666 last = EBB_LAST_BB (bb)->next_bb; 4667 4668 do 4669 { 4670 note_list = BB_NOTE_LIST (first); 4671 restore_other_notes (NULL, first); 4672 BB_NOTE_LIST (first) = NULL; 4673 4674 FOR_BB_INSNS (first, insn) 4675 if (NONDEBUG_INSN_P (insn)) 4676 reemit_notes (insn); 4677 4678 first = first->next_bb; 4679 } 4680 while (first != last); 4681 } 4682 } 4683 4684 /* Free per-bb data structures. */ 4685 void 4686 sel_finish_bbs (void) 4687 { 4688 sel_restore_notes (); 4689 4690 /* Remove current loop preheader from this loop. */ 4691 if (current_loop_nest) 4692 sel_remove_loop_preheader (); 4693 4694 finish_region_bb_info (); 4695 } 4696 4697 /* Return true if INSN has a single successor of type FLAGS. */ 4698 bool 4699 sel_insn_has_single_succ_p (insn_t insn, int flags) 4700 { 4701 insn_t succ; 4702 succ_iterator si; 4703 bool first_p = true; 4704 4705 FOR_EACH_SUCC_1 (succ, si, insn, flags) 4706 { 4707 if (first_p) 4708 first_p = false; 4709 else 4710 return false; 4711 } 4712 4713 return true; 4714 } 4715 4716 /* Allocate successor's info. */ 4717 static struct succs_info * 4718 alloc_succs_info (void) 4719 { 4720 if (succs_info_pool.top == succs_info_pool.max_top) 4721 { 4722 int i; 4723 4724 if (++succs_info_pool.max_top >= succs_info_pool.size) 4725 gcc_unreachable (); 4726 4727 i = ++succs_info_pool.top; 4728 succs_info_pool.stack[i].succs_ok.create (10); 4729 succs_info_pool.stack[i].succs_other.create (10); 4730 succs_info_pool.stack[i].probs_ok.create (10); 4731 } 4732 else 4733 succs_info_pool.top++; 4734 4735 return &succs_info_pool.stack[succs_info_pool.top]; 4736 } 4737 4738 /* Free successor's info. */ 4739 void 4740 free_succs_info (struct succs_info * sinfo) 4741 { 4742 gcc_assert (succs_info_pool.top >= 0 4743 && &succs_info_pool.stack[succs_info_pool.top] == sinfo); 4744 succs_info_pool.top--; 4745 4746 /* Clear stale info. */ 4747 sinfo->succs_ok.block_remove (0, sinfo->succs_ok.length ()); 4748 sinfo->succs_other.block_remove (0, sinfo->succs_other.length ()); 4749 sinfo->probs_ok.block_remove (0, sinfo->probs_ok.length ()); 4750 sinfo->all_prob = 0; 4751 sinfo->succs_ok_n = 0; 4752 sinfo->all_succs_n = 0; 4753 } 4754 4755 /* Compute successor info for INSN. FLAGS are the flags passed 4756 to the FOR_EACH_SUCC_1 iterator. */ 4757 struct succs_info * 4758 compute_succs_info (insn_t insn, short flags) 4759 { 4760 succ_iterator si; 4761 insn_t succ; 4762 struct succs_info *sinfo = alloc_succs_info (); 4763 4764 /* Traverse *all* successors and decide what to do with each. */ 4765 FOR_EACH_SUCC_1 (succ, si, insn, SUCCS_ALL) 4766 { 4767 /* FIXME: this doesn't work for skipping to loop exits, as we don't 4768 perform code motion through inner loops. */ 4769 short current_flags = si.current_flags & ~SUCCS_SKIP_TO_LOOP_EXITS; 4770 4771 if (current_flags & flags) 4772 { 4773 sinfo->succs_ok.safe_push (succ); 4774 sinfo->probs_ok.safe_push ( 4775 /* FIXME: Improve calculation when skipping 4776 inner loop to exits. */ 4777 si.bb_end 4778 ? (si.e1->probability.initialized_p () 4779 ? si.e1->probability.to_reg_br_prob_base () 4780 : 0) 4781 : REG_BR_PROB_BASE); 4782 sinfo->succs_ok_n++; 4783 } 4784 else 4785 sinfo->succs_other.safe_push (succ); 4786 4787 /* Compute all_prob. */ 4788 if (!si.bb_end) 4789 sinfo->all_prob = REG_BR_PROB_BASE; 4790 else if (si.e1->probability.initialized_p ()) 4791 sinfo->all_prob += si.e1->probability.to_reg_br_prob_base (); 4792 4793 sinfo->all_succs_n++; 4794 } 4795 4796 return sinfo; 4797 } 4798 4799 /* Return the predecessors of BB in PREDS and their number in N. 4800 Empty blocks are skipped. SIZE is used to allocate PREDS. */ 4801 static void 4802 cfg_preds_1 (basic_block bb, insn_t **preds, int *n, int *size) 4803 { 4804 edge e; 4805 edge_iterator ei; 4806 4807 gcc_assert (BLOCK_TO_BB (bb->index) != 0); 4808 4809 FOR_EACH_EDGE (e, ei, bb->preds) 4810 { 4811 basic_block pred_bb = e->src; 4812 insn_t bb_end = BB_END (pred_bb); 4813 4814 if (!in_current_region_p (pred_bb)) 4815 { 4816 gcc_assert (flag_sel_sched_pipelining_outer_loops 4817 && current_loop_nest); 4818 continue; 4819 } 4820 4821 if (sel_bb_empty_p (pred_bb)) 4822 cfg_preds_1 (pred_bb, preds, n, size); 4823 else 4824 { 4825 if (*n == *size) 4826 *preds = XRESIZEVEC (insn_t, *preds, 4827 (*size = 2 * *size + 1)); 4828 (*preds)[(*n)++] = bb_end; 4829 } 4830 } 4831 4832 gcc_assert (*n != 0 4833 || (flag_sel_sched_pipelining_outer_loops 4834 && current_loop_nest)); 4835 } 4836 4837 /* Find all predecessors of BB and record them in PREDS and their number 4838 in N. Empty blocks are skipped, and only normal (forward in-region) 4839 edges are processed. */ 4840 static void 4841 cfg_preds (basic_block bb, insn_t **preds, int *n) 4842 { 4843 int size = 0; 4844 4845 *preds = NULL; 4846 *n = 0; 4847 cfg_preds_1 (bb, preds, n, &size); 4848 } 4849 4850 /* Returns true if we are moving INSN through join point. */ 4851 bool 4852 sel_num_cfg_preds_gt_1 (insn_t insn) 4853 { 4854 basic_block bb; 4855 4856 if (!sel_bb_head_p (insn) || INSN_BB (insn) == 0) 4857 return false; 4858 4859 bb = BLOCK_FOR_INSN (insn); 4860 4861 while (1) 4862 { 4863 if (EDGE_COUNT (bb->preds) > 1) 4864 return true; 4865 4866 gcc_assert (EDGE_PRED (bb, 0)->dest == bb); 4867 bb = EDGE_PRED (bb, 0)->src; 4868 4869 if (!sel_bb_empty_p (bb)) 4870 break; 4871 } 4872 4873 return false; 4874 } 4875 4876 /* Returns true when BB should be the end of an ebb. Adapted from the 4877 code in sched-ebb.c. */ 4878 bool 4879 bb_ends_ebb_p (basic_block bb) 4880 { 4881 basic_block next_bb = bb_next_bb (bb); 4882 edge e; 4883 4884 if (next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun) 4885 || bitmap_bit_p (forced_ebb_heads, next_bb->index) 4886 || (LABEL_P (BB_HEAD (next_bb)) 4887 /* NB: LABEL_NUSES () is not maintained outside of jump.c. 4888 Work around that. */ 4889 && !single_pred_p (next_bb))) 4890 return true; 4891 4892 if (!in_current_region_p (next_bb)) 4893 return true; 4894 4895 e = find_fallthru_edge (bb->succs); 4896 if (e) 4897 { 4898 gcc_assert (e->dest == next_bb); 4899 4900 return false; 4901 } 4902 4903 return true; 4904 } 4905 4906 /* Returns true when INSN and SUCC are in the same EBB, given that SUCC is a 4907 successor of INSN. */ 4908 bool 4909 in_same_ebb_p (insn_t insn, insn_t succ) 4910 { 4911 basic_block ptr = BLOCK_FOR_INSN (insn); 4912 4913 for (;;) 4914 { 4915 if (ptr == BLOCK_FOR_INSN (succ)) 4916 return true; 4917 4918 if (bb_ends_ebb_p (ptr)) 4919 return false; 4920 4921 ptr = bb_next_bb (ptr); 4922 } 4923 4924 gcc_unreachable (); 4925 return false; 4926 } 4927 4928 /* Recomputes the reverse topological order for the function and 4929 saves it in REV_TOP_ORDER_INDEX. REV_TOP_ORDER_INDEX_LEN is also 4930 modified appropriately. */ 4931 static void 4932 recompute_rev_top_order (void) 4933 { 4934 int *postorder; 4935 int n_blocks, i; 4936 4937 if (!rev_top_order_index 4938 || rev_top_order_index_len < last_basic_block_for_fn (cfun)) 4939 { 4940 rev_top_order_index_len = last_basic_block_for_fn (cfun); 4941 rev_top_order_index = XRESIZEVEC (int, rev_top_order_index, 4942 rev_top_order_index_len); 4943 } 4944 4945 postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun)); 4946 4947 n_blocks = post_order_compute (postorder, true, false); 4948 gcc_assert (n_basic_blocks_for_fn (cfun) == n_blocks); 4949 4950 /* Build reverse function: for each basic block with BB->INDEX == K 4951 rev_top_order_index[K] is it's reverse topological sort number. */ 4952 for (i = 0; i < n_blocks; i++) 4953 { 4954 gcc_assert (postorder[i] < rev_top_order_index_len); 4955 rev_top_order_index[postorder[i]] = i; 4956 } 4957 4958 free (postorder); 4959 } 4960 4961 /* Clear all flags from insns in BB that could spoil its rescheduling. */ 4962 void 4963 clear_outdated_rtx_info (basic_block bb) 4964 { 4965 rtx_insn *insn; 4966 4967 FOR_BB_INSNS (bb, insn) 4968 if (INSN_P (insn)) 4969 { 4970 SCHED_GROUP_P (insn) = 0; 4971 INSN_AFTER_STALL_P (insn) = 0; 4972 INSN_SCHED_TIMES (insn) = 0; 4973 EXPR_PRIORITY_ADJ (INSN_EXPR (insn)) = 0; 4974 4975 /* We cannot use the changed caches, as previously we could ignore 4976 the LHS dependence due to enabled renaming and transform 4977 the expression, and currently we'll be unable to do this. */ 4978 htab_empty (INSN_TRANSFORMED_INSNS (insn)); 4979 } 4980 } 4981 4982 /* Add BB_NOTE to the pool of available basic block notes. */ 4983 static void 4984 return_bb_to_pool (basic_block bb) 4985 { 4986 rtx_note *note = bb_note (bb); 4987 4988 gcc_assert (NOTE_BASIC_BLOCK (note) == bb 4989 && bb->aux == NULL); 4990 4991 /* It turns out that current cfg infrastructure does not support 4992 reuse of basic blocks. Don't bother for now. */ 4993 /*bb_note_pool.safe_push (note);*/ 4994 } 4995 4996 /* Get a bb_note from pool or return NULL_RTX if pool is empty. */ 4997 static rtx_note * 4998 get_bb_note_from_pool (void) 4999 { 5000 if (bb_note_pool.is_empty ()) 5001 return NULL; 5002 else 5003 { 5004 rtx_note *note = bb_note_pool.pop (); 5005 5006 SET_PREV_INSN (note) = NULL_RTX; 5007 SET_NEXT_INSN (note) = NULL_RTX; 5008 5009 return note; 5010 } 5011 } 5012 5013 /* Free bb_note_pool. */ 5014 void 5015 free_bb_note_pool (void) 5016 { 5017 bb_note_pool.release (); 5018 } 5019 5020 /* Setup scheduler pool and successor structure. */ 5021 void 5022 alloc_sched_pools (void) 5023 { 5024 int succs_size; 5025 5026 succs_size = MAX_WS + 1; 5027 succs_info_pool.stack = XCNEWVEC (struct succs_info, succs_size); 5028 succs_info_pool.size = succs_size; 5029 succs_info_pool.top = -1; 5030 succs_info_pool.max_top = -1; 5031 } 5032 5033 /* Free the pools. */ 5034 void 5035 free_sched_pools (void) 5036 { 5037 int i; 5038 5039 sched_lists_pool.release (); 5040 gcc_assert (succs_info_pool.top == -1); 5041 for (i = 0; i <= succs_info_pool.max_top; i++) 5042 { 5043 succs_info_pool.stack[i].succs_ok.release (); 5044 succs_info_pool.stack[i].succs_other.release (); 5045 succs_info_pool.stack[i].probs_ok.release (); 5046 } 5047 free (succs_info_pool.stack); 5048 } 5049 5050 5051 /* Returns a position in RGN where BB can be inserted retaining 5052 topological order. */ 5053 static int 5054 find_place_to_insert_bb (basic_block bb, int rgn) 5055 { 5056 bool has_preds_outside_rgn = false; 5057 edge e; 5058 edge_iterator ei; 5059 5060 /* Find whether we have preds outside the region. */ 5061 FOR_EACH_EDGE (e, ei, bb->preds) 5062 if (!in_current_region_p (e->src)) 5063 { 5064 has_preds_outside_rgn = true; 5065 break; 5066 } 5067 5068 /* Recompute the top order -- needed when we have > 1 pred 5069 and in case we don't have preds outside. */ 5070 if (flag_sel_sched_pipelining_outer_loops 5071 && (has_preds_outside_rgn || EDGE_COUNT (bb->preds) > 1)) 5072 { 5073 int i, bbi = bb->index, cur_bbi; 5074 5075 recompute_rev_top_order (); 5076 for (i = RGN_NR_BLOCKS (rgn) - 1; i >= 0; i--) 5077 { 5078 cur_bbi = BB_TO_BLOCK (i); 5079 if (rev_top_order_index[bbi] 5080 < rev_top_order_index[cur_bbi]) 5081 break; 5082 } 5083 5084 /* We skipped the right block, so we increase i. We accommodate 5085 it for increasing by step later, so we decrease i. */ 5086 return (i + 1) - 1; 5087 } 5088 else if (has_preds_outside_rgn) 5089 { 5090 /* This is the case when we generate an extra empty block 5091 to serve as region head during pipelining. */ 5092 e = EDGE_SUCC (bb, 0); 5093 gcc_assert (EDGE_COUNT (bb->succs) == 1 5094 && in_current_region_p (EDGE_SUCC (bb, 0)->dest) 5095 && (BLOCK_TO_BB (e->dest->index) == 0)); 5096 return -1; 5097 } 5098 5099 /* We don't have preds outside the region. We should have 5100 the only pred, because the multiple preds case comes from 5101 the pipelining of outer loops, and that is handled above. 5102 Just take the bbi of this single pred. */ 5103 if (EDGE_COUNT (bb->succs) > 0) 5104 { 5105 int pred_bbi; 5106 5107 gcc_assert (EDGE_COUNT (bb->preds) == 1); 5108 5109 pred_bbi = EDGE_PRED (bb, 0)->src->index; 5110 return BLOCK_TO_BB (pred_bbi); 5111 } 5112 else 5113 /* BB has no successors. It is safe to put it in the end. */ 5114 return current_nr_blocks - 1; 5115 } 5116 5117 /* Deletes an empty basic block freeing its data. */ 5118 static void 5119 delete_and_free_basic_block (basic_block bb) 5120 { 5121 gcc_assert (sel_bb_empty_p (bb)); 5122 5123 if (BB_LV_SET (bb)) 5124 free_lv_set (bb); 5125 5126 bitmap_clear_bit (blocks_to_reschedule, bb->index); 5127 5128 /* Can't assert av_set properties because we use sel_aremove_bb 5129 when removing loop preheader from the region. At the point of 5130 removing the preheader we already have deallocated sel_region_bb_info. */ 5131 gcc_assert (BB_LV_SET (bb) == NULL 5132 && !BB_LV_SET_VALID_P (bb) 5133 && BB_AV_LEVEL (bb) == 0 5134 && BB_AV_SET (bb) == NULL); 5135 5136 delete_basic_block (bb); 5137 } 5138 5139 /* Add BB to the current region and update the region data. */ 5140 static void 5141 add_block_to_current_region (basic_block bb) 5142 { 5143 int i, pos, bbi = -2, rgn; 5144 5145 rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); 5146 bbi = find_place_to_insert_bb (bb, rgn); 5147 bbi += 1; 5148 pos = RGN_BLOCKS (rgn) + bbi; 5149 5150 gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0 5151 && ebb_head[bbi] == pos); 5152 5153 /* Make a place for the new block. */ 5154 extend_regions (); 5155 5156 for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--) 5157 BLOCK_TO_BB (rgn_bb_table[i])++; 5158 5159 memmove (rgn_bb_table + pos + 1, 5160 rgn_bb_table + pos, 5161 (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table)); 5162 5163 /* Initialize data for BB. */ 5164 rgn_bb_table[pos] = bb->index; 5165 BLOCK_TO_BB (bb->index) = bbi; 5166 CONTAINING_RGN (bb->index) = rgn; 5167 5168 RGN_NR_BLOCKS (rgn)++; 5169 5170 for (i = rgn + 1; i <= nr_regions; i++) 5171 RGN_BLOCKS (i)++; 5172 } 5173 5174 /* Remove BB from the current region and update the region data. */ 5175 static void 5176 remove_bb_from_region (basic_block bb) 5177 { 5178 int i, pos, bbi = -2, rgn; 5179 5180 rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); 5181 bbi = BLOCK_TO_BB (bb->index); 5182 pos = RGN_BLOCKS (rgn) + bbi; 5183 5184 gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0 5185 && ebb_head[bbi] == pos); 5186 5187 for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--) 5188 BLOCK_TO_BB (rgn_bb_table[i])--; 5189 5190 memmove (rgn_bb_table + pos, 5191 rgn_bb_table + pos + 1, 5192 (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table)); 5193 5194 RGN_NR_BLOCKS (rgn)--; 5195 for (i = rgn + 1; i <= nr_regions; i++) 5196 RGN_BLOCKS (i)--; 5197 } 5198 5199 /* Add BB to the current region and update all data. If BB is NULL, add all 5200 blocks from last_added_blocks vector. */ 5201 static void 5202 sel_add_bb (basic_block bb) 5203 { 5204 /* Extend luids so that new notes will receive zero luids. */ 5205 sched_extend_luids (); 5206 sched_init_bbs (); 5207 sel_init_bbs (last_added_blocks); 5208 5209 /* When bb is passed explicitly, the vector should contain 5210 the only element that equals to bb; otherwise, the vector 5211 should not be NULL. */ 5212 gcc_assert (last_added_blocks.exists ()); 5213 5214 if (bb != NULL) 5215 { 5216 gcc_assert (last_added_blocks.length () == 1 5217 && last_added_blocks[0] == bb); 5218 add_block_to_current_region (bb); 5219 5220 /* We associate creating/deleting data sets with the first insn 5221 appearing / disappearing in the bb. */ 5222 if (!sel_bb_empty_p (bb) && BB_LV_SET (bb) == NULL) 5223 create_initial_data_sets (bb); 5224 5225 last_added_blocks.release (); 5226 } 5227 else 5228 /* BB is NULL - process LAST_ADDED_BLOCKS instead. */ 5229 { 5230 int i; 5231 basic_block temp_bb = NULL; 5232 5233 for (i = 0; 5234 last_added_blocks.iterate (i, &bb); i++) 5235 { 5236 add_block_to_current_region (bb); 5237 temp_bb = bb; 5238 } 5239 5240 /* We need to fetch at least one bb so we know the region 5241 to update. */ 5242 gcc_assert (temp_bb != NULL); 5243 bb = temp_bb; 5244 5245 last_added_blocks.release (); 5246 } 5247 5248 rgn_setup_region (CONTAINING_RGN (bb->index)); 5249 } 5250 5251 /* Remove BB from the current region and update all data. 5252 If REMOVE_FROM_CFG_PBB is true, also remove the block cfom cfg. */ 5253 static void 5254 sel_remove_bb (basic_block bb, bool remove_from_cfg_p) 5255 { 5256 unsigned idx = bb->index; 5257 5258 gcc_assert (bb != NULL && BB_NOTE_LIST (bb) == NULL_RTX); 5259 5260 remove_bb_from_region (bb); 5261 return_bb_to_pool (bb); 5262 bitmap_clear_bit (blocks_to_reschedule, idx); 5263 5264 if (remove_from_cfg_p) 5265 { 5266 basic_block succ = single_succ (bb); 5267 delete_and_free_basic_block (bb); 5268 set_immediate_dominator (CDI_DOMINATORS, succ, 5269 recompute_dominator (CDI_DOMINATORS, succ)); 5270 } 5271 5272 rgn_setup_region (CONTAINING_RGN (idx)); 5273 } 5274 5275 /* Concatenate info of EMPTY_BB to info of MERGE_BB. */ 5276 static void 5277 move_bb_info (basic_block merge_bb, basic_block empty_bb) 5278 { 5279 if (in_current_region_p (merge_bb)) 5280 concat_note_lists (BB_NOTE_LIST (empty_bb), 5281 &BB_NOTE_LIST (merge_bb)); 5282 BB_NOTE_LIST (empty_bb) = NULL; 5283 5284 } 5285 5286 /* Remove EMPTY_BB. If REMOVE_FROM_CFG_P is false, remove EMPTY_BB from 5287 region, but keep it in CFG. */ 5288 static void 5289 remove_empty_bb (basic_block empty_bb, bool remove_from_cfg_p) 5290 { 5291 /* The block should contain just a note or a label. 5292 We try to check whether it is unused below. */ 5293 gcc_assert (BB_HEAD (empty_bb) == BB_END (empty_bb) 5294 || LABEL_P (BB_HEAD (empty_bb))); 5295 5296 /* If basic block has predecessors or successors, redirect them. */ 5297 if (remove_from_cfg_p 5298 && (EDGE_COUNT (empty_bb->preds) > 0 5299 || EDGE_COUNT (empty_bb->succs) > 0)) 5300 { 5301 basic_block pred; 5302 basic_block succ; 5303 5304 /* We need to init PRED and SUCC before redirecting edges. */ 5305 if (EDGE_COUNT (empty_bb->preds) > 0) 5306 { 5307 edge e; 5308 5309 gcc_assert (EDGE_COUNT (empty_bb->preds) == 1); 5310 5311 e = EDGE_PRED (empty_bb, 0); 5312 gcc_assert (e->src == empty_bb->prev_bb 5313 && (e->flags & EDGE_FALLTHRU)); 5314 5315 pred = empty_bb->prev_bb; 5316 } 5317 else 5318 pred = NULL; 5319 5320 if (EDGE_COUNT (empty_bb->succs) > 0) 5321 { 5322 /* We do not check fallthruness here as above, because 5323 after removing a jump the edge may actually be not fallthru. */ 5324 gcc_assert (EDGE_COUNT (empty_bb->succs) == 1); 5325 succ = EDGE_SUCC (empty_bb, 0)->dest; 5326 } 5327 else 5328 succ = NULL; 5329 5330 if (EDGE_COUNT (empty_bb->preds) > 0 && succ != NULL) 5331 { 5332 edge e = EDGE_PRED (empty_bb, 0); 5333 5334 if (e->flags & EDGE_FALLTHRU) 5335 redirect_edge_succ_nodup (e, succ); 5336 else 5337 sel_redirect_edge_and_branch (EDGE_PRED (empty_bb, 0), succ); 5338 } 5339 5340 if (EDGE_COUNT (empty_bb->succs) > 0 && pred != NULL) 5341 { 5342 edge e = EDGE_SUCC (empty_bb, 0); 5343 5344 if (find_edge (pred, e->dest) == NULL) 5345 redirect_edge_pred (e, pred); 5346 } 5347 } 5348 5349 /* Finish removing. */ 5350 sel_remove_bb (empty_bb, remove_from_cfg_p); 5351 } 5352 5353 /* An implementation of create_basic_block hook, which additionally updates 5354 per-bb data structures. */ 5355 static basic_block 5356 sel_create_basic_block (void *headp, void *endp, basic_block after) 5357 { 5358 basic_block new_bb; 5359 rtx_note *new_bb_note; 5360 5361 gcc_assert (flag_sel_sched_pipelining_outer_loops 5362 || !last_added_blocks.exists ()); 5363 5364 new_bb_note = get_bb_note_from_pool (); 5365 5366 if (new_bb_note == NULL_RTX) 5367 new_bb = orig_cfg_hooks.create_basic_block (headp, endp, after); 5368 else 5369 { 5370 new_bb = create_basic_block_structure ((rtx_insn *) headp, 5371 (rtx_insn *) endp, 5372 new_bb_note, after); 5373 new_bb->aux = NULL; 5374 } 5375 5376 last_added_blocks.safe_push (new_bb); 5377 5378 return new_bb; 5379 } 5380 5381 /* Implement sched_init_only_bb (). */ 5382 static void 5383 sel_init_only_bb (basic_block bb, basic_block after) 5384 { 5385 gcc_assert (after == NULL); 5386 5387 extend_regions (); 5388 rgn_make_new_region_out_of_new_block (bb); 5389 } 5390 5391 /* Update the latch when we've splitted or merged it from FROM block to TO. 5392 This should be checked for all outer loops, too. */ 5393 static void 5394 change_loops_latches (basic_block from, basic_block to) 5395 { 5396 gcc_assert (from != to); 5397 5398 if (current_loop_nest) 5399 { 5400 struct loop *loop; 5401 5402 for (loop = current_loop_nest; loop; loop = loop_outer (loop)) 5403 if (considered_for_pipelining_p (loop) && loop->latch == from) 5404 { 5405 gcc_assert (loop == current_loop_nest); 5406 loop->latch = to; 5407 gcc_assert (loop_latch_edge (loop)); 5408 } 5409 } 5410 } 5411 5412 /* Splits BB on two basic blocks, adding it to the region and extending 5413 per-bb data structures. Returns the newly created bb. */ 5414 static basic_block 5415 sel_split_block (basic_block bb, rtx after) 5416 { 5417 basic_block new_bb; 5418 insn_t insn; 5419 5420 new_bb = sched_split_block_1 (bb, after); 5421 sel_add_bb (new_bb); 5422 5423 /* This should be called after sel_add_bb, because this uses 5424 CONTAINING_RGN for the new block, which is not yet initialized. 5425 FIXME: this function may be a no-op now. */ 5426 change_loops_latches (bb, new_bb); 5427 5428 /* Update ORIG_BB_INDEX for insns moved into the new block. */ 5429 FOR_BB_INSNS (new_bb, insn) 5430 if (INSN_P (insn)) 5431 EXPR_ORIG_BB_INDEX (INSN_EXPR (insn)) = new_bb->index; 5432 5433 if (sel_bb_empty_p (bb)) 5434 { 5435 gcc_assert (!sel_bb_empty_p (new_bb)); 5436 5437 /* NEW_BB has data sets that need to be updated and BB holds 5438 data sets that should be removed. Exchange these data sets 5439 so that we won't lose BB's valid data sets. */ 5440 exchange_data_sets (new_bb, bb); 5441 free_data_sets (bb); 5442 } 5443 5444 if (!sel_bb_empty_p (new_bb) 5445 && bitmap_bit_p (blocks_to_reschedule, bb->index)) 5446 bitmap_set_bit (blocks_to_reschedule, new_bb->index); 5447 5448 return new_bb; 5449 } 5450 5451 /* If BB ends with a jump insn whose ID is bigger then PREV_MAX_UID, return it. 5452 Otherwise returns NULL. */ 5453 static rtx_insn * 5454 check_for_new_jump (basic_block bb, int prev_max_uid) 5455 { 5456 rtx_insn *end; 5457 5458 end = sel_bb_end (bb); 5459 if (end && INSN_UID (end) >= prev_max_uid) 5460 return end; 5461 return NULL; 5462 } 5463 5464 /* Look for a new jump either in FROM_BB block or in newly created JUMP_BB block. 5465 New means having UID at least equal to PREV_MAX_UID. */ 5466 static rtx_insn * 5467 find_new_jump (basic_block from, basic_block jump_bb, int prev_max_uid) 5468 { 5469 rtx_insn *jump; 5470 5471 /* Return immediately if no new insns were emitted. */ 5472 if (get_max_uid () == prev_max_uid) 5473 return NULL; 5474 5475 /* Now check both blocks for new jumps. It will ever be only one. */ 5476 if ((jump = check_for_new_jump (from, prev_max_uid))) 5477 return jump; 5478 5479 if (jump_bb != NULL 5480 && (jump = check_for_new_jump (jump_bb, prev_max_uid))) 5481 return jump; 5482 return NULL; 5483 } 5484 5485 /* Splits E and adds the newly created basic block to the current region. 5486 Returns this basic block. */ 5487 basic_block 5488 sel_split_edge (edge e) 5489 { 5490 basic_block new_bb, src, other_bb = NULL; 5491 int prev_max_uid; 5492 rtx_insn *jump; 5493 5494 src = e->src; 5495 prev_max_uid = get_max_uid (); 5496 new_bb = split_edge (e); 5497 5498 if (flag_sel_sched_pipelining_outer_loops 5499 && current_loop_nest) 5500 { 5501 int i; 5502 basic_block bb; 5503 5504 /* Some of the basic blocks might not have been added to the loop. 5505 Add them here, until this is fixed in force_fallthru. */ 5506 for (i = 0; 5507 last_added_blocks.iterate (i, &bb); i++) 5508 if (!bb->loop_father) 5509 { 5510 add_bb_to_loop (bb, e->dest->loop_father); 5511 5512 gcc_assert (!other_bb && (new_bb->index != bb->index)); 5513 other_bb = bb; 5514 } 5515 } 5516 5517 /* Add all last_added_blocks to the region. */ 5518 sel_add_bb (NULL); 5519 5520 jump = find_new_jump (src, new_bb, prev_max_uid); 5521 if (jump) 5522 sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); 5523 5524 /* Put the correct lv set on this block. */ 5525 if (other_bb && !sel_bb_empty_p (other_bb)) 5526 compute_live (sel_bb_head (other_bb)); 5527 5528 return new_bb; 5529 } 5530 5531 /* Implement sched_create_empty_bb (). */ 5532 static basic_block 5533 sel_create_empty_bb (basic_block after) 5534 { 5535 basic_block new_bb; 5536 5537 new_bb = sched_create_empty_bb_1 (after); 5538 5539 /* We'll explicitly initialize NEW_BB via sel_init_only_bb () a bit 5540 later. */ 5541 gcc_assert (last_added_blocks.length () == 1 5542 && last_added_blocks[0] == new_bb); 5543 5544 last_added_blocks.release (); 5545 return new_bb; 5546 } 5547 5548 /* Implement sched_create_recovery_block. ORIG_INSN is where block 5549 will be splitted to insert a check. */ 5550 basic_block 5551 sel_create_recovery_block (insn_t orig_insn) 5552 { 5553 basic_block first_bb, second_bb, recovery_block; 5554 basic_block before_recovery = NULL; 5555 rtx_insn *jump; 5556 5557 first_bb = BLOCK_FOR_INSN (orig_insn); 5558 if (sel_bb_end_p (orig_insn)) 5559 { 5560 /* Avoid introducing an empty block while splitting. */ 5561 gcc_assert (single_succ_p (first_bb)); 5562 second_bb = single_succ (first_bb); 5563 } 5564 else 5565 second_bb = sched_split_block (first_bb, orig_insn); 5566 5567 recovery_block = sched_create_recovery_block (&before_recovery); 5568 if (before_recovery) 5569 copy_lv_set_from (before_recovery, EXIT_BLOCK_PTR_FOR_FN (cfun)); 5570 5571 gcc_assert (sel_bb_empty_p (recovery_block)); 5572 sched_create_recovery_edges (first_bb, recovery_block, second_bb); 5573 if (current_loops != NULL) 5574 add_bb_to_loop (recovery_block, first_bb->loop_father); 5575 5576 sel_add_bb (recovery_block); 5577 5578 jump = BB_END (recovery_block); 5579 gcc_assert (sel_bb_head (recovery_block) == jump); 5580 sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP); 5581 5582 return recovery_block; 5583 } 5584 5585 /* Merge basic block B into basic block A. */ 5586 static void 5587 sel_merge_blocks (basic_block a, basic_block b) 5588 { 5589 gcc_assert (sel_bb_empty_p (b) 5590 && EDGE_COUNT (b->preds) == 1 5591 && EDGE_PRED (b, 0)->src == b->prev_bb); 5592 5593 move_bb_info (b->prev_bb, b); 5594 remove_empty_bb (b, false); 5595 merge_blocks (a, b); 5596 change_loops_latches (b, a); 5597 } 5598 5599 /* A wrapper for redirect_edge_and_branch_force, which also initializes 5600 data structures for possibly created bb and insns. */ 5601 void 5602 sel_redirect_edge_and_branch_force (edge e, basic_block to) 5603 { 5604 basic_block jump_bb, src, orig_dest = e->dest; 5605 int prev_max_uid; 5606 rtx_insn *jump; 5607 int old_seqno = -1; 5608 5609 /* This function is now used only for bookkeeping code creation, where 5610 we'll never get the single pred of orig_dest block and thus will not 5611 hit unreachable blocks when updating dominator info. */ 5612 gcc_assert (!sel_bb_empty_p (e->src) 5613 && !single_pred_p (orig_dest)); 5614 src = e->src; 5615 prev_max_uid = get_max_uid (); 5616 /* Compute and pass old_seqno down to sel_init_new_insn only for the case 5617 when the conditional jump being redirected may become unconditional. */ 5618 if (any_condjump_p (BB_END (src)) 5619 && INSN_SEQNO (BB_END (src)) >= 0) 5620 old_seqno = INSN_SEQNO (BB_END (src)); 5621 5622 jump_bb = redirect_edge_and_branch_force (e, to); 5623 if (jump_bb != NULL) 5624 sel_add_bb (jump_bb); 5625 5626 /* This function could not be used to spoil the loop structure by now, 5627 thus we don't care to update anything. But check it to be sure. */ 5628 if (current_loop_nest 5629 && pipelining_p) 5630 gcc_assert (loop_latch_edge (current_loop_nest)); 5631 5632 jump = find_new_jump (src, jump_bb, prev_max_uid); 5633 if (jump) 5634 sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP, 5635 old_seqno); 5636 set_immediate_dominator (CDI_DOMINATORS, to, 5637 recompute_dominator (CDI_DOMINATORS, to)); 5638 set_immediate_dominator (CDI_DOMINATORS, orig_dest, 5639 recompute_dominator (CDI_DOMINATORS, orig_dest)); 5640 if (jump && sel_bb_head_p (jump)) 5641 compute_live (jump); 5642 } 5643 5644 /* A wrapper for redirect_edge_and_branch. Return TRUE if blocks connected by 5645 redirected edge are in reverse topological order. */ 5646 bool 5647 sel_redirect_edge_and_branch (edge e, basic_block to) 5648 { 5649 bool latch_edge_p; 5650 basic_block src, orig_dest = e->dest; 5651 int prev_max_uid; 5652 rtx_insn *jump; 5653 edge redirected; 5654 bool recompute_toporder_p = false; 5655 bool maybe_unreachable = single_pred_p (orig_dest); 5656 int old_seqno = -1; 5657 5658 latch_edge_p = (pipelining_p 5659 && current_loop_nest 5660 && e == loop_latch_edge (current_loop_nest)); 5661 5662 src = e->src; 5663 prev_max_uid = get_max_uid (); 5664 5665 /* Compute and pass old_seqno down to sel_init_new_insn only for the case 5666 when the conditional jump being redirected may become unconditional. */ 5667 if (any_condjump_p (BB_END (src)) 5668 && INSN_SEQNO (BB_END (src)) >= 0) 5669 old_seqno = INSN_SEQNO (BB_END (src)); 5670 5671 redirected = redirect_edge_and_branch (e, to); 5672 5673 gcc_assert (redirected && !last_added_blocks.exists ()); 5674 5675 /* When we've redirected a latch edge, update the header. */ 5676 if (latch_edge_p) 5677 { 5678 current_loop_nest->header = to; 5679 gcc_assert (loop_latch_edge (current_loop_nest)); 5680 } 5681 5682 /* In rare situations, the topological relation between the blocks connected 5683 by the redirected edge can change (see PR42245 for an example). Update 5684 block_to_bb/bb_to_block. */ 5685 if (CONTAINING_RGN (e->src->index) == CONTAINING_RGN (to->index) 5686 && BLOCK_TO_BB (e->src->index) > BLOCK_TO_BB (to->index)) 5687 recompute_toporder_p = true; 5688 5689 jump = find_new_jump (src, NULL, prev_max_uid); 5690 if (jump) 5691 sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP, old_seqno); 5692 5693 /* Only update dominator info when we don't have unreachable blocks. 5694 Otherwise we'll update in maybe_tidy_empty_bb. */ 5695 if (!maybe_unreachable) 5696 { 5697 set_immediate_dominator (CDI_DOMINATORS, to, 5698 recompute_dominator (CDI_DOMINATORS, to)); 5699 set_immediate_dominator (CDI_DOMINATORS, orig_dest, 5700 recompute_dominator (CDI_DOMINATORS, orig_dest)); 5701 } 5702 if (jump && sel_bb_head_p (jump)) 5703 compute_live (jump); 5704 return recompute_toporder_p; 5705 } 5706 5707 /* This variable holds the cfg hooks used by the selective scheduler. */ 5708 static struct cfg_hooks sel_cfg_hooks; 5709 5710 /* Register sel-sched cfg hooks. */ 5711 void 5712 sel_register_cfg_hooks (void) 5713 { 5714 sched_split_block = sel_split_block; 5715 5716 orig_cfg_hooks = get_cfg_hooks (); 5717 sel_cfg_hooks = orig_cfg_hooks; 5718 5719 sel_cfg_hooks.create_basic_block = sel_create_basic_block; 5720 5721 set_cfg_hooks (sel_cfg_hooks); 5722 5723 sched_init_only_bb = sel_init_only_bb; 5724 sched_split_block = sel_split_block; 5725 sched_create_empty_bb = sel_create_empty_bb; 5726 } 5727 5728 /* Unregister sel-sched cfg hooks. */ 5729 void 5730 sel_unregister_cfg_hooks (void) 5731 { 5732 sched_create_empty_bb = NULL; 5733 sched_split_block = NULL; 5734 sched_init_only_bb = NULL; 5735 5736 set_cfg_hooks (orig_cfg_hooks); 5737 } 5738 5739 5740 /* Emit an insn rtx based on PATTERN. If a jump insn is wanted, 5741 LABEL is where this jump should be directed. */ 5742 rtx_insn * 5743 create_insn_rtx_from_pattern (rtx pattern, rtx label) 5744 { 5745 rtx_insn *insn_rtx; 5746 5747 gcc_assert (!INSN_P (pattern)); 5748 5749 start_sequence (); 5750 5751 if (label == NULL_RTX) 5752 insn_rtx = emit_insn (pattern); 5753 else if (DEBUG_INSN_P (label)) 5754 insn_rtx = emit_debug_insn (pattern); 5755 else 5756 { 5757 insn_rtx = emit_jump_insn (pattern); 5758 JUMP_LABEL (insn_rtx) = label; 5759 ++LABEL_NUSES (label); 5760 } 5761 5762 end_sequence (); 5763 5764 sched_extend_luids (); 5765 sched_extend_target (); 5766 sched_deps_init (false); 5767 5768 /* Initialize INSN_CODE now. */ 5769 recog_memoized (insn_rtx); 5770 return insn_rtx; 5771 } 5772 5773 /* Create a new vinsn for INSN_RTX. FORCE_UNIQUE_P is true when the vinsn 5774 must not be clonable. */ 5775 vinsn_t 5776 create_vinsn_from_insn_rtx (rtx_insn *insn_rtx, bool force_unique_p) 5777 { 5778 gcc_assert (INSN_P (insn_rtx) && !INSN_IN_STREAM_P (insn_rtx)); 5779 5780 /* If VINSN_TYPE is not USE, retain its uniqueness. */ 5781 return vinsn_create (insn_rtx, force_unique_p); 5782 } 5783 5784 /* Create a copy of INSN_RTX. */ 5785 rtx_insn * 5786 create_copy_of_insn_rtx (rtx insn_rtx) 5787 { 5788 rtx_insn *res; 5789 rtx link; 5790 5791 if (DEBUG_INSN_P (insn_rtx)) 5792 return create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)), 5793 insn_rtx); 5794 5795 gcc_assert (NONJUMP_INSN_P (insn_rtx)); 5796 5797 res = create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)), 5798 NULL_RTX); 5799 5800 /* Locate the end of existing REG_NOTES in NEW_RTX. */ 5801 rtx *ptail = ®_NOTES (res); 5802 while (*ptail != NULL_RTX) 5803 ptail = &XEXP (*ptail, 1); 5804 5805 /* Copy all REG_NOTES except REG_EQUAL/REG_EQUIV and REG_LABEL_OPERAND 5806 since mark_jump_label will make them. REG_LABEL_TARGETs are created 5807 there too, but are supposed to be sticky, so we copy them. */ 5808 for (link = REG_NOTES (insn_rtx); link; link = XEXP (link, 1)) 5809 if (REG_NOTE_KIND (link) != REG_LABEL_OPERAND 5810 && REG_NOTE_KIND (link) != REG_EQUAL 5811 && REG_NOTE_KIND (link) != REG_EQUIV) 5812 { 5813 *ptail = duplicate_reg_note (link); 5814 ptail = &XEXP (*ptail, 1); 5815 } 5816 5817 return res; 5818 } 5819 5820 /* Change vinsn field of EXPR to hold NEW_VINSN. */ 5821 void 5822 change_vinsn_in_expr (expr_t expr, vinsn_t new_vinsn) 5823 { 5824 vinsn_detach (EXPR_VINSN (expr)); 5825 5826 EXPR_VINSN (expr) = new_vinsn; 5827 vinsn_attach (new_vinsn); 5828 } 5829 5830 /* Helpers for global init. */ 5831 /* This structure is used to be able to call existing bundling mechanism 5832 and calculate insn priorities. */ 5833 static struct haifa_sched_info sched_sel_haifa_sched_info = 5834 { 5835 NULL, /* init_ready_list */ 5836 NULL, /* can_schedule_ready_p */ 5837 NULL, /* schedule_more_p */ 5838 NULL, /* new_ready */ 5839 NULL, /* rgn_rank */ 5840 sel_print_insn, /* rgn_print_insn */ 5841 contributes_to_priority, 5842 NULL, /* insn_finishes_block_p */ 5843 5844 NULL, NULL, 5845 NULL, NULL, 5846 0, 0, 5847 5848 NULL, /* add_remove_insn */ 5849 NULL, /* begin_schedule_ready */ 5850 NULL, /* begin_move_insn */ 5851 NULL, /* advance_target_bb */ 5852 5853 NULL, 5854 NULL, 5855 5856 SEL_SCHED | NEW_BBS 5857 }; 5858 5859 /* Setup special insns used in the scheduler. */ 5860 void 5861 setup_nop_and_exit_insns (void) 5862 { 5863 gcc_assert (nop_pattern == NULL_RTX 5864 && exit_insn == NULL_RTX); 5865 5866 nop_pattern = constm1_rtx; 5867 5868 start_sequence (); 5869 emit_insn (nop_pattern); 5870 exit_insn = get_insns (); 5871 end_sequence (); 5872 set_block_for_insn (exit_insn, EXIT_BLOCK_PTR_FOR_FN (cfun)); 5873 } 5874 5875 /* Free special insns used in the scheduler. */ 5876 void 5877 free_nop_and_exit_insns (void) 5878 { 5879 exit_insn = NULL; 5880 nop_pattern = NULL_RTX; 5881 } 5882 5883 /* Setup a special vinsn used in new insns initialization. */ 5884 void 5885 setup_nop_vinsn (void) 5886 { 5887 nop_vinsn = vinsn_create (exit_insn, false); 5888 vinsn_attach (nop_vinsn); 5889 } 5890 5891 /* Free a special vinsn used in new insns initialization. */ 5892 void 5893 free_nop_vinsn (void) 5894 { 5895 gcc_assert (VINSN_COUNT (nop_vinsn) == 1); 5896 vinsn_detach (nop_vinsn); 5897 nop_vinsn = NULL; 5898 } 5899 5900 /* Call a set_sched_flags hook. */ 5901 void 5902 sel_set_sched_flags (void) 5903 { 5904 /* ??? This means that set_sched_flags were called, and we decided to 5905 support speculation. However, set_sched_flags also modifies flags 5906 on current_sched_info, doing this only at global init. And we 5907 sometimes change c_s_i later. So put the correct flags again. */ 5908 if (spec_info && targetm.sched.set_sched_flags) 5909 targetm.sched.set_sched_flags (spec_info); 5910 } 5911 5912 /* Setup pointers to global sched info structures. */ 5913 void 5914 sel_setup_sched_infos (void) 5915 { 5916 rgn_setup_common_sched_info (); 5917 5918 memcpy (&sel_common_sched_info, common_sched_info, 5919 sizeof (sel_common_sched_info)); 5920 5921 sel_common_sched_info.fix_recovery_cfg = NULL; 5922 sel_common_sched_info.add_block = NULL; 5923 sel_common_sched_info.estimate_number_of_insns 5924 = sel_estimate_number_of_insns; 5925 sel_common_sched_info.luid_for_non_insn = sel_luid_for_non_insn; 5926 sel_common_sched_info.sched_pass_id = SCHED_SEL_PASS; 5927 5928 common_sched_info = &sel_common_sched_info; 5929 5930 current_sched_info = &sched_sel_haifa_sched_info; 5931 current_sched_info->sched_max_insns_priority = 5932 get_rgn_sched_max_insns_priority (); 5933 5934 sel_set_sched_flags (); 5935 } 5936 5937 5938 /* Adds basic block BB to region RGN at the position *BB_ORD_INDEX, 5939 *BB_ORD_INDEX after that is increased. */ 5940 static void 5941 sel_add_block_to_region (basic_block bb, int *bb_ord_index, int rgn) 5942 { 5943 RGN_NR_BLOCKS (rgn) += 1; 5944 RGN_DONT_CALC_DEPS (rgn) = 0; 5945 RGN_HAS_REAL_EBB (rgn) = 0; 5946 CONTAINING_RGN (bb->index) = rgn; 5947 BLOCK_TO_BB (bb->index) = *bb_ord_index; 5948 rgn_bb_table[RGN_BLOCKS (rgn) + *bb_ord_index] = bb->index; 5949 (*bb_ord_index)++; 5950 5951 /* FIXME: it is true only when not scheduling ebbs. */ 5952 RGN_BLOCKS (rgn + 1) = RGN_BLOCKS (rgn) + RGN_NR_BLOCKS (rgn); 5953 } 5954 5955 /* Functions to support pipelining of outer loops. */ 5956 5957 /* Creates a new empty region and returns it's number. */ 5958 static int 5959 sel_create_new_region (void) 5960 { 5961 int new_rgn_number = nr_regions; 5962 5963 RGN_NR_BLOCKS (new_rgn_number) = 0; 5964 5965 /* FIXME: This will work only when EBBs are not created. */ 5966 if (new_rgn_number != 0) 5967 RGN_BLOCKS (new_rgn_number) = RGN_BLOCKS (new_rgn_number - 1) + 5968 RGN_NR_BLOCKS (new_rgn_number - 1); 5969 else 5970 RGN_BLOCKS (new_rgn_number) = 0; 5971 5972 /* Set the blocks of the next region so the other functions may 5973 calculate the number of blocks in the region. */ 5974 RGN_BLOCKS (new_rgn_number + 1) = RGN_BLOCKS (new_rgn_number) + 5975 RGN_NR_BLOCKS (new_rgn_number); 5976 5977 nr_regions++; 5978 5979 return new_rgn_number; 5980 } 5981 5982 /* If X has a smaller topological sort number than Y, returns -1; 5983 if greater, returns 1. */ 5984 static int 5985 bb_top_order_comparator (const void *x, const void *y) 5986 { 5987 basic_block bb1 = *(const basic_block *) x; 5988 basic_block bb2 = *(const basic_block *) y; 5989 5990 gcc_assert (bb1 == bb2 5991 || rev_top_order_index[bb1->index] 5992 != rev_top_order_index[bb2->index]); 5993 5994 /* It's a reverse topological order in REV_TOP_ORDER_INDEX, so 5995 bbs with greater number should go earlier. */ 5996 if (rev_top_order_index[bb1->index] > rev_top_order_index[bb2->index]) 5997 return -1; 5998 else 5999 return 1; 6000 } 6001 6002 /* Create a region for LOOP and return its number. If we don't want 6003 to pipeline LOOP, return -1. */ 6004 static int 6005 make_region_from_loop (struct loop *loop) 6006 { 6007 unsigned int i; 6008 int new_rgn_number = -1; 6009 struct loop *inner; 6010 6011 /* Basic block index, to be assigned to BLOCK_TO_BB. */ 6012 int bb_ord_index = 0; 6013 basic_block *loop_blocks; 6014 basic_block preheader_block; 6015 6016 if (loop->num_nodes 6017 > (unsigned) PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_BLOCKS)) 6018 return -1; 6019 6020 /* Don't pipeline loops whose latch belongs to some of its inner loops. */ 6021 for (inner = loop->inner; inner; inner = inner->inner) 6022 if (flow_bb_inside_loop_p (inner, loop->latch)) 6023 return -1; 6024 6025 loop->ninsns = num_loop_insns (loop); 6026 if ((int) loop->ninsns > PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_INSNS)) 6027 return -1; 6028 6029 loop_blocks = get_loop_body_in_custom_order (loop, bb_top_order_comparator); 6030 6031 for (i = 0; i < loop->num_nodes; i++) 6032 if (loop_blocks[i]->flags & BB_IRREDUCIBLE_LOOP) 6033 { 6034 free (loop_blocks); 6035 return -1; 6036 } 6037 6038 preheader_block = loop_preheader_edge (loop)->src; 6039 gcc_assert (preheader_block); 6040 gcc_assert (loop_blocks[0] == loop->header); 6041 6042 new_rgn_number = sel_create_new_region (); 6043 6044 sel_add_block_to_region (preheader_block, &bb_ord_index, new_rgn_number); 6045 bitmap_set_bit (bbs_in_loop_rgns, preheader_block->index); 6046 6047 for (i = 0; i < loop->num_nodes; i++) 6048 { 6049 /* Add only those blocks that haven't been scheduled in the inner loop. 6050 The exception is the basic blocks with bookkeeping code - they should 6051 be added to the region (and they actually don't belong to the loop 6052 body, but to the region containing that loop body). */ 6053 6054 gcc_assert (new_rgn_number >= 0); 6055 6056 if (! bitmap_bit_p (bbs_in_loop_rgns, loop_blocks[i]->index)) 6057 { 6058 sel_add_block_to_region (loop_blocks[i], &bb_ord_index, 6059 new_rgn_number); 6060 bitmap_set_bit (bbs_in_loop_rgns, loop_blocks[i]->index); 6061 } 6062 } 6063 6064 free (loop_blocks); 6065 MARK_LOOP_FOR_PIPELINING (loop); 6066 6067 return new_rgn_number; 6068 } 6069 6070 /* Create a new region from preheader blocks LOOP_BLOCKS. */ 6071 void 6072 make_region_from_loop_preheader (vec<basic_block> *&loop_blocks) 6073 { 6074 unsigned int i; 6075 int new_rgn_number = -1; 6076 basic_block bb; 6077 6078 /* Basic block index, to be assigned to BLOCK_TO_BB. */ 6079 int bb_ord_index = 0; 6080 6081 new_rgn_number = sel_create_new_region (); 6082 6083 FOR_EACH_VEC_ELT (*loop_blocks, i, bb) 6084 { 6085 gcc_assert (new_rgn_number >= 0); 6086 6087 sel_add_block_to_region (bb, &bb_ord_index, new_rgn_number); 6088 } 6089 6090 vec_free (loop_blocks); 6091 } 6092 6093 6094 /* Create region(s) from loop nest LOOP, such that inner loops will be 6095 pipelined before outer loops. Returns true when a region for LOOP 6096 is created. */ 6097 static bool 6098 make_regions_from_loop_nest (struct loop *loop) 6099 { 6100 struct loop *cur_loop; 6101 int rgn_number; 6102 6103 /* Traverse all inner nodes of the loop. */ 6104 for (cur_loop = loop->inner; cur_loop; cur_loop = cur_loop->next) 6105 if (! bitmap_bit_p (bbs_in_loop_rgns, cur_loop->header->index)) 6106 return false; 6107 6108 /* At this moment all regular inner loops should have been pipelined. 6109 Try to create a region from this loop. */ 6110 rgn_number = make_region_from_loop (loop); 6111 6112 if (rgn_number < 0) 6113 return false; 6114 6115 loop_nests.safe_push (loop); 6116 return true; 6117 } 6118 6119 /* Initalize data structures needed. */ 6120 void 6121 sel_init_pipelining (void) 6122 { 6123 /* Collect loop information to be used in outer loops pipelining. */ 6124 loop_optimizer_init (LOOPS_HAVE_PREHEADERS 6125 | LOOPS_HAVE_FALLTHRU_PREHEADERS 6126 | LOOPS_HAVE_RECORDED_EXITS 6127 | LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS); 6128 current_loop_nest = NULL; 6129 6130 bbs_in_loop_rgns = sbitmap_alloc (last_basic_block_for_fn (cfun)); 6131 bitmap_clear (bbs_in_loop_rgns); 6132 6133 recompute_rev_top_order (); 6134 } 6135 6136 /* Returns a struct loop for region RGN. */ 6137 loop_p 6138 get_loop_nest_for_rgn (unsigned int rgn) 6139 { 6140 /* Regions created with extend_rgns don't have corresponding loop nests, 6141 because they don't represent loops. */ 6142 if (rgn < loop_nests.length ()) 6143 return loop_nests[rgn]; 6144 else 6145 return NULL; 6146 } 6147 6148 /* True when LOOP was included into pipelining regions. */ 6149 bool 6150 considered_for_pipelining_p (struct loop *loop) 6151 { 6152 if (loop_depth (loop) == 0) 6153 return false; 6154 6155 /* Now, the loop could be too large or irreducible. Check whether its 6156 region is in LOOP_NESTS. 6157 We determine the region number of LOOP as the region number of its 6158 latch. We can't use header here, because this header could be 6159 just removed preheader and it will give us the wrong region number. 6160 Latch can't be used because it could be in the inner loop too. */ 6161 if (LOOP_MARKED_FOR_PIPELINING_P (loop)) 6162 { 6163 int rgn = CONTAINING_RGN (loop->latch->index); 6164 6165 gcc_assert ((unsigned) rgn < loop_nests.length ()); 6166 return true; 6167 } 6168 6169 return false; 6170 } 6171 6172 /* Makes regions from the rest of the blocks, after loops are chosen 6173 for pipelining. */ 6174 static void 6175 make_regions_from_the_rest (void) 6176 { 6177 int cur_rgn_blocks; 6178 int *loop_hdr; 6179 int i; 6180 6181 basic_block bb; 6182 edge e; 6183 edge_iterator ei; 6184 int *degree; 6185 6186 /* Index in rgn_bb_table where to start allocating new regions. */ 6187 cur_rgn_blocks = nr_regions ? RGN_BLOCKS (nr_regions) : 0; 6188 6189 /* Make regions from all the rest basic blocks - those that don't belong to 6190 any loop or belong to irreducible loops. Prepare the data structures 6191 for extend_rgns. */ 6192 6193 /* LOOP_HDR[I] == -1 if I-th bb doesn't belong to any loop, 6194 LOOP_HDR[I] == LOOP_HDR[J] iff basic blocks I and J reside within the same 6195 loop. */ 6196 loop_hdr = XNEWVEC (int, last_basic_block_for_fn (cfun)); 6197 degree = XCNEWVEC (int, last_basic_block_for_fn (cfun)); 6198 6199 6200 /* For each basic block that belongs to some loop assign the number 6201 of innermost loop it belongs to. */ 6202 for (i = 0; i < last_basic_block_for_fn (cfun); i++) 6203 loop_hdr[i] = -1; 6204 6205 FOR_EACH_BB_FN (bb, cfun) 6206 { 6207 if (bb->loop_father && bb->loop_father->num != 0 6208 && !(bb->flags & BB_IRREDUCIBLE_LOOP)) 6209 loop_hdr[bb->index] = bb->loop_father->num; 6210 } 6211 6212 /* For each basic block degree is calculated as the number of incoming 6213 edges, that are going out of bbs that are not yet scheduled. 6214 The basic blocks that are scheduled have degree value of zero. */ 6215 FOR_EACH_BB_FN (bb, cfun) 6216 { 6217 degree[bb->index] = 0; 6218 6219 if (!bitmap_bit_p (bbs_in_loop_rgns, bb->index)) 6220 { 6221 FOR_EACH_EDGE (e, ei, bb->preds) 6222 if (!bitmap_bit_p (bbs_in_loop_rgns, e->src->index)) 6223 degree[bb->index]++; 6224 } 6225 else 6226 degree[bb->index] = -1; 6227 } 6228 6229 extend_rgns (degree, &cur_rgn_blocks, bbs_in_loop_rgns, loop_hdr); 6230 6231 /* Any block that did not end up in a region is placed into a region 6232 by itself. */ 6233 FOR_EACH_BB_FN (bb, cfun) 6234 if (degree[bb->index] >= 0) 6235 { 6236 rgn_bb_table[cur_rgn_blocks] = bb->index; 6237 RGN_NR_BLOCKS (nr_regions) = 1; 6238 RGN_BLOCKS (nr_regions) = cur_rgn_blocks++; 6239 RGN_DONT_CALC_DEPS (nr_regions) = 0; 6240 RGN_HAS_REAL_EBB (nr_regions) = 0; 6241 CONTAINING_RGN (bb->index) = nr_regions++; 6242 BLOCK_TO_BB (bb->index) = 0; 6243 } 6244 6245 free (degree); 6246 free (loop_hdr); 6247 } 6248 6249 /* Free data structures used in pipelining of loops. */ 6250 void sel_finish_pipelining (void) 6251 { 6252 struct loop *loop; 6253 6254 /* Release aux fields so we don't free them later by mistake. */ 6255 FOR_EACH_LOOP (loop, 0) 6256 loop->aux = NULL; 6257 6258 loop_optimizer_finalize (); 6259 6260 loop_nests.release (); 6261 6262 free (rev_top_order_index); 6263 rev_top_order_index = NULL; 6264 } 6265 6266 /* This function replaces the find_rgns when 6267 FLAG_SEL_SCHED_PIPELINING_OUTER_LOOPS is set. */ 6268 void 6269 sel_find_rgns (void) 6270 { 6271 sel_init_pipelining (); 6272 extend_regions (); 6273 6274 if (current_loops) 6275 { 6276 loop_p loop; 6277 6278 FOR_EACH_LOOP (loop, (flag_sel_sched_pipelining_outer_loops 6279 ? LI_FROM_INNERMOST 6280 : LI_ONLY_INNERMOST)) 6281 make_regions_from_loop_nest (loop); 6282 } 6283 6284 /* Make regions from all the rest basic blocks and schedule them. 6285 These blocks include blocks that don't belong to any loop or belong 6286 to irreducible loops. */ 6287 make_regions_from_the_rest (); 6288 6289 /* We don't need bbs_in_loop_rgns anymore. */ 6290 sbitmap_free (bbs_in_loop_rgns); 6291 bbs_in_loop_rgns = NULL; 6292 } 6293 6294 /* Add the preheader blocks from previous loop to current region taking 6295 it from LOOP_PREHEADER_BLOCKS (current_loop_nest) and record them in *BBS. 6296 This function is only used with -fsel-sched-pipelining-outer-loops. */ 6297 void 6298 sel_add_loop_preheaders (bb_vec_t *bbs) 6299 { 6300 int i; 6301 basic_block bb; 6302 vec<basic_block> *preheader_blocks 6303 = LOOP_PREHEADER_BLOCKS (current_loop_nest); 6304 6305 if (!preheader_blocks) 6306 return; 6307 6308 for (i = 0; preheader_blocks->iterate (i, &bb); i++) 6309 { 6310 bbs->safe_push (bb); 6311 last_added_blocks.safe_push (bb); 6312 sel_add_bb (bb); 6313 } 6314 6315 vec_free (preheader_blocks); 6316 } 6317 6318 /* While pipelining outer loops, returns TRUE if BB is a loop preheader. 6319 Please note that the function should also work when pipelining_p is 6320 false, because it is used when deciding whether we should or should 6321 not reschedule pipelined code. */ 6322 bool 6323 sel_is_loop_preheader_p (basic_block bb) 6324 { 6325 if (current_loop_nest) 6326 { 6327 struct loop *outer; 6328 6329 if (preheader_removed) 6330 return false; 6331 6332 /* Preheader is the first block in the region. */ 6333 if (BLOCK_TO_BB (bb->index) == 0) 6334 return true; 6335 6336 /* We used to find a preheader with the topological information. 6337 Check that the above code is equivalent to what we did before. */ 6338 6339 if (in_current_region_p (current_loop_nest->header)) 6340 gcc_assert (!(BLOCK_TO_BB (bb->index) 6341 < BLOCK_TO_BB (current_loop_nest->header->index))); 6342 6343 /* Support the situation when the latch block of outer loop 6344 could be from here. */ 6345 for (outer = loop_outer (current_loop_nest); 6346 outer; 6347 outer = loop_outer (outer)) 6348 if (considered_for_pipelining_p (outer) && outer->latch == bb) 6349 gcc_unreachable (); 6350 } 6351 6352 return false; 6353 } 6354 6355 /* Check whether JUMP_BB ends with a jump insn that leads only to DEST_BB and 6356 can be removed, making the corresponding edge fallthrough (assuming that 6357 all basic blocks between JUMP_BB and DEST_BB are empty). */ 6358 static bool 6359 bb_has_removable_jump_to_p (basic_block jump_bb, basic_block dest_bb) 6360 { 6361 if (!onlyjump_p (BB_END (jump_bb)) 6362 || tablejump_p (BB_END (jump_bb), NULL, NULL)) 6363 return false; 6364 6365 /* Several outgoing edges, abnormal edge or destination of jump is 6366 not DEST_BB. */ 6367 if (EDGE_COUNT (jump_bb->succs) != 1 6368 || EDGE_SUCC (jump_bb, 0)->flags & (EDGE_ABNORMAL | EDGE_CROSSING) 6369 || EDGE_SUCC (jump_bb, 0)->dest != dest_bb) 6370 return false; 6371 6372 /* If not anything of the upper. */ 6373 return true; 6374 } 6375 6376 /* Removes the loop preheader from the current region and saves it in 6377 PREHEADER_BLOCKS of the father loop, so they will be added later to 6378 region that represents an outer loop. */ 6379 static void 6380 sel_remove_loop_preheader (void) 6381 { 6382 int i, old_len; 6383 int cur_rgn = CONTAINING_RGN (BB_TO_BLOCK (0)); 6384 basic_block bb; 6385 bool all_empty_p = true; 6386 vec<basic_block> *preheader_blocks 6387 = LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest)); 6388 6389 vec_check_alloc (preheader_blocks, 0); 6390 6391 gcc_assert (current_loop_nest); 6392 old_len = preheader_blocks->length (); 6393 6394 /* Add blocks that aren't within the current loop to PREHEADER_BLOCKS. */ 6395 for (i = 0; i < RGN_NR_BLOCKS (cur_rgn); i++) 6396 { 6397 bb = BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i)); 6398 6399 /* If the basic block belongs to region, but doesn't belong to 6400 corresponding loop, then it should be a preheader. */ 6401 if (sel_is_loop_preheader_p (bb)) 6402 { 6403 preheader_blocks->safe_push (bb); 6404 if (BB_END (bb) != bb_note (bb)) 6405 all_empty_p = false; 6406 } 6407 } 6408 6409 /* Remove these blocks only after iterating over the whole region. */ 6410 for (i = preheader_blocks->length () - 1; i >= old_len; i--) 6411 { 6412 bb = (*preheader_blocks)[i]; 6413 sel_remove_bb (bb, false); 6414 } 6415 6416 if (!considered_for_pipelining_p (loop_outer (current_loop_nest))) 6417 { 6418 if (!all_empty_p) 6419 /* Immediately create new region from preheader. */ 6420 make_region_from_loop_preheader (preheader_blocks); 6421 else 6422 { 6423 /* If all preheader blocks are empty - dont create new empty region. 6424 Instead, remove them completely. */ 6425 FOR_EACH_VEC_ELT (*preheader_blocks, i, bb) 6426 { 6427 edge e; 6428 edge_iterator ei; 6429 basic_block prev_bb = bb->prev_bb, next_bb = bb->next_bb; 6430 6431 /* Redirect all incoming edges to next basic block. */ 6432 for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei)); ) 6433 { 6434 if (! (e->flags & EDGE_FALLTHRU)) 6435 redirect_edge_and_branch (e, bb->next_bb); 6436 else 6437 redirect_edge_succ (e, bb->next_bb); 6438 } 6439 gcc_assert (BB_NOTE_LIST (bb) == NULL); 6440 delete_and_free_basic_block (bb); 6441 6442 /* Check if after deleting preheader there is a nonconditional 6443 jump in PREV_BB that leads to the next basic block NEXT_BB. 6444 If it is so - delete this jump and clear data sets of its 6445 basic block if it becomes empty. */ 6446 if (next_bb->prev_bb == prev_bb 6447 && prev_bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) 6448 && bb_has_removable_jump_to_p (prev_bb, next_bb)) 6449 { 6450 redirect_edge_and_branch (EDGE_SUCC (prev_bb, 0), next_bb); 6451 if (BB_END (prev_bb) == bb_note (prev_bb)) 6452 free_data_sets (prev_bb); 6453 } 6454 6455 set_immediate_dominator (CDI_DOMINATORS, next_bb, 6456 recompute_dominator (CDI_DOMINATORS, 6457 next_bb)); 6458 } 6459 } 6460 vec_free (preheader_blocks); 6461 } 6462 else 6463 /* Store preheader within the father's loop structure. */ 6464 SET_LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest), 6465 preheader_blocks); 6466 } 6467 6468 #endif 6469