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