1 /* Instruction scheduling pass. This file contains definitions used 2 internally in the scheduler. 3 Copyright (C) 1992-2017 Free Software Foundation, Inc. 4 5 This file is part of GCC. 6 7 GCC is free software; you can redistribute it and/or modify it under 8 the terms of the GNU General Public License as published by the Free 9 Software Foundation; either version 3, or (at your option) any later 10 version. 11 12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY 13 WARRANTY; without even the implied warranty of MERCHANTABILITY or 14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 15 for more details. 16 17 You should have received a copy of the GNU General Public License 18 along with GCC; see the file COPYING3. If not see 19 <http://www.gnu.org/licenses/>. */ 20 21 #ifndef GCC_SCHED_INT_H 22 #define GCC_SCHED_INT_H 23 24 #ifdef INSN_SCHEDULING 25 26 /* Identificator of a scheduler pass. */ 27 enum sched_pass_id_t { SCHED_PASS_UNKNOWN, SCHED_RGN_PASS, SCHED_EBB_PASS, 28 SCHED_SMS_PASS, SCHED_SEL_PASS }; 29 30 /* The algorithm used to implement -fsched-pressure. */ 31 enum sched_pressure_algorithm 32 { 33 SCHED_PRESSURE_NONE, 34 SCHED_PRESSURE_WEIGHTED, 35 SCHED_PRESSURE_MODEL 36 }; 37 38 typedef vec<basic_block> bb_vec_t; 39 typedef vec<rtx_insn *> insn_vec_t; 40 typedef vec<rtx_insn *> rtx_vec_t; 41 42 extern void sched_init_bbs (void); 43 44 extern void sched_extend_luids (void); 45 extern void sched_init_insn_luid (rtx_insn *); 46 extern void sched_init_luids (bb_vec_t); 47 extern void sched_finish_luids (void); 48 49 extern void sched_extend_target (void); 50 51 extern void haifa_init_h_i_d (bb_vec_t); 52 extern void haifa_finish_h_i_d (void); 53 54 /* Hooks that are common to all the schedulers. */ 55 struct common_sched_info_def 56 { 57 /* Called after blocks were rearranged due to movement of jump instruction. 58 The first parameter - index of basic block, in which jump currently is. 59 The second parameter - index of basic block, in which jump used 60 to be. 61 The third parameter - index of basic block, that follows the second 62 parameter. */ 63 void (*fix_recovery_cfg) (int, int, int); 64 65 /* Called to notify frontend, that new basic block is being added. 66 The first parameter - new basic block. 67 The second parameter - block, after which new basic block is being added, 68 or the exit block, if recovery block is being added, 69 or NULL, if standalone block is being added. */ 70 void (*add_block) (basic_block, basic_block); 71 72 /* Estimate number of insns in the basic block. */ 73 int (*estimate_number_of_insns) (basic_block); 74 75 /* Given a non-insn (!INSN_P (x)) return 76 -1 - if this rtx don't need a luid. 77 0 - if it should have the same luid as the previous insn. 78 1 - if it needs a separate luid. */ 79 int (*luid_for_non_insn) (rtx); 80 81 /* Scheduler pass identifier. It is preferably used in assertions. */ 82 enum sched_pass_id_t sched_pass_id; 83 }; 84 85 extern struct common_sched_info_def *common_sched_info; 86 87 extern const struct common_sched_info_def haifa_common_sched_info; 88 89 /* Return true if selective scheduling pass is working. */ 90 static inline bool 91 sel_sched_p (void) 92 { 93 return common_sched_info->sched_pass_id == SCHED_SEL_PASS; 94 } 95 96 /* Returns maximum priority that an insn was assigned to. */ 97 extern int get_rgn_sched_max_insns_priority (void); 98 99 /* Increases effective priority for INSN by AMOUNT. */ 100 extern void sel_add_to_insn_priority (rtx, int); 101 102 /* True if during selective scheduling we need to emulate some of haifa 103 scheduler behavior. */ 104 extern int sched_emulate_haifa_p; 105 106 /* Mapping from INSN_UID to INSN_LUID. In the end all other per insn data 107 structures should be indexed by luid. */ 108 extern vec<int> sched_luids; 109 #define INSN_LUID(INSN) (sched_luids[INSN_UID (INSN)]) 110 #define LUID_BY_UID(UID) (sched_luids[UID]) 111 112 #define SET_INSN_LUID(INSN, LUID) \ 113 (sched_luids[INSN_UID (INSN)] = (LUID)) 114 115 /* The highest INSN_LUID. */ 116 extern int sched_max_luid; 117 118 extern int insn_luid (rtx); 119 120 /* This list holds ripped off notes from the current block. These notes will 121 be attached to the beginning of the block when its scheduling is 122 finished. */ 123 extern rtx_insn *note_list; 124 125 extern void remove_notes (rtx_insn *, rtx_insn *); 126 extern rtx_insn *restore_other_notes (rtx_insn *, basic_block); 127 extern void sched_insns_init (rtx); 128 extern void sched_insns_finish (void); 129 130 extern void *xrecalloc (void *, size_t, size_t, size_t); 131 132 extern void reemit_notes (rtx_insn *); 133 134 /* Functions in haifa-sched.c. */ 135 extern int haifa_classify_insn (const_rtx); 136 137 /* Functions in sel-sched-ir.c. */ 138 extern void sel_find_rgns (void); 139 extern void sel_mark_hard_insn (rtx); 140 141 extern size_t dfa_state_size; 142 143 extern void advance_state (state_t); 144 145 extern void setup_sched_dump (void); 146 extern void sched_init (void); 147 extern void sched_finish (void); 148 149 extern bool sel_insn_is_speculation_check (rtx); 150 151 /* Describe the ready list of the scheduler. 152 VEC holds space enough for all insns in the current region. VECLEN 153 says how many exactly. 154 FIRST is the index of the element with the highest priority; i.e. the 155 last one in the ready list, since elements are ordered by ascending 156 priority. 157 N_READY determines how many insns are on the ready list. 158 N_DEBUG determines how many debug insns are on the ready list. */ 159 struct ready_list 160 { 161 rtx_insn **vec; 162 int veclen; 163 int first; 164 int n_ready; 165 int n_debug; 166 }; 167 168 extern signed char *ready_try; 169 extern struct ready_list ready; 170 171 extern int max_issue (struct ready_list *, int, state_t, bool, int *); 172 173 extern void ebb_compute_jump_reg_dependencies (rtx, regset); 174 175 extern edge find_fallthru_edge_from (basic_block); 176 177 extern void (* sched_init_only_bb) (basic_block, basic_block); 178 extern basic_block (* sched_split_block) (basic_block, rtx); 179 extern basic_block sched_split_block_1 (basic_block, rtx); 180 extern basic_block (* sched_create_empty_bb) (basic_block); 181 extern basic_block sched_create_empty_bb_1 (basic_block); 182 183 extern basic_block sched_create_recovery_block (basic_block *); 184 extern void sched_create_recovery_edges (basic_block, basic_block, 185 basic_block); 186 187 /* Pointer to data describing the current DFA state. */ 188 extern state_t curr_state; 189 190 /* Type to represent status of a dependence. */ 191 typedef unsigned int ds_t; 192 #define BITS_PER_DEP_STATUS HOST_BITS_PER_INT 193 194 /* Type to represent weakness of speculative dependence. */ 195 typedef unsigned int dw_t; 196 197 extern enum reg_note ds_to_dk (ds_t); 198 extern ds_t dk_to_ds (enum reg_note); 199 200 /* Describe a dependency that can be broken by making a replacement 201 in one of the patterns. LOC is the location, ORIG and NEWVAL the 202 two alternative contents, and INSN the instruction that must be 203 changed. */ 204 struct dep_replacement 205 { 206 rtx *loc; 207 rtx orig; 208 rtx newval; 209 rtx_insn *insn; 210 }; 211 212 /* Information about the dependency. */ 213 struct _dep 214 { 215 /* Producer. */ 216 rtx_insn *pro; 217 218 /* Consumer. */ 219 rtx_insn *con; 220 221 /* If nonnull, holds a pointer to information about how to break the 222 dependency by making a replacement in one of the insns. There is 223 only one such dependency for each insn that must be modified in 224 order to break such a dependency. */ 225 struct dep_replacement *replace; 226 227 /* Dependency status. This field holds all dependency types and additional 228 information for speculative dependencies. */ 229 ds_t status; 230 231 /* Dependency major type. This field is superseded by STATUS above. 232 Though, it is still in place because some targets use it. */ 233 ENUM_BITFIELD(reg_note) type:6; 234 235 unsigned nonreg:1; 236 unsigned multiple:1; 237 238 /* Cached cost of the dependency. Make sure to update UNKNOWN_DEP_COST 239 when changing the size of this field. */ 240 int cost:20; 241 }; 242 243 #define UNKNOWN_DEP_COST ((int) ((unsigned int) -1 << 19)) 244 245 typedef struct _dep dep_def; 246 typedef dep_def *dep_t; 247 248 #define DEP_PRO(D) ((D)->pro) 249 #define DEP_CON(D) ((D)->con) 250 #define DEP_TYPE(D) ((D)->type) 251 #define DEP_STATUS(D) ((D)->status) 252 #define DEP_COST(D) ((D)->cost) 253 #define DEP_NONREG(D) ((D)->nonreg) 254 #define DEP_MULTIPLE(D) ((D)->multiple) 255 #define DEP_REPLACE(D) ((D)->replace) 256 257 /* Functions to work with dep. */ 258 259 extern void init_dep_1 (dep_t, rtx_insn *, rtx_insn *, enum reg_note, ds_t); 260 extern void init_dep (dep_t, rtx_insn *, rtx_insn *, enum reg_note); 261 262 extern void sd_debug_dep (dep_t); 263 264 /* Definition of this struct resides below. */ 265 struct _dep_node; 266 typedef struct _dep_node *dep_node_t; 267 268 /* A link in the dependency list. This is essentially an equivalent of a 269 single {INSN, DEPS}_LIST rtx. */ 270 struct _dep_link 271 { 272 /* Dep node with all the data. */ 273 dep_node_t node; 274 275 /* Next link in the list. For the last one it is NULL. */ 276 struct _dep_link *next; 277 278 /* Pointer to the next field of the previous link in the list. 279 For the first link this points to the deps_list->first. 280 281 With help of this field it is easy to remove and insert links to the 282 list. */ 283 struct _dep_link **prev_nextp; 284 }; 285 typedef struct _dep_link *dep_link_t; 286 287 #define DEP_LINK_NODE(N) ((N)->node) 288 #define DEP_LINK_NEXT(N) ((N)->next) 289 #define DEP_LINK_PREV_NEXTP(N) ((N)->prev_nextp) 290 291 /* Macros to work dep_link. For most usecases only part of the dependency 292 information is need. These macros conveniently provide that piece of 293 information. */ 294 295 #define DEP_LINK_DEP(N) (DEP_NODE_DEP (DEP_LINK_NODE (N))) 296 #define DEP_LINK_PRO(N) (DEP_PRO (DEP_LINK_DEP (N))) 297 #define DEP_LINK_CON(N) (DEP_CON (DEP_LINK_DEP (N))) 298 #define DEP_LINK_TYPE(N) (DEP_TYPE (DEP_LINK_DEP (N))) 299 #define DEP_LINK_STATUS(N) (DEP_STATUS (DEP_LINK_DEP (N))) 300 301 /* A list of dep_links. */ 302 struct _deps_list 303 { 304 /* First element. */ 305 dep_link_t first; 306 307 /* Total number of elements in the list. */ 308 int n_links; 309 }; 310 typedef struct _deps_list *deps_list_t; 311 312 #define DEPS_LIST_FIRST(L) ((L)->first) 313 #define DEPS_LIST_N_LINKS(L) ((L)->n_links) 314 315 /* Suppose we have a dependence Y between insn pro1 and con1, where pro1 has 316 additional dependents con0 and con2, and con1 is dependent on additional 317 insns pro0 and pro1: 318 319 .con0 pro0 320 . ^ | 321 . | | 322 . | | 323 . X A 324 . | | 325 . | | 326 . | V 327 .pro1--Y-->con1 328 . | ^ 329 . | | 330 . | | 331 . Z B 332 . | | 333 . | | 334 . V | 335 .con2 pro2 336 337 This is represented using a "dep_node" for each dependence arc, which are 338 connected as follows (diagram is centered around Y which is fully shown; 339 other dep_nodes shown partially): 340 341 . +------------+ +--------------+ +------------+ 342 . : dep_node X : | dep_node Y | : dep_node Z : 343 . : : | | : : 344 . : : | | : : 345 . : forw : | forw | : forw : 346 . : +--------+ : | +--------+ | : +--------+ : 347 forw_deps : |dep_link| : | |dep_link| | : |dep_link| : 348 +-----+ : | +----+ | : | | +----+ | | : | +----+ | : 349 |first|----->| |next|-+------+->| |next|-+--+----->| |next|-+--->NULL 350 +-----+ : | +----+ | : | | +----+ | | : | +----+ | : 351 . ^ ^ : | ^ | : | | ^ | | : | | : 352 . | | : | | | : | | | | | : | | : 353 . | +--<----+--+ +--+---<--+--+--+ +--+--+--<---+--+ | : 354 . | : | | | : | | | | | : | | | : 355 . | : | +----+ | : | | +----+ | | : | +----+ | : 356 . | : | |prev| | : | | |prev| | | : | |prev| | : 357 . | : | |next| | : | | |next| | | : | |next| | : 358 . | : | +----+ | : | | +----+ | | : | +----+ | : 359 . | : | | :<-+ | | | |<-+ : | | :<-+ 360 . | : | +----+ | : | | | +----+ | | | : | +----+ | : | 361 . | : | |node|-+----+ | | |node|-+--+--+ : | |node|-+----+ 362 . | : | +----+ | : | | +----+ | | : | +----+ | : 363 . | : | | : | | | | : | | : 364 . | : +--------+ : | +--------+ | : +--------+ : 365 . | : : | | : : 366 . | : SAME pro1 : | +--------+ | : SAME pro1 : 367 . | : DIFF con0 : | |dep | | : DIFF con2 : 368 . | : : | | | | : : 369 . | | | +----+ | | 370 .RTX<------------------------+--+-|pro1| | | 371 .pro1 | | +----+ | | 372 . | | | | 373 . | | +----+ | | 374 .RTX<------------------------+--+-|con1| | | 375 .con1 | | +----+ | | 376 . | | | | | 377 . | | | +----+ | | 378 . | | | |kind| | | 379 . | | | +----+ | | 380 . | : : | | |stat| | | : : 381 . | : DIFF pro0 : | | +----+ | | : DIFF pro2 : 382 . | : SAME con1 : | | | | : SAME con1 : 383 . | : : | +--------+ | : : 384 . | : : | | : : 385 . | : back : | back | : back : 386 . v : +--------+ : | +--------+ | : +--------+ : 387 back_deps : |dep_link| : | |dep_link| | : |dep_link| : 388 +-----+ : | +----+ | : | | +----+ | | : | +----+ | : 389 |first|----->| |next|-+------+->| |next|-+--+----->| |next|-+--->NULL 390 +-----+ : | +----+ | : | | +----+ | | : | +----+ | : 391 . ^ : | ^ | : | | ^ | | : | | : 392 . | : | | | : | | | | | : | | : 393 . +--<----+--+ +--+---<--+--+--+ +--+--+--<---+--+ | : 394 . : | | | : | | | | | : | | | : 395 . : | +----+ | : | | +----+ | | : | +----+ | : 396 . : | |prev| | : | | |prev| | | : | |prev| | : 397 . : | |next| | : | | |next| | | : | |next| | : 398 . : | +----+ | : | | +----+ | | : | +----+ | : 399 . : | | :<-+ | | | |<-+ : | | :<-+ 400 . : | +----+ | : | | | +----+ | | | : | +----+ | : | 401 . : | |node|-+----+ | | |node|-+--+--+ : | |node|-+----+ 402 . : | +----+ | : | | +----+ | | : | +----+ | : 403 . : | | : | | | | : | | : 404 . : +--------+ : | +--------+ | : +--------+ : 405 . : : | | : : 406 . : dep_node A : | dep_node Y | : dep_node B : 407 . +------------+ +--------------+ +------------+ 408 */ 409 410 struct _dep_node 411 { 412 /* Backward link. */ 413 struct _dep_link back; 414 415 /* The dep. */ 416 struct _dep dep; 417 418 /* Forward link. */ 419 struct _dep_link forw; 420 }; 421 422 #define DEP_NODE_BACK(N) (&(N)->back) 423 #define DEP_NODE_DEP(N) (&(N)->dep) 424 #define DEP_NODE_FORW(N) (&(N)->forw) 425 426 /* The following enumeration values tell us what dependencies we 427 should use to implement the barrier. We use true-dependencies for 428 TRUE_BARRIER and anti-dependencies for MOVE_BARRIER. */ 429 enum reg_pending_barrier_mode 430 { 431 NOT_A_BARRIER = 0, 432 MOVE_BARRIER, 433 TRUE_BARRIER 434 }; 435 436 /* Whether a register movement is associated with a call. */ 437 enum post_call_group 438 { 439 not_post_call, 440 post_call, 441 post_call_initial 442 }; 443 444 /* Insns which affect pseudo-registers. */ 445 struct deps_reg 446 { 447 rtx_insn_list *uses; 448 rtx_insn_list *sets; 449 rtx_insn_list *implicit_sets; 450 rtx_insn_list *control_uses; 451 rtx_insn_list *clobbers; 452 int uses_length; 453 int clobbers_length; 454 }; 455 456 /* Describe state of dependencies used during sched_analyze phase. */ 457 struct deps_desc 458 { 459 /* The *_insns and *_mems are paired lists. Each pending memory operation 460 will have a pointer to the MEM rtx on one list and a pointer to the 461 containing insn on the other list in the same place in the list. */ 462 463 /* We can't use add_dependence like the old code did, because a single insn 464 may have multiple memory accesses, and hence needs to be on the list 465 once for each memory access. Add_dependence won't let you add an insn 466 to a list more than once. */ 467 468 /* An INSN_LIST containing all insns with pending read operations. */ 469 rtx_insn_list *pending_read_insns; 470 471 /* An EXPR_LIST containing all MEM rtx's which are pending reads. */ 472 rtx_expr_list *pending_read_mems; 473 474 /* An INSN_LIST containing all insns with pending write operations. */ 475 rtx_insn_list *pending_write_insns; 476 477 /* An EXPR_LIST containing all MEM rtx's which are pending writes. */ 478 rtx_expr_list *pending_write_mems; 479 480 /* An INSN_LIST containing all jump insns. */ 481 rtx_insn_list *pending_jump_insns; 482 483 /* We must prevent the above lists from ever growing too large since 484 the number of dependencies produced is at least O(N*N), 485 and execution time is at least O(4*N*N), as a function of the 486 length of these pending lists. */ 487 488 /* Indicates the length of the pending_read list. */ 489 int pending_read_list_length; 490 491 /* Indicates the length of the pending_write list. */ 492 int pending_write_list_length; 493 494 /* Length of the pending memory flush list plus the length of the pending 495 jump insn list. Large functions with no calls may build up extremely 496 large lists. */ 497 int pending_flush_length; 498 499 /* The last insn upon which all memory references must depend. 500 This is an insn which flushed the pending lists, creating a dependency 501 between it and all previously pending memory references. This creates 502 a barrier (or a checkpoint) which no memory reference is allowed to cross. 503 504 This includes all non constant CALL_INSNs. When we do interprocedural 505 alias analysis, this restriction can be relaxed. 506 This may also be an INSN that writes memory if the pending lists grow 507 too large. */ 508 rtx_insn_list *last_pending_memory_flush; 509 510 /* A list of the last function calls we have seen. We use a list to 511 represent last function calls from multiple predecessor blocks. 512 Used to prevent register lifetimes from expanding unnecessarily. */ 513 rtx_insn_list *last_function_call; 514 515 /* A list of the last function calls that may not return normally 516 we have seen. We use a list to represent last function calls from 517 multiple predecessor blocks. Used to prevent moving trapping insns 518 across such calls. */ 519 rtx_insn_list *last_function_call_may_noreturn; 520 521 /* A list of insns which use a pseudo register that does not already 522 cross a call. We create dependencies between each of those insn 523 and the next call insn, to ensure that they won't cross a call after 524 scheduling is done. */ 525 rtx_insn_list *sched_before_next_call; 526 527 /* Similarly, a list of insns which should not cross a branch. */ 528 rtx_insn_list *sched_before_next_jump; 529 530 /* Used to keep post-call pseudo/hard reg movements together with 531 the call. */ 532 enum post_call_group in_post_call_group_p; 533 534 /* The last debug insn we've seen. */ 535 rtx_insn *last_debug_insn; 536 537 /* The last insn bearing REG_ARGS_SIZE that we've seen. */ 538 rtx_insn *last_args_size; 539 540 /* A list of all prologue insns we have seen without intervening epilogue 541 insns, and one of all epilogue insns we have seen without intervening 542 prologue insns. This is used to prevent mixing prologue and epilogue 543 insns. See PR78029. */ 544 rtx_insn_list *last_prologue; 545 rtx_insn_list *last_epilogue; 546 547 /* Whether the last *logue insn was an epilogue insn or a prologue insn 548 instead. */ 549 bool last_logue_was_epilogue; 550 551 /* The maximum register number for the following arrays. Before reload 552 this is max_reg_num; after reload it is FIRST_PSEUDO_REGISTER. */ 553 int max_reg; 554 555 /* Element N is the next insn that sets (hard or pseudo) register 556 N within the current basic block; or zero, if there is no 557 such insn. Needed for new registers which may be introduced 558 by splitting insns. */ 559 struct deps_reg *reg_last; 560 561 /* Element N is set for each register that has any nonzero element 562 in reg_last[N].{uses,sets,clobbers}. */ 563 regset_head reg_last_in_use; 564 565 /* Shows the last value of reg_pending_barrier associated with the insn. */ 566 enum reg_pending_barrier_mode last_reg_pending_barrier; 567 568 /* True when this context should be treated as a readonly by 569 the analysis. */ 570 BOOL_BITFIELD readonly : 1; 571 }; 572 573 typedef struct deps_desc *deps_t; 574 575 /* This structure holds some state of the current scheduling pass, and 576 contains some function pointers that abstract out some of the non-generic 577 functionality from functions such as schedule_block or schedule_insn. 578 There is one global variable, current_sched_info, which points to the 579 sched_info structure currently in use. */ 580 struct haifa_sched_info 581 { 582 /* Add all insns that are initially ready to the ready list. Called once 583 before scheduling a set of insns. */ 584 void (*init_ready_list) (void); 585 /* Called after taking an insn from the ready list. Returns nonzero if 586 this insn can be scheduled, nonzero if we should silently discard it. */ 587 int (*can_schedule_ready_p) (rtx_insn *); 588 /* Return nonzero if there are more insns that should be scheduled. */ 589 int (*schedule_more_p) (void); 590 /* Called after an insn has all its hard dependencies resolved. 591 Adjusts status of instruction (which is passed through second parameter) 592 to indicate if instruction should be moved to the ready list or the 593 queue, or if it should silently discard it (until next resolved 594 dependence). */ 595 ds_t (*new_ready) (rtx_insn *, ds_t); 596 /* Compare priority of two insns. Return a positive number if the second 597 insn is to be preferred for scheduling, and a negative one if the first 598 is to be preferred. Zero if they are equally good. */ 599 int (*rank) (rtx_insn *, rtx_insn *); 600 /* Return a string that contains the insn uid and optionally anything else 601 necessary to identify this insn in an output. It's valid to use a 602 static buffer for this. The ALIGNED parameter should cause the string 603 to be formatted so that multiple output lines will line up nicely. */ 604 const char *(*print_insn) (const rtx_insn *, int); 605 /* Return nonzero if an insn should be included in priority 606 calculations. */ 607 int (*contributes_to_priority) (rtx_insn *, rtx_insn *); 608 609 /* Return true if scheduling insn (passed as the parameter) will trigger 610 finish of scheduling current block. */ 611 bool (*insn_finishes_block_p) (rtx_insn *); 612 613 /* The boundaries of the set of insns to be scheduled. */ 614 rtx_insn *prev_head, *next_tail; 615 616 /* Filled in after the schedule is finished; the first and last scheduled 617 insns. */ 618 rtx_insn *head, *tail; 619 620 /* If nonzero, enables an additional sanity check in schedule_block. */ 621 unsigned int queue_must_finish_empty:1; 622 623 /* Maximum priority that has been assigned to an insn. */ 624 int sched_max_insns_priority; 625 626 /* Hooks to support speculative scheduling. */ 627 628 /* Called to notify frontend that instruction is being added (second 629 parameter == 0) or removed (second parameter == 1). */ 630 void (*add_remove_insn) (rtx_insn *, int); 631 632 /* Called to notify the frontend that instruction INSN is being 633 scheduled. */ 634 void (*begin_schedule_ready) (rtx_insn *insn); 635 636 /* Called to notify the frontend that an instruction INSN is about to be 637 moved to its correct place in the final schedule. This is done for all 638 insns in order of the schedule. LAST indicates the last scheduled 639 instruction. */ 640 void (*begin_move_insn) (rtx_insn *insn, rtx_insn *last); 641 642 /* If the second parameter is not NULL, return nonnull value, if the 643 basic block should be advanced. 644 If the second parameter is NULL, return the next basic block in EBB. 645 The first parameter is the current basic block in EBB. */ 646 basic_block (*advance_target_bb) (basic_block, rtx_insn *); 647 648 /* Allocate memory, store the frontend scheduler state in it, and 649 return it. */ 650 void *(*save_state) (void); 651 /* Restore frontend scheduler state from the argument, and free the 652 memory. */ 653 void (*restore_state) (void *); 654 655 /* ??? FIXME: should use straight bitfields inside sched_info instead of 656 this flag field. */ 657 unsigned int flags; 658 }; 659 660 /* This structure holds description of the properties for speculative 661 scheduling. */ 662 struct spec_info_def 663 { 664 /* Holds types of allowed speculations: BEGIN_{DATA|CONTROL}, 665 BE_IN_{DATA_CONTROL}. */ 666 int mask; 667 668 /* A dump file for additional information on speculative scheduling. */ 669 FILE *dump; 670 671 /* Minimal cumulative weakness of speculative instruction's 672 dependencies, so that insn will be scheduled. */ 673 dw_t data_weakness_cutoff; 674 675 /* Minimal usefulness of speculative instruction to be considered for 676 scheduling. */ 677 int control_weakness_cutoff; 678 679 /* Flags from the enum SPEC_SCHED_FLAGS. */ 680 int flags; 681 }; 682 typedef struct spec_info_def *spec_info_t; 683 684 extern spec_info_t spec_info; 685 686 extern struct haifa_sched_info *current_sched_info; 687 688 /* Do register pressure sensitive insn scheduling if the flag is set 689 up. */ 690 extern enum sched_pressure_algorithm sched_pressure; 691 692 /* Map regno -> its pressure class. The map defined only when 693 SCHED_PRESSURE_P is true. */ 694 extern enum reg_class *sched_regno_pressure_class; 695 696 /* Indexed by INSN_UID, the collection of all data associated with 697 a single instruction. */ 698 699 struct _haifa_deps_insn_data 700 { 701 /* The number of incoming edges in the forward dependency graph. 702 As scheduling proceeds, counts are decreased. An insn moves to 703 the ready queue when its counter reaches zero. */ 704 int dep_count; 705 706 /* Nonzero if instruction has internal dependence 707 (e.g. add_dependence was invoked with (insn == elem)). */ 708 unsigned int has_internal_dep; 709 710 /* NB: We can't place 'struct _deps_list' here instead of deps_list_t into 711 h_i_d because when h_i_d extends, addresses of the deps_list->first 712 change without updating deps_list->first->next->prev_nextp. Thus 713 BACK_DEPS and RESOLVED_BACK_DEPS are allocated on the heap and FORW_DEPS 714 list is allocated on the obstack. */ 715 716 /* A list of hard backward dependencies. The insn is a consumer of all the 717 deps mentioned here. */ 718 deps_list_t hard_back_deps; 719 720 /* A list of speculative (weak) dependencies. The insn is a consumer of all 721 the deps mentioned here. */ 722 deps_list_t spec_back_deps; 723 724 /* A list of insns which depend on the instruction. Unlike 'back_deps', 725 it represents forward dependencies. */ 726 deps_list_t forw_deps; 727 728 /* A list of scheduled producers of the instruction. Links are being moved 729 from 'back_deps' to 'resolved_back_deps' while scheduling. */ 730 deps_list_t resolved_back_deps; 731 732 /* A list of scheduled consumers of the instruction. Links are being moved 733 from 'forw_deps' to 'resolved_forw_deps' while scheduling to fasten the 734 search in 'forw_deps'. */ 735 deps_list_t resolved_forw_deps; 736 737 /* If the insn is conditional (either through COND_EXEC, or because 738 it is a conditional branch), this records the condition. NULL 739 for insns that haven't been seen yet or don't have a condition; 740 const_true_rtx to mark an insn without a condition, or with a 741 condition that has been clobbered by a subsequent insn. */ 742 rtx cond; 743 744 /* For a conditional insn, a list of insns that could set the condition 745 register. Used when generating control dependencies. */ 746 rtx_insn_list *cond_deps; 747 748 /* True if the condition in 'cond' should be reversed to get the actual 749 condition. */ 750 unsigned int reverse_cond : 1; 751 752 /* Some insns (e.g. call) are not allowed to move across blocks. */ 753 unsigned int cant_move : 1; 754 }; 755 756 757 /* Bits used for storing values of the fields in the following 758 structure. */ 759 #define INCREASE_BITS 8 760 761 /* The structure describes how the corresponding insn increases the 762 register pressure for each pressure class. */ 763 struct reg_pressure_data 764 { 765 /* Pressure increase for given class because of clobber. */ 766 unsigned int clobber_increase : INCREASE_BITS; 767 /* Increase in register pressure for given class because of register 768 sets. */ 769 unsigned int set_increase : INCREASE_BITS; 770 /* Pressure increase for given class because of unused register 771 set. */ 772 unsigned int unused_set_increase : INCREASE_BITS; 773 /* Pressure change: #sets - #deaths. */ 774 int change : INCREASE_BITS; 775 }; 776 777 /* The following structure describes usage of registers by insns. */ 778 struct reg_use_data 779 { 780 /* Regno used in the insn. */ 781 int regno; 782 /* Insn using the regno. */ 783 rtx_insn *insn; 784 /* Cyclic list of elements with the same regno. */ 785 struct reg_use_data *next_regno_use; 786 /* List of elements with the same insn. */ 787 struct reg_use_data *next_insn_use; 788 }; 789 790 /* The following structure describes used sets of registers by insns. 791 Registers are pseudos whose pressure class is not NO_REGS or hard 792 registers available for allocations. */ 793 struct reg_set_data 794 { 795 /* Regno used in the insn. */ 796 int regno; 797 /* Insn setting the regno. */ 798 rtx insn; 799 /* List of elements with the same insn. */ 800 struct reg_set_data *next_insn_set; 801 }; 802 803 enum autopref_multipass_data_status { 804 /* Entry is irrelevant for auto-prefetcher. */ 805 AUTOPREF_MULTIPASS_DATA_IRRELEVANT = -2, 806 /* Entry is uninitialized. */ 807 AUTOPREF_MULTIPASS_DATA_UNINITIALIZED = -1, 808 /* Entry is relevant for auto-prefetcher and insn can be delayed 809 to allow another insn through. */ 810 AUTOPREF_MULTIPASS_DATA_NORMAL = 0, 811 /* Entry is relevant for auto-prefetcher, but insn should not be 812 delayed as that will break scheduling. */ 813 AUTOPREF_MULTIPASS_DATA_DONT_DELAY = 1 814 }; 815 816 /* Data for modeling cache auto-prefetcher. */ 817 struct autopref_multipass_data_ 818 { 819 /* Base part of memory address. */ 820 rtx base; 821 822 /* Memory offsets from the base. For single simple sets 823 only min_offset is valid. For multi-set insns min_offset 824 and max_offset record the minimum and maximum offsets from the same 825 base among the sets inside the PARALLEL. */ 826 int min_offset; 827 int max_offset; 828 829 /* True if this is a load/store-multiple instruction. */ 830 bool multi_mem_insn_p; 831 832 /* Entry status. */ 833 enum autopref_multipass_data_status status; 834 }; 835 typedef struct autopref_multipass_data_ autopref_multipass_data_def; 836 typedef autopref_multipass_data_def *autopref_multipass_data_t; 837 838 struct _haifa_insn_data 839 { 840 /* We can't place 'struct _deps_list' into h_i_d instead of deps_list_t 841 because when h_i_d extends, addresses of the deps_list->first 842 change without updating deps_list->first->next->prev_nextp. */ 843 844 /* Logical uid gives the original ordering of the insns. */ 845 int luid; 846 847 /* A priority for each insn. */ 848 int priority; 849 850 /* The fusion priority for each insn. */ 851 int fusion_priority; 852 853 /* The minimum clock tick at which the insn becomes ready. This is 854 used to note timing constraints for the insns in the pending list. */ 855 int tick; 856 857 /* For insns that are scheduled at a fixed difference from another, 858 this records the tick in which they must be ready. */ 859 int exact_tick; 860 861 /* INTER_TICK is used to adjust INSN_TICKs of instructions from the 862 subsequent blocks in a region. */ 863 int inter_tick; 864 865 /* Used temporarily to estimate an INSN_TICK value for an insn given 866 current knowledge. */ 867 int tick_estimate; 868 869 /* See comment on QUEUE_INDEX macro in haifa-sched.c. */ 870 int queue_index; 871 872 short cost; 873 874 /* '> 0' if priority is valid, 875 '== 0' if priority was not yet computed, 876 '< 0' if priority in invalid and should be recomputed. */ 877 signed char priority_status; 878 879 /* Set if there's DEF-USE dependence between some speculatively 880 moved load insn and this one. */ 881 unsigned int fed_by_spec_load : 1; 882 unsigned int is_load_insn : 1; 883 /* Nonzero if this insn has negative-cost forward dependencies against 884 an already scheduled insn. */ 885 unsigned int feeds_backtrack_insn : 1; 886 887 /* Nonzero if this insn is a shadow of another, scheduled after a fixed 888 delay. We only emit shadows at the end of a cycle, with no other 889 real insns following them. */ 890 unsigned int shadow_p : 1; 891 892 /* Used internally in unschedule_insns_until to mark insns that must have 893 their TODO_SPEC recomputed. */ 894 unsigned int must_recompute_spec : 1; 895 896 /* What speculations are necessary to apply to schedule the instruction. */ 897 ds_t todo_spec; 898 899 /* What speculations were already applied. */ 900 ds_t done_spec; 901 902 /* What speculations are checked by this instruction. */ 903 ds_t check_spec; 904 905 /* Recovery block for speculation checks. */ 906 basic_block recovery_block; 907 908 /* Original pattern of the instruction. */ 909 rtx orig_pat; 910 911 /* For insns with DEP_CONTROL dependencies, the predicated pattern if it 912 was ever successfully constructed. */ 913 rtx predicated_pat; 914 915 /* The following array contains info how the insn increases register 916 pressure. There is an element for each cover class of pseudos 917 referenced in insns. */ 918 struct reg_pressure_data *reg_pressure; 919 /* The following array contains maximal reg pressure between last 920 scheduled insn and given insn. There is an element for each 921 pressure class of pseudos referenced in insns. This info updated 922 after scheduling each insn for each insn between the two 923 mentioned insns. */ 924 int *max_reg_pressure; 925 /* The following list contains info about used pseudos and hard 926 registers available for allocation. */ 927 struct reg_use_data *reg_use_list; 928 /* The following list contains info about set pseudos and hard 929 registers available for allocation. */ 930 struct reg_set_data *reg_set_list; 931 /* Info about how scheduling the insn changes cost of register 932 pressure excess (between source and target). */ 933 int reg_pressure_excess_cost_change; 934 int model_index; 935 936 /* Original order of insns in the ready list. */ 937 int rfs_debug_orig_order; 938 939 /* The deciding reason for INSN's place in the ready list. */ 940 int last_rfs_win; 941 942 /* Two entries for cache auto-prefetcher model: one for mem reads, 943 and one for mem writes. */ 944 autopref_multipass_data_def autopref_multipass_data[2]; 945 }; 946 947 typedef struct _haifa_insn_data haifa_insn_data_def; 948 typedef haifa_insn_data_def *haifa_insn_data_t; 949 950 951 extern vec<haifa_insn_data_def> h_i_d; 952 953 #define HID(INSN) (&h_i_d[INSN_UID (INSN)]) 954 955 /* Accessor macros for h_i_d. There are more in haifa-sched.c and 956 sched-rgn.c. */ 957 #define INSN_PRIORITY(INSN) (HID (INSN)->priority) 958 #define INSN_FUSION_PRIORITY(INSN) (HID (INSN)->fusion_priority) 959 #define INSN_REG_PRESSURE(INSN) (HID (INSN)->reg_pressure) 960 #define INSN_MAX_REG_PRESSURE(INSN) (HID (INSN)->max_reg_pressure) 961 #define INSN_REG_USE_LIST(INSN) (HID (INSN)->reg_use_list) 962 #define INSN_REG_SET_LIST(INSN) (HID (INSN)->reg_set_list) 963 #define INSN_REG_PRESSURE_EXCESS_COST_CHANGE(INSN) \ 964 (HID (INSN)->reg_pressure_excess_cost_change) 965 #define INSN_PRIORITY_STATUS(INSN) (HID (INSN)->priority_status) 966 #define INSN_MODEL_INDEX(INSN) (HID (INSN)->model_index) 967 #define INSN_AUTOPREF_MULTIPASS_DATA(INSN) \ 968 (HID (INSN)->autopref_multipass_data) 969 970 typedef struct _haifa_deps_insn_data haifa_deps_insn_data_def; 971 typedef haifa_deps_insn_data_def *haifa_deps_insn_data_t; 972 973 974 extern vec<haifa_deps_insn_data_def> h_d_i_d; 975 976 #define HDID(INSN) (&h_d_i_d[INSN_LUID (INSN)]) 977 #define INSN_DEP_COUNT(INSN) (HDID (INSN)->dep_count) 978 #define HAS_INTERNAL_DEP(INSN) (HDID (INSN)->has_internal_dep) 979 #define INSN_FORW_DEPS(INSN) (HDID (INSN)->forw_deps) 980 #define INSN_RESOLVED_BACK_DEPS(INSN) (HDID (INSN)->resolved_back_deps) 981 #define INSN_RESOLVED_FORW_DEPS(INSN) (HDID (INSN)->resolved_forw_deps) 982 #define INSN_HARD_BACK_DEPS(INSN) (HDID (INSN)->hard_back_deps) 983 #define INSN_SPEC_BACK_DEPS(INSN) (HDID (INSN)->spec_back_deps) 984 #define INSN_CACHED_COND(INSN) (HDID (INSN)->cond) 985 #define INSN_REVERSE_COND(INSN) (HDID (INSN)->reverse_cond) 986 #define INSN_COND_DEPS(INSN) (HDID (INSN)->cond_deps) 987 #define CANT_MOVE(INSN) (HDID (INSN)->cant_move) 988 #define CANT_MOVE_BY_LUID(LUID) (h_d_i_d[LUID].cant_move) 989 990 991 #define INSN_PRIORITY(INSN) (HID (INSN)->priority) 992 #define INSN_PRIORITY_STATUS(INSN) (HID (INSN)->priority_status) 993 #define INSN_PRIORITY_KNOWN(INSN) (INSN_PRIORITY_STATUS (INSN) > 0) 994 #define TODO_SPEC(INSN) (HID (INSN)->todo_spec) 995 #define DONE_SPEC(INSN) (HID (INSN)->done_spec) 996 #define CHECK_SPEC(INSN) (HID (INSN)->check_spec) 997 #define RECOVERY_BLOCK(INSN) (HID (INSN)->recovery_block) 998 #define ORIG_PAT(INSN) (HID (INSN)->orig_pat) 999 #define PREDICATED_PAT(INSN) (HID (INSN)->predicated_pat) 1000 1001 /* INSN is either a simple or a branchy speculation check. */ 1002 #define IS_SPECULATION_CHECK_P(INSN) \ 1003 (sel_sched_p () ? sel_insn_is_speculation_check (INSN) : RECOVERY_BLOCK (INSN) != NULL) 1004 1005 /* INSN is a speculation check that will simply reexecute the speculatively 1006 scheduled instruction if the speculation fails. */ 1007 #define IS_SPECULATION_SIMPLE_CHECK_P(INSN) \ 1008 (RECOVERY_BLOCK (INSN) == EXIT_BLOCK_PTR_FOR_FN (cfun)) 1009 1010 /* INSN is a speculation check that will branch to RECOVERY_BLOCK if the 1011 speculation fails. Insns in that block will reexecute the speculatively 1012 scheduled code and then will return immediately after INSN thus preserving 1013 semantics of the program. */ 1014 #define IS_SPECULATION_BRANCHY_CHECK_P(INSN) \ 1015 (RECOVERY_BLOCK (INSN) != NULL \ 1016 && RECOVERY_BLOCK (INSN) != EXIT_BLOCK_PTR_FOR_FN (cfun)) 1017 1018 1019 /* Dep status (aka ds_t) of the link encapsulates all information for a given 1020 dependency, including everything that is needed for speculative scheduling. 1021 1022 The lay-out of a ds_t is as follows: 1023 1024 1. Integers corresponding to the probability of the dependence to *not* 1025 exist. This is the probability that overcoming this dependence will 1026 not be followed by execution of the recovery code. Note that however 1027 high this probability is, the recovery code should still always be 1028 generated to preserve semantics of the program. 1029 1030 The probability values can be set or retrieved using the functions 1031 the set_dep_weak() and get_dep_weak() in sched-deps.c. The values 1032 are always in the range [0, MAX_DEP_WEAK]. 1033 1034 BEGIN_DATA : BITS_PER_DEP_WEAK 1035 BE_IN_DATA : BITS_PER_DEP_WEAK 1036 BEGIN_CONTROL : BITS_PER_DEP_WEAK 1037 BE_IN_CONTROL : BITS_PER_DEP_WEAK 1038 1039 The basic type of DS_T is a host int. For a 32-bits int, the values 1040 will each take 6 bits. 1041 1042 2. The type of dependence. This supercedes the old-style REG_NOTE_KIND 1043 values. TODO: Use this field instead of DEP_TYPE, or make DEP_TYPE 1044 extract the dependence type from here. 1045 1046 dep_type : 4 => DEP_{TRUE|OUTPUT|ANTI|CONTROL} 1047 1048 3. Various flags: 1049 1050 HARD_DEP : 1 => Set if an instruction has a non-speculative 1051 dependence. This is an instruction property 1052 so this bit can only appear in the TODO_SPEC 1053 field of an instruction. 1054 DEP_POSTPONED : 1 => Like HARD_DEP, but the hard dependence may 1055 still be broken by adjusting the instruction. 1056 DEP_CANCELLED : 1 => Set if a dependency has been broken using 1057 some form of speculation. 1058 RESERVED : 1 => Reserved for use in the delay slot scheduler. 1059 1060 See also: check_dep_status () in sched-deps.c . */ 1061 1062 /* The number of bits per weakness probability. There are 4 weakness types 1063 and we need 8 bits for other data in a DS_T. */ 1064 #define BITS_PER_DEP_WEAK ((BITS_PER_DEP_STATUS - 8) / 4) 1065 1066 /* Mask of speculative weakness in dep_status. */ 1067 #define DEP_WEAK_MASK ((1 << BITS_PER_DEP_WEAK) - 1) 1068 1069 /* This constant means that dependence is fake with 99.999...% probability. 1070 This is the maximum value, that can appear in dep_status. 1071 Note, that we don't want MAX_DEP_WEAK to be the same as DEP_WEAK_MASK for 1072 debugging reasons. Though, it can be set to DEP_WEAK_MASK, and, when 1073 done so, we'll get fast (mul for)/(div by) NO_DEP_WEAK. */ 1074 #define MAX_DEP_WEAK (DEP_WEAK_MASK - 1) 1075 1076 /* This constant means that dependence is 99.999...% real and it is a really 1077 bad idea to overcome it (though this can be done, preserving program 1078 semantics). */ 1079 #define MIN_DEP_WEAK 1 1080 1081 /* This constant represents 100% probability. 1082 E.g. it is used to represent weakness of dependence, that doesn't exist. 1083 This value never appears in a ds_t, it is only used for computing the 1084 weakness of a dependence. */ 1085 #define NO_DEP_WEAK (MAX_DEP_WEAK + MIN_DEP_WEAK) 1086 1087 /* Default weakness of speculative dependence. Used when we can't say 1088 neither bad nor good about the dependence. */ 1089 #define UNCERTAIN_DEP_WEAK (MAX_DEP_WEAK - MAX_DEP_WEAK / 4) 1090 1091 /* Offset for speculative weaknesses in dep_status. */ 1092 enum SPEC_TYPES_OFFSETS { 1093 BEGIN_DATA_BITS_OFFSET = 0, 1094 BE_IN_DATA_BITS_OFFSET = BEGIN_DATA_BITS_OFFSET + BITS_PER_DEP_WEAK, 1095 BEGIN_CONTROL_BITS_OFFSET = BE_IN_DATA_BITS_OFFSET + BITS_PER_DEP_WEAK, 1096 BE_IN_CONTROL_BITS_OFFSET = BEGIN_CONTROL_BITS_OFFSET + BITS_PER_DEP_WEAK 1097 }; 1098 1099 /* The following defines provide numerous constants used to distinguish 1100 between different types of speculative dependencies. They are also 1101 used as masks to clear/preserve the bits corresponding to the type 1102 of dependency weakness. */ 1103 1104 /* Dependence can be overcome with generation of new data speculative 1105 instruction. */ 1106 #define BEGIN_DATA (((ds_t) DEP_WEAK_MASK) << BEGIN_DATA_BITS_OFFSET) 1107 1108 /* This dependence is to the instruction in the recovery block, that was 1109 formed to recover after data-speculation failure. 1110 Thus, this dependence can overcome with generating of the copy of 1111 this instruction in the recovery block. */ 1112 #define BE_IN_DATA (((ds_t) DEP_WEAK_MASK) << BE_IN_DATA_BITS_OFFSET) 1113 1114 /* Dependence can be overcome with generation of new control speculative 1115 instruction. */ 1116 #define BEGIN_CONTROL (((ds_t) DEP_WEAK_MASK) << BEGIN_CONTROL_BITS_OFFSET) 1117 1118 /* This dependence is to the instruction in the recovery block, that was 1119 formed to recover after control-speculation failure. 1120 Thus, this dependence can be overcome with generating of the copy of 1121 this instruction in the recovery block. */ 1122 #define BE_IN_CONTROL (((ds_t) DEP_WEAK_MASK) << BE_IN_CONTROL_BITS_OFFSET) 1123 1124 /* A few convenient combinations. */ 1125 #define BEGIN_SPEC (BEGIN_DATA | BEGIN_CONTROL) 1126 #define DATA_SPEC (BEGIN_DATA | BE_IN_DATA) 1127 #define CONTROL_SPEC (BEGIN_CONTROL | BE_IN_CONTROL) 1128 #define SPECULATIVE (DATA_SPEC | CONTROL_SPEC) 1129 #define BE_IN_SPEC (BE_IN_DATA | BE_IN_CONTROL) 1130 1131 /* Constants, that are helpful in iterating through dep_status. */ 1132 #define FIRST_SPEC_TYPE BEGIN_DATA 1133 #define LAST_SPEC_TYPE BE_IN_CONTROL 1134 #define SPEC_TYPE_SHIFT BITS_PER_DEP_WEAK 1135 1136 /* Dependence on instruction can be of multiple types 1137 (e.g. true and output). This fields enhance REG_NOTE_KIND information 1138 of the dependence. */ 1139 #define DEP_TRUE (((ds_t) 1) << (BE_IN_CONTROL_BITS_OFFSET + BITS_PER_DEP_WEAK)) 1140 #define DEP_OUTPUT (DEP_TRUE << 1) 1141 #define DEP_ANTI (DEP_OUTPUT << 1) 1142 #define DEP_CONTROL (DEP_ANTI << 1) 1143 1144 #define DEP_TYPES (DEP_TRUE | DEP_OUTPUT | DEP_ANTI | DEP_CONTROL) 1145 1146 /* Instruction has non-speculative dependence. This bit represents the 1147 property of an instruction - not the one of a dependence. 1148 Therefore, it can appear only in the TODO_SPEC field of an instruction. */ 1149 #define HARD_DEP (DEP_CONTROL << 1) 1150 1151 /* Like HARD_DEP, but dependencies can perhaps be broken by modifying 1152 the instructions. This is used for example to change: 1153 1154 rn++ => rm=[rn + 4] 1155 rm=[rn] rn++ 1156 1157 For instructions that have this bit set, one of the dependencies of 1158 the instructions will have a non-NULL REPLACE field in its DEP_T. 1159 Just like HARD_DEP, this bit is only ever set in TODO_SPEC. */ 1160 #define DEP_POSTPONED (HARD_DEP << 1) 1161 1162 /* Set if a dependency is cancelled via speculation. */ 1163 #define DEP_CANCELLED (DEP_POSTPONED << 1) 1164 1165 1166 /* This represents the results of calling sched-deps.c functions, 1167 which modify dependencies. */ 1168 enum DEPS_ADJUST_RESULT { 1169 /* No dependence needed (e.g. producer == consumer). */ 1170 DEP_NODEP, 1171 /* Dependence is already present and wasn't modified. */ 1172 DEP_PRESENT, 1173 /* Existing dependence was modified to include additional information. */ 1174 DEP_CHANGED, 1175 /* New dependence has been created. */ 1176 DEP_CREATED 1177 }; 1178 1179 /* Represents the bits that can be set in the flags field of the 1180 sched_info structure. */ 1181 enum SCHED_FLAGS { 1182 /* If set, generate links between instruction as DEPS_LIST. 1183 Otherwise, generate usual INSN_LIST links. */ 1184 USE_DEPS_LIST = 1, 1185 /* Perform data or control (or both) speculation. 1186 Results in generation of data and control speculative dependencies. 1187 Requires USE_DEPS_LIST set. */ 1188 DO_SPECULATION = USE_DEPS_LIST << 1, 1189 DO_BACKTRACKING = DO_SPECULATION << 1, 1190 DO_PREDICATION = DO_BACKTRACKING << 1, 1191 DONT_BREAK_DEPENDENCIES = DO_PREDICATION << 1, 1192 SCHED_RGN = DONT_BREAK_DEPENDENCIES << 1, 1193 SCHED_EBB = SCHED_RGN << 1, 1194 /* Scheduler can possibly create new basic blocks. Used for assertions. */ 1195 NEW_BBS = SCHED_EBB << 1, 1196 SEL_SCHED = NEW_BBS << 1 1197 }; 1198 1199 enum SPEC_SCHED_FLAGS { 1200 COUNT_SPEC_IN_CRITICAL_PATH = 1, 1201 SEL_SCHED_SPEC_DONT_CHECK_CONTROL = COUNT_SPEC_IN_CRITICAL_PATH << 1 1202 }; 1203 1204 #define NOTE_NOT_BB_P(NOTE) (NOTE_P (NOTE) && (NOTE_KIND (NOTE) \ 1205 != NOTE_INSN_BASIC_BLOCK)) 1206 1207 extern FILE *sched_dump; 1208 extern int sched_verbose; 1209 1210 extern spec_info_t spec_info; 1211 extern bool haifa_recovery_bb_ever_added_p; 1212 1213 /* Exception Free Loads: 1214 1215 We define five classes of speculative loads: IFREE, IRISKY, 1216 PFREE, PRISKY, and MFREE. 1217 1218 IFREE loads are loads that are proved to be exception-free, just 1219 by examining the load insn. Examples for such loads are loads 1220 from TOC and loads of global data. 1221 1222 IRISKY loads are loads that are proved to be exception-risky, 1223 just by examining the load insn. Examples for such loads are 1224 volatile loads and loads from shared memory. 1225 1226 PFREE loads are loads for which we can prove, by examining other 1227 insns, that they are exception-free. Currently, this class consists 1228 of loads for which we are able to find a "similar load", either in 1229 the target block, or, if only one split-block exists, in that split 1230 block. Load2 is similar to load1 if both have same single base 1231 register. We identify only part of the similar loads, by finding 1232 an insn upon which both load1 and load2 have a DEF-USE dependence. 1233 1234 PRISKY loads are loads for which we can prove, by examining other 1235 insns, that they are exception-risky. Currently we have two proofs for 1236 such loads. The first proof detects loads that are probably guarded by a 1237 test on the memory address. This proof is based on the 1238 backward and forward data dependence information for the region. 1239 Let load-insn be the examined load. 1240 Load-insn is PRISKY iff ALL the following hold: 1241 1242 - insn1 is not in the same block as load-insn 1243 - there is a DEF-USE dependence chain (insn1, ..., load-insn) 1244 - test-insn is either a compare or a branch, not in the same block 1245 as load-insn 1246 - load-insn is reachable from test-insn 1247 - there is a DEF-USE dependence chain (insn1, ..., test-insn) 1248 1249 This proof might fail when the compare and the load are fed 1250 by an insn not in the region. To solve this, we will add to this 1251 group all loads that have no input DEF-USE dependence. 1252 1253 The second proof detects loads that are directly or indirectly 1254 fed by a speculative load. This proof is affected by the 1255 scheduling process. We will use the flag fed_by_spec_load. 1256 Initially, all insns have this flag reset. After a speculative 1257 motion of an insn, if insn is either a load, or marked as 1258 fed_by_spec_load, we will also mark as fed_by_spec_load every 1259 insn1 for which a DEF-USE dependence (insn, insn1) exists. A 1260 load which is fed_by_spec_load is also PRISKY. 1261 1262 MFREE (maybe-free) loads are all the remaining loads. They may be 1263 exception-free, but we cannot prove it. 1264 1265 Now, all loads in IFREE and PFREE classes are considered 1266 exception-free, while all loads in IRISKY and PRISKY classes are 1267 considered exception-risky. As for loads in the MFREE class, 1268 these are considered either exception-free or exception-risky, 1269 depending on whether we are pessimistic or optimistic. We have 1270 to take the pessimistic approach to assure the safety of 1271 speculative scheduling, but we can take the optimistic approach 1272 by invoking the -fsched_spec_load_dangerous option. */ 1273 1274 enum INSN_TRAP_CLASS 1275 { 1276 TRAP_FREE = 0, IFREE = 1, PFREE_CANDIDATE = 2, 1277 PRISKY_CANDIDATE = 3, IRISKY = 4, TRAP_RISKY = 5 1278 }; 1279 1280 #define WORST_CLASS(class1, class2) \ 1281 ((class1 > class2) ? class1 : class2) 1282 1283 #ifndef __GNUC__ 1284 #define __inline 1285 #endif 1286 1287 #ifndef HAIFA_INLINE 1288 #define HAIFA_INLINE __inline 1289 #endif 1290 1291 struct sched_deps_info_def 1292 { 1293 /* Called when computing dependencies for a JUMP_INSN. This function 1294 should store the set of registers that must be considered as set by 1295 the jump in the regset. */ 1296 void (*compute_jump_reg_dependencies) (rtx, regset); 1297 1298 /* Start analyzing insn. */ 1299 void (*start_insn) (rtx_insn *); 1300 1301 /* Finish analyzing insn. */ 1302 void (*finish_insn) (void); 1303 1304 /* Start analyzing insn LHS (Left Hand Side). */ 1305 void (*start_lhs) (rtx); 1306 1307 /* Finish analyzing insn LHS. */ 1308 void (*finish_lhs) (void); 1309 1310 /* Start analyzing insn RHS (Right Hand Side). */ 1311 void (*start_rhs) (rtx); 1312 1313 /* Finish analyzing insn RHS. */ 1314 void (*finish_rhs) (void); 1315 1316 /* Note set of the register. */ 1317 void (*note_reg_set) (int); 1318 1319 /* Note clobber of the register. */ 1320 void (*note_reg_clobber) (int); 1321 1322 /* Note use of the register. */ 1323 void (*note_reg_use) (int); 1324 1325 /* Note memory dependence of type DS between MEM1 and MEM2 (which is 1326 in the INSN2). */ 1327 void (*note_mem_dep) (rtx mem1, rtx mem2, rtx_insn *insn2, ds_t ds); 1328 1329 /* Note a dependence of type DS from the INSN. */ 1330 void (*note_dep) (rtx_insn *, ds_t ds); 1331 1332 /* Nonzero if we should use cselib for better alias analysis. This 1333 must be 0 if the dependency information is used after sched_analyze 1334 has completed, e.g. if we're using it to initialize state for successor 1335 blocks in region scheduling. */ 1336 unsigned int use_cselib : 1; 1337 1338 /* If set, generate links between instruction as DEPS_LIST. 1339 Otherwise, generate usual INSN_LIST links. */ 1340 unsigned int use_deps_list : 1; 1341 1342 /* Generate data and control speculative dependencies. 1343 Requires USE_DEPS_LIST set. */ 1344 unsigned int generate_spec_deps : 1; 1345 }; 1346 1347 extern struct sched_deps_info_def *sched_deps_info; 1348 1349 1350 /* Functions in sched-deps.c. */ 1351 extern rtx sched_get_reverse_condition_uncached (const rtx_insn *); 1352 extern bool sched_insns_conditions_mutex_p (const rtx_insn *, 1353 const rtx_insn *); 1354 extern bool sched_insn_is_legitimate_for_speculation_p (const rtx_insn *, ds_t); 1355 extern void add_dependence (rtx_insn *, rtx_insn *, enum reg_note); 1356 extern void sched_analyze (struct deps_desc *, rtx_insn *, rtx_insn *); 1357 extern void init_deps (struct deps_desc *, bool); 1358 extern void init_deps_reg_last (struct deps_desc *); 1359 extern void free_deps (struct deps_desc *); 1360 extern void init_deps_global (void); 1361 extern void finish_deps_global (void); 1362 extern void deps_analyze_insn (struct deps_desc *, rtx_insn *); 1363 extern void remove_from_deps (struct deps_desc *, rtx_insn *); 1364 extern void init_insn_reg_pressure_info (rtx_insn *); 1365 extern void get_implicit_reg_pending_clobbers (HARD_REG_SET *, rtx_insn *); 1366 1367 extern dw_t get_dep_weak (ds_t, ds_t); 1368 extern ds_t set_dep_weak (ds_t, ds_t, dw_t); 1369 extern dw_t estimate_dep_weak (rtx, rtx); 1370 extern ds_t ds_merge (ds_t, ds_t); 1371 extern ds_t ds_full_merge (ds_t, ds_t, rtx, rtx); 1372 extern ds_t ds_max_merge (ds_t, ds_t); 1373 extern dw_t ds_weak (ds_t); 1374 extern ds_t ds_get_speculation_types (ds_t); 1375 extern ds_t ds_get_max_dep_weak (ds_t); 1376 1377 extern void sched_deps_init (bool); 1378 extern void sched_deps_finish (void); 1379 1380 extern void haifa_note_reg_set (int); 1381 extern void haifa_note_reg_clobber (int); 1382 extern void haifa_note_reg_use (int); 1383 1384 extern void maybe_extend_reg_info_p (void); 1385 1386 extern void deps_start_bb (struct deps_desc *, rtx_insn *); 1387 extern enum reg_note ds_to_dt (ds_t); 1388 1389 extern bool deps_pools_are_empty_p (void); 1390 extern void sched_free_deps (rtx_insn *, rtx_insn *, bool); 1391 extern void extend_dependency_caches (int, bool); 1392 1393 extern void debug_ds (ds_t); 1394 1395 1396 /* Functions in haifa-sched.c. */ 1397 extern void initialize_live_range_shrinkage (void); 1398 extern void finish_live_range_shrinkage (void); 1399 extern void sched_init_region_reg_pressure_info (void); 1400 extern void free_global_sched_pressure_data (void); 1401 extern int haifa_classify_insn (const_rtx); 1402 extern void get_ebb_head_tail (basic_block, basic_block, 1403 rtx_insn **, rtx_insn **); 1404 extern int no_real_insns_p (const rtx_insn *, const rtx_insn *); 1405 1406 extern int insn_cost (rtx_insn *); 1407 extern int dep_cost_1 (dep_t, dw_t); 1408 extern int dep_cost (dep_t); 1409 extern int set_priorities (rtx_insn *, rtx_insn *); 1410 1411 extern void sched_setup_bb_reg_pressure_info (basic_block, rtx_insn *); 1412 extern bool schedule_block (basic_block *, state_t); 1413 1414 extern int cycle_issued_insns; 1415 extern int issue_rate; 1416 extern int dfa_lookahead; 1417 1418 extern int autopref_multipass_dfa_lookahead_guard (rtx_insn *, int); 1419 1420 extern rtx_insn *ready_element (struct ready_list *, int); 1421 extern rtx_insn **ready_lastpos (struct ready_list *); 1422 1423 extern int try_ready (rtx_insn *); 1424 extern void sched_extend_ready_list (int); 1425 extern void sched_finish_ready_list (void); 1426 extern void sched_change_pattern (rtx, rtx); 1427 extern int sched_speculate_insn (rtx_insn *, ds_t, rtx *); 1428 extern void unlink_bb_notes (basic_block, basic_block); 1429 extern void add_block (basic_block, basic_block); 1430 extern rtx_note *bb_note (basic_block); 1431 extern void concat_note_lists (rtx_insn *, rtx_insn **); 1432 extern rtx_insn *sched_emit_insn (rtx); 1433 extern rtx_insn *get_ready_element (int); 1434 extern int number_in_ready (void); 1435 1436 /* Types and functions in sched-ebb.c. */ 1437 1438 extern basic_block schedule_ebb (rtx_insn *, rtx_insn *, bool); 1439 extern void schedule_ebbs_init (void); 1440 extern void schedule_ebbs_finish (void); 1441 1442 /* Types and functions in sched-rgn.c. */ 1443 1444 /* A region is the main entity for interblock scheduling: insns 1445 are allowed to move between blocks in the same region, along 1446 control flow graph edges, in the 'up' direction. */ 1447 struct region 1448 { 1449 /* Number of extended basic blocks in region. */ 1450 int rgn_nr_blocks; 1451 /* cblocks in the region (actually index in rgn_bb_table). */ 1452 int rgn_blocks; 1453 /* Dependencies for this region are already computed. Basically, indicates, 1454 that this is a recovery block. */ 1455 unsigned int dont_calc_deps : 1; 1456 /* This region has at least one non-trivial ebb. */ 1457 unsigned int has_real_ebb : 1; 1458 }; 1459 1460 extern int nr_regions; 1461 extern region *rgn_table; 1462 extern int *rgn_bb_table; 1463 extern int *block_to_bb; 1464 extern int *containing_rgn; 1465 1466 /* Often used short-hand in the scheduler. The rest of the compiler uses 1467 BLOCK_FOR_INSN(INSN) and an indirect reference to get the basic block 1468 number ("index"). For historical reasons, the scheduler does not. */ 1469 #define BLOCK_NUM(INSN) (BLOCK_FOR_INSN (INSN)->index + 0) 1470 1471 #define RGN_NR_BLOCKS(rgn) (rgn_table[rgn].rgn_nr_blocks) 1472 #define RGN_BLOCKS(rgn) (rgn_table[rgn].rgn_blocks) 1473 #define RGN_DONT_CALC_DEPS(rgn) (rgn_table[rgn].dont_calc_deps) 1474 #define RGN_HAS_REAL_EBB(rgn) (rgn_table[rgn].has_real_ebb) 1475 #define BLOCK_TO_BB(block) (block_to_bb[block]) 1476 #define CONTAINING_RGN(block) (containing_rgn[block]) 1477 1478 /* The mapping from ebb to block. */ 1479 extern int *ebb_head; 1480 #define BB_TO_BLOCK(ebb) (rgn_bb_table[ebb_head[ebb]]) 1481 #define EBB_FIRST_BB(ebb) BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (ebb)) 1482 #define EBB_LAST_BB(ebb) \ 1483 BASIC_BLOCK_FOR_FN (cfun, rgn_bb_table[ebb_head[ebb + 1] - 1]) 1484 #define INSN_BB(INSN) (BLOCK_TO_BB (BLOCK_NUM (INSN))) 1485 1486 extern int current_nr_blocks; 1487 extern int current_blocks; 1488 extern int target_bb; 1489 extern bool sched_no_dce; 1490 1491 extern void set_modulo_params (int, int, int, int); 1492 extern void record_delay_slot_pair (rtx_insn *, rtx_insn *, int, int); 1493 extern rtx_insn *real_insn_for_shadow (rtx_insn *); 1494 extern void discard_delay_pairs_above (int); 1495 extern void free_delay_pairs (void); 1496 extern void add_delay_dependencies (rtx_insn *); 1497 extern bool sched_is_disabled_for_current_region_p (void); 1498 extern void sched_rgn_init (bool); 1499 extern void sched_rgn_finish (void); 1500 extern void rgn_setup_region (int); 1501 extern void sched_rgn_compute_dependencies (int); 1502 extern void sched_rgn_local_init (int); 1503 extern void sched_rgn_local_finish (void); 1504 extern void sched_rgn_local_free (void); 1505 extern void extend_regions (void); 1506 extern void rgn_make_new_region_out_of_new_block (basic_block); 1507 1508 extern void compute_priorities (void); 1509 extern void increase_insn_priority (rtx_insn *, int); 1510 extern void debug_rgn_dependencies (int); 1511 extern void debug_dependencies (rtx_insn *, rtx_insn *); 1512 extern void dump_rgn_dependencies_dot (FILE *); 1513 extern void dump_rgn_dependencies_dot (const char *); 1514 1515 extern void free_rgn_deps (void); 1516 extern int contributes_to_priority (rtx_insn *, rtx_insn *); 1517 extern void extend_rgns (int *, int *, sbitmap, int *); 1518 extern void deps_join (struct deps_desc *, struct deps_desc *); 1519 1520 extern void rgn_setup_common_sched_info (void); 1521 extern void rgn_setup_sched_infos (void); 1522 1523 extern void debug_regions (void); 1524 extern void debug_region (int); 1525 extern void dump_region_dot (FILE *, int); 1526 extern void dump_region_dot_file (const char *, int); 1527 1528 extern void haifa_sched_init (void); 1529 extern void haifa_sched_finish (void); 1530 1531 extern void find_modifiable_mems (rtx_insn *, rtx_insn *); 1532 1533 /* sched-deps.c interface to walk, add, search, update, resolve, delete 1534 and debug instruction dependencies. */ 1535 1536 /* Constants defining dependences lists. */ 1537 1538 /* No list. */ 1539 #define SD_LIST_NONE (0) 1540 1541 /* hard_back_deps. */ 1542 #define SD_LIST_HARD_BACK (1) 1543 1544 /* spec_back_deps. */ 1545 #define SD_LIST_SPEC_BACK (2) 1546 1547 /* forw_deps. */ 1548 #define SD_LIST_FORW (4) 1549 1550 /* resolved_back_deps. */ 1551 #define SD_LIST_RES_BACK (8) 1552 1553 /* resolved_forw_deps. */ 1554 #define SD_LIST_RES_FORW (16) 1555 1556 #define SD_LIST_BACK (SD_LIST_HARD_BACK | SD_LIST_SPEC_BACK) 1557 1558 /* A type to hold above flags. */ 1559 typedef int sd_list_types_def; 1560 1561 extern void sd_next_list (const_rtx, sd_list_types_def *, deps_list_t *, bool *); 1562 1563 /* Iterator to walk through, resolve and delete dependencies. */ 1564 struct _sd_iterator 1565 { 1566 /* What lists to walk. Can be any combination of SD_LIST_* flags. */ 1567 sd_list_types_def types; 1568 1569 /* Instruction dependencies lists of which will be walked. */ 1570 rtx insn; 1571 1572 /* Pointer to the next field of the previous element. This is not 1573 simply a pointer to the next element to allow easy deletion from the 1574 list. When a dep is being removed from the list the iterator 1575 will automatically advance because the value in *linkp will start 1576 referring to the next element. */ 1577 dep_link_t *linkp; 1578 1579 /* True if the current list is a resolved one. */ 1580 bool resolved_p; 1581 }; 1582 1583 typedef struct _sd_iterator sd_iterator_def; 1584 1585 /* ??? We can move some definitions that are used in below inline functions 1586 out of sched-int.h to sched-deps.c provided that the below functions will 1587 become global externals. 1588 These definitions include: 1589 * struct _deps_list: opaque pointer is needed at global scope. 1590 * struct _dep_link: opaque pointer is needed at scope of sd_iterator_def. 1591 * struct _dep_node: opaque pointer is needed at scope of 1592 struct _deps_link. */ 1593 1594 /* Return initialized iterator. */ 1595 static inline sd_iterator_def 1596 sd_iterator_start (rtx insn, sd_list_types_def types) 1597 { 1598 /* Some dep_link a pointer to which will return NULL. */ 1599 static dep_link_t null_link = NULL; 1600 1601 sd_iterator_def i; 1602 1603 i.types = types; 1604 i.insn = insn; 1605 i.linkp = &null_link; 1606 1607 /* Avoid 'uninitialized warning'. */ 1608 i.resolved_p = false; 1609 1610 return i; 1611 } 1612 1613 /* Return the current element. */ 1614 static inline bool 1615 sd_iterator_cond (sd_iterator_def *it_ptr, dep_t *dep_ptr) 1616 { 1617 while (true) 1618 { 1619 dep_link_t link = *it_ptr->linkp; 1620 1621 if (link != NULL) 1622 { 1623 *dep_ptr = DEP_LINK_DEP (link); 1624 return true; 1625 } 1626 else 1627 { 1628 sd_list_types_def types = it_ptr->types; 1629 1630 if (types != SD_LIST_NONE) 1631 /* Switch to next list. */ 1632 { 1633 deps_list_t list; 1634 1635 sd_next_list (it_ptr->insn, 1636 &it_ptr->types, &list, &it_ptr->resolved_p); 1637 1638 if (list) 1639 { 1640 it_ptr->linkp = &DEPS_LIST_FIRST (list); 1641 continue; 1642 } 1643 } 1644 1645 *dep_ptr = NULL; 1646 return false; 1647 } 1648 } 1649 } 1650 1651 /* Advance iterator. */ 1652 static inline void 1653 sd_iterator_next (sd_iterator_def *it_ptr) 1654 { 1655 it_ptr->linkp = &DEP_LINK_NEXT (*it_ptr->linkp); 1656 } 1657 1658 /* A cycle wrapper. */ 1659 #define FOR_EACH_DEP(INSN, LIST_TYPES, ITER, DEP) \ 1660 for ((ITER) = sd_iterator_start ((INSN), (LIST_TYPES)); \ 1661 sd_iterator_cond (&(ITER), &(DEP)); \ 1662 sd_iterator_next (&(ITER))) 1663 1664 #define IS_DISPATCH_ON 1 1665 #define IS_CMP 2 1666 #define DISPATCH_VIOLATION 3 1667 #define FITS_DISPATCH_WINDOW 4 1668 #define DISPATCH_INIT 5 1669 #define ADD_TO_DISPATCH_WINDOW 6 1670 1671 extern int sd_lists_size (const_rtx, sd_list_types_def); 1672 extern bool sd_lists_empty_p (const_rtx, sd_list_types_def); 1673 extern void sd_init_insn (rtx_insn *); 1674 extern void sd_finish_insn (rtx_insn *); 1675 extern dep_t sd_find_dep_between (rtx, rtx, bool); 1676 extern void sd_add_dep (dep_t, bool); 1677 extern enum DEPS_ADJUST_RESULT sd_add_or_update_dep (dep_t, bool); 1678 extern void sd_resolve_dep (sd_iterator_def); 1679 extern void sd_unresolve_dep (sd_iterator_def); 1680 extern void sd_copy_back_deps (rtx_insn *, rtx_insn *, bool); 1681 extern void sd_delete_dep (sd_iterator_def); 1682 extern void sd_debug_lists (rtx, sd_list_types_def); 1683 1684 /* Macros and declarations for scheduling fusion. */ 1685 #define FUSION_MAX_PRIORITY (INT_MAX) 1686 extern bool sched_fusion; 1687 1688 #endif /* INSN_SCHEDULING */ 1689 1690 #endif /* GCC_SCHED_INT_H */ 1691 1692