2010-05-18 Vladimir Makarov <vmakarov@redhat.com>
[official-gcc.git] / gcc / haifa-sched.c
blob8bea5e3811e60df698e6461860e4fc5f23c78b6e
1 /* Instruction scheduling pass.
2 Copyright (C) 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000,
3 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
4 Free Software Foundation, Inc.
5 Contributed by Michael Tiemann (tiemann@cygnus.com) Enhanced by,
6 and currently maintained by, Jim Wilson (wilson@cygnus.com)
8 This file is part of GCC.
10 GCC is free software; you can redistribute it and/or modify it under
11 the terms of the GNU General Public License as published by the Free
12 Software Foundation; either version 3, or (at your option) any later
13 version.
15 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
16 WARRANTY; without even the implied warranty of MERCHANTABILITY or
17 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 for more details.
20 You should have received a copy of the GNU General Public License
21 along with GCC; see the file COPYING3. If not see
22 <http://www.gnu.org/licenses/>. */
24 /* Instruction scheduling pass. This file, along with sched-deps.c,
25 contains the generic parts. The actual entry point is found for
26 the normal instruction scheduling pass is found in sched-rgn.c.
28 We compute insn priorities based on data dependencies. Flow
29 analysis only creates a fraction of the data-dependencies we must
30 observe: namely, only those dependencies which the combiner can be
31 expected to use. For this pass, we must therefore create the
32 remaining dependencies we need to observe: register dependencies,
33 memory dependencies, dependencies to keep function calls in order,
34 and the dependence between a conditional branch and the setting of
35 condition codes are all dealt with here.
37 The scheduler first traverses the data flow graph, starting with
38 the last instruction, and proceeding to the first, assigning values
39 to insn_priority as it goes. This sorts the instructions
40 topologically by data dependence.
42 Once priorities have been established, we order the insns using
43 list scheduling. This works as follows: starting with a list of
44 all the ready insns, and sorted according to priority number, we
45 schedule the insn from the end of the list by placing its
46 predecessors in the list according to their priority order. We
47 consider this insn scheduled by setting the pointer to the "end" of
48 the list to point to the previous insn. When an insn has no
49 predecessors, we either queue it until sufficient time has elapsed
50 or add it to the ready list. As the instructions are scheduled or
51 when stalls are introduced, the queue advances and dumps insns into
52 the ready list. When all insns down to the lowest priority have
53 been scheduled, the critical path of the basic block has been made
54 as short as possible. The remaining insns are then scheduled in
55 remaining slots.
57 The following list shows the order in which we want to break ties
58 among insns in the ready list:
60 1. choose insn with the longest path to end of bb, ties
61 broken by
62 2. choose insn with least contribution to register pressure,
63 ties broken by
64 3. prefer in-block upon interblock motion, ties broken by
65 4. prefer useful upon speculative motion, ties broken by
66 5. choose insn with largest control flow probability, ties
67 broken by
68 6. choose insn with the least dependences upon the previously
69 scheduled insn, or finally
70 7 choose the insn which has the most insns dependent on it.
71 8. choose insn with lowest UID.
73 Memory references complicate matters. Only if we can be certain
74 that memory references are not part of the data dependency graph
75 (via true, anti, or output dependence), can we move operations past
76 memory references. To first approximation, reads can be done
77 independently, while writes introduce dependencies. Better
78 approximations will yield fewer dependencies.
80 Before reload, an extended analysis of interblock data dependences
81 is required for interblock scheduling. This is performed in
82 compute_block_backward_dependences ().
84 Dependencies set up by memory references are treated in exactly the
85 same way as other dependencies, by using insn backward dependences
86 INSN_BACK_DEPS. INSN_BACK_DEPS are translated into forward dependences
87 INSN_FORW_DEPS the purpose of forward list scheduling.
89 Having optimized the critical path, we may have also unduly
90 extended the lifetimes of some registers. If an operation requires
91 that constants be loaded into registers, it is certainly desirable
92 to load those constants as early as necessary, but no earlier.
93 I.e., it will not do to load up a bunch of registers at the
94 beginning of a basic block only to use them at the end, if they
95 could be loaded later, since this may result in excessive register
96 utilization.
98 Note that since branches are never in basic blocks, but only end
99 basic blocks, this pass will not move branches. But that is ok,
100 since we can use GNU's delayed branch scheduling pass to take care
101 of this case.
103 Also note that no further optimizations based on algebraic
104 identities are performed, so this pass would be a good one to
105 perform instruction splitting, such as breaking up a multiply
106 instruction into shifts and adds where that is profitable.
108 Given the memory aliasing analysis that this pass should perform,
109 it should be possible to remove redundant stores to memory, and to
110 load values from registers instead of hitting memory.
112 Before reload, speculative insns are moved only if a 'proof' exists
113 that no exception will be caused by this, and if no live registers
114 exist that inhibit the motion (live registers constraints are not
115 represented by data dependence edges).
117 This pass must update information that subsequent passes expect to
118 be correct. Namely: reg_n_refs, reg_n_sets, reg_n_deaths,
119 reg_n_calls_crossed, and reg_live_length. Also, BB_HEAD, BB_END.
121 The information in the line number notes is carefully retained by
122 this pass. Notes that refer to the starting and ending of
123 exception regions are also carefully retained by this pass. All
124 other NOTE insns are grouped in their same relative order at the
125 beginning of basic blocks and regions that have been scheduled. */
127 #include "config.h"
128 #include "system.h"
129 #include "coretypes.h"
130 #include "tm.h"
131 #include "toplev.h"
132 #include "rtl.h"
133 #include "tm_p.h"
134 #include "hard-reg-set.h"
135 #include "regs.h"
136 #include "function.h"
137 #include "flags.h"
138 #include "insn-config.h"
139 #include "insn-attr.h"
140 #include "except.h"
141 #include "toplev.h"
142 #include "recog.h"
143 #include "sched-int.h"
144 #include "target.h"
145 #include "output.h"
146 #include "params.h"
147 #include "vecprim.h"
148 #include "dbgcnt.h"
149 #include "cfgloop.h"
150 #include "ira.h"
152 #ifdef INSN_SCHEDULING
154 /* issue_rate is the number of insns that can be scheduled in the same
155 machine cycle. It can be defined in the config/mach/mach.h file,
156 otherwise we set it to 1. */
158 int issue_rate;
160 /* sched-verbose controls the amount of debugging output the
161 scheduler prints. It is controlled by -fsched-verbose=N:
162 N>0 and no -DSR : the output is directed to stderr.
163 N>=10 will direct the printouts to stderr (regardless of -dSR).
164 N=1: same as -dSR.
165 N=2: bb's probabilities, detailed ready list info, unit/insn info.
166 N=3: rtl at abort point, control-flow, regions info.
167 N=5: dependences info. */
169 static int sched_verbose_param = 0;
170 int sched_verbose = 0;
172 /* Debugging file. All printouts are sent to dump, which is always set,
173 either to stderr, or to the dump listing file (-dRS). */
174 FILE *sched_dump = 0;
176 /* fix_sched_param() is called from toplev.c upon detection
177 of the -fsched-verbose=N option. */
179 void
180 fix_sched_param (const char *param, const char *val)
182 if (!strcmp (param, "verbose"))
183 sched_verbose_param = atoi (val);
184 else
185 warning (0, "fix_sched_param: unknown param: %s", param);
188 /* This is a placeholder for the scheduler parameters common
189 to all schedulers. */
190 struct common_sched_info_def *common_sched_info;
192 #define INSN_TICK(INSN) (HID (INSN)->tick)
193 #define INTER_TICK(INSN) (HID (INSN)->inter_tick)
195 /* If INSN_TICK of an instruction is equal to INVALID_TICK,
196 then it should be recalculated from scratch. */
197 #define INVALID_TICK (-(max_insn_queue_index + 1))
198 /* The minimal value of the INSN_TICK of an instruction. */
199 #define MIN_TICK (-max_insn_queue_index)
201 /* Issue points are used to distinguish between instructions in max_issue ().
202 For now, all instructions are equally good. */
203 #define ISSUE_POINTS(INSN) 1
205 /* List of important notes we must keep around. This is a pointer to the
206 last element in the list. */
207 rtx note_list;
209 static struct spec_info_def spec_info_var;
210 /* Description of the speculative part of the scheduling.
211 If NULL - no speculation. */
212 spec_info_t spec_info = NULL;
214 /* True, if recovery block was added during scheduling of current block.
215 Used to determine, if we need to fix INSN_TICKs. */
216 static bool haifa_recovery_bb_recently_added_p;
218 /* True, if recovery block was added during this scheduling pass.
219 Used to determine if we should have empty memory pools of dependencies
220 after finishing current region. */
221 bool haifa_recovery_bb_ever_added_p;
223 /* Counters of different types of speculative instructions. */
224 static int nr_begin_data, nr_be_in_data, nr_begin_control, nr_be_in_control;
226 /* Array used in {unlink, restore}_bb_notes. */
227 static rtx *bb_header = 0;
229 /* Basic block after which recovery blocks will be created. */
230 static basic_block before_recovery;
232 /* Basic block just before the EXIT_BLOCK and after recovery, if we have
233 created it. */
234 basic_block after_recovery;
236 /* FALSE if we add bb to another region, so we don't need to initialize it. */
237 bool adding_bb_to_current_region_p = true;
239 /* Queues, etc. */
241 /* An instruction is ready to be scheduled when all insns preceding it
242 have already been scheduled. It is important to ensure that all
243 insns which use its result will not be executed until its result
244 has been computed. An insn is maintained in one of four structures:
246 (P) the "Pending" set of insns which cannot be scheduled until
247 their dependencies have been satisfied.
248 (Q) the "Queued" set of insns that can be scheduled when sufficient
249 time has passed.
250 (R) the "Ready" list of unscheduled, uncommitted insns.
251 (S) the "Scheduled" list of insns.
253 Initially, all insns are either "Pending" or "Ready" depending on
254 whether their dependencies are satisfied.
256 Insns move from the "Ready" list to the "Scheduled" list as they
257 are committed to the schedule. As this occurs, the insns in the
258 "Pending" list have their dependencies satisfied and move to either
259 the "Ready" list or the "Queued" set depending on whether
260 sufficient time has passed to make them ready. As time passes,
261 insns move from the "Queued" set to the "Ready" list.
263 The "Pending" list (P) are the insns in the INSN_FORW_DEPS of the
264 unscheduled insns, i.e., those that are ready, queued, and pending.
265 The "Queued" set (Q) is implemented by the variable `insn_queue'.
266 The "Ready" list (R) is implemented by the variables `ready' and
267 `n_ready'.
268 The "Scheduled" list (S) is the new insn chain built by this pass.
270 The transition (R->S) is implemented in the scheduling loop in
271 `schedule_block' when the best insn to schedule is chosen.
272 The transitions (P->R and P->Q) are implemented in `schedule_insn' as
273 insns move from the ready list to the scheduled list.
274 The transition (Q->R) is implemented in 'queue_to_insn' as time
275 passes or stalls are introduced. */
277 /* Implement a circular buffer to delay instructions until sufficient
278 time has passed. For the new pipeline description interface,
279 MAX_INSN_QUEUE_INDEX is a power of two minus one which is not less
280 than maximal time of instruction execution computed by genattr.c on
281 the base maximal time of functional unit reservations and getting a
282 result. This is the longest time an insn may be queued. */
284 static rtx *insn_queue;
285 static int q_ptr = 0;
286 static int q_size = 0;
287 #define NEXT_Q(X) (((X)+1) & max_insn_queue_index)
288 #define NEXT_Q_AFTER(X, C) (((X)+C) & max_insn_queue_index)
290 #define QUEUE_SCHEDULED (-3)
291 #define QUEUE_NOWHERE (-2)
292 #define QUEUE_READY (-1)
293 /* QUEUE_SCHEDULED - INSN is scheduled.
294 QUEUE_NOWHERE - INSN isn't scheduled yet and is neither in
295 queue or ready list.
296 QUEUE_READY - INSN is in ready list.
297 N >= 0 - INSN queued for X [where NEXT_Q_AFTER (q_ptr, X) == N] cycles. */
299 #define QUEUE_INDEX(INSN) (HID (INSN)->queue_index)
301 /* The following variable value refers for all current and future
302 reservations of the processor units. */
303 state_t curr_state;
305 /* The following variable value is size of memory representing all
306 current and future reservations of the processor units. */
307 size_t dfa_state_size;
309 /* The following array is used to find the best insn from ready when
310 the automaton pipeline interface is used. */
311 char *ready_try = NULL;
313 /* The ready list. */
314 struct ready_list ready = {NULL, 0, 0, 0, 0};
316 /* The pointer to the ready list (to be removed). */
317 static struct ready_list *readyp = &ready;
319 /* Scheduling clock. */
320 static int clock_var;
322 static int may_trap_exp (const_rtx, int);
324 /* Nonzero iff the address is comprised from at most 1 register. */
325 #define CONST_BASED_ADDRESS_P(x) \
326 (REG_P (x) \
327 || ((GET_CODE (x) == PLUS || GET_CODE (x) == MINUS \
328 || (GET_CODE (x) == LO_SUM)) \
329 && (CONSTANT_P (XEXP (x, 0)) \
330 || CONSTANT_P (XEXP (x, 1)))))
332 /* Returns a class that insn with GET_DEST(insn)=x may belong to,
333 as found by analyzing insn's expression. */
336 static int haifa_luid_for_non_insn (rtx x);
338 /* Haifa version of sched_info hooks common to all headers. */
339 const struct common_sched_info_def haifa_common_sched_info =
341 NULL, /* fix_recovery_cfg */
342 NULL, /* add_block */
343 NULL, /* estimate_number_of_insns */
344 haifa_luid_for_non_insn, /* luid_for_non_insn */
345 SCHED_PASS_UNKNOWN /* sched_pass_id */
348 const struct sched_scan_info_def *sched_scan_info;
350 /* Mapping from instruction UID to its Logical UID. */
351 VEC (int, heap) *sched_luids = NULL;
353 /* Next LUID to assign to an instruction. */
354 int sched_max_luid = 1;
356 /* Haifa Instruction Data. */
357 VEC (haifa_insn_data_def, heap) *h_i_d = NULL;
359 void (* sched_init_only_bb) (basic_block, basic_block);
361 /* Split block function. Different schedulers might use different functions
362 to handle their internal data consistent. */
363 basic_block (* sched_split_block) (basic_block, rtx);
365 /* Create empty basic block after the specified block. */
366 basic_block (* sched_create_empty_bb) (basic_block);
368 static int
369 may_trap_exp (const_rtx x, int is_store)
371 enum rtx_code code;
373 if (x == 0)
374 return TRAP_FREE;
375 code = GET_CODE (x);
376 if (is_store)
378 if (code == MEM && may_trap_p (x))
379 return TRAP_RISKY;
380 else
381 return TRAP_FREE;
383 if (code == MEM)
385 /* The insn uses memory: a volatile load. */
386 if (MEM_VOLATILE_P (x))
387 return IRISKY;
388 /* An exception-free load. */
389 if (!may_trap_p (x))
390 return IFREE;
391 /* A load with 1 base register, to be further checked. */
392 if (CONST_BASED_ADDRESS_P (XEXP (x, 0)))
393 return PFREE_CANDIDATE;
394 /* No info on the load, to be further checked. */
395 return PRISKY_CANDIDATE;
397 else
399 const char *fmt;
400 int i, insn_class = TRAP_FREE;
402 /* Neither store nor load, check if it may cause a trap. */
403 if (may_trap_p (x))
404 return TRAP_RISKY;
405 /* Recursive step: walk the insn... */
406 fmt = GET_RTX_FORMAT (code);
407 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
409 if (fmt[i] == 'e')
411 int tmp_class = may_trap_exp (XEXP (x, i), is_store);
412 insn_class = WORST_CLASS (insn_class, tmp_class);
414 else if (fmt[i] == 'E')
416 int j;
417 for (j = 0; j < XVECLEN (x, i); j++)
419 int tmp_class = may_trap_exp (XVECEXP (x, i, j), is_store);
420 insn_class = WORST_CLASS (insn_class, tmp_class);
421 if (insn_class == TRAP_RISKY || insn_class == IRISKY)
422 break;
425 if (insn_class == TRAP_RISKY || insn_class == IRISKY)
426 break;
428 return insn_class;
432 /* Classifies rtx X of an insn for the purpose of verifying that X can be
433 executed speculatively (and consequently the insn can be moved
434 speculatively), by examining X, returning:
435 TRAP_RISKY: store, or risky non-load insn (e.g. division by variable).
436 TRAP_FREE: non-load insn.
437 IFREE: load from a globally safe location.
438 IRISKY: volatile load.
439 PFREE_CANDIDATE, PRISKY_CANDIDATE: load that need to be checked for
440 being either PFREE or PRISKY. */
442 static int
443 haifa_classify_rtx (const_rtx x)
445 int tmp_class = TRAP_FREE;
446 int insn_class = TRAP_FREE;
447 enum rtx_code code;
449 if (GET_CODE (x) == PARALLEL)
451 int i, len = XVECLEN (x, 0);
453 for (i = len - 1; i >= 0; i--)
455 tmp_class = haifa_classify_rtx (XVECEXP (x, 0, i));
456 insn_class = WORST_CLASS (insn_class, tmp_class);
457 if (insn_class == TRAP_RISKY || insn_class == IRISKY)
458 break;
461 else
463 code = GET_CODE (x);
464 switch (code)
466 case CLOBBER:
467 /* Test if it is a 'store'. */
468 tmp_class = may_trap_exp (XEXP (x, 0), 1);
469 break;
470 case SET:
471 /* Test if it is a store. */
472 tmp_class = may_trap_exp (SET_DEST (x), 1);
473 if (tmp_class == TRAP_RISKY)
474 break;
475 /* Test if it is a load. */
476 tmp_class =
477 WORST_CLASS (tmp_class,
478 may_trap_exp (SET_SRC (x), 0));
479 break;
480 case COND_EXEC:
481 tmp_class = haifa_classify_rtx (COND_EXEC_CODE (x));
482 if (tmp_class == TRAP_RISKY)
483 break;
484 tmp_class = WORST_CLASS (tmp_class,
485 may_trap_exp (COND_EXEC_TEST (x), 0));
486 break;
487 case TRAP_IF:
488 tmp_class = TRAP_RISKY;
489 break;
490 default:;
492 insn_class = tmp_class;
495 return insn_class;
499 haifa_classify_insn (const_rtx insn)
501 return haifa_classify_rtx (PATTERN (insn));
504 /* Forward declarations. */
506 static int priority (rtx);
507 static int rank_for_schedule (const void *, const void *);
508 static void swap_sort (rtx *, int);
509 static void queue_insn (rtx, int);
510 static int schedule_insn (rtx);
511 static void adjust_priority (rtx);
512 static void advance_one_cycle (void);
513 static void extend_h_i_d (void);
516 /* Notes handling mechanism:
517 =========================
518 Generally, NOTES are saved before scheduling and restored after scheduling.
519 The scheduler distinguishes between two types of notes:
521 (1) LOOP_BEGIN, LOOP_END, SETJMP, EHREGION_BEG, EHREGION_END notes:
522 Before scheduling a region, a pointer to the note is added to the insn
523 that follows or precedes it. (This happens as part of the data dependence
524 computation). After scheduling an insn, the pointer contained in it is
525 used for regenerating the corresponding note (in reemit_notes).
527 (2) All other notes (e.g. INSN_DELETED): Before scheduling a block,
528 these notes are put in a list (in rm_other_notes() and
529 unlink_other_notes ()). After scheduling the block, these notes are
530 inserted at the beginning of the block (in schedule_block()). */
532 static void ready_add (struct ready_list *, rtx, bool);
533 static rtx ready_remove_first (struct ready_list *);
535 static void queue_to_ready (struct ready_list *);
536 static int early_queue_to_ready (state_t, struct ready_list *);
538 static void debug_ready_list (struct ready_list *);
540 /* The following functions are used to implement multi-pass scheduling
541 on the first cycle. */
542 static rtx ready_remove (struct ready_list *, int);
543 static void ready_remove_insn (rtx);
545 static int choose_ready (struct ready_list *, rtx *);
547 static void fix_inter_tick (rtx, rtx);
548 static int fix_tick_ready (rtx);
549 static void change_queue_index (rtx, int);
551 /* The following functions are used to implement scheduling of data/control
552 speculative instructions. */
554 static void extend_h_i_d (void);
555 static void init_h_i_d (rtx);
556 static void generate_recovery_code (rtx);
557 static void process_insn_forw_deps_be_in_spec (rtx, rtx, ds_t);
558 static void begin_speculative_block (rtx);
559 static void add_to_speculative_block (rtx);
560 static void init_before_recovery (basic_block *);
561 static void create_check_block_twin (rtx, bool);
562 static void fix_recovery_deps (basic_block);
563 static void haifa_change_pattern (rtx, rtx);
564 static void dump_new_block_header (int, basic_block, rtx, rtx);
565 static void restore_bb_notes (basic_block);
566 static void fix_jump_move (rtx);
567 static void move_block_after_check (rtx);
568 static void move_succs (VEC(edge,gc) **, basic_block);
569 static void sched_remove_insn (rtx);
570 static void clear_priorities (rtx, rtx_vec_t *);
571 static void calc_priorities (rtx_vec_t);
572 static void add_jump_dependencies (rtx, rtx);
573 #ifdef ENABLE_CHECKING
574 static int has_edge_p (VEC(edge,gc) *, int);
575 static void check_cfg (rtx, rtx);
576 #endif
578 #endif /* INSN_SCHEDULING */
580 /* Point to state used for the current scheduling pass. */
581 struct haifa_sched_info *current_sched_info;
583 #ifndef INSN_SCHEDULING
584 void
585 schedule_insns (void)
588 #else
590 /* Do register pressure sensitive insn scheduling if the flag is set
591 up. */
592 bool sched_pressure_p;
594 /* Map regno -> its cover class. The map defined only when
595 SCHED_PRESSURE_P is true. */
596 enum reg_class *sched_regno_cover_class;
598 /* The current register pressure. Only elements corresponding cover
599 classes are defined. */
600 static int curr_reg_pressure[N_REG_CLASSES];
602 /* Saved value of the previous array. */
603 static int saved_reg_pressure[N_REG_CLASSES];
605 /* Register living at given scheduling point. */
606 static bitmap curr_reg_live;
608 /* Saved value of the previous array. */
609 static bitmap saved_reg_live;
611 /* Registers mentioned in the current region. */
612 static bitmap region_ref_regs;
614 /* Initiate register pressure relative info for scheduling the current
615 region. Currently it is only clearing register mentioned in the
616 current region. */
617 void
618 sched_init_region_reg_pressure_info (void)
620 bitmap_clear (region_ref_regs);
623 /* Update current register pressure related info after birth (if
624 BIRTH_P) or death of register REGNO. */
625 static void
626 mark_regno_birth_or_death (int regno, bool birth_p)
628 enum reg_class cover_class;
630 cover_class = sched_regno_cover_class[regno];
631 if (regno >= FIRST_PSEUDO_REGISTER)
633 if (cover_class != NO_REGS)
635 if (birth_p)
637 bitmap_set_bit (curr_reg_live, regno);
638 curr_reg_pressure[cover_class]
639 += ira_reg_class_nregs[cover_class][PSEUDO_REGNO_MODE (regno)];
641 else
643 bitmap_clear_bit (curr_reg_live, regno);
644 curr_reg_pressure[cover_class]
645 -= ira_reg_class_nregs[cover_class][PSEUDO_REGNO_MODE (regno)];
649 else if (cover_class != NO_REGS
650 && ! TEST_HARD_REG_BIT (ira_no_alloc_regs, regno))
652 if (birth_p)
654 bitmap_set_bit (curr_reg_live, regno);
655 curr_reg_pressure[cover_class]++;
657 else
659 bitmap_clear_bit (curr_reg_live, regno);
660 curr_reg_pressure[cover_class]--;
665 /* Initiate current register pressure related info from living
666 registers given by LIVE. */
667 static void
668 initiate_reg_pressure_info (bitmap live)
670 int i;
671 unsigned int j;
672 bitmap_iterator bi;
674 for (i = 0; i < ira_reg_class_cover_size; i++)
675 curr_reg_pressure[ira_reg_class_cover[i]] = 0;
676 bitmap_clear (curr_reg_live);
677 EXECUTE_IF_SET_IN_BITMAP (live, 0, j, bi)
678 if (current_nr_blocks == 1 || bitmap_bit_p (region_ref_regs, j))
679 mark_regno_birth_or_death (j, true);
682 /* Mark registers in X as mentioned in the current region. */
683 static void
684 setup_ref_regs (rtx x)
686 int i, j, regno;
687 const RTX_CODE code = GET_CODE (x);
688 const char *fmt;
690 if (REG_P (x))
692 regno = REGNO (x);
693 if (regno >= FIRST_PSEUDO_REGISTER)
694 bitmap_set_bit (region_ref_regs, REGNO (x));
695 else
696 for (i = hard_regno_nregs[regno][GET_MODE (x)] - 1; i >= 0; i--)
697 bitmap_set_bit (region_ref_regs, regno + i);
698 return;
700 fmt = GET_RTX_FORMAT (code);
701 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
702 if (fmt[i] == 'e')
703 setup_ref_regs (XEXP (x, i));
704 else if (fmt[i] == 'E')
706 for (j = 0; j < XVECLEN (x, i); j++)
707 setup_ref_regs (XVECEXP (x, i, j));
711 /* Initiate current register pressure related info at the start of
712 basic block BB. */
713 static void
714 initiate_bb_reg_pressure_info (basic_block bb)
716 unsigned int i;
717 rtx insn;
719 if (current_nr_blocks > 1)
720 FOR_BB_INSNS (bb, insn)
721 if (INSN_P (insn))
722 setup_ref_regs (PATTERN (insn));
723 initiate_reg_pressure_info (df_get_live_in (bb));
724 #ifdef EH_RETURN_DATA_REGNO
725 if (bb_has_eh_pred (bb))
726 for (i = 0; ; ++i)
728 unsigned int regno = EH_RETURN_DATA_REGNO (i);
730 if (regno == INVALID_REGNUM)
731 break;
732 if (! bitmap_bit_p (df_get_live_in (bb), regno))
733 mark_regno_birth_or_death (regno, true);
735 #endif
738 /* Save current register pressure related info. */
739 static void
740 save_reg_pressure (void)
742 int i;
744 for (i = 0; i < ira_reg_class_cover_size; i++)
745 saved_reg_pressure[ira_reg_class_cover[i]]
746 = curr_reg_pressure[ira_reg_class_cover[i]];
747 bitmap_copy (saved_reg_live, curr_reg_live);
750 /* Restore saved register pressure related info. */
751 static void
752 restore_reg_pressure (void)
754 int i;
756 for (i = 0; i < ira_reg_class_cover_size; i++)
757 curr_reg_pressure[ira_reg_class_cover[i]]
758 = saved_reg_pressure[ira_reg_class_cover[i]];
759 bitmap_copy (curr_reg_live, saved_reg_live);
762 /* Return TRUE if the register is dying after its USE. */
763 static bool
764 dying_use_p (struct reg_use_data *use)
766 struct reg_use_data *next;
768 for (next = use->next_regno_use; next != use; next = next->next_regno_use)
769 if (NONDEBUG_INSN_P (next->insn)
770 && QUEUE_INDEX (next->insn) != QUEUE_SCHEDULED)
771 return false;
772 return true;
775 /* Print info about the current register pressure and its excess for
776 each cover class. */
777 static void
778 print_curr_reg_pressure (void)
780 int i;
781 enum reg_class cl;
783 fprintf (sched_dump, ";;\t");
784 for (i = 0; i < ira_reg_class_cover_size; i++)
786 cl = ira_reg_class_cover[i];
787 gcc_assert (curr_reg_pressure[cl] >= 0);
788 fprintf (sched_dump, " %s:%d(%d)", reg_class_names[cl],
789 curr_reg_pressure[cl],
790 curr_reg_pressure[cl] - ira_available_class_regs[cl]);
792 fprintf (sched_dump, "\n");
795 /* Pointer to the last instruction scheduled. Used by rank_for_schedule,
796 so that insns independent of the last scheduled insn will be preferred
797 over dependent instructions. */
799 static rtx last_scheduled_insn;
801 /* Cached cost of the instruction. Use below function to get cost of the
802 insn. -1 here means that the field is not initialized. */
803 #define INSN_COST(INSN) (HID (INSN)->cost)
805 /* Compute cost of executing INSN.
806 This is the number of cycles between instruction issue and
807 instruction results. */
809 insn_cost (rtx insn)
811 int cost;
813 if (sel_sched_p ())
815 if (recog_memoized (insn) < 0)
816 return 0;
818 cost = insn_default_latency (insn);
819 if (cost < 0)
820 cost = 0;
822 return cost;
825 cost = INSN_COST (insn);
827 if (cost < 0)
829 /* A USE insn, or something else we don't need to
830 understand. We can't pass these directly to
831 result_ready_cost or insn_default_latency because it will
832 trigger a fatal error for unrecognizable insns. */
833 if (recog_memoized (insn) < 0)
835 INSN_COST (insn) = 0;
836 return 0;
838 else
840 cost = insn_default_latency (insn);
841 if (cost < 0)
842 cost = 0;
844 INSN_COST (insn) = cost;
848 return cost;
851 /* Compute cost of dependence LINK.
852 This is the number of cycles between instruction issue and
853 instruction results.
854 ??? We also use this function to call recog_memoized on all insns. */
856 dep_cost_1 (dep_t link, dw_t dw)
858 rtx insn = DEP_PRO (link);
859 rtx used = DEP_CON (link);
860 int cost;
862 /* A USE insn should never require the value used to be computed.
863 This allows the computation of a function's result and parameter
864 values to overlap the return and call. We don't care about the
865 the dependence cost when only decreasing register pressure. */
866 if (recog_memoized (used) < 0)
868 cost = 0;
869 recog_memoized (insn);
871 else
873 enum reg_note dep_type = DEP_TYPE (link);
875 cost = insn_cost (insn);
877 if (INSN_CODE (insn) >= 0)
879 if (dep_type == REG_DEP_ANTI)
880 cost = 0;
881 else if (dep_type == REG_DEP_OUTPUT)
883 cost = (insn_default_latency (insn)
884 - insn_default_latency (used));
885 if (cost <= 0)
886 cost = 1;
888 else if (bypass_p (insn))
889 cost = insn_latency (insn, used);
893 if (targetm.sched.adjust_cost_2)
894 cost = targetm.sched.adjust_cost_2 (used, (int) dep_type, insn, cost,
895 dw);
896 else if (targetm.sched.adjust_cost != NULL)
898 /* This variable is used for backward compatibility with the
899 targets. */
900 rtx dep_cost_rtx_link = alloc_INSN_LIST (NULL_RTX, NULL_RTX);
902 /* Make it self-cycled, so that if some tries to walk over this
903 incomplete list he/she will be caught in an endless loop. */
904 XEXP (dep_cost_rtx_link, 1) = dep_cost_rtx_link;
906 /* Targets use only REG_NOTE_KIND of the link. */
907 PUT_REG_NOTE_KIND (dep_cost_rtx_link, DEP_TYPE (link));
909 cost = targetm.sched.adjust_cost (used, dep_cost_rtx_link,
910 insn, cost);
912 free_INSN_LIST_node (dep_cost_rtx_link);
915 if (cost < 0)
916 cost = 0;
919 return cost;
922 /* Compute cost of dependence LINK.
923 This is the number of cycles between instruction issue and
924 instruction results. */
926 dep_cost (dep_t link)
928 return dep_cost_1 (link, 0);
931 /* Use this sel-sched.c friendly function in reorder2 instead of increasing
932 INSN_PRIORITY explicitly. */
933 void
934 increase_insn_priority (rtx insn, int amount)
936 if (!sel_sched_p ())
938 /* We're dealing with haifa-sched.c INSN_PRIORITY. */
939 if (INSN_PRIORITY_KNOWN (insn))
940 INSN_PRIORITY (insn) += amount;
942 else
944 /* In sel-sched.c INSN_PRIORITY is not kept up to date.
945 Use EXPR_PRIORITY instead. */
946 sel_add_to_insn_priority (insn, amount);
950 /* Return 'true' if DEP should be included in priority calculations. */
951 static bool
952 contributes_to_priority_p (dep_t dep)
954 if (DEBUG_INSN_P (DEP_CON (dep))
955 || DEBUG_INSN_P (DEP_PRO (dep)))
956 return false;
958 /* Critical path is meaningful in block boundaries only. */
959 if (!current_sched_info->contributes_to_priority (DEP_CON (dep),
960 DEP_PRO (dep)))
961 return false;
963 /* If flag COUNT_SPEC_IN_CRITICAL_PATH is set,
964 then speculative instructions will less likely be
965 scheduled. That is because the priority of
966 their producers will increase, and, thus, the
967 producers will more likely be scheduled, thus,
968 resolving the dependence. */
969 if (sched_deps_info->generate_spec_deps
970 && !(spec_info->flags & COUNT_SPEC_IN_CRITICAL_PATH)
971 && (DEP_STATUS (dep) & SPECULATIVE))
972 return false;
974 return true;
977 /* Compute the number of nondebug forward deps of an insn. */
979 static int
980 dep_list_size (rtx insn)
982 sd_iterator_def sd_it;
983 dep_t dep;
984 int dbgcount = 0, nodbgcount = 0;
986 if (!MAY_HAVE_DEBUG_INSNS)
987 return sd_lists_size (insn, SD_LIST_FORW);
989 FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep)
991 if (DEBUG_INSN_P (DEP_CON (dep)))
992 dbgcount++;
993 else if (!DEBUG_INSN_P (DEP_PRO (dep)))
994 nodbgcount++;
997 gcc_assert (dbgcount + nodbgcount == sd_lists_size (insn, SD_LIST_FORW));
999 return nodbgcount;
1002 /* Compute the priority number for INSN. */
1003 static int
1004 priority (rtx insn)
1006 if (! INSN_P (insn))
1007 return 0;
1009 /* We should not be interested in priority of an already scheduled insn. */
1010 gcc_assert (QUEUE_INDEX (insn) != QUEUE_SCHEDULED);
1012 if (!INSN_PRIORITY_KNOWN (insn))
1014 int this_priority = -1;
1016 if (dep_list_size (insn) == 0)
1017 /* ??? We should set INSN_PRIORITY to insn_cost when and insn has
1018 some forward deps but all of them are ignored by
1019 contributes_to_priority hook. At the moment we set priority of
1020 such insn to 0. */
1021 this_priority = insn_cost (insn);
1022 else
1024 rtx prev_first, twin;
1025 basic_block rec;
1027 /* For recovery check instructions we calculate priority slightly
1028 different than that of normal instructions. Instead of walking
1029 through INSN_FORW_DEPS (check) list, we walk through
1030 INSN_FORW_DEPS list of each instruction in the corresponding
1031 recovery block. */
1033 /* Selective scheduling does not define RECOVERY_BLOCK macro. */
1034 rec = sel_sched_p () ? NULL : RECOVERY_BLOCK (insn);
1035 if (!rec || rec == EXIT_BLOCK_PTR)
1037 prev_first = PREV_INSN (insn);
1038 twin = insn;
1040 else
1042 prev_first = NEXT_INSN (BB_HEAD (rec));
1043 twin = PREV_INSN (BB_END (rec));
1048 sd_iterator_def sd_it;
1049 dep_t dep;
1051 FOR_EACH_DEP (twin, SD_LIST_FORW, sd_it, dep)
1053 rtx next;
1054 int next_priority;
1056 next = DEP_CON (dep);
1058 if (BLOCK_FOR_INSN (next) != rec)
1060 int cost;
1062 if (!contributes_to_priority_p (dep))
1063 continue;
1065 if (twin == insn)
1066 cost = dep_cost (dep);
1067 else
1069 struct _dep _dep1, *dep1 = &_dep1;
1071 init_dep (dep1, insn, next, REG_DEP_ANTI);
1073 cost = dep_cost (dep1);
1076 next_priority = cost + priority (next);
1078 if (next_priority > this_priority)
1079 this_priority = next_priority;
1083 twin = PREV_INSN (twin);
1085 while (twin != prev_first);
1088 if (this_priority < 0)
1090 gcc_assert (this_priority == -1);
1092 this_priority = insn_cost (insn);
1095 INSN_PRIORITY (insn) = this_priority;
1096 INSN_PRIORITY_STATUS (insn) = 1;
1099 return INSN_PRIORITY (insn);
1102 /* Macros and functions for keeping the priority queue sorted, and
1103 dealing with queuing and dequeuing of instructions. */
1105 #define SCHED_SORT(READY, N_READY) \
1106 do { if ((N_READY) == 2) \
1107 swap_sort (READY, N_READY); \
1108 else if ((N_READY) > 2) \
1109 qsort (READY, N_READY, sizeof (rtx), rank_for_schedule); } \
1110 while (0)
1112 /* Setup info about the current register pressure impact of scheduling
1113 INSN at the current scheduling point. */
1114 static void
1115 setup_insn_reg_pressure_info (rtx insn)
1117 int i, change, before, after, hard_regno;
1118 int excess_cost_change;
1119 enum machine_mode mode;
1120 enum reg_class cl;
1121 struct reg_pressure_data *pressure_info;
1122 int *max_reg_pressure;
1123 struct reg_use_data *use;
1124 static int death[N_REG_CLASSES];
1126 excess_cost_change = 0;
1127 for (i = 0; i < ira_reg_class_cover_size; i++)
1128 death[ira_reg_class_cover[i]] = 0;
1129 for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
1130 if (dying_use_p (use))
1132 cl = sched_regno_cover_class[use->regno];
1133 if (use->regno < FIRST_PSEUDO_REGISTER)
1134 death[cl]++;
1135 else
1136 death[cl] += ira_reg_class_nregs[cl][PSEUDO_REGNO_MODE (use->regno)];
1138 pressure_info = INSN_REG_PRESSURE (insn);
1139 max_reg_pressure = INSN_MAX_REG_PRESSURE (insn);
1140 gcc_assert (pressure_info != NULL && max_reg_pressure != NULL);
1141 for (i = 0; i < ira_reg_class_cover_size; i++)
1143 cl = ira_reg_class_cover[i];
1144 gcc_assert (curr_reg_pressure[cl] >= 0);
1145 change = (int) pressure_info[i].set_increase - death[cl];
1146 before = MAX (0, max_reg_pressure[i] - ira_available_class_regs[cl]);
1147 after = MAX (0, max_reg_pressure[i] + change
1148 - ira_available_class_regs[cl]);
1149 hard_regno = ira_class_hard_regs[cl][0];
1150 gcc_assert (hard_regno >= 0);
1151 mode = reg_raw_mode[hard_regno];
1152 excess_cost_change += ((after - before)
1153 * (ira_memory_move_cost[mode][cl][0]
1154 + ira_memory_move_cost[mode][cl][1]));
1156 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insn) = excess_cost_change;
1159 /* Returns a positive value if x is preferred; returns a negative value if
1160 y is preferred. Should never return 0, since that will make the sort
1161 unstable. */
1163 static int
1164 rank_for_schedule (const void *x, const void *y)
1166 rtx tmp = *(const rtx *) y;
1167 rtx tmp2 = *(const rtx *) x;
1168 rtx last;
1169 int tmp_class, tmp2_class;
1170 int val, priority_val, info_val;
1172 if (MAY_HAVE_DEBUG_INSNS)
1174 /* Schedule debug insns as early as possible. */
1175 if (DEBUG_INSN_P (tmp) && !DEBUG_INSN_P (tmp2))
1176 return -1;
1177 else if (DEBUG_INSN_P (tmp2))
1178 return 1;
1181 /* The insn in a schedule group should be issued the first. */
1182 if (flag_sched_group_heuristic &&
1183 SCHED_GROUP_P (tmp) != SCHED_GROUP_P (tmp2))
1184 return SCHED_GROUP_P (tmp2) ? 1 : -1;
1186 /* Make sure that priority of TMP and TMP2 are initialized. */
1187 gcc_assert (INSN_PRIORITY_KNOWN (tmp) && INSN_PRIORITY_KNOWN (tmp2));
1189 if (sched_pressure_p)
1191 int diff;
1193 /* Prefer insn whose scheduling results in the smallest register
1194 pressure excess. */
1195 if ((diff = (INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp)
1196 + (INSN_TICK (tmp) > clock_var
1197 ? INSN_TICK (tmp) - clock_var : 0)
1198 - INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2)
1199 - (INSN_TICK (tmp2) > clock_var
1200 ? INSN_TICK (tmp2) - clock_var : 0))) != 0)
1201 return diff;
1205 if (sched_pressure_p
1206 && (INSN_TICK (tmp2) > clock_var || INSN_TICK (tmp) > clock_var))
1208 if (INSN_TICK (tmp) <= clock_var)
1209 return -1;
1210 else if (INSN_TICK (tmp2) <= clock_var)
1211 return 1;
1212 else
1213 return INSN_TICK (tmp) - INSN_TICK (tmp2);
1215 /* Prefer insn with higher priority. */
1216 priority_val = INSN_PRIORITY (tmp2) - INSN_PRIORITY (tmp);
1218 if (flag_sched_critical_path_heuristic && priority_val)
1219 return priority_val;
1221 /* Prefer speculative insn with greater dependencies weakness. */
1222 if (flag_sched_spec_insn_heuristic && spec_info)
1224 ds_t ds1, ds2;
1225 dw_t dw1, dw2;
1226 int dw;
1228 ds1 = TODO_SPEC (tmp) & SPECULATIVE;
1229 if (ds1)
1230 dw1 = ds_weak (ds1);
1231 else
1232 dw1 = NO_DEP_WEAK;
1234 ds2 = TODO_SPEC (tmp2) & SPECULATIVE;
1235 if (ds2)
1236 dw2 = ds_weak (ds2);
1237 else
1238 dw2 = NO_DEP_WEAK;
1240 dw = dw2 - dw1;
1241 if (dw > (NO_DEP_WEAK / 8) || dw < -(NO_DEP_WEAK / 8))
1242 return dw;
1245 info_val = (*current_sched_info->rank) (tmp, tmp2);
1246 if(flag_sched_rank_heuristic && info_val)
1247 return info_val;
1249 if (flag_sched_last_insn_heuristic)
1251 last = last_scheduled_insn;
1253 if (DEBUG_INSN_P (last) && last != current_sched_info->prev_head)
1255 last = PREV_INSN (last);
1256 while (!NONDEBUG_INSN_P (last)
1257 && last != current_sched_info->prev_head);
1260 /* Compare insns based on their relation to the last scheduled
1261 non-debug insn. */
1262 if (flag_sched_last_insn_heuristic && NONDEBUG_INSN_P (last))
1264 dep_t dep1;
1265 dep_t dep2;
1267 /* Classify the instructions into three classes:
1268 1) Data dependent on last schedule insn.
1269 2) Anti/Output dependent on last scheduled insn.
1270 3) Independent of last scheduled insn, or has latency of one.
1271 Choose the insn from the highest numbered class if different. */
1272 dep1 = sd_find_dep_between (last, tmp, true);
1274 if (dep1 == NULL || dep_cost (dep1) == 1)
1275 tmp_class = 3;
1276 else if (/* Data dependence. */
1277 DEP_TYPE (dep1) == REG_DEP_TRUE)
1278 tmp_class = 1;
1279 else
1280 tmp_class = 2;
1282 dep2 = sd_find_dep_between (last, tmp2, true);
1284 if (dep2 == NULL || dep_cost (dep2) == 1)
1285 tmp2_class = 3;
1286 else if (/* Data dependence. */
1287 DEP_TYPE (dep2) == REG_DEP_TRUE)
1288 tmp2_class = 1;
1289 else
1290 tmp2_class = 2;
1292 if ((val = tmp2_class - tmp_class))
1293 return val;
1296 /* Prefer the insn which has more later insns that depend on it.
1297 This gives the scheduler more freedom when scheduling later
1298 instructions at the expense of added register pressure. */
1300 val = (dep_list_size (tmp2) - dep_list_size (tmp));
1302 if (flag_sched_dep_count_heuristic && val != 0)
1303 return val;
1305 /* If insns are equally good, sort by INSN_LUID (original insn order),
1306 so that we make the sort stable. This minimizes instruction movement,
1307 thus minimizing sched's effect on debugging and cross-jumping. */
1308 return INSN_LUID (tmp) - INSN_LUID (tmp2);
1311 /* Resort the array A in which only element at index N may be out of order. */
1313 HAIFA_INLINE static void
1314 swap_sort (rtx *a, int n)
1316 rtx insn = a[n - 1];
1317 int i = n - 2;
1319 while (i >= 0 && rank_for_schedule (a + i, &insn) >= 0)
1321 a[i + 1] = a[i];
1322 i -= 1;
1324 a[i + 1] = insn;
1327 /* Add INSN to the insn queue so that it can be executed at least
1328 N_CYCLES after the currently executing insn. Preserve insns
1329 chain for debugging purposes. */
1331 HAIFA_INLINE static void
1332 queue_insn (rtx insn, int n_cycles)
1334 int next_q = NEXT_Q_AFTER (q_ptr, n_cycles);
1335 rtx link = alloc_INSN_LIST (insn, insn_queue[next_q]);
1337 gcc_assert (n_cycles <= max_insn_queue_index);
1338 gcc_assert (!DEBUG_INSN_P (insn));
1340 insn_queue[next_q] = link;
1341 q_size += 1;
1343 if (sched_verbose >= 2)
1345 fprintf (sched_dump, ";;\t\tReady-->Q: insn %s: ",
1346 (*current_sched_info->print_insn) (insn, 0));
1348 fprintf (sched_dump, "queued for %d cycles.\n", n_cycles);
1351 QUEUE_INDEX (insn) = next_q;
1354 /* Remove INSN from queue. */
1355 static void
1356 queue_remove (rtx insn)
1358 gcc_assert (QUEUE_INDEX (insn) >= 0);
1359 remove_free_INSN_LIST_elem (insn, &insn_queue[QUEUE_INDEX (insn)]);
1360 q_size--;
1361 QUEUE_INDEX (insn) = QUEUE_NOWHERE;
1364 /* Return a pointer to the bottom of the ready list, i.e. the insn
1365 with the lowest priority. */
1367 rtx *
1368 ready_lastpos (struct ready_list *ready)
1370 gcc_assert (ready->n_ready >= 1);
1371 return ready->vec + ready->first - ready->n_ready + 1;
1374 /* Add an element INSN to the ready list so that it ends up with the
1375 lowest/highest priority depending on FIRST_P. */
1377 HAIFA_INLINE static void
1378 ready_add (struct ready_list *ready, rtx insn, bool first_p)
1380 if (!first_p)
1382 if (ready->first == ready->n_ready)
1384 memmove (ready->vec + ready->veclen - ready->n_ready,
1385 ready_lastpos (ready),
1386 ready->n_ready * sizeof (rtx));
1387 ready->first = ready->veclen - 1;
1389 ready->vec[ready->first - ready->n_ready] = insn;
1391 else
1393 if (ready->first == ready->veclen - 1)
1395 if (ready->n_ready)
1396 /* ready_lastpos() fails when called with (ready->n_ready == 0). */
1397 memmove (ready->vec + ready->veclen - ready->n_ready - 1,
1398 ready_lastpos (ready),
1399 ready->n_ready * sizeof (rtx));
1400 ready->first = ready->veclen - 2;
1402 ready->vec[++(ready->first)] = insn;
1405 ready->n_ready++;
1406 if (DEBUG_INSN_P (insn))
1407 ready->n_debug++;
1409 gcc_assert (QUEUE_INDEX (insn) != QUEUE_READY);
1410 QUEUE_INDEX (insn) = QUEUE_READY;
1413 /* Remove the element with the highest priority from the ready list and
1414 return it. */
1416 HAIFA_INLINE static rtx
1417 ready_remove_first (struct ready_list *ready)
1419 rtx t;
1421 gcc_assert (ready->n_ready);
1422 t = ready->vec[ready->first--];
1423 ready->n_ready--;
1424 if (DEBUG_INSN_P (t))
1425 ready->n_debug--;
1426 /* If the queue becomes empty, reset it. */
1427 if (ready->n_ready == 0)
1428 ready->first = ready->veclen - 1;
1430 gcc_assert (QUEUE_INDEX (t) == QUEUE_READY);
1431 QUEUE_INDEX (t) = QUEUE_NOWHERE;
1433 return t;
1436 /* The following code implements multi-pass scheduling for the first
1437 cycle. In other words, we will try to choose ready insn which
1438 permits to start maximum number of insns on the same cycle. */
1440 /* Return a pointer to the element INDEX from the ready. INDEX for
1441 insn with the highest priority is 0, and the lowest priority has
1442 N_READY - 1. */
1445 ready_element (struct ready_list *ready, int index)
1447 gcc_assert (ready->n_ready && index < ready->n_ready);
1449 return ready->vec[ready->first - index];
1452 /* Remove the element INDEX from the ready list and return it. INDEX
1453 for insn with the highest priority is 0, and the lowest priority
1454 has N_READY - 1. */
1456 HAIFA_INLINE static rtx
1457 ready_remove (struct ready_list *ready, int index)
1459 rtx t;
1460 int i;
1462 if (index == 0)
1463 return ready_remove_first (ready);
1464 gcc_assert (ready->n_ready && index < ready->n_ready);
1465 t = ready->vec[ready->first - index];
1466 ready->n_ready--;
1467 if (DEBUG_INSN_P (t))
1468 ready->n_debug--;
1469 for (i = index; i < ready->n_ready; i++)
1470 ready->vec[ready->first - i] = ready->vec[ready->first - i - 1];
1471 QUEUE_INDEX (t) = QUEUE_NOWHERE;
1472 return t;
1475 /* Remove INSN from the ready list. */
1476 static void
1477 ready_remove_insn (rtx insn)
1479 int i;
1481 for (i = 0; i < readyp->n_ready; i++)
1482 if (ready_element (readyp, i) == insn)
1484 ready_remove (readyp, i);
1485 return;
1487 gcc_unreachable ();
1490 /* Sort the ready list READY by ascending priority, using the SCHED_SORT
1491 macro. */
1493 void
1494 ready_sort (struct ready_list *ready)
1496 int i;
1497 rtx *first = ready_lastpos (ready);
1499 if (sched_pressure_p)
1501 for (i = 0; i < ready->n_ready; i++)
1502 setup_insn_reg_pressure_info (first[i]);
1504 SCHED_SORT (first, ready->n_ready);
1507 /* PREV is an insn that is ready to execute. Adjust its priority if that
1508 will help shorten or lengthen register lifetimes as appropriate. Also
1509 provide a hook for the target to tweak itself. */
1511 HAIFA_INLINE static void
1512 adjust_priority (rtx prev)
1514 /* ??? There used to be code here to try and estimate how an insn
1515 affected register lifetimes, but it did it by looking at REG_DEAD
1516 notes, which we removed in schedule_region. Nor did it try to
1517 take into account register pressure or anything useful like that.
1519 Revisit when we have a machine model to work with and not before. */
1521 if (targetm.sched.adjust_priority)
1522 INSN_PRIORITY (prev) =
1523 targetm.sched.adjust_priority (prev, INSN_PRIORITY (prev));
1526 /* Advance DFA state STATE on one cycle. */
1527 void
1528 advance_state (state_t state)
1530 if (targetm.sched.dfa_pre_advance_cycle)
1531 targetm.sched.dfa_pre_advance_cycle ();
1533 if (targetm.sched.dfa_pre_cycle_insn)
1534 state_transition (state,
1535 targetm.sched.dfa_pre_cycle_insn ());
1537 state_transition (state, NULL);
1539 if (targetm.sched.dfa_post_cycle_insn)
1540 state_transition (state,
1541 targetm.sched.dfa_post_cycle_insn ());
1543 if (targetm.sched.dfa_post_advance_cycle)
1544 targetm.sched.dfa_post_advance_cycle ();
1547 /* Advance time on one cycle. */
1548 HAIFA_INLINE static void
1549 advance_one_cycle (void)
1551 advance_state (curr_state);
1552 if (sched_verbose >= 6)
1553 fprintf (sched_dump, ";;\tAdvanced a state.\n");
1556 /* Clock at which the previous instruction was issued. */
1557 static int last_clock_var;
1559 /* Update register pressure after scheduling INSN. */
1560 static void
1561 update_register_pressure (rtx insn)
1563 struct reg_use_data *use;
1564 struct reg_set_data *set;
1566 for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
1567 if (dying_use_p (use) && bitmap_bit_p (curr_reg_live, use->regno))
1568 mark_regno_birth_or_death (use->regno, false);
1569 for (set = INSN_REG_SET_LIST (insn); set != NULL; set = set->next_insn_set)
1570 mark_regno_birth_or_death (set->regno, true);
1573 /* Set up or update (if UPDATE_P) max register pressure (see its
1574 meaning in sched-int.h::_haifa_insn_data) for all current BB insns
1575 after insn AFTER. */
1576 static void
1577 setup_insn_max_reg_pressure (rtx after, bool update_p)
1579 int i, p;
1580 bool eq_p;
1581 rtx insn;
1582 static int max_reg_pressure[N_REG_CLASSES];
1584 save_reg_pressure ();
1585 for (i = 0; i < ira_reg_class_cover_size; i++)
1586 max_reg_pressure[ira_reg_class_cover[i]]
1587 = curr_reg_pressure[ira_reg_class_cover[i]];
1588 for (insn = NEXT_INSN (after);
1589 insn != NULL_RTX && ! BARRIER_P (insn)
1590 && BLOCK_FOR_INSN (insn) == BLOCK_FOR_INSN (after);
1591 insn = NEXT_INSN (insn))
1592 if (NONDEBUG_INSN_P (insn))
1594 eq_p = true;
1595 for (i = 0; i < ira_reg_class_cover_size; i++)
1597 p = max_reg_pressure[ira_reg_class_cover[i]];
1598 if (INSN_MAX_REG_PRESSURE (insn)[i] != p)
1600 eq_p = false;
1601 INSN_MAX_REG_PRESSURE (insn)[i]
1602 = max_reg_pressure[ira_reg_class_cover[i]];
1605 if (update_p && eq_p)
1606 break;
1607 update_register_pressure (insn);
1608 for (i = 0; i < ira_reg_class_cover_size; i++)
1609 if (max_reg_pressure[ira_reg_class_cover[i]]
1610 < curr_reg_pressure[ira_reg_class_cover[i]])
1611 max_reg_pressure[ira_reg_class_cover[i]]
1612 = curr_reg_pressure[ira_reg_class_cover[i]];
1614 restore_reg_pressure ();
1617 /* Update the current register pressure after scheduling INSN. Update
1618 also max register pressure for unscheduled insns of the current
1619 BB. */
1620 static void
1621 update_reg_and_insn_max_reg_pressure (rtx insn)
1623 int i;
1624 int before[N_REG_CLASSES];
1626 for (i = 0; i < ira_reg_class_cover_size; i++)
1627 before[i] = curr_reg_pressure[ira_reg_class_cover[i]];
1628 update_register_pressure (insn);
1629 for (i = 0; i < ira_reg_class_cover_size; i++)
1630 if (curr_reg_pressure[ira_reg_class_cover[i]] != before[i])
1631 break;
1632 if (i < ira_reg_class_cover_size)
1633 setup_insn_max_reg_pressure (insn, true);
1636 /* Set up register pressure at the beginning of basic block BB whose
1637 insns starting after insn AFTER. Set up also max register pressure
1638 for all insns of the basic block. */
1639 void
1640 sched_setup_bb_reg_pressure_info (basic_block bb, rtx after)
1642 gcc_assert (sched_pressure_p);
1643 initiate_bb_reg_pressure_info (bb);
1644 setup_insn_max_reg_pressure (after, false);
1647 /* INSN is the "currently executing insn". Launch each insn which was
1648 waiting on INSN. READY is the ready list which contains the insns
1649 that are ready to fire. CLOCK is the current cycle. The function
1650 returns necessary cycle advance after issuing the insn (it is not
1651 zero for insns in a schedule group). */
1653 static int
1654 schedule_insn (rtx insn)
1656 sd_iterator_def sd_it;
1657 dep_t dep;
1658 int i;
1659 int advance = 0;
1661 if (sched_verbose >= 1)
1663 struct reg_pressure_data *pressure_info;
1664 char buf[2048];
1666 print_insn (buf, insn, 0);
1667 buf[40] = 0;
1668 fprintf (sched_dump, ";;\t%3i--> %-40s:", clock_var, buf);
1670 if (recog_memoized (insn) < 0)
1671 fprintf (sched_dump, "nothing");
1672 else
1673 print_reservation (sched_dump, insn);
1674 pressure_info = INSN_REG_PRESSURE (insn);
1675 if (pressure_info != NULL)
1677 fputc (':', sched_dump);
1678 for (i = 0; i < ira_reg_class_cover_size; i++)
1679 fprintf (sched_dump, "%s%+d(%d)",
1680 reg_class_names[ira_reg_class_cover[i]],
1681 pressure_info[i].set_increase, pressure_info[i].change);
1683 fputc ('\n', sched_dump);
1686 if (sched_pressure_p)
1687 update_reg_and_insn_max_reg_pressure (insn);
1689 /* Scheduling instruction should have all its dependencies resolved and
1690 should have been removed from the ready list. */
1691 gcc_assert (sd_lists_empty_p (insn, SD_LIST_BACK));
1693 /* Reset debug insns invalidated by moving this insn. */
1694 if (MAY_HAVE_DEBUG_INSNS && !DEBUG_INSN_P (insn))
1695 for (sd_it = sd_iterator_start (insn, SD_LIST_BACK);
1696 sd_iterator_cond (&sd_it, &dep);)
1698 rtx dbg = DEP_PRO (dep);
1699 struct reg_use_data *use, *next;
1701 gcc_assert (DEBUG_INSN_P (dbg));
1703 if (sched_verbose >= 6)
1704 fprintf (sched_dump, ";;\t\tresetting: debug insn %d\n",
1705 INSN_UID (dbg));
1707 /* ??? Rather than resetting the debug insn, we might be able
1708 to emit a debug temp before the just-scheduled insn, but
1709 this would involve checking that the expression at the
1710 point of the debug insn is equivalent to the expression
1711 before the just-scheduled insn. They might not be: the
1712 expression in the debug insn may depend on other insns not
1713 yet scheduled that set MEMs, REGs or even other debug
1714 insns. It's not clear that attempting to preserve debug
1715 information in these cases is worth the effort, given how
1716 uncommon these resets are and the likelihood that the debug
1717 temps introduced won't survive the schedule change. */
1718 INSN_VAR_LOCATION_LOC (dbg) = gen_rtx_UNKNOWN_VAR_LOC ();
1719 df_insn_rescan (dbg);
1721 /* Unknown location doesn't use any registers. */
1722 for (use = INSN_REG_USE_LIST (dbg); use != NULL; use = next)
1724 next = use->next_insn_use;
1725 free (use);
1727 INSN_REG_USE_LIST (dbg) = NULL;
1729 /* We delete rather than resolve these deps, otherwise we
1730 crash in sched_free_deps(), because forward deps are
1731 expected to be released before backward deps. */
1732 sd_delete_dep (sd_it);
1735 gcc_assert (QUEUE_INDEX (insn) == QUEUE_NOWHERE);
1736 QUEUE_INDEX (insn) = QUEUE_SCHEDULED;
1738 gcc_assert (INSN_TICK (insn) >= MIN_TICK);
1739 if (INSN_TICK (insn) > clock_var)
1740 /* INSN has been prematurely moved from the queue to the ready list.
1741 This is possible only if following flag is set. */
1742 gcc_assert (flag_sched_stalled_insns);
1744 /* ??? Probably, if INSN is scheduled prematurely, we should leave
1745 INSN_TICK untouched. This is a machine-dependent issue, actually. */
1746 INSN_TICK (insn) = clock_var;
1748 /* Update dependent instructions. */
1749 for (sd_it = sd_iterator_start (insn, SD_LIST_FORW);
1750 sd_iterator_cond (&sd_it, &dep);)
1752 rtx next = DEP_CON (dep);
1754 /* Resolve the dependence between INSN and NEXT.
1755 sd_resolve_dep () moves current dep to another list thus
1756 advancing the iterator. */
1757 sd_resolve_dep (sd_it);
1759 /* Don't bother trying to mark next as ready if insn is a debug
1760 insn. If insn is the last hard dependency, it will have
1761 already been discounted. */
1762 if (DEBUG_INSN_P (insn) && !DEBUG_INSN_P (next))
1763 continue;
1765 if (!IS_SPECULATION_BRANCHY_CHECK_P (insn))
1767 int effective_cost;
1769 effective_cost = try_ready (next);
1771 if (effective_cost >= 0
1772 && SCHED_GROUP_P (next)
1773 && advance < effective_cost)
1774 advance = effective_cost;
1776 else
1777 /* Check always has only one forward dependence (to the first insn in
1778 the recovery block), therefore, this will be executed only once. */
1780 gcc_assert (sd_lists_empty_p (insn, SD_LIST_FORW));
1781 fix_recovery_deps (RECOVERY_BLOCK (insn));
1785 /* This is the place where scheduler doesn't *basically* need backward and
1786 forward dependencies for INSN anymore. Nevertheless they are used in
1787 heuristics in rank_for_schedule (), early_queue_to_ready () and in
1788 some targets (e.g. rs6000). Thus the earliest place where we *can*
1789 remove dependencies is after targetm.sched.md_finish () call in
1790 schedule_block (). But, on the other side, the safest place to remove
1791 dependencies is when we are finishing scheduling entire region. As we
1792 don't generate [many] dependencies during scheduling itself, we won't
1793 need memory until beginning of next region.
1794 Bottom line: Dependencies are removed for all insns in the end of
1795 scheduling the region. */
1797 /* Annotate the instruction with issue information -- TImode
1798 indicates that the instruction is expected not to be able
1799 to issue on the same cycle as the previous insn. A machine
1800 may use this information to decide how the instruction should
1801 be aligned. */
1802 if (issue_rate > 1
1803 && GET_CODE (PATTERN (insn)) != USE
1804 && GET_CODE (PATTERN (insn)) != CLOBBER
1805 && !DEBUG_INSN_P (insn))
1807 if (reload_completed)
1808 PUT_MODE (insn, clock_var > last_clock_var ? TImode : VOIDmode);
1809 last_clock_var = clock_var;
1812 return advance;
1815 /* Functions for handling of notes. */
1817 /* Add note list that ends on FROM_END to the end of TO_ENDP. */
1818 void
1819 concat_note_lists (rtx from_end, rtx *to_endp)
1821 rtx from_start;
1823 /* It's easy when have nothing to concat. */
1824 if (from_end == NULL)
1825 return;
1827 /* It's also easy when destination is empty. */
1828 if (*to_endp == NULL)
1830 *to_endp = from_end;
1831 return;
1834 from_start = from_end;
1835 while (PREV_INSN (from_start) != NULL)
1836 from_start = PREV_INSN (from_start);
1838 PREV_INSN (from_start) = *to_endp;
1839 NEXT_INSN (*to_endp) = from_start;
1840 *to_endp = from_end;
1843 /* Delete notes between HEAD and TAIL and put them in the chain
1844 of notes ended by NOTE_LIST. */
1845 void
1846 remove_notes (rtx head, rtx tail)
1848 rtx next_tail, insn, next;
1850 note_list = 0;
1851 if (head == tail && !INSN_P (head))
1852 return;
1854 next_tail = NEXT_INSN (tail);
1855 for (insn = head; insn != next_tail; insn = next)
1857 next = NEXT_INSN (insn);
1858 if (!NOTE_P (insn))
1859 continue;
1861 switch (NOTE_KIND (insn))
1863 case NOTE_INSN_BASIC_BLOCK:
1864 continue;
1866 case NOTE_INSN_EPILOGUE_BEG:
1867 if (insn != tail)
1869 remove_insn (insn);
1870 add_reg_note (next, REG_SAVE_NOTE,
1871 GEN_INT (NOTE_INSN_EPILOGUE_BEG));
1872 break;
1874 /* FALLTHRU */
1876 default:
1877 remove_insn (insn);
1879 /* Add the note to list that ends at NOTE_LIST. */
1880 PREV_INSN (insn) = note_list;
1881 NEXT_INSN (insn) = NULL_RTX;
1882 if (note_list)
1883 NEXT_INSN (note_list) = insn;
1884 note_list = insn;
1885 break;
1888 gcc_assert ((sel_sched_p () || insn != tail) && insn != head);
1893 /* Return the head and tail pointers of ebb starting at BEG and ending
1894 at END. */
1895 void
1896 get_ebb_head_tail (basic_block beg, basic_block end, rtx *headp, rtx *tailp)
1898 rtx beg_head = BB_HEAD (beg);
1899 rtx beg_tail = BB_END (beg);
1900 rtx end_head = BB_HEAD (end);
1901 rtx end_tail = BB_END (end);
1903 /* Don't include any notes or labels at the beginning of the BEG
1904 basic block, or notes at the end of the END basic blocks. */
1906 if (LABEL_P (beg_head))
1907 beg_head = NEXT_INSN (beg_head);
1909 while (beg_head != beg_tail)
1910 if (NOTE_P (beg_head) || BOUNDARY_DEBUG_INSN_P (beg_head))
1911 beg_head = NEXT_INSN (beg_head);
1912 else
1913 break;
1915 *headp = beg_head;
1917 if (beg == end)
1918 end_head = beg_head;
1919 else if (LABEL_P (end_head))
1920 end_head = NEXT_INSN (end_head);
1922 while (end_head != end_tail)
1923 if (NOTE_P (end_tail) || BOUNDARY_DEBUG_INSN_P (end_tail))
1924 end_tail = PREV_INSN (end_tail);
1925 else
1926 break;
1928 *tailp = end_tail;
1931 /* Return nonzero if there are no real insns in the range [ HEAD, TAIL ]. */
1934 no_real_insns_p (const_rtx head, const_rtx tail)
1936 while (head != NEXT_INSN (tail))
1938 if (!NOTE_P (head) && !LABEL_P (head)
1939 && !BOUNDARY_DEBUG_INSN_P (head))
1940 return 0;
1941 head = NEXT_INSN (head);
1943 return 1;
1946 /* Restore-other-notes: NOTE_LIST is the end of a chain of notes
1947 previously found among the insns. Insert them just before HEAD. */
1949 restore_other_notes (rtx head, basic_block head_bb)
1951 if (note_list != 0)
1953 rtx note_head = note_list;
1955 if (head)
1956 head_bb = BLOCK_FOR_INSN (head);
1957 else
1958 head = NEXT_INSN (bb_note (head_bb));
1960 while (PREV_INSN (note_head))
1962 set_block_for_insn (note_head, head_bb);
1963 note_head = PREV_INSN (note_head);
1965 /* In the above cycle we've missed this note. */
1966 set_block_for_insn (note_head, head_bb);
1968 PREV_INSN (note_head) = PREV_INSN (head);
1969 NEXT_INSN (PREV_INSN (head)) = note_head;
1970 PREV_INSN (head) = note_list;
1971 NEXT_INSN (note_list) = head;
1973 if (BLOCK_FOR_INSN (head) != head_bb)
1974 BB_END (head_bb) = note_list;
1976 head = note_head;
1979 return head;
1982 /* Move insns that became ready to fire from queue to ready list. */
1984 static void
1985 queue_to_ready (struct ready_list *ready)
1987 rtx insn;
1988 rtx link;
1989 rtx skip_insn;
1991 q_ptr = NEXT_Q (q_ptr);
1993 if (dbg_cnt (sched_insn) == false)
1995 /* If debug counter is activated do not requeue insn next after
1996 last_scheduled_insn. */
1997 skip_insn = next_nonnote_insn (last_scheduled_insn);
1998 while (skip_insn && DEBUG_INSN_P (skip_insn))
1999 skip_insn = next_nonnote_insn (skip_insn);
2001 else
2002 skip_insn = NULL_RTX;
2004 /* Add all pending insns that can be scheduled without stalls to the
2005 ready list. */
2006 for (link = insn_queue[q_ptr]; link; link = XEXP (link, 1))
2008 insn = XEXP (link, 0);
2009 q_size -= 1;
2011 if (sched_verbose >= 2)
2012 fprintf (sched_dump, ";;\t\tQ-->Ready: insn %s: ",
2013 (*current_sched_info->print_insn) (insn, 0));
2015 /* If the ready list is full, delay the insn for 1 cycle.
2016 See the comment in schedule_block for the rationale. */
2017 if (!reload_completed
2018 && ready->n_ready - ready->n_debug > MAX_SCHED_READY_INSNS
2019 && !SCHED_GROUP_P (insn)
2020 && insn != skip_insn)
2022 if (sched_verbose >= 2)
2023 fprintf (sched_dump, "requeued because ready full\n");
2024 queue_insn (insn, 1);
2026 else
2028 ready_add (ready, insn, false);
2029 if (sched_verbose >= 2)
2030 fprintf (sched_dump, "moving to ready without stalls\n");
2033 free_INSN_LIST_list (&insn_queue[q_ptr]);
2035 /* If there are no ready insns, stall until one is ready and add all
2036 of the pending insns at that point to the ready list. */
2037 if (ready->n_ready == 0)
2039 int stalls;
2041 for (stalls = 1; stalls <= max_insn_queue_index; stalls++)
2043 if ((link = insn_queue[NEXT_Q_AFTER (q_ptr, stalls)]))
2045 for (; link; link = XEXP (link, 1))
2047 insn = XEXP (link, 0);
2048 q_size -= 1;
2050 if (sched_verbose >= 2)
2051 fprintf (sched_dump, ";;\t\tQ-->Ready: insn %s: ",
2052 (*current_sched_info->print_insn) (insn, 0));
2054 ready_add (ready, insn, false);
2055 if (sched_verbose >= 2)
2056 fprintf (sched_dump, "moving to ready with %d stalls\n", stalls);
2058 free_INSN_LIST_list (&insn_queue[NEXT_Q_AFTER (q_ptr, stalls)]);
2060 advance_one_cycle ();
2062 break;
2065 advance_one_cycle ();
2068 q_ptr = NEXT_Q_AFTER (q_ptr, stalls);
2069 clock_var += stalls;
2073 /* Used by early_queue_to_ready. Determines whether it is "ok" to
2074 prematurely move INSN from the queue to the ready list. Currently,
2075 if a target defines the hook 'is_costly_dependence', this function
2076 uses the hook to check whether there exist any dependences which are
2077 considered costly by the target, between INSN and other insns that
2078 have already been scheduled. Dependences are checked up to Y cycles
2079 back, with default Y=1; The flag -fsched-stalled-insns-dep=Y allows
2080 controlling this value.
2081 (Other considerations could be taken into account instead (or in
2082 addition) depending on user flags and target hooks. */
2084 static bool
2085 ok_for_early_queue_removal (rtx insn)
2087 int n_cycles;
2088 rtx prev_insn = last_scheduled_insn;
2090 if (targetm.sched.is_costly_dependence)
2092 for (n_cycles = flag_sched_stalled_insns_dep; n_cycles; n_cycles--)
2094 for ( ; prev_insn; prev_insn = PREV_INSN (prev_insn))
2096 int cost;
2098 if (prev_insn == current_sched_info->prev_head)
2100 prev_insn = NULL;
2101 break;
2104 if (!NOTE_P (prev_insn))
2106 dep_t dep;
2108 dep = sd_find_dep_between (prev_insn, insn, true);
2110 if (dep != NULL)
2112 cost = dep_cost (dep);
2114 if (targetm.sched.is_costly_dependence (dep, cost,
2115 flag_sched_stalled_insns_dep - n_cycles))
2116 return false;
2120 if (GET_MODE (prev_insn) == TImode) /* end of dispatch group */
2121 break;
2124 if (!prev_insn)
2125 break;
2126 prev_insn = PREV_INSN (prev_insn);
2130 return true;
2134 /* Remove insns from the queue, before they become "ready" with respect
2135 to FU latency considerations. */
2137 static int
2138 early_queue_to_ready (state_t state, struct ready_list *ready)
2140 rtx insn;
2141 rtx link;
2142 rtx next_link;
2143 rtx prev_link;
2144 bool move_to_ready;
2145 int cost;
2146 state_t temp_state = alloca (dfa_state_size);
2147 int stalls;
2148 int insns_removed = 0;
2151 Flag '-fsched-stalled-insns=X' determines the aggressiveness of this
2152 function:
2154 X == 0: There is no limit on how many queued insns can be removed
2155 prematurely. (flag_sched_stalled_insns = -1).
2157 X >= 1: Only X queued insns can be removed prematurely in each
2158 invocation. (flag_sched_stalled_insns = X).
2160 Otherwise: Early queue removal is disabled.
2161 (flag_sched_stalled_insns = 0)
2164 if (! flag_sched_stalled_insns)
2165 return 0;
2167 for (stalls = 0; stalls <= max_insn_queue_index; stalls++)
2169 if ((link = insn_queue[NEXT_Q_AFTER (q_ptr, stalls)]))
2171 if (sched_verbose > 6)
2172 fprintf (sched_dump, ";; look at index %d + %d\n", q_ptr, stalls);
2174 prev_link = 0;
2175 while (link)
2177 next_link = XEXP (link, 1);
2178 insn = XEXP (link, 0);
2179 if (insn && sched_verbose > 6)
2180 print_rtl_single (sched_dump, insn);
2182 memcpy (temp_state, state, dfa_state_size);
2183 if (recog_memoized (insn) < 0)
2184 /* non-negative to indicate that it's not ready
2185 to avoid infinite Q->R->Q->R... */
2186 cost = 0;
2187 else
2188 cost = state_transition (temp_state, insn);
2190 if (sched_verbose >= 6)
2191 fprintf (sched_dump, "transition cost = %d\n", cost);
2193 move_to_ready = false;
2194 if (cost < 0)
2196 move_to_ready = ok_for_early_queue_removal (insn);
2197 if (move_to_ready == true)
2199 /* move from Q to R */
2200 q_size -= 1;
2201 ready_add (ready, insn, false);
2203 if (prev_link)
2204 XEXP (prev_link, 1) = next_link;
2205 else
2206 insn_queue[NEXT_Q_AFTER (q_ptr, stalls)] = next_link;
2208 free_INSN_LIST_node (link);
2210 if (sched_verbose >= 2)
2211 fprintf (sched_dump, ";;\t\tEarly Q-->Ready: insn %s\n",
2212 (*current_sched_info->print_insn) (insn, 0));
2214 insns_removed++;
2215 if (insns_removed == flag_sched_stalled_insns)
2216 /* Remove no more than flag_sched_stalled_insns insns
2217 from Q at a time. */
2218 return insns_removed;
2222 if (move_to_ready == false)
2223 prev_link = link;
2225 link = next_link;
2226 } /* while link */
2227 } /* if link */
2229 } /* for stalls.. */
2231 return insns_removed;
2235 /* Print the ready list for debugging purposes. Callable from debugger. */
2237 static void
2238 debug_ready_list (struct ready_list *ready)
2240 rtx *p;
2241 int i;
2243 if (ready->n_ready == 0)
2245 fprintf (sched_dump, "\n");
2246 return;
2249 p = ready_lastpos (ready);
2250 for (i = 0; i < ready->n_ready; i++)
2252 fprintf (sched_dump, " %s:%d",
2253 (*current_sched_info->print_insn) (p[i], 0),
2254 INSN_LUID (p[i]));
2255 if (sched_pressure_p)
2256 fprintf (sched_dump, "(cost=%d",
2257 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (p[i]));
2258 if (INSN_TICK (p[i]) > clock_var)
2259 fprintf (sched_dump, ":delay=%d", INSN_TICK (p[i]) - clock_var);
2260 if (sched_pressure_p)
2261 fprintf (sched_dump, ")");
2263 fprintf (sched_dump, "\n");
2266 /* Search INSN for REG_SAVE_NOTE notes and convert them back into insn
2267 NOTEs. This is used for NOTE_INSN_EPILOGUE_BEG, so that sched-ebb
2268 replaces the epilogue note in the correct basic block. */
2269 void
2270 reemit_notes (rtx insn)
2272 rtx note, last = insn;
2274 for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
2276 if (REG_NOTE_KIND (note) == REG_SAVE_NOTE)
2278 enum insn_note note_type = (enum insn_note) INTVAL (XEXP (note, 0));
2280 last = emit_note_before (note_type, last);
2281 remove_note (insn, note);
2286 /* Move INSN. Reemit notes if needed. Update CFG, if needed. */
2287 static void
2288 move_insn (rtx insn, rtx last, rtx nt)
2290 if (PREV_INSN (insn) != last)
2292 basic_block bb;
2293 rtx note;
2294 int jump_p = 0;
2296 bb = BLOCK_FOR_INSN (insn);
2298 /* BB_HEAD is either LABEL or NOTE. */
2299 gcc_assert (BB_HEAD (bb) != insn);
2301 if (BB_END (bb) == insn)
2302 /* If this is last instruction in BB, move end marker one
2303 instruction up. */
2305 /* Jumps are always placed at the end of basic block. */
2306 jump_p = control_flow_insn_p (insn);
2308 gcc_assert (!jump_p
2309 || ((common_sched_info->sched_pass_id == SCHED_RGN_PASS)
2310 && IS_SPECULATION_BRANCHY_CHECK_P (insn))
2311 || (common_sched_info->sched_pass_id
2312 == SCHED_EBB_PASS));
2314 gcc_assert (BLOCK_FOR_INSN (PREV_INSN (insn)) == bb);
2316 BB_END (bb) = PREV_INSN (insn);
2319 gcc_assert (BB_END (bb) != last);
2321 if (jump_p)
2322 /* We move the block note along with jump. */
2324 gcc_assert (nt);
2326 note = NEXT_INSN (insn);
2327 while (NOTE_NOT_BB_P (note) && note != nt)
2328 note = NEXT_INSN (note);
2330 if (note != nt
2331 && (LABEL_P (note)
2332 || BARRIER_P (note)))
2333 note = NEXT_INSN (note);
2335 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
2337 else
2338 note = insn;
2340 NEXT_INSN (PREV_INSN (insn)) = NEXT_INSN (note);
2341 PREV_INSN (NEXT_INSN (note)) = PREV_INSN (insn);
2343 NEXT_INSN (note) = NEXT_INSN (last);
2344 PREV_INSN (NEXT_INSN (last)) = note;
2346 NEXT_INSN (last) = insn;
2347 PREV_INSN (insn) = last;
2349 bb = BLOCK_FOR_INSN (last);
2351 if (jump_p)
2353 fix_jump_move (insn);
2355 if (BLOCK_FOR_INSN (insn) != bb)
2356 move_block_after_check (insn);
2358 gcc_assert (BB_END (bb) == last);
2361 df_insn_change_bb (insn, bb);
2363 /* Update BB_END, if needed. */
2364 if (BB_END (bb) == last)
2365 BB_END (bb) = insn;
2368 SCHED_GROUP_P (insn) = 0;
2371 /* Return true if scheduling INSN will finish current clock cycle. */
2372 static bool
2373 insn_finishes_cycle_p (rtx insn)
2375 if (SCHED_GROUP_P (insn))
2376 /* After issuing INSN, rest of the sched_group will be forced to issue
2377 in order. Don't make any plans for the rest of cycle. */
2378 return true;
2380 /* Finishing the block will, apparently, finish the cycle. */
2381 if (current_sched_info->insn_finishes_block_p
2382 && current_sched_info->insn_finishes_block_p (insn))
2383 return true;
2385 return false;
2388 /* The following structure describe an entry of the stack of choices. */
2389 struct choice_entry
2391 /* Ordinal number of the issued insn in the ready queue. */
2392 int index;
2393 /* The number of the rest insns whose issues we should try. */
2394 int rest;
2395 /* The number of issued essential insns. */
2396 int n;
2397 /* State after issuing the insn. */
2398 state_t state;
2401 /* The following array is used to implement a stack of choices used in
2402 function max_issue. */
2403 static struct choice_entry *choice_stack;
2405 /* The following variable value is number of essential insns issued on
2406 the current cycle. An insn is essential one if it changes the
2407 processors state. */
2408 int cycle_issued_insns;
2410 /* This holds the value of the target dfa_lookahead hook. */
2411 int dfa_lookahead;
2413 /* The following variable value is maximal number of tries of issuing
2414 insns for the first cycle multipass insn scheduling. We define
2415 this value as constant*(DFA_LOOKAHEAD**ISSUE_RATE). We would not
2416 need this constraint if all real insns (with non-negative codes)
2417 had reservations because in this case the algorithm complexity is
2418 O(DFA_LOOKAHEAD**ISSUE_RATE). Unfortunately, the dfa descriptions
2419 might be incomplete and such insn might occur. For such
2420 descriptions, the complexity of algorithm (without the constraint)
2421 could achieve DFA_LOOKAHEAD ** N , where N is the queue length. */
2422 static int max_lookahead_tries;
2424 /* The following value is value of hook
2425 `first_cycle_multipass_dfa_lookahead' at the last call of
2426 `max_issue'. */
2427 static int cached_first_cycle_multipass_dfa_lookahead = 0;
2429 /* The following value is value of `issue_rate' at the last call of
2430 `sched_init'. */
2431 static int cached_issue_rate = 0;
2433 /* The following function returns maximal (or close to maximal) number
2434 of insns which can be issued on the same cycle and one of which
2435 insns is insns with the best rank (the first insn in READY). To
2436 make this function tries different samples of ready insns. READY
2437 is current queue `ready'. Global array READY_TRY reflects what
2438 insns are already issued in this try. MAX_POINTS is the sum of points
2439 of all instructions in READY. The function stops immediately,
2440 if it reached the such a solution, that all instruction can be issued.
2441 INDEX will contain index of the best insn in READY. The following
2442 function is used only for first cycle multipass scheduling.
2444 PRIVILEGED_N >= 0
2446 This function expects recognized insns only. All USEs,
2447 CLOBBERs, etc must be filtered elsewhere. */
2449 max_issue (struct ready_list *ready, int privileged_n, state_t state,
2450 int *index)
2452 int n, i, all, n_ready, best, delay, tries_num, max_points;
2453 int more_issue;
2454 struct choice_entry *top;
2455 rtx insn;
2457 n_ready = ready->n_ready;
2458 gcc_assert (dfa_lookahead >= 1 && privileged_n >= 0
2459 && privileged_n <= n_ready);
2461 /* Init MAX_LOOKAHEAD_TRIES. */
2462 if (cached_first_cycle_multipass_dfa_lookahead != dfa_lookahead)
2464 cached_first_cycle_multipass_dfa_lookahead = dfa_lookahead;
2465 max_lookahead_tries = 100;
2466 for (i = 0; i < issue_rate; i++)
2467 max_lookahead_tries *= dfa_lookahead;
2470 /* Init max_points. */
2471 max_points = 0;
2472 more_issue = issue_rate - cycle_issued_insns;
2474 /* ??? We used to assert here that we never issue more insns than issue_rate.
2475 However, some targets (e.g. MIPS/SB1) claim lower issue rate than can be
2476 achieved to get better performance. Until these targets are fixed to use
2477 scheduler hooks to manipulate insns priority instead, the assert should
2478 be disabled.
2480 gcc_assert (more_issue >= 0); */
2482 for (i = 0; i < n_ready; i++)
2483 if (!ready_try [i])
2485 if (more_issue-- > 0)
2486 max_points += ISSUE_POINTS (ready_element (ready, i));
2487 else
2488 break;
2491 /* The number of the issued insns in the best solution. */
2492 best = 0;
2494 top = choice_stack;
2496 /* Set initial state of the search. */
2497 memcpy (top->state, state, dfa_state_size);
2498 top->rest = dfa_lookahead;
2499 top->n = 0;
2501 /* Count the number of the insns to search among. */
2502 for (all = i = 0; i < n_ready; i++)
2503 if (!ready_try [i])
2504 all++;
2506 /* I is the index of the insn to try next. */
2507 i = 0;
2508 tries_num = 0;
2509 for (;;)
2511 if (/* If we've reached a dead end or searched enough of what we have
2512 been asked... */
2513 top->rest == 0
2514 /* Or have nothing else to try. */
2515 || i >= n_ready)
2517 /* ??? (... || i == n_ready). */
2518 gcc_assert (i <= n_ready);
2520 if (top == choice_stack)
2521 break;
2523 if (best < top - choice_stack)
2525 if (privileged_n)
2527 n = privileged_n;
2528 /* Try to find issued privileged insn. */
2529 while (n && !ready_try[--n]);
2532 if (/* If all insns are equally good... */
2533 privileged_n == 0
2534 /* Or a privileged insn will be issued. */
2535 || ready_try[n])
2536 /* Then we have a solution. */
2538 best = top - choice_stack;
2539 /* This is the index of the insn issued first in this
2540 solution. */
2541 *index = choice_stack [1].index;
2542 if (top->n == max_points || best == all)
2543 break;
2547 /* Set ready-list index to point to the last insn
2548 ('i++' below will advance it to the next insn). */
2549 i = top->index;
2551 /* Backtrack. */
2552 ready_try [i] = 0;
2553 top--;
2554 memcpy (state, top->state, dfa_state_size);
2556 else if (!ready_try [i])
2558 tries_num++;
2559 if (tries_num > max_lookahead_tries)
2560 break;
2561 insn = ready_element (ready, i);
2562 delay = state_transition (state, insn);
2563 if (delay < 0)
2565 if (state_dead_lock_p (state)
2566 || insn_finishes_cycle_p (insn))
2567 /* We won't issue any more instructions in the next
2568 choice_state. */
2569 top->rest = 0;
2570 else
2571 top->rest--;
2573 n = top->n;
2574 if (memcmp (top->state, state, dfa_state_size) != 0)
2575 n += ISSUE_POINTS (insn);
2577 /* Advance to the next choice_entry. */
2578 top++;
2579 /* Initialize it. */
2580 top->rest = dfa_lookahead;
2581 top->index = i;
2582 top->n = n;
2583 memcpy (top->state, state, dfa_state_size);
2585 ready_try [i] = 1;
2586 i = -1;
2590 /* Increase ready-list index. */
2591 i++;
2594 /* Restore the original state of the DFA. */
2595 memcpy (state, choice_stack->state, dfa_state_size);
2597 return best;
2600 /* The following function chooses insn from READY and modifies
2601 READY. The following function is used only for first
2602 cycle multipass scheduling.
2603 Return:
2604 -1 if cycle should be advanced,
2605 0 if INSN_PTR is set to point to the desirable insn,
2606 1 if choose_ready () should be restarted without advancing the cycle. */
2607 static int
2608 choose_ready (struct ready_list *ready, rtx *insn_ptr)
2610 int lookahead;
2612 if (dbg_cnt (sched_insn) == false)
2614 rtx insn;
2616 insn = next_nonnote_insn (last_scheduled_insn);
2618 if (QUEUE_INDEX (insn) == QUEUE_READY)
2619 /* INSN is in the ready_list. */
2621 ready_remove_insn (insn);
2622 *insn_ptr = insn;
2623 return 0;
2626 /* INSN is in the queue. Advance cycle to move it to the ready list. */
2627 return -1;
2630 lookahead = 0;
2632 if (targetm.sched.first_cycle_multipass_dfa_lookahead)
2633 lookahead = targetm.sched.first_cycle_multipass_dfa_lookahead ();
2634 if (lookahead <= 0 || SCHED_GROUP_P (ready_element (ready, 0))
2635 || DEBUG_INSN_P (ready_element (ready, 0)))
2637 *insn_ptr = ready_remove_first (ready);
2638 return 0;
2640 else
2642 /* Try to choose the better insn. */
2643 int index = 0, i, n;
2644 rtx insn;
2645 int try_data = 1, try_control = 1;
2646 ds_t ts;
2648 insn = ready_element (ready, 0);
2649 if (INSN_CODE (insn) < 0)
2651 *insn_ptr = ready_remove_first (ready);
2652 return 0;
2655 if (spec_info
2656 && spec_info->flags & (PREFER_NON_DATA_SPEC
2657 | PREFER_NON_CONTROL_SPEC))
2659 for (i = 0, n = ready->n_ready; i < n; i++)
2661 rtx x;
2662 ds_t s;
2664 x = ready_element (ready, i);
2665 s = TODO_SPEC (x);
2667 if (spec_info->flags & PREFER_NON_DATA_SPEC
2668 && !(s & DATA_SPEC))
2670 try_data = 0;
2671 if (!(spec_info->flags & PREFER_NON_CONTROL_SPEC)
2672 || !try_control)
2673 break;
2676 if (spec_info->flags & PREFER_NON_CONTROL_SPEC
2677 && !(s & CONTROL_SPEC))
2679 try_control = 0;
2680 if (!(spec_info->flags & PREFER_NON_DATA_SPEC) || !try_data)
2681 break;
2686 ts = TODO_SPEC (insn);
2687 if ((ts & SPECULATIVE)
2688 && (((!try_data && (ts & DATA_SPEC))
2689 || (!try_control && (ts & CONTROL_SPEC)))
2690 || (targetm.sched.first_cycle_multipass_dfa_lookahead_guard_spec
2691 && !targetm.sched
2692 .first_cycle_multipass_dfa_lookahead_guard_spec (insn))))
2693 /* Discard speculative instruction that stands first in the ready
2694 list. */
2696 change_queue_index (insn, 1);
2697 return 1;
2700 ready_try[0] = 0;
2702 for (i = 1; i < ready->n_ready; i++)
2704 insn = ready_element (ready, i);
2706 ready_try [i]
2707 = ((!try_data && (TODO_SPEC (insn) & DATA_SPEC))
2708 || (!try_control && (TODO_SPEC (insn) & CONTROL_SPEC)));
2711 /* Let the target filter the search space. */
2712 for (i = 1; i < ready->n_ready; i++)
2713 if (!ready_try[i])
2715 insn = ready_element (ready, i);
2717 #ifdef ENABLE_CHECKING
2718 /* If this insn is recognizable we should have already
2719 recognized it earlier.
2720 ??? Not very clear where this is supposed to be done.
2721 See dep_cost_1. */
2722 gcc_assert (INSN_CODE (insn) >= 0
2723 || recog_memoized (insn) < 0);
2724 #endif
2726 ready_try [i]
2727 = (/* INSN_CODE check can be omitted here as it is also done later
2728 in max_issue (). */
2729 INSN_CODE (insn) < 0
2730 || (targetm.sched.first_cycle_multipass_dfa_lookahead_guard
2731 && !targetm.sched.first_cycle_multipass_dfa_lookahead_guard
2732 (insn)));
2735 if (max_issue (ready, 1, curr_state, &index) == 0)
2737 *insn_ptr = ready_remove_first (ready);
2738 if (sched_verbose >= 4)
2739 fprintf (sched_dump, ";;\t\tChosen insn (but can't issue) : %s \n",
2740 (*current_sched_info->print_insn) (*insn_ptr, 0));
2741 return 0;
2743 else
2745 if (sched_verbose >= 4)
2746 fprintf (sched_dump, ";;\t\tChosen insn : %s\n",
2747 (*current_sched_info->print_insn)
2748 (ready_element (ready, index), 0));
2750 *insn_ptr = ready_remove (ready, index);
2751 return 0;
2756 /* Use forward list scheduling to rearrange insns of block pointed to by
2757 TARGET_BB, possibly bringing insns from subsequent blocks in the same
2758 region. */
2760 void
2761 schedule_block (basic_block *target_bb)
2763 int i, first_cycle_insn_p;
2764 int can_issue_more;
2765 state_t temp_state = NULL; /* It is used for multipass scheduling. */
2766 int sort_p, advance, start_clock_var;
2768 /* Head/tail info for this block. */
2769 rtx prev_head = current_sched_info->prev_head;
2770 rtx next_tail = current_sched_info->next_tail;
2771 rtx head = NEXT_INSN (prev_head);
2772 rtx tail = PREV_INSN (next_tail);
2774 /* We used to have code to avoid getting parameters moved from hard
2775 argument registers into pseudos.
2777 However, it was removed when it proved to be of marginal benefit
2778 and caused problems because schedule_block and compute_forward_dependences
2779 had different notions of what the "head" insn was. */
2781 gcc_assert (head != tail || INSN_P (head));
2783 haifa_recovery_bb_recently_added_p = false;
2785 /* Debug info. */
2786 if (sched_verbose)
2787 dump_new_block_header (0, *target_bb, head, tail);
2789 state_reset (curr_state);
2791 /* Clear the ready list. */
2792 ready.first = ready.veclen - 1;
2793 ready.n_ready = 0;
2794 ready.n_debug = 0;
2796 /* It is used for first cycle multipass scheduling. */
2797 temp_state = alloca (dfa_state_size);
2799 if (targetm.sched.md_init)
2800 targetm.sched.md_init (sched_dump, sched_verbose, ready.veclen);
2802 /* We start inserting insns after PREV_HEAD. */
2803 last_scheduled_insn = prev_head;
2805 gcc_assert ((NOTE_P (last_scheduled_insn)
2806 || BOUNDARY_DEBUG_INSN_P (last_scheduled_insn))
2807 && BLOCK_FOR_INSN (last_scheduled_insn) == *target_bb);
2809 /* Initialize INSN_QUEUE. Q_SIZE is the total number of insns in the
2810 queue. */
2811 q_ptr = 0;
2812 q_size = 0;
2814 insn_queue = XALLOCAVEC (rtx, max_insn_queue_index + 1);
2815 memset (insn_queue, 0, (max_insn_queue_index + 1) * sizeof (rtx));
2817 /* Start just before the beginning of time. */
2818 clock_var = -1;
2820 /* We need queue and ready lists and clock_var be initialized
2821 in try_ready () (which is called through init_ready_list ()). */
2822 (*current_sched_info->init_ready_list) ();
2824 /* The algorithm is O(n^2) in the number of ready insns at any given
2825 time in the worst case. Before reload we are more likely to have
2826 big lists so truncate them to a reasonable size. */
2827 if (!reload_completed
2828 && ready.n_ready - ready.n_debug > MAX_SCHED_READY_INSNS)
2830 ready_sort (&ready);
2832 /* Find first free-standing insn past MAX_SCHED_READY_INSNS.
2833 If there are debug insns, we know they're first. */
2834 for (i = MAX_SCHED_READY_INSNS + ready.n_debug; i < ready.n_ready; i++)
2835 if (!SCHED_GROUP_P (ready_element (&ready, i)))
2836 break;
2838 if (sched_verbose >= 2)
2840 fprintf (sched_dump,
2841 ";;\t\tReady list on entry: %d insns\n", ready.n_ready);
2842 fprintf (sched_dump,
2843 ";;\t\t before reload => truncated to %d insns\n", i);
2846 /* Delay all insns past it for 1 cycle. If debug counter is
2847 activated make an exception for the insn right after
2848 last_scheduled_insn. */
2850 rtx skip_insn;
2852 if (dbg_cnt (sched_insn) == false)
2853 skip_insn = next_nonnote_insn (last_scheduled_insn);
2854 else
2855 skip_insn = NULL_RTX;
2857 while (i < ready.n_ready)
2859 rtx insn;
2861 insn = ready_remove (&ready, i);
2863 if (insn != skip_insn)
2864 queue_insn (insn, 1);
2869 /* Now we can restore basic block notes and maintain precise cfg. */
2870 restore_bb_notes (*target_bb);
2872 last_clock_var = -1;
2874 advance = 0;
2876 sort_p = TRUE;
2877 /* Loop until all the insns in BB are scheduled. */
2878 while ((*current_sched_info->schedule_more_p) ())
2882 start_clock_var = clock_var;
2884 clock_var++;
2886 advance_one_cycle ();
2888 /* Add to the ready list all pending insns that can be issued now.
2889 If there are no ready insns, increment clock until one
2890 is ready and add all pending insns at that point to the ready
2891 list. */
2892 queue_to_ready (&ready);
2894 gcc_assert (ready.n_ready);
2896 if (sched_verbose >= 2)
2898 fprintf (sched_dump, ";;\t\tReady list after queue_to_ready: ");
2899 debug_ready_list (&ready);
2901 advance -= clock_var - start_clock_var;
2903 while (advance > 0);
2905 if (sort_p)
2907 /* Sort the ready list based on priority. */
2908 ready_sort (&ready);
2910 if (sched_verbose >= 2)
2912 fprintf (sched_dump, ";;\t\tReady list after ready_sort: ");
2913 debug_ready_list (&ready);
2917 /* We don't want md sched reorder to even see debug isns, so put
2918 them out right away. */
2919 if (ready.n_ready && DEBUG_INSN_P (ready_element (&ready, 0)))
2921 if (control_flow_insn_p (last_scheduled_insn))
2923 *target_bb = current_sched_info->advance_target_bb
2924 (*target_bb, 0);
2926 if (sched_verbose)
2928 rtx x;
2930 x = next_real_insn (last_scheduled_insn);
2931 gcc_assert (x);
2932 dump_new_block_header (1, *target_bb, x, tail);
2935 last_scheduled_insn = bb_note (*target_bb);
2938 while (ready.n_ready && DEBUG_INSN_P (ready_element (&ready, 0)))
2940 rtx insn = ready_remove_first (&ready);
2941 gcc_assert (DEBUG_INSN_P (insn));
2942 (*current_sched_info->begin_schedule_ready) (insn,
2943 last_scheduled_insn);
2944 move_insn (insn, last_scheduled_insn,
2945 current_sched_info->next_tail);
2946 last_scheduled_insn = insn;
2947 advance = schedule_insn (insn);
2948 gcc_assert (advance == 0);
2949 if (ready.n_ready > 0)
2950 ready_sort (&ready);
2953 if (!ready.n_ready)
2954 continue;
2957 /* Allow the target to reorder the list, typically for
2958 better instruction bundling. */
2959 if (sort_p && targetm.sched.reorder
2960 && (ready.n_ready == 0
2961 || !SCHED_GROUP_P (ready_element (&ready, 0))))
2962 can_issue_more =
2963 targetm.sched.reorder (sched_dump, sched_verbose,
2964 ready_lastpos (&ready),
2965 &ready.n_ready, clock_var);
2966 else
2967 can_issue_more = issue_rate;
2969 first_cycle_insn_p = 1;
2970 cycle_issued_insns = 0;
2971 for (;;)
2973 rtx insn;
2974 int cost;
2975 bool asm_p = false;
2977 if (sched_verbose >= 2)
2979 fprintf (sched_dump, ";;\tReady list (t = %3d): ",
2980 clock_var);
2981 debug_ready_list (&ready);
2982 if (sched_pressure_p)
2983 print_curr_reg_pressure ();
2986 if (ready.n_ready == 0
2987 && can_issue_more
2988 && reload_completed)
2990 /* Allow scheduling insns directly from the queue in case
2991 there's nothing better to do (ready list is empty) but
2992 there are still vacant dispatch slots in the current cycle. */
2993 if (sched_verbose >= 6)
2994 fprintf (sched_dump,";;\t\tSecond chance\n");
2995 memcpy (temp_state, curr_state, dfa_state_size);
2996 if (early_queue_to_ready (temp_state, &ready))
2997 ready_sort (&ready);
3000 if (ready.n_ready == 0
3001 || !can_issue_more
3002 || state_dead_lock_p (curr_state)
3003 || !(*current_sched_info->schedule_more_p) ())
3004 break;
3006 /* Select and remove the insn from the ready list. */
3007 if (sort_p)
3009 int res;
3011 insn = NULL_RTX;
3012 res = choose_ready (&ready, &insn);
3014 if (res < 0)
3015 /* Finish cycle. */
3016 break;
3017 if (res > 0)
3018 /* Restart choose_ready (). */
3019 continue;
3021 gcc_assert (insn != NULL_RTX);
3023 else
3024 insn = ready_remove_first (&ready);
3026 if (sched_pressure_p && INSN_TICK (insn) > clock_var)
3028 ready_add (&ready, insn, true);
3029 advance = 1;
3030 break;
3033 if (targetm.sched.dfa_new_cycle
3034 && targetm.sched.dfa_new_cycle (sched_dump, sched_verbose,
3035 insn, last_clock_var,
3036 clock_var, &sort_p))
3037 /* SORT_P is used by the target to override sorting
3038 of the ready list. This is needed when the target
3039 has modified its internal structures expecting that
3040 the insn will be issued next. As we need the insn
3041 to have the highest priority (so it will be returned by
3042 the ready_remove_first call above), we invoke
3043 ready_add (&ready, insn, true).
3044 But, still, there is one issue: INSN can be later
3045 discarded by scheduler's front end through
3046 current_sched_info->can_schedule_ready_p, hence, won't
3047 be issued next. */
3049 ready_add (&ready, insn, true);
3050 break;
3053 sort_p = TRUE;
3054 memcpy (temp_state, curr_state, dfa_state_size);
3055 if (recog_memoized (insn) < 0)
3057 asm_p = (GET_CODE (PATTERN (insn)) == ASM_INPUT
3058 || asm_noperands (PATTERN (insn)) >= 0);
3059 if (!first_cycle_insn_p && asm_p)
3060 /* This is asm insn which is tried to be issued on the
3061 cycle not first. Issue it on the next cycle. */
3062 cost = 1;
3063 else
3064 /* A USE insn, or something else we don't need to
3065 understand. We can't pass these directly to
3066 state_transition because it will trigger a
3067 fatal error for unrecognizable insns. */
3068 cost = 0;
3070 else if (sched_pressure_p)
3071 cost = 0;
3072 else
3074 cost = state_transition (temp_state, insn);
3075 if (cost < 0)
3076 cost = 0;
3077 else if (cost == 0)
3078 cost = 1;
3081 if (cost >= 1)
3083 queue_insn (insn, cost);
3084 if (SCHED_GROUP_P (insn))
3086 advance = cost;
3087 break;
3090 continue;
3093 if (current_sched_info->can_schedule_ready_p
3094 && ! (*current_sched_info->can_schedule_ready_p) (insn))
3095 /* We normally get here only if we don't want to move
3096 insn from the split block. */
3098 TODO_SPEC (insn) = (TODO_SPEC (insn) & ~SPECULATIVE) | HARD_DEP;
3099 continue;
3102 /* DECISION is made. */
3104 if (TODO_SPEC (insn) & SPECULATIVE)
3105 generate_recovery_code (insn);
3107 if (control_flow_insn_p (last_scheduled_insn)
3108 /* This is used to switch basic blocks by request
3109 from scheduler front-end (actually, sched-ebb.c only).
3110 This is used to process blocks with single fallthru
3111 edge. If succeeding block has jump, it [jump] will try
3112 move at the end of current bb, thus corrupting CFG. */
3113 || current_sched_info->advance_target_bb (*target_bb, insn))
3115 *target_bb = current_sched_info->advance_target_bb
3116 (*target_bb, 0);
3118 if (sched_verbose)
3120 rtx x;
3122 x = next_real_insn (last_scheduled_insn);
3123 gcc_assert (x);
3124 dump_new_block_header (1, *target_bb, x, tail);
3127 last_scheduled_insn = bb_note (*target_bb);
3130 /* Update counters, etc in the scheduler's front end. */
3131 (*current_sched_info->begin_schedule_ready) (insn,
3132 last_scheduled_insn);
3134 move_insn (insn, last_scheduled_insn, current_sched_info->next_tail);
3135 reemit_notes (insn);
3136 last_scheduled_insn = insn;
3138 if (memcmp (curr_state, temp_state, dfa_state_size) != 0)
3140 cycle_issued_insns++;
3141 memcpy (curr_state, temp_state, dfa_state_size);
3144 if (targetm.sched.variable_issue)
3145 can_issue_more =
3146 targetm.sched.variable_issue (sched_dump, sched_verbose,
3147 insn, can_issue_more);
3148 /* A naked CLOBBER or USE generates no instruction, so do
3149 not count them against the issue rate. */
3150 else if (GET_CODE (PATTERN (insn)) != USE
3151 && GET_CODE (PATTERN (insn)) != CLOBBER)
3152 can_issue_more--;
3153 advance = schedule_insn (insn);
3155 /* After issuing an asm insn we should start a new cycle. */
3156 if (advance == 0 && asm_p)
3157 advance = 1;
3158 if (advance != 0)
3159 break;
3161 first_cycle_insn_p = 0;
3163 /* Sort the ready list based on priority. This must be
3164 redone here, as schedule_insn may have readied additional
3165 insns that will not be sorted correctly. */
3166 if (ready.n_ready > 0)
3167 ready_sort (&ready);
3169 /* Quickly go through debug insns such that md sched
3170 reorder2 doesn't have to deal with debug insns. */
3171 if (ready.n_ready && DEBUG_INSN_P (ready_element (&ready, 0))
3172 && (*current_sched_info->schedule_more_p) ())
3174 if (control_flow_insn_p (last_scheduled_insn))
3176 *target_bb = current_sched_info->advance_target_bb
3177 (*target_bb, 0);
3179 if (sched_verbose)
3181 rtx x;
3183 x = next_real_insn (last_scheduled_insn);
3184 gcc_assert (x);
3185 dump_new_block_header (1, *target_bb, x, tail);
3188 last_scheduled_insn = bb_note (*target_bb);
3191 while (ready.n_ready && DEBUG_INSN_P (ready_element (&ready, 0)))
3193 insn = ready_remove_first (&ready);
3194 gcc_assert (DEBUG_INSN_P (insn));
3195 (*current_sched_info->begin_schedule_ready)
3196 (insn, last_scheduled_insn);
3197 move_insn (insn, last_scheduled_insn,
3198 current_sched_info->next_tail);
3199 advance = schedule_insn (insn);
3200 last_scheduled_insn = insn;
3201 gcc_assert (advance == 0);
3202 if (ready.n_ready > 0)
3203 ready_sort (&ready);
3207 if (targetm.sched.reorder2
3208 && (ready.n_ready == 0
3209 || !SCHED_GROUP_P (ready_element (&ready, 0))))
3211 can_issue_more =
3212 targetm.sched.reorder2 (sched_dump, sched_verbose,
3213 ready.n_ready
3214 ? ready_lastpos (&ready) : NULL,
3215 &ready.n_ready, clock_var);
3220 /* Debug info. */
3221 if (sched_verbose)
3223 fprintf (sched_dump, ";;\tReady list (final): ");
3224 debug_ready_list (&ready);
3227 if (current_sched_info->queue_must_finish_empty)
3228 /* Sanity check -- queue must be empty now. Meaningless if region has
3229 multiple bbs. */
3230 gcc_assert (!q_size && !ready.n_ready && !ready.n_debug);
3231 else
3233 /* We must maintain QUEUE_INDEX between blocks in region. */
3234 for (i = ready.n_ready - 1; i >= 0; i--)
3236 rtx x;
3238 x = ready_element (&ready, i);
3239 QUEUE_INDEX (x) = QUEUE_NOWHERE;
3240 TODO_SPEC (x) = (TODO_SPEC (x) & ~SPECULATIVE) | HARD_DEP;
3243 if (q_size)
3244 for (i = 0; i <= max_insn_queue_index; i++)
3246 rtx link;
3247 for (link = insn_queue[i]; link; link = XEXP (link, 1))
3249 rtx x;
3251 x = XEXP (link, 0);
3252 QUEUE_INDEX (x) = QUEUE_NOWHERE;
3253 TODO_SPEC (x) = (TODO_SPEC (x) & ~SPECULATIVE) | HARD_DEP;
3255 free_INSN_LIST_list (&insn_queue[i]);
3259 if (sched_verbose)
3260 fprintf (sched_dump, ";; total time = %d\n", clock_var);
3262 if (!current_sched_info->queue_must_finish_empty
3263 || haifa_recovery_bb_recently_added_p)
3265 /* INSN_TICK (minimum clock tick at which the insn becomes
3266 ready) may be not correct for the insn in the subsequent
3267 blocks of the region. We should use a correct value of
3268 `clock_var' or modify INSN_TICK. It is better to keep
3269 clock_var value equal to 0 at the start of a basic block.
3270 Therefore we modify INSN_TICK here. */
3271 fix_inter_tick (NEXT_INSN (prev_head), last_scheduled_insn);
3274 if (targetm.sched.md_finish)
3276 targetm.sched.md_finish (sched_dump, sched_verbose);
3277 /* Target might have added some instructions to the scheduled block
3278 in its md_finish () hook. These new insns don't have any data
3279 initialized and to identify them we extend h_i_d so that they'll
3280 get zero luids. */
3281 sched_init_luids (NULL, NULL, NULL, NULL);
3284 if (sched_verbose)
3285 fprintf (sched_dump, ";; new head = %d\n;; new tail = %d\n\n",
3286 INSN_UID (head), INSN_UID (tail));
3288 /* Update head/tail boundaries. */
3289 head = NEXT_INSN (prev_head);
3290 tail = last_scheduled_insn;
3292 head = restore_other_notes (head, NULL);
3294 current_sched_info->head = head;
3295 current_sched_info->tail = tail;
3298 /* Set_priorities: compute priority of each insn in the block. */
3301 set_priorities (rtx head, rtx tail)
3303 rtx insn;
3304 int n_insn;
3305 int sched_max_insns_priority =
3306 current_sched_info->sched_max_insns_priority;
3307 rtx prev_head;
3309 if (head == tail && (! INSN_P (head) || BOUNDARY_DEBUG_INSN_P (head)))
3310 gcc_unreachable ();
3312 n_insn = 0;
3314 prev_head = PREV_INSN (head);
3315 for (insn = tail; insn != prev_head; insn = PREV_INSN (insn))
3317 if (!INSN_P (insn))
3318 continue;
3320 n_insn++;
3321 (void) priority (insn);
3323 gcc_assert (INSN_PRIORITY_KNOWN (insn));
3325 sched_max_insns_priority = MAX (sched_max_insns_priority,
3326 INSN_PRIORITY (insn));
3329 current_sched_info->sched_max_insns_priority = sched_max_insns_priority;
3331 return n_insn;
3334 /* Set dump and sched_verbose for the desired debugging output. If no
3335 dump-file was specified, but -fsched-verbose=N (any N), print to stderr.
3336 For -fsched-verbose=N, N>=10, print everything to stderr. */
3337 void
3338 setup_sched_dump (void)
3340 sched_verbose = sched_verbose_param;
3341 if (sched_verbose_param == 0 && dump_file)
3342 sched_verbose = 1;
3343 sched_dump = ((sched_verbose_param >= 10 || !dump_file)
3344 ? stderr : dump_file);
3347 /* Initialize some global state for the scheduler. This function works
3348 with the common data shared between all the schedulers. It is called
3349 from the scheduler specific initialization routine. */
3351 void
3352 sched_init (void)
3354 /* Disable speculative loads in their presence if cc0 defined. */
3355 #ifdef HAVE_cc0
3356 flag_schedule_speculative_load = 0;
3357 #endif
3359 sched_pressure_p = (flag_sched_pressure && ! reload_completed
3360 && common_sched_info->sched_pass_id == SCHED_RGN_PASS);
3361 if (sched_pressure_p)
3362 ira_setup_eliminable_regset ();
3364 /* Initialize SPEC_INFO. */
3365 if (targetm.sched.set_sched_flags)
3367 spec_info = &spec_info_var;
3368 targetm.sched.set_sched_flags (spec_info);
3370 if (spec_info->mask != 0)
3372 spec_info->data_weakness_cutoff =
3373 (PARAM_VALUE (PARAM_SCHED_SPEC_PROB_CUTOFF) * MAX_DEP_WEAK) / 100;
3374 spec_info->control_weakness_cutoff =
3375 (PARAM_VALUE (PARAM_SCHED_SPEC_PROB_CUTOFF)
3376 * REG_BR_PROB_BASE) / 100;
3378 else
3379 /* So we won't read anything accidentally. */
3380 spec_info = NULL;
3383 else
3384 /* So we won't read anything accidentally. */
3385 spec_info = 0;
3387 /* Initialize issue_rate. */
3388 if (targetm.sched.issue_rate)
3389 issue_rate = targetm.sched.issue_rate ();
3390 else
3391 issue_rate = 1;
3393 if (cached_issue_rate != issue_rate)
3395 cached_issue_rate = issue_rate;
3396 /* To invalidate max_lookahead_tries: */
3397 cached_first_cycle_multipass_dfa_lookahead = 0;
3400 if (targetm.sched.first_cycle_multipass_dfa_lookahead)
3401 dfa_lookahead = targetm.sched.first_cycle_multipass_dfa_lookahead ();
3402 else
3403 dfa_lookahead = 0;
3405 if (targetm.sched.init_dfa_pre_cycle_insn)
3406 targetm.sched.init_dfa_pre_cycle_insn ();
3408 if (targetm.sched.init_dfa_post_cycle_insn)
3409 targetm.sched.init_dfa_post_cycle_insn ();
3411 dfa_start ();
3412 dfa_state_size = state_size ();
3414 init_alias_analysis ();
3416 df_set_flags (DF_LR_RUN_DCE);
3417 df_note_add_problem ();
3419 /* More problems needed for interloop dep calculation in SMS. */
3420 if (common_sched_info->sched_pass_id == SCHED_SMS_PASS)
3422 df_rd_add_problem ();
3423 df_chain_add_problem (DF_DU_CHAIN + DF_UD_CHAIN);
3426 df_analyze ();
3428 /* Do not run DCE after reload, as this can kill nops inserted
3429 by bundling. */
3430 if (reload_completed)
3431 df_clear_flags (DF_LR_RUN_DCE);
3433 regstat_compute_calls_crossed ();
3435 if (targetm.sched.md_init_global)
3436 targetm.sched.md_init_global (sched_dump, sched_verbose,
3437 get_max_uid () + 1);
3439 if (sched_pressure_p)
3441 int i, max_regno = max_reg_num ();
3443 ira_set_pseudo_classes (sched_verbose ? sched_dump : NULL);
3444 sched_regno_cover_class
3445 = (enum reg_class *) xmalloc (max_regno * sizeof (enum reg_class));
3446 for (i = 0; i < max_regno; i++)
3447 sched_regno_cover_class[i]
3448 = (i < FIRST_PSEUDO_REGISTER
3449 ? ira_class_translate[REGNO_REG_CLASS (i)]
3450 : reg_cover_class (i));
3451 curr_reg_live = BITMAP_ALLOC (NULL);
3452 saved_reg_live = BITMAP_ALLOC (NULL);
3453 region_ref_regs = BITMAP_ALLOC (NULL);
3456 curr_state = xmalloc (dfa_state_size);
3459 static void haifa_init_only_bb (basic_block, basic_block);
3461 /* Initialize data structures specific to the Haifa scheduler. */
3462 void
3463 haifa_sched_init (void)
3465 setup_sched_dump ();
3466 sched_init ();
3468 if (spec_info != NULL)
3470 sched_deps_info->use_deps_list = 1;
3471 sched_deps_info->generate_spec_deps = 1;
3474 /* Initialize luids, dependency caches, target and h_i_d for the
3475 whole function. */
3477 bb_vec_t bbs = VEC_alloc (basic_block, heap, n_basic_blocks);
3478 basic_block bb;
3480 sched_init_bbs ();
3482 FOR_EACH_BB (bb)
3483 VEC_quick_push (basic_block, bbs, bb);
3484 sched_init_luids (bbs, NULL, NULL, NULL);
3485 sched_deps_init (true);
3486 sched_extend_target ();
3487 haifa_init_h_i_d (bbs, NULL, NULL, NULL);
3489 VEC_free (basic_block, heap, bbs);
3492 sched_init_only_bb = haifa_init_only_bb;
3493 sched_split_block = sched_split_block_1;
3494 sched_create_empty_bb = sched_create_empty_bb_1;
3495 haifa_recovery_bb_ever_added_p = false;
3497 #ifdef ENABLE_CHECKING
3498 /* This is used preferably for finding bugs in check_cfg () itself.
3499 We must call sched_bbs_init () before check_cfg () because check_cfg ()
3500 assumes that the last insn in the last bb has a non-null successor. */
3501 check_cfg (0, 0);
3502 #endif
3504 nr_begin_data = nr_begin_control = nr_be_in_data = nr_be_in_control = 0;
3505 before_recovery = 0;
3506 after_recovery = 0;
3509 /* Finish work with the data specific to the Haifa scheduler. */
3510 void
3511 haifa_sched_finish (void)
3513 sched_create_empty_bb = NULL;
3514 sched_split_block = NULL;
3515 sched_init_only_bb = NULL;
3517 if (spec_info && spec_info->dump)
3519 char c = reload_completed ? 'a' : 'b';
3521 fprintf (spec_info->dump,
3522 ";; %s:\n", current_function_name ());
3524 fprintf (spec_info->dump,
3525 ";; Procedure %cr-begin-data-spec motions == %d\n",
3526 c, nr_begin_data);
3527 fprintf (spec_info->dump,
3528 ";; Procedure %cr-be-in-data-spec motions == %d\n",
3529 c, nr_be_in_data);
3530 fprintf (spec_info->dump,
3531 ";; Procedure %cr-begin-control-spec motions == %d\n",
3532 c, nr_begin_control);
3533 fprintf (spec_info->dump,
3534 ";; Procedure %cr-be-in-control-spec motions == %d\n",
3535 c, nr_be_in_control);
3538 /* Finalize h_i_d, dependency caches, and luids for the whole
3539 function. Target will be finalized in md_global_finish (). */
3540 sched_deps_finish ();
3541 sched_finish_luids ();
3542 current_sched_info = NULL;
3543 sched_finish ();
3546 /* Free global data used during insn scheduling. This function works with
3547 the common data shared between the schedulers. */
3549 void
3550 sched_finish (void)
3552 haifa_finish_h_i_d ();
3553 if (sched_pressure_p)
3555 free (sched_regno_cover_class);
3556 BITMAP_FREE (region_ref_regs);
3557 BITMAP_FREE (saved_reg_live);
3558 BITMAP_FREE (curr_reg_live);
3560 free (curr_state);
3562 if (targetm.sched.md_finish_global)
3563 targetm.sched.md_finish_global (sched_dump, sched_verbose);
3565 end_alias_analysis ();
3567 regstat_free_calls_crossed ();
3569 dfa_finish ();
3571 #ifdef ENABLE_CHECKING
3572 /* After reload ia64 backend clobbers CFG, so can't check anything. */
3573 if (!reload_completed)
3574 check_cfg (0, 0);
3575 #endif
3578 /* Fix INSN_TICKs of the instructions in the current block as well as
3579 INSN_TICKs of their dependents.
3580 HEAD and TAIL are the begin and the end of the current scheduled block. */
3581 static void
3582 fix_inter_tick (rtx head, rtx tail)
3584 /* Set of instructions with corrected INSN_TICK. */
3585 bitmap_head processed;
3586 /* ??? It is doubtful if we should assume that cycle advance happens on
3587 basic block boundaries. Basically insns that are unconditionally ready
3588 on the start of the block are more preferable then those which have
3589 a one cycle dependency over insn from the previous block. */
3590 int next_clock = clock_var + 1;
3592 bitmap_initialize (&processed, 0);
3594 /* Iterates over scheduled instructions and fix their INSN_TICKs and
3595 INSN_TICKs of dependent instructions, so that INSN_TICKs are consistent
3596 across different blocks. */
3597 for (tail = NEXT_INSN (tail); head != tail; head = NEXT_INSN (head))
3599 if (INSN_P (head))
3601 int tick;
3602 sd_iterator_def sd_it;
3603 dep_t dep;
3605 tick = INSN_TICK (head);
3606 gcc_assert (tick >= MIN_TICK);
3608 /* Fix INSN_TICK of instruction from just scheduled block. */
3609 if (!bitmap_bit_p (&processed, INSN_LUID (head)))
3611 bitmap_set_bit (&processed, INSN_LUID (head));
3612 tick -= next_clock;
3614 if (tick < MIN_TICK)
3615 tick = MIN_TICK;
3617 INSN_TICK (head) = tick;
3620 FOR_EACH_DEP (head, SD_LIST_RES_FORW, sd_it, dep)
3622 rtx next;
3624 next = DEP_CON (dep);
3625 tick = INSN_TICK (next);
3627 if (tick != INVALID_TICK
3628 /* If NEXT has its INSN_TICK calculated, fix it.
3629 If not - it will be properly calculated from
3630 scratch later in fix_tick_ready. */
3631 && !bitmap_bit_p (&processed, INSN_LUID (next)))
3633 bitmap_set_bit (&processed, INSN_LUID (next));
3634 tick -= next_clock;
3636 if (tick < MIN_TICK)
3637 tick = MIN_TICK;
3639 if (tick > INTER_TICK (next))
3640 INTER_TICK (next) = tick;
3641 else
3642 tick = INTER_TICK (next);
3644 INSN_TICK (next) = tick;
3649 bitmap_clear (&processed);
3652 static int haifa_speculate_insn (rtx, ds_t, rtx *);
3654 /* Check if NEXT is ready to be added to the ready or queue list.
3655 If "yes", add it to the proper list.
3656 Returns:
3657 -1 - is not ready yet,
3658 0 - added to the ready list,
3659 0 < N - queued for N cycles. */
3661 try_ready (rtx next)
3663 ds_t old_ts, *ts;
3665 ts = &TODO_SPEC (next);
3666 old_ts = *ts;
3668 gcc_assert (!(old_ts & ~(SPECULATIVE | HARD_DEP))
3669 && ((old_ts & HARD_DEP)
3670 || (old_ts & SPECULATIVE)));
3672 if (sd_lists_empty_p (next, SD_LIST_BACK))
3673 /* NEXT has all its dependencies resolved. */
3675 /* Remove HARD_DEP bit from NEXT's status. */
3676 *ts &= ~HARD_DEP;
3678 if (current_sched_info->flags & DO_SPECULATION)
3679 /* Remove all speculative bits from NEXT's status. */
3680 *ts &= ~SPECULATIVE;
3682 else
3684 /* One of the NEXT's dependencies has been resolved.
3685 Recalculate NEXT's status. */
3687 *ts &= ~SPECULATIVE & ~HARD_DEP;
3689 if (sd_lists_empty_p (next, SD_LIST_HARD_BACK))
3690 /* Now we've got NEXT with speculative deps only.
3691 1. Look at the deps to see what we have to do.
3692 2. Check if we can do 'todo'. */
3694 sd_iterator_def sd_it;
3695 dep_t dep;
3696 bool first_p = true;
3698 FOR_EACH_DEP (next, SD_LIST_BACK, sd_it, dep)
3700 ds_t ds = DEP_STATUS (dep) & SPECULATIVE;
3702 if (DEBUG_INSN_P (DEP_PRO (dep))
3703 && !DEBUG_INSN_P (next))
3704 continue;
3706 if (first_p)
3708 first_p = false;
3710 *ts = ds;
3712 else
3713 *ts = ds_merge (*ts, ds);
3716 if (ds_weak (*ts) < spec_info->data_weakness_cutoff)
3717 /* Too few points. */
3718 *ts = (*ts & ~SPECULATIVE) | HARD_DEP;
3720 else
3721 *ts |= HARD_DEP;
3724 if (*ts & HARD_DEP)
3725 gcc_assert (*ts == old_ts
3726 && QUEUE_INDEX (next) == QUEUE_NOWHERE);
3727 else if (current_sched_info->new_ready)
3728 *ts = current_sched_info->new_ready (next, *ts);
3730 /* * if !(old_ts & SPECULATIVE) (e.g. HARD_DEP or 0), then insn might
3731 have its original pattern or changed (speculative) one. This is due
3732 to changing ebb in region scheduling.
3733 * But if (old_ts & SPECULATIVE), then we are pretty sure that insn
3734 has speculative pattern.
3736 We can't assert (!(*ts & HARD_DEP) || *ts == old_ts) here because
3737 control-speculative NEXT could have been discarded by sched-rgn.c
3738 (the same case as when discarded by can_schedule_ready_p ()). */
3740 if ((*ts & SPECULATIVE)
3741 /* If (old_ts == *ts), then (old_ts & SPECULATIVE) and we don't
3742 need to change anything. */
3743 && *ts != old_ts)
3745 int res;
3746 rtx new_pat;
3748 gcc_assert ((*ts & SPECULATIVE) && !(*ts & ~SPECULATIVE));
3750 res = haifa_speculate_insn (next, *ts, &new_pat);
3752 switch (res)
3754 case -1:
3755 /* It would be nice to change DEP_STATUS of all dependences,
3756 which have ((DEP_STATUS & SPECULATIVE) == *ts) to HARD_DEP,
3757 so we won't reanalyze anything. */
3758 *ts = (*ts & ~SPECULATIVE) | HARD_DEP;
3759 break;
3761 case 0:
3762 /* We follow the rule, that every speculative insn
3763 has non-null ORIG_PAT. */
3764 if (!ORIG_PAT (next))
3765 ORIG_PAT (next) = PATTERN (next);
3766 break;
3768 case 1:
3769 if (!ORIG_PAT (next))
3770 /* If we gonna to overwrite the original pattern of insn,
3771 save it. */
3772 ORIG_PAT (next) = PATTERN (next);
3774 haifa_change_pattern (next, new_pat);
3775 break;
3777 default:
3778 gcc_unreachable ();
3782 /* We need to restore pattern only if (*ts == 0), because otherwise it is
3783 either correct (*ts & SPECULATIVE),
3784 or we simply don't care (*ts & HARD_DEP). */
3786 gcc_assert (!ORIG_PAT (next)
3787 || !IS_SPECULATION_BRANCHY_CHECK_P (next));
3789 if (*ts & HARD_DEP)
3791 /* We can't assert (QUEUE_INDEX (next) == QUEUE_NOWHERE) here because
3792 control-speculative NEXT could have been discarded by sched-rgn.c
3793 (the same case as when discarded by can_schedule_ready_p ()). */
3794 /*gcc_assert (QUEUE_INDEX (next) == QUEUE_NOWHERE);*/
3796 change_queue_index (next, QUEUE_NOWHERE);
3797 return -1;
3799 else if (!(*ts & BEGIN_SPEC) && ORIG_PAT (next) && !IS_SPECULATION_CHECK_P (next))
3800 /* We should change pattern of every previously speculative
3801 instruction - and we determine if NEXT was speculative by using
3802 ORIG_PAT field. Except one case - speculation checks have ORIG_PAT
3803 pat too, so skip them. */
3805 haifa_change_pattern (next, ORIG_PAT (next));
3806 ORIG_PAT (next) = 0;
3809 if (sched_verbose >= 2)
3811 int s = TODO_SPEC (next);
3813 fprintf (sched_dump, ";;\t\tdependencies resolved: insn %s",
3814 (*current_sched_info->print_insn) (next, 0));
3816 if (spec_info && spec_info->dump)
3818 if (s & BEGIN_DATA)
3819 fprintf (spec_info->dump, "; data-spec;");
3820 if (s & BEGIN_CONTROL)
3821 fprintf (spec_info->dump, "; control-spec;");
3822 if (s & BE_IN_CONTROL)
3823 fprintf (spec_info->dump, "; in-control-spec;");
3826 fprintf (sched_dump, "\n");
3829 adjust_priority (next);
3831 return fix_tick_ready (next);
3834 /* Calculate INSN_TICK of NEXT and add it to either ready or queue list. */
3835 static int
3836 fix_tick_ready (rtx next)
3838 int tick, delay;
3840 if (!sd_lists_empty_p (next, SD_LIST_RES_BACK))
3842 int full_p;
3843 sd_iterator_def sd_it;
3844 dep_t dep;
3846 tick = INSN_TICK (next);
3847 /* if tick is not equal to INVALID_TICK, then update
3848 INSN_TICK of NEXT with the most recent resolved dependence
3849 cost. Otherwise, recalculate from scratch. */
3850 full_p = (tick == INVALID_TICK);
3852 FOR_EACH_DEP (next, SD_LIST_RES_BACK, sd_it, dep)
3854 rtx pro = DEP_PRO (dep);
3855 int tick1;
3857 gcc_assert (INSN_TICK (pro) >= MIN_TICK);
3859 tick1 = INSN_TICK (pro) + dep_cost (dep);
3860 if (tick1 > tick)
3861 tick = tick1;
3863 if (!full_p)
3864 break;
3867 else
3868 tick = -1;
3870 INSN_TICK (next) = tick;
3872 delay = tick - clock_var;
3873 if (delay <= 0 || sched_pressure_p)
3874 delay = QUEUE_READY;
3876 change_queue_index (next, delay);
3878 return delay;
3881 /* Move NEXT to the proper queue list with (DELAY >= 1),
3882 or add it to the ready list (DELAY == QUEUE_READY),
3883 or remove it from ready and queue lists at all (DELAY == QUEUE_NOWHERE). */
3884 static void
3885 change_queue_index (rtx next, int delay)
3887 int i = QUEUE_INDEX (next);
3889 gcc_assert (QUEUE_NOWHERE <= delay && delay <= max_insn_queue_index
3890 && delay != 0);
3891 gcc_assert (i != QUEUE_SCHEDULED);
3893 if ((delay > 0 && NEXT_Q_AFTER (q_ptr, delay) == i)
3894 || (delay < 0 && delay == i))
3895 /* We have nothing to do. */
3896 return;
3898 /* Remove NEXT from wherever it is now. */
3899 if (i == QUEUE_READY)
3900 ready_remove_insn (next);
3901 else if (i >= 0)
3902 queue_remove (next);
3904 /* Add it to the proper place. */
3905 if (delay == QUEUE_READY)
3906 ready_add (readyp, next, false);
3907 else if (delay >= 1)
3908 queue_insn (next, delay);
3910 if (sched_verbose >= 2)
3912 fprintf (sched_dump, ";;\t\ttick updated: insn %s",
3913 (*current_sched_info->print_insn) (next, 0));
3915 if (delay == QUEUE_READY)
3916 fprintf (sched_dump, " into ready\n");
3917 else if (delay >= 1)
3918 fprintf (sched_dump, " into queue with cost=%d\n", delay);
3919 else
3920 fprintf (sched_dump, " removed from ready or queue lists\n");
3924 static int sched_ready_n_insns = -1;
3926 /* Initialize per region data structures. */
3927 void
3928 sched_extend_ready_list (int new_sched_ready_n_insns)
3930 int i;
3932 if (sched_ready_n_insns == -1)
3933 /* At the first call we need to initialize one more choice_stack
3934 entry. */
3936 i = 0;
3937 sched_ready_n_insns = 0;
3939 else
3940 i = sched_ready_n_insns + 1;
3942 ready.veclen = new_sched_ready_n_insns + issue_rate;
3943 ready.vec = XRESIZEVEC (rtx, ready.vec, ready.veclen);
3945 gcc_assert (new_sched_ready_n_insns >= sched_ready_n_insns);
3947 ready_try = (char *) xrecalloc (ready_try, new_sched_ready_n_insns,
3948 sched_ready_n_insns, sizeof (*ready_try));
3950 /* We allocate +1 element to save initial state in the choice_stack[0]
3951 entry. */
3952 choice_stack = XRESIZEVEC (struct choice_entry, choice_stack,
3953 new_sched_ready_n_insns + 1);
3955 for (; i <= new_sched_ready_n_insns; i++)
3956 choice_stack[i].state = xmalloc (dfa_state_size);
3958 sched_ready_n_insns = new_sched_ready_n_insns;
3961 /* Free per region data structures. */
3962 void
3963 sched_finish_ready_list (void)
3965 int i;
3967 free (ready.vec);
3968 ready.vec = NULL;
3969 ready.veclen = 0;
3971 free (ready_try);
3972 ready_try = NULL;
3974 for (i = 0; i <= sched_ready_n_insns; i++)
3975 free (choice_stack [i].state);
3976 free (choice_stack);
3977 choice_stack = NULL;
3979 sched_ready_n_insns = -1;
3982 static int
3983 haifa_luid_for_non_insn (rtx x)
3985 gcc_assert (NOTE_P (x) || LABEL_P (x));
3987 return 0;
3990 /* Generates recovery code for INSN. */
3991 static void
3992 generate_recovery_code (rtx insn)
3994 if (TODO_SPEC (insn) & BEGIN_SPEC)
3995 begin_speculative_block (insn);
3997 /* Here we have insn with no dependencies to
3998 instructions other then CHECK_SPEC ones. */
4000 if (TODO_SPEC (insn) & BE_IN_SPEC)
4001 add_to_speculative_block (insn);
4004 /* Helper function.
4005 Tries to add speculative dependencies of type FS between instructions
4006 in deps_list L and TWIN. */
4007 static void
4008 process_insn_forw_deps_be_in_spec (rtx insn, rtx twin, ds_t fs)
4010 sd_iterator_def sd_it;
4011 dep_t dep;
4013 FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep)
4015 ds_t ds;
4016 rtx consumer;
4018 consumer = DEP_CON (dep);
4020 ds = DEP_STATUS (dep);
4022 if (/* If we want to create speculative dep. */
4024 /* And we can do that because this is a true dep. */
4025 && (ds & DEP_TYPES) == DEP_TRUE)
4027 gcc_assert (!(ds & BE_IN_SPEC));
4029 if (/* If this dep can be overcome with 'begin speculation'. */
4030 ds & BEGIN_SPEC)
4031 /* Then we have a choice: keep the dep 'begin speculative'
4032 or transform it into 'be in speculative'. */
4034 if (/* In try_ready we assert that if insn once became ready
4035 it can be removed from the ready (or queue) list only
4036 due to backend decision. Hence we can't let the
4037 probability of the speculative dep to decrease. */
4038 ds_weak (ds) <= ds_weak (fs))
4040 ds_t new_ds;
4042 new_ds = (ds & ~BEGIN_SPEC) | fs;
4044 if (/* consumer can 'be in speculative'. */
4045 sched_insn_is_legitimate_for_speculation_p (consumer,
4046 new_ds))
4047 /* Transform it to be in speculative. */
4048 ds = new_ds;
4051 else
4052 /* Mark the dep as 'be in speculative'. */
4053 ds |= fs;
4057 dep_def _new_dep, *new_dep = &_new_dep;
4059 init_dep_1 (new_dep, twin, consumer, DEP_TYPE (dep), ds);
4060 sd_add_dep (new_dep, false);
4065 /* Generates recovery code for BEGIN speculative INSN. */
4066 static void
4067 begin_speculative_block (rtx insn)
4069 if (TODO_SPEC (insn) & BEGIN_DATA)
4070 nr_begin_data++;
4071 if (TODO_SPEC (insn) & BEGIN_CONTROL)
4072 nr_begin_control++;
4074 create_check_block_twin (insn, false);
4076 TODO_SPEC (insn) &= ~BEGIN_SPEC;
4079 static void haifa_init_insn (rtx);
4081 /* Generates recovery code for BE_IN speculative INSN. */
4082 static void
4083 add_to_speculative_block (rtx insn)
4085 ds_t ts;
4086 sd_iterator_def sd_it;
4087 dep_t dep;
4088 rtx twins = NULL;
4089 rtx_vec_t priorities_roots;
4091 ts = TODO_SPEC (insn);
4092 gcc_assert (!(ts & ~BE_IN_SPEC));
4094 if (ts & BE_IN_DATA)
4095 nr_be_in_data++;
4096 if (ts & BE_IN_CONTROL)
4097 nr_be_in_control++;
4099 TODO_SPEC (insn) &= ~BE_IN_SPEC;
4100 gcc_assert (!TODO_SPEC (insn));
4102 DONE_SPEC (insn) |= ts;
4104 /* First we convert all simple checks to branchy. */
4105 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
4106 sd_iterator_cond (&sd_it, &dep);)
4108 rtx check = DEP_PRO (dep);
4110 if (IS_SPECULATION_SIMPLE_CHECK_P (check))
4112 create_check_block_twin (check, true);
4114 /* Restart search. */
4115 sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
4117 else
4118 /* Continue search. */
4119 sd_iterator_next (&sd_it);
4122 priorities_roots = NULL;
4123 clear_priorities (insn, &priorities_roots);
4125 while (1)
4127 rtx check, twin;
4128 basic_block rec;
4130 /* Get the first backward dependency of INSN. */
4131 sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
4132 if (!sd_iterator_cond (&sd_it, &dep))
4133 /* INSN has no backward dependencies left. */
4134 break;
4136 gcc_assert ((DEP_STATUS (dep) & BEGIN_SPEC) == 0
4137 && (DEP_STATUS (dep) & BE_IN_SPEC) != 0
4138 && (DEP_STATUS (dep) & DEP_TYPES) == DEP_TRUE);
4140 check = DEP_PRO (dep);
4142 gcc_assert (!IS_SPECULATION_CHECK_P (check) && !ORIG_PAT (check)
4143 && QUEUE_INDEX (check) == QUEUE_NOWHERE);
4145 rec = BLOCK_FOR_INSN (check);
4147 twin = emit_insn_before (copy_insn (PATTERN (insn)), BB_END (rec));
4148 haifa_init_insn (twin);
4150 sd_copy_back_deps (twin, insn, true);
4152 if (sched_verbose && spec_info->dump)
4153 /* INSN_BB (insn) isn't determined for twin insns yet.
4154 So we can't use current_sched_info->print_insn. */
4155 fprintf (spec_info->dump, ";;\t\tGenerated twin insn : %d/rec%d\n",
4156 INSN_UID (twin), rec->index);
4158 twins = alloc_INSN_LIST (twin, twins);
4160 /* Add dependences between TWIN and all appropriate
4161 instructions from REC. */
4162 FOR_EACH_DEP (insn, SD_LIST_SPEC_BACK, sd_it, dep)
4164 rtx pro = DEP_PRO (dep);
4166 gcc_assert (DEP_TYPE (dep) == REG_DEP_TRUE);
4168 /* INSN might have dependencies from the instructions from
4169 several recovery blocks. At this iteration we process those
4170 producers that reside in REC. */
4171 if (BLOCK_FOR_INSN (pro) == rec)
4173 dep_def _new_dep, *new_dep = &_new_dep;
4175 init_dep (new_dep, pro, twin, REG_DEP_TRUE);
4176 sd_add_dep (new_dep, false);
4180 process_insn_forw_deps_be_in_spec (insn, twin, ts);
4182 /* Remove all dependencies between INSN and insns in REC. */
4183 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
4184 sd_iterator_cond (&sd_it, &dep);)
4186 rtx pro = DEP_PRO (dep);
4188 if (BLOCK_FOR_INSN (pro) == rec)
4189 sd_delete_dep (sd_it);
4190 else
4191 sd_iterator_next (&sd_it);
4195 /* We couldn't have added the dependencies between INSN and TWINS earlier
4196 because that would make TWINS appear in the INSN_BACK_DEPS (INSN). */
4197 while (twins)
4199 rtx twin;
4201 twin = XEXP (twins, 0);
4204 dep_def _new_dep, *new_dep = &_new_dep;
4206 init_dep (new_dep, insn, twin, REG_DEP_OUTPUT);
4207 sd_add_dep (new_dep, false);
4210 twin = XEXP (twins, 1);
4211 free_INSN_LIST_node (twins);
4212 twins = twin;
4215 calc_priorities (priorities_roots);
4216 VEC_free (rtx, heap, priorities_roots);
4219 /* Extends and fills with zeros (only the new part) array pointed to by P. */
4220 void *
4221 xrecalloc (void *p, size_t new_nmemb, size_t old_nmemb, size_t size)
4223 gcc_assert (new_nmemb >= old_nmemb);
4224 p = XRESIZEVAR (void, p, new_nmemb * size);
4225 memset (((char *) p) + old_nmemb * size, 0, (new_nmemb - old_nmemb) * size);
4226 return p;
4229 /* Helper function.
4230 Find fallthru edge from PRED. */
4231 edge
4232 find_fallthru_edge (basic_block pred)
4234 edge e;
4235 edge_iterator ei;
4236 basic_block succ;
4238 succ = pred->next_bb;
4239 gcc_assert (succ->prev_bb == pred);
4241 if (EDGE_COUNT (pred->succs) <= EDGE_COUNT (succ->preds))
4243 FOR_EACH_EDGE (e, ei, pred->succs)
4244 if (e->flags & EDGE_FALLTHRU)
4246 gcc_assert (e->dest == succ);
4247 return e;
4250 else
4252 FOR_EACH_EDGE (e, ei, succ->preds)
4253 if (e->flags & EDGE_FALLTHRU)
4255 gcc_assert (e->src == pred);
4256 return e;
4260 return NULL;
4263 /* Extend per basic block data structures. */
4264 static void
4265 sched_extend_bb (void)
4267 rtx insn;
4269 /* The following is done to keep current_sched_info->next_tail non null. */
4270 insn = BB_END (EXIT_BLOCK_PTR->prev_bb);
4271 if (NEXT_INSN (insn) == 0
4272 || (!NOTE_P (insn)
4273 && !LABEL_P (insn)
4274 /* Don't emit a NOTE if it would end up before a BARRIER. */
4275 && !BARRIER_P (NEXT_INSN (insn))))
4277 rtx note = emit_note_after (NOTE_INSN_DELETED, insn);
4278 /* Make insn appear outside BB. */
4279 set_block_for_insn (note, NULL);
4280 BB_END (EXIT_BLOCK_PTR->prev_bb) = insn;
4284 /* Init per basic block data structures. */
4285 void
4286 sched_init_bbs (void)
4288 sched_extend_bb ();
4291 /* Initialize BEFORE_RECOVERY variable. */
4292 static void
4293 init_before_recovery (basic_block *before_recovery_ptr)
4295 basic_block last;
4296 edge e;
4298 last = EXIT_BLOCK_PTR->prev_bb;
4299 e = find_fallthru_edge (last);
4301 if (e)
4303 /* We create two basic blocks:
4304 1. Single instruction block is inserted right after E->SRC
4305 and has jump to
4306 2. Empty block right before EXIT_BLOCK.
4307 Between these two blocks recovery blocks will be emitted. */
4309 basic_block single, empty;
4310 rtx x, label;
4312 /* If the fallthrough edge to exit we've found is from the block we've
4313 created before, don't do anything more. */
4314 if (last == after_recovery)
4315 return;
4317 adding_bb_to_current_region_p = false;
4319 single = sched_create_empty_bb (last);
4320 empty = sched_create_empty_bb (single);
4322 /* Add new blocks to the root loop. */
4323 if (current_loops != NULL)
4325 add_bb_to_loop (single, VEC_index (loop_p, current_loops->larray, 0));
4326 add_bb_to_loop (empty, VEC_index (loop_p, current_loops->larray, 0));
4329 single->count = last->count;
4330 empty->count = last->count;
4331 single->frequency = last->frequency;
4332 empty->frequency = last->frequency;
4333 BB_COPY_PARTITION (single, last);
4334 BB_COPY_PARTITION (empty, last);
4336 redirect_edge_succ (e, single);
4337 make_single_succ_edge (single, empty, 0);
4338 make_single_succ_edge (empty, EXIT_BLOCK_PTR,
4339 EDGE_FALLTHRU | EDGE_CAN_FALLTHRU);
4341 label = block_label (empty);
4342 x = emit_jump_insn_after (gen_jump (label), BB_END (single));
4343 JUMP_LABEL (x) = label;
4344 LABEL_NUSES (label)++;
4345 haifa_init_insn (x);
4347 emit_barrier_after (x);
4349 sched_init_only_bb (empty, NULL);
4350 sched_init_only_bb (single, NULL);
4351 sched_extend_bb ();
4353 adding_bb_to_current_region_p = true;
4354 before_recovery = single;
4355 after_recovery = empty;
4357 if (before_recovery_ptr)
4358 *before_recovery_ptr = before_recovery;
4360 if (sched_verbose >= 2 && spec_info->dump)
4361 fprintf (spec_info->dump,
4362 ";;\t\tFixed fallthru to EXIT : %d->>%d->%d->>EXIT\n",
4363 last->index, single->index, empty->index);
4365 else
4366 before_recovery = last;
4369 /* Returns new recovery block. */
4370 basic_block
4371 sched_create_recovery_block (basic_block *before_recovery_ptr)
4373 rtx label;
4374 rtx barrier;
4375 basic_block rec;
4377 haifa_recovery_bb_recently_added_p = true;
4378 haifa_recovery_bb_ever_added_p = true;
4380 init_before_recovery (before_recovery_ptr);
4382 barrier = get_last_bb_insn (before_recovery);
4383 gcc_assert (BARRIER_P (barrier));
4385 label = emit_label_after (gen_label_rtx (), barrier);
4387 rec = create_basic_block (label, label, before_recovery);
4389 /* A recovery block always ends with an unconditional jump. */
4390 emit_barrier_after (BB_END (rec));
4392 if (BB_PARTITION (before_recovery) != BB_UNPARTITIONED)
4393 BB_SET_PARTITION (rec, BB_COLD_PARTITION);
4395 if (sched_verbose && spec_info->dump)
4396 fprintf (spec_info->dump, ";;\t\tGenerated recovery block rec%d\n",
4397 rec->index);
4399 return rec;
4402 /* Create edges: FIRST_BB -> REC; FIRST_BB -> SECOND_BB; REC -> SECOND_BB
4403 and emit necessary jumps. */
4404 void
4405 sched_create_recovery_edges (basic_block first_bb, basic_block rec,
4406 basic_block second_bb)
4408 rtx label;
4409 rtx jump;
4410 int edge_flags;
4412 /* This is fixing of incoming edge. */
4413 /* ??? Which other flags should be specified? */
4414 if (BB_PARTITION (first_bb) != BB_PARTITION (rec))
4415 /* Partition type is the same, if it is "unpartitioned". */
4416 edge_flags = EDGE_CROSSING;
4417 else
4418 edge_flags = 0;
4420 make_edge (first_bb, rec, edge_flags);
4421 label = block_label (second_bb);
4422 jump = emit_jump_insn_after (gen_jump (label), BB_END (rec));
4423 JUMP_LABEL (jump) = label;
4424 LABEL_NUSES (label)++;
4426 if (BB_PARTITION (second_bb) != BB_PARTITION (rec))
4427 /* Partition type is the same, if it is "unpartitioned". */
4429 /* Rewritten from cfgrtl.c. */
4430 if (flag_reorder_blocks_and_partition
4431 && targetm.have_named_sections)
4433 /* We don't need the same note for the check because
4434 any_condjump_p (check) == true. */
4435 add_reg_note (jump, REG_CROSSING_JUMP, NULL_RTX);
4437 edge_flags = EDGE_CROSSING;
4439 else
4440 edge_flags = 0;
4442 make_single_succ_edge (rec, second_bb, edge_flags);
4445 /* This function creates recovery code for INSN. If MUTATE_P is nonzero,
4446 INSN is a simple check, that should be converted to branchy one. */
4447 static void
4448 create_check_block_twin (rtx insn, bool mutate_p)
4450 basic_block rec;
4451 rtx label, check, twin;
4452 ds_t fs;
4453 sd_iterator_def sd_it;
4454 dep_t dep;
4455 dep_def _new_dep, *new_dep = &_new_dep;
4456 ds_t todo_spec;
4458 gcc_assert (ORIG_PAT (insn) != NULL_RTX);
4460 if (!mutate_p)
4461 todo_spec = TODO_SPEC (insn);
4462 else
4464 gcc_assert (IS_SPECULATION_SIMPLE_CHECK_P (insn)
4465 && (TODO_SPEC (insn) & SPECULATIVE) == 0);
4467 todo_spec = CHECK_SPEC (insn);
4470 todo_spec &= SPECULATIVE;
4472 /* Create recovery block. */
4473 if (mutate_p || targetm.sched.needs_block_p (todo_spec))
4475 rec = sched_create_recovery_block (NULL);
4476 label = BB_HEAD (rec);
4478 else
4480 rec = EXIT_BLOCK_PTR;
4481 label = NULL_RTX;
4484 /* Emit CHECK. */
4485 check = targetm.sched.gen_spec_check (insn, label, todo_spec);
4487 if (rec != EXIT_BLOCK_PTR)
4489 /* To have mem_reg alive at the beginning of second_bb,
4490 we emit check BEFORE insn, so insn after splitting
4491 insn will be at the beginning of second_bb, which will
4492 provide us with the correct life information. */
4493 check = emit_jump_insn_before (check, insn);
4494 JUMP_LABEL (check) = label;
4495 LABEL_NUSES (label)++;
4497 else
4498 check = emit_insn_before (check, insn);
4500 /* Extend data structures. */
4501 haifa_init_insn (check);
4503 /* CHECK is being added to current region. Extend ready list. */
4504 gcc_assert (sched_ready_n_insns != -1);
4505 sched_extend_ready_list (sched_ready_n_insns + 1);
4507 if (current_sched_info->add_remove_insn)
4508 current_sched_info->add_remove_insn (insn, 0);
4510 RECOVERY_BLOCK (check) = rec;
4512 if (sched_verbose && spec_info->dump)
4513 fprintf (spec_info->dump, ";;\t\tGenerated check insn : %s\n",
4514 (*current_sched_info->print_insn) (check, 0));
4516 gcc_assert (ORIG_PAT (insn));
4518 /* Initialize TWIN (twin is a duplicate of original instruction
4519 in the recovery block). */
4520 if (rec != EXIT_BLOCK_PTR)
4522 sd_iterator_def sd_it;
4523 dep_t dep;
4525 FOR_EACH_DEP (insn, SD_LIST_RES_BACK, sd_it, dep)
4526 if ((DEP_STATUS (dep) & DEP_OUTPUT) != 0)
4528 struct _dep _dep2, *dep2 = &_dep2;
4530 init_dep (dep2, DEP_PRO (dep), check, REG_DEP_TRUE);
4532 sd_add_dep (dep2, true);
4535 twin = emit_insn_after (ORIG_PAT (insn), BB_END (rec));
4536 haifa_init_insn (twin);
4538 if (sched_verbose && spec_info->dump)
4539 /* INSN_BB (insn) isn't determined for twin insns yet.
4540 So we can't use current_sched_info->print_insn. */
4541 fprintf (spec_info->dump, ";;\t\tGenerated twin insn : %d/rec%d\n",
4542 INSN_UID (twin), rec->index);
4544 else
4546 ORIG_PAT (check) = ORIG_PAT (insn);
4547 HAS_INTERNAL_DEP (check) = 1;
4548 twin = check;
4549 /* ??? We probably should change all OUTPUT dependencies to
4550 (TRUE | OUTPUT). */
4553 /* Copy all resolved back dependencies of INSN to TWIN. This will
4554 provide correct value for INSN_TICK (TWIN). */
4555 sd_copy_back_deps (twin, insn, true);
4557 if (rec != EXIT_BLOCK_PTR)
4558 /* In case of branchy check, fix CFG. */
4560 basic_block first_bb, second_bb;
4561 rtx jump;
4563 first_bb = BLOCK_FOR_INSN (check);
4564 second_bb = sched_split_block (first_bb, check);
4566 sched_create_recovery_edges (first_bb, rec, second_bb);
4568 sched_init_only_bb (second_bb, first_bb);
4569 sched_init_only_bb (rec, EXIT_BLOCK_PTR);
4571 jump = BB_END (rec);
4572 haifa_init_insn (jump);
4575 /* Move backward dependences from INSN to CHECK and
4576 move forward dependences from INSN to TWIN. */
4578 /* First, create dependencies between INSN's producers and CHECK & TWIN. */
4579 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
4581 rtx pro = DEP_PRO (dep);
4582 ds_t ds;
4584 /* If BEGIN_DATA: [insn ~~TRUE~~> producer]:
4585 check --TRUE--> producer ??? or ANTI ???
4586 twin --TRUE--> producer
4587 twin --ANTI--> check
4589 If BEGIN_CONTROL: [insn ~~ANTI~~> producer]:
4590 check --ANTI--> producer
4591 twin --ANTI--> producer
4592 twin --ANTI--> check
4594 If BE_IN_SPEC: [insn ~~TRUE~~> producer]:
4595 check ~~TRUE~~> producer
4596 twin ~~TRUE~~> producer
4597 twin --ANTI--> check */
4599 ds = DEP_STATUS (dep);
4601 if (ds & BEGIN_SPEC)
4603 gcc_assert (!mutate_p);
4604 ds &= ~BEGIN_SPEC;
4607 init_dep_1 (new_dep, pro, check, DEP_TYPE (dep), ds);
4608 sd_add_dep (new_dep, false);
4610 if (rec != EXIT_BLOCK_PTR)
4612 DEP_CON (new_dep) = twin;
4613 sd_add_dep (new_dep, false);
4617 /* Second, remove backward dependencies of INSN. */
4618 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
4619 sd_iterator_cond (&sd_it, &dep);)
4621 if ((DEP_STATUS (dep) & BEGIN_SPEC)
4622 || mutate_p)
4623 /* We can delete this dep because we overcome it with
4624 BEGIN_SPECULATION. */
4625 sd_delete_dep (sd_it);
4626 else
4627 sd_iterator_next (&sd_it);
4630 /* Future Speculations. Determine what BE_IN speculations will be like. */
4631 fs = 0;
4633 /* Fields (DONE_SPEC (x) & BEGIN_SPEC) and CHECK_SPEC (x) are set only
4634 here. */
4636 gcc_assert (!DONE_SPEC (insn));
4638 if (!mutate_p)
4640 ds_t ts = TODO_SPEC (insn);
4642 DONE_SPEC (insn) = ts & BEGIN_SPEC;
4643 CHECK_SPEC (check) = ts & BEGIN_SPEC;
4645 /* Luckiness of future speculations solely depends upon initial
4646 BEGIN speculation. */
4647 if (ts & BEGIN_DATA)
4648 fs = set_dep_weak (fs, BE_IN_DATA, get_dep_weak (ts, BEGIN_DATA));
4649 if (ts & BEGIN_CONTROL)
4650 fs = set_dep_weak (fs, BE_IN_CONTROL,
4651 get_dep_weak (ts, BEGIN_CONTROL));
4653 else
4654 CHECK_SPEC (check) = CHECK_SPEC (insn);
4656 /* Future speculations: call the helper. */
4657 process_insn_forw_deps_be_in_spec (insn, twin, fs);
4659 if (rec != EXIT_BLOCK_PTR)
4661 /* Which types of dependencies should we use here is,
4662 generally, machine-dependent question... But, for now,
4663 it is not. */
4665 if (!mutate_p)
4667 init_dep (new_dep, insn, check, REG_DEP_TRUE);
4668 sd_add_dep (new_dep, false);
4670 init_dep (new_dep, insn, twin, REG_DEP_OUTPUT);
4671 sd_add_dep (new_dep, false);
4673 else
4675 if (spec_info->dump)
4676 fprintf (spec_info->dump, ";;\t\tRemoved simple check : %s\n",
4677 (*current_sched_info->print_insn) (insn, 0));
4679 /* Remove all dependencies of the INSN. */
4681 sd_it = sd_iterator_start (insn, (SD_LIST_FORW
4682 | SD_LIST_BACK
4683 | SD_LIST_RES_BACK));
4684 while (sd_iterator_cond (&sd_it, &dep))
4685 sd_delete_dep (sd_it);
4688 /* If former check (INSN) already was moved to the ready (or queue)
4689 list, add new check (CHECK) there too. */
4690 if (QUEUE_INDEX (insn) != QUEUE_NOWHERE)
4691 try_ready (check);
4693 /* Remove old check from instruction stream and free its
4694 data. */
4695 sched_remove_insn (insn);
4698 init_dep (new_dep, check, twin, REG_DEP_ANTI);
4699 sd_add_dep (new_dep, false);
4701 else
4703 init_dep_1 (new_dep, insn, check, REG_DEP_TRUE, DEP_TRUE | DEP_OUTPUT);
4704 sd_add_dep (new_dep, false);
4707 if (!mutate_p)
4708 /* Fix priorities. If MUTATE_P is nonzero, this is not necessary,
4709 because it'll be done later in add_to_speculative_block. */
4711 rtx_vec_t priorities_roots = NULL;
4713 clear_priorities (twin, &priorities_roots);
4714 calc_priorities (priorities_roots);
4715 VEC_free (rtx, heap, priorities_roots);
4719 /* Removes dependency between instructions in the recovery block REC
4720 and usual region instructions. It keeps inner dependences so it
4721 won't be necessary to recompute them. */
4722 static void
4723 fix_recovery_deps (basic_block rec)
4725 rtx note, insn, jump, ready_list = 0;
4726 bitmap_head in_ready;
4727 rtx link;
4729 bitmap_initialize (&in_ready, 0);
4731 /* NOTE - a basic block note. */
4732 note = NEXT_INSN (BB_HEAD (rec));
4733 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
4734 insn = BB_END (rec);
4735 gcc_assert (JUMP_P (insn));
4736 insn = PREV_INSN (insn);
4740 sd_iterator_def sd_it;
4741 dep_t dep;
4743 for (sd_it = sd_iterator_start (insn, SD_LIST_FORW);
4744 sd_iterator_cond (&sd_it, &dep);)
4746 rtx consumer = DEP_CON (dep);
4748 if (BLOCK_FOR_INSN (consumer) != rec)
4750 sd_delete_dep (sd_it);
4752 if (!bitmap_bit_p (&in_ready, INSN_LUID (consumer)))
4754 ready_list = alloc_INSN_LIST (consumer, ready_list);
4755 bitmap_set_bit (&in_ready, INSN_LUID (consumer));
4758 else
4760 gcc_assert ((DEP_STATUS (dep) & DEP_TYPES) == DEP_TRUE);
4762 sd_iterator_next (&sd_it);
4766 insn = PREV_INSN (insn);
4768 while (insn != note);
4770 bitmap_clear (&in_ready);
4772 /* Try to add instructions to the ready or queue list. */
4773 for (link = ready_list; link; link = XEXP (link, 1))
4774 try_ready (XEXP (link, 0));
4775 free_INSN_LIST_list (&ready_list);
4777 /* Fixing jump's dependences. */
4778 insn = BB_HEAD (rec);
4779 jump = BB_END (rec);
4781 gcc_assert (LABEL_P (insn));
4782 insn = NEXT_INSN (insn);
4784 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (insn));
4785 add_jump_dependencies (insn, jump);
4788 /* Change pattern of INSN to NEW_PAT. */
4789 void
4790 sched_change_pattern (rtx insn, rtx new_pat)
4792 int t;
4794 t = validate_change (insn, &PATTERN (insn), new_pat, 0);
4795 gcc_assert (t);
4796 dfa_clear_single_insn_cache (insn);
4799 /* Change pattern of INSN to NEW_PAT. Invalidate cached haifa
4800 instruction data. */
4801 static void
4802 haifa_change_pattern (rtx insn, rtx new_pat)
4804 sched_change_pattern (insn, new_pat);
4806 /* Invalidate INSN_COST, so it'll be recalculated. */
4807 INSN_COST (insn) = -1;
4808 /* Invalidate INSN_TICK, so it'll be recalculated. */
4809 INSN_TICK (insn) = INVALID_TICK;
4812 /* -1 - can't speculate,
4813 0 - for speculation with REQUEST mode it is OK to use
4814 current instruction pattern,
4815 1 - need to change pattern for *NEW_PAT to be speculative. */
4817 sched_speculate_insn (rtx insn, ds_t request, rtx *new_pat)
4819 gcc_assert (current_sched_info->flags & DO_SPECULATION
4820 && (request & SPECULATIVE)
4821 && sched_insn_is_legitimate_for_speculation_p (insn, request));
4823 if ((request & spec_info->mask) != request)
4824 return -1;
4826 if (request & BE_IN_SPEC
4827 && !(request & BEGIN_SPEC))
4828 return 0;
4830 return targetm.sched.speculate_insn (insn, request, new_pat);
4833 static int
4834 haifa_speculate_insn (rtx insn, ds_t request, rtx *new_pat)
4836 gcc_assert (sched_deps_info->generate_spec_deps
4837 && !IS_SPECULATION_CHECK_P (insn));
4839 if (HAS_INTERNAL_DEP (insn)
4840 || SCHED_GROUP_P (insn))
4841 return -1;
4843 return sched_speculate_insn (insn, request, new_pat);
4846 /* Print some information about block BB, which starts with HEAD and
4847 ends with TAIL, before scheduling it.
4848 I is zero, if scheduler is about to start with the fresh ebb. */
4849 static void
4850 dump_new_block_header (int i, basic_block bb, rtx head, rtx tail)
4852 if (!i)
4853 fprintf (sched_dump,
4854 ";; ======================================================\n");
4855 else
4856 fprintf (sched_dump,
4857 ";; =====================ADVANCING TO=====================\n");
4858 fprintf (sched_dump,
4859 ";; -- basic block %d from %d to %d -- %s reload\n",
4860 bb->index, INSN_UID (head), INSN_UID (tail),
4861 (reload_completed ? "after" : "before"));
4862 fprintf (sched_dump,
4863 ";; ======================================================\n");
4864 fprintf (sched_dump, "\n");
4867 /* Unlink basic block notes and labels and saves them, so they
4868 can be easily restored. We unlink basic block notes in EBB to
4869 provide back-compatibility with the previous code, as target backends
4870 assume, that there'll be only instructions between
4871 current_sched_info->{head and tail}. We restore these notes as soon
4872 as we can.
4873 FIRST (LAST) is the first (last) basic block in the ebb.
4874 NB: In usual case (FIRST == LAST) nothing is really done. */
4875 void
4876 unlink_bb_notes (basic_block first, basic_block last)
4878 /* We DON'T unlink basic block notes of the first block in the ebb. */
4879 if (first == last)
4880 return;
4882 bb_header = XNEWVEC (rtx, last_basic_block);
4884 /* Make a sentinel. */
4885 if (last->next_bb != EXIT_BLOCK_PTR)
4886 bb_header[last->next_bb->index] = 0;
4888 first = first->next_bb;
4891 rtx prev, label, note, next;
4893 label = BB_HEAD (last);
4894 if (LABEL_P (label))
4895 note = NEXT_INSN (label);
4896 else
4897 note = label;
4898 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
4900 prev = PREV_INSN (label);
4901 next = NEXT_INSN (note);
4902 gcc_assert (prev && next);
4904 NEXT_INSN (prev) = next;
4905 PREV_INSN (next) = prev;
4907 bb_header[last->index] = label;
4909 if (last == first)
4910 break;
4912 last = last->prev_bb;
4914 while (1);
4917 /* Restore basic block notes.
4918 FIRST is the first basic block in the ebb. */
4919 static void
4920 restore_bb_notes (basic_block first)
4922 if (!bb_header)
4923 return;
4925 /* We DON'T unlink basic block notes of the first block in the ebb. */
4926 first = first->next_bb;
4927 /* Remember: FIRST is actually a second basic block in the ebb. */
4929 while (first != EXIT_BLOCK_PTR
4930 && bb_header[first->index])
4932 rtx prev, label, note, next;
4934 label = bb_header[first->index];
4935 prev = PREV_INSN (label);
4936 next = NEXT_INSN (prev);
4938 if (LABEL_P (label))
4939 note = NEXT_INSN (label);
4940 else
4941 note = label;
4942 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
4944 bb_header[first->index] = 0;
4946 NEXT_INSN (prev) = label;
4947 NEXT_INSN (note) = next;
4948 PREV_INSN (next) = note;
4950 first = first->next_bb;
4953 free (bb_header);
4954 bb_header = 0;
4957 /* Helper function.
4958 Fix CFG after both in- and inter-block movement of
4959 control_flow_insn_p JUMP. */
4960 static void
4961 fix_jump_move (rtx jump)
4963 basic_block bb, jump_bb, jump_bb_next;
4965 bb = BLOCK_FOR_INSN (PREV_INSN (jump));
4966 jump_bb = BLOCK_FOR_INSN (jump);
4967 jump_bb_next = jump_bb->next_bb;
4969 gcc_assert (common_sched_info->sched_pass_id == SCHED_EBB_PASS
4970 || IS_SPECULATION_BRANCHY_CHECK_P (jump));
4972 if (!NOTE_INSN_BASIC_BLOCK_P (BB_END (jump_bb_next)))
4973 /* if jump_bb_next is not empty. */
4974 BB_END (jump_bb) = BB_END (jump_bb_next);
4976 if (BB_END (bb) != PREV_INSN (jump))
4977 /* Then there are instruction after jump that should be placed
4978 to jump_bb_next. */
4979 BB_END (jump_bb_next) = BB_END (bb);
4980 else
4981 /* Otherwise jump_bb_next is empty. */
4982 BB_END (jump_bb_next) = NEXT_INSN (BB_HEAD (jump_bb_next));
4984 /* To make assertion in move_insn happy. */
4985 BB_END (bb) = PREV_INSN (jump);
4987 update_bb_for_insn (jump_bb_next);
4990 /* Fix CFG after interblock movement of control_flow_insn_p JUMP. */
4991 static void
4992 move_block_after_check (rtx jump)
4994 basic_block bb, jump_bb, jump_bb_next;
4995 VEC(edge,gc) *t;
4997 bb = BLOCK_FOR_INSN (PREV_INSN (jump));
4998 jump_bb = BLOCK_FOR_INSN (jump);
4999 jump_bb_next = jump_bb->next_bb;
5001 update_bb_for_insn (jump_bb);
5003 gcc_assert (IS_SPECULATION_CHECK_P (jump)
5004 || IS_SPECULATION_CHECK_P (BB_END (jump_bb_next)));
5006 unlink_block (jump_bb_next);
5007 link_block (jump_bb_next, bb);
5009 t = bb->succs;
5010 bb->succs = 0;
5011 move_succs (&(jump_bb->succs), bb);
5012 move_succs (&(jump_bb_next->succs), jump_bb);
5013 move_succs (&t, jump_bb_next);
5015 df_mark_solutions_dirty ();
5017 common_sched_info->fix_recovery_cfg
5018 (bb->index, jump_bb->index, jump_bb_next->index);
5021 /* Helper function for move_block_after_check.
5022 This functions attaches edge vector pointed to by SUCCSP to
5023 block TO. */
5024 static void
5025 move_succs (VEC(edge,gc) **succsp, basic_block to)
5027 edge e;
5028 edge_iterator ei;
5030 gcc_assert (to->succs == 0);
5032 to->succs = *succsp;
5034 FOR_EACH_EDGE (e, ei, to->succs)
5035 e->src = to;
5037 *succsp = 0;
5040 /* Remove INSN from the instruction stream.
5041 INSN should have any dependencies. */
5042 static void
5043 sched_remove_insn (rtx insn)
5045 sd_finish_insn (insn);
5047 change_queue_index (insn, QUEUE_NOWHERE);
5048 current_sched_info->add_remove_insn (insn, 1);
5049 remove_insn (insn);
5052 /* Clear priorities of all instructions, that are forward dependent on INSN.
5053 Store in vector pointed to by ROOTS_PTR insns on which priority () should
5054 be invoked to initialize all cleared priorities. */
5055 static void
5056 clear_priorities (rtx insn, rtx_vec_t *roots_ptr)
5058 sd_iterator_def sd_it;
5059 dep_t dep;
5060 bool insn_is_root_p = true;
5062 gcc_assert (QUEUE_INDEX (insn) != QUEUE_SCHEDULED);
5064 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
5066 rtx pro = DEP_PRO (dep);
5068 if (INSN_PRIORITY_STATUS (pro) >= 0
5069 && QUEUE_INDEX (insn) != QUEUE_SCHEDULED)
5071 /* If DEP doesn't contribute to priority then INSN itself should
5072 be added to priority roots. */
5073 if (contributes_to_priority_p (dep))
5074 insn_is_root_p = false;
5076 INSN_PRIORITY_STATUS (pro) = -1;
5077 clear_priorities (pro, roots_ptr);
5081 if (insn_is_root_p)
5082 VEC_safe_push (rtx, heap, *roots_ptr, insn);
5085 /* Recompute priorities of instructions, whose priorities might have been
5086 changed. ROOTS is a vector of instructions whose priority computation will
5087 trigger initialization of all cleared priorities. */
5088 static void
5089 calc_priorities (rtx_vec_t roots)
5091 int i;
5092 rtx insn;
5094 for (i = 0; VEC_iterate (rtx, roots, i, insn); i++)
5095 priority (insn);
5099 /* Add dependences between JUMP and other instructions in the recovery
5100 block. INSN is the first insn the recovery block. */
5101 static void
5102 add_jump_dependencies (rtx insn, rtx jump)
5106 insn = NEXT_INSN (insn);
5107 if (insn == jump)
5108 break;
5110 if (dep_list_size (insn) == 0)
5112 dep_def _new_dep, *new_dep = &_new_dep;
5114 init_dep (new_dep, insn, jump, REG_DEP_ANTI);
5115 sd_add_dep (new_dep, false);
5118 while (1);
5120 gcc_assert (!sd_lists_empty_p (jump, SD_LIST_BACK));
5123 /* Return the NOTE_INSN_BASIC_BLOCK of BB. */
5125 bb_note (basic_block bb)
5127 rtx note;
5129 note = BB_HEAD (bb);
5130 if (LABEL_P (note))
5131 note = NEXT_INSN (note);
5133 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
5134 return note;
5137 #ifdef ENABLE_CHECKING
5138 /* Helper function for check_cfg.
5139 Return nonzero, if edge vector pointed to by EL has edge with TYPE in
5140 its flags. */
5141 static int
5142 has_edge_p (VEC(edge,gc) *el, int type)
5144 edge e;
5145 edge_iterator ei;
5147 FOR_EACH_EDGE (e, ei, el)
5148 if (e->flags & type)
5149 return 1;
5150 return 0;
5153 /* Search back, starting at INSN, for an insn that is not a
5154 NOTE_INSN_VAR_LOCATION. Don't search beyond HEAD, and return it if
5155 no such insn can be found. */
5156 static inline rtx
5157 prev_non_location_insn (rtx insn, rtx head)
5159 while (insn != head && NOTE_P (insn)
5160 && NOTE_KIND (insn) == NOTE_INSN_VAR_LOCATION)
5161 insn = PREV_INSN (insn);
5163 return insn;
5166 /* Check few properties of CFG between HEAD and TAIL.
5167 If HEAD (TAIL) is NULL check from the beginning (till the end) of the
5168 instruction stream. */
5169 static void
5170 check_cfg (rtx head, rtx tail)
5172 rtx next_tail;
5173 basic_block bb = 0;
5174 int not_first = 0, not_last;
5176 if (head == NULL)
5177 head = get_insns ();
5178 if (tail == NULL)
5179 tail = get_last_insn ();
5180 next_tail = NEXT_INSN (tail);
5184 not_last = head != tail;
5186 if (not_first)
5187 gcc_assert (NEXT_INSN (PREV_INSN (head)) == head);
5188 if (not_last)
5189 gcc_assert (PREV_INSN (NEXT_INSN (head)) == head);
5191 if (LABEL_P (head)
5192 || (NOTE_INSN_BASIC_BLOCK_P (head)
5193 && (!not_first
5194 || (not_first && !LABEL_P (PREV_INSN (head))))))
5196 gcc_assert (bb == 0);
5197 bb = BLOCK_FOR_INSN (head);
5198 if (bb != 0)
5199 gcc_assert (BB_HEAD (bb) == head);
5200 else
5201 /* This is the case of jump table. See inside_basic_block_p (). */
5202 gcc_assert (LABEL_P (head) && !inside_basic_block_p (head));
5205 if (bb == 0)
5207 gcc_assert (!inside_basic_block_p (head));
5208 head = NEXT_INSN (head);
5210 else
5212 gcc_assert (inside_basic_block_p (head)
5213 || NOTE_P (head));
5214 gcc_assert (BLOCK_FOR_INSN (head) == bb);
5216 if (LABEL_P (head))
5218 head = NEXT_INSN (head);
5219 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (head));
5221 else
5223 if (control_flow_insn_p (head))
5225 gcc_assert (prev_non_location_insn (BB_END (bb), head)
5226 == head);
5228 if (any_uncondjump_p (head))
5229 gcc_assert (EDGE_COUNT (bb->succs) == 1
5230 && BARRIER_P (NEXT_INSN (head)));
5231 else if (any_condjump_p (head))
5232 gcc_assert (/* Usual case. */
5233 (EDGE_COUNT (bb->succs) > 1
5234 && !BARRIER_P (NEXT_INSN (head)))
5235 /* Or jump to the next instruction. */
5236 || (EDGE_COUNT (bb->succs) == 1
5237 && (BB_HEAD (EDGE_I (bb->succs, 0)->dest)
5238 == JUMP_LABEL (head))));
5240 if (BB_END (bb) == head)
5242 if (EDGE_COUNT (bb->succs) > 1)
5243 gcc_assert (control_flow_insn_p (prev_non_location_insn
5244 (head, BB_HEAD (bb)))
5245 || has_edge_p (bb->succs, EDGE_COMPLEX));
5246 bb = 0;
5249 head = NEXT_INSN (head);
5253 not_first = 1;
5255 while (head != next_tail);
5257 gcc_assert (bb == 0);
5260 #endif /* ENABLE_CHECKING */
5262 /* Extend per basic block data structures. */
5263 static void
5264 extend_bb (void)
5266 if (sched_scan_info->extend_bb)
5267 sched_scan_info->extend_bb ();
5270 /* Init data for BB. */
5271 static void
5272 init_bb (basic_block bb)
5274 if (sched_scan_info->init_bb)
5275 sched_scan_info->init_bb (bb);
5278 /* Extend per insn data structures. */
5279 static void
5280 extend_insn (void)
5282 if (sched_scan_info->extend_insn)
5283 sched_scan_info->extend_insn ();
5286 /* Init data structures for INSN. */
5287 static void
5288 init_insn (rtx insn)
5290 if (sched_scan_info->init_insn)
5291 sched_scan_info->init_insn (insn);
5294 /* Init all insns in BB. */
5295 static void
5296 init_insns_in_bb (basic_block bb)
5298 rtx insn;
5300 FOR_BB_INSNS (bb, insn)
5301 init_insn (insn);
5304 /* A driver function to add a set of basic blocks (BBS),
5305 a single basic block (BB), a set of insns (INSNS) or a single insn (INSN)
5306 to the scheduling region. */
5307 void
5308 sched_scan (const struct sched_scan_info_def *ssi,
5309 bb_vec_t bbs, basic_block bb, insn_vec_t insns, rtx insn)
5311 sched_scan_info = ssi;
5313 if (bbs != NULL || bb != NULL)
5315 extend_bb ();
5317 if (bbs != NULL)
5319 unsigned i;
5320 basic_block x;
5322 for (i = 0; VEC_iterate (basic_block, bbs, i, x); i++)
5323 init_bb (x);
5326 if (bb != NULL)
5327 init_bb (bb);
5330 extend_insn ();
5332 if (bbs != NULL)
5334 unsigned i;
5335 basic_block x;
5337 for (i = 0; VEC_iterate (basic_block, bbs, i, x); i++)
5338 init_insns_in_bb (x);
5341 if (bb != NULL)
5342 init_insns_in_bb (bb);
5344 if (insns != NULL)
5346 unsigned i;
5347 rtx x;
5349 for (i = 0; VEC_iterate (rtx, insns, i, x); i++)
5350 init_insn (x);
5353 if (insn != NULL)
5354 init_insn (insn);
5358 /* Extend data structures for logical insn UID. */
5359 static void
5360 luids_extend_insn (void)
5362 int new_luids_max_uid = get_max_uid () + 1;
5364 VEC_safe_grow_cleared (int, heap, sched_luids, new_luids_max_uid);
5367 /* Initialize LUID for INSN. */
5368 static void
5369 luids_init_insn (rtx insn)
5371 int i = INSN_P (insn) ? 1 : common_sched_info->luid_for_non_insn (insn);
5372 int luid;
5374 if (i >= 0)
5376 luid = sched_max_luid;
5377 sched_max_luid += i;
5379 else
5380 luid = -1;
5382 SET_INSN_LUID (insn, luid);
5385 /* Initialize luids for BBS, BB, INSNS and INSN.
5386 The hook common_sched_info->luid_for_non_insn () is used to determine
5387 if notes, labels, etc. need luids. */
5388 void
5389 sched_init_luids (bb_vec_t bbs, basic_block bb, insn_vec_t insns, rtx insn)
5391 const struct sched_scan_info_def ssi =
5393 NULL, /* extend_bb */
5394 NULL, /* init_bb */
5395 luids_extend_insn, /* extend_insn */
5396 luids_init_insn /* init_insn */
5399 sched_scan (&ssi, bbs, bb, insns, insn);
5402 /* Free LUIDs. */
5403 void
5404 sched_finish_luids (void)
5406 VEC_free (int, heap, sched_luids);
5407 sched_max_luid = 1;
5410 /* Return logical uid of INSN. Helpful while debugging. */
5412 insn_luid (rtx insn)
5414 return INSN_LUID (insn);
5417 /* Extend per insn data in the target. */
5418 void
5419 sched_extend_target (void)
5421 if (targetm.sched.h_i_d_extended)
5422 targetm.sched.h_i_d_extended ();
5425 /* Extend global scheduler structures (those, that live across calls to
5426 schedule_block) to include information about just emitted INSN. */
5427 static void
5428 extend_h_i_d (void)
5430 int reserve = (get_max_uid () + 1
5431 - VEC_length (haifa_insn_data_def, h_i_d));
5432 if (reserve > 0
5433 && ! VEC_space (haifa_insn_data_def, h_i_d, reserve))
5435 VEC_safe_grow_cleared (haifa_insn_data_def, heap, h_i_d,
5436 3 * get_max_uid () / 2);
5437 sched_extend_target ();
5441 /* Initialize h_i_d entry of the INSN with default values.
5442 Values, that are not explicitly initialized here, hold zero. */
5443 static void
5444 init_h_i_d (rtx insn)
5446 if (INSN_LUID (insn) > 0)
5448 INSN_COST (insn) = -1;
5449 QUEUE_INDEX (insn) = QUEUE_NOWHERE;
5450 INSN_TICK (insn) = INVALID_TICK;
5451 INTER_TICK (insn) = INVALID_TICK;
5452 TODO_SPEC (insn) = HARD_DEP;
5456 /* Initialize haifa_insn_data for BBS, BB, INSNS and INSN. */
5457 void
5458 haifa_init_h_i_d (bb_vec_t bbs, basic_block bb, insn_vec_t insns, rtx insn)
5460 const struct sched_scan_info_def ssi =
5462 NULL, /* extend_bb */
5463 NULL, /* init_bb */
5464 extend_h_i_d, /* extend_insn */
5465 init_h_i_d /* init_insn */
5468 sched_scan (&ssi, bbs, bb, insns, insn);
5471 /* Finalize haifa_insn_data. */
5472 void
5473 haifa_finish_h_i_d (void)
5475 int i;
5476 haifa_insn_data_t data;
5477 struct reg_use_data *use, *next;
5479 for (i = 0; VEC_iterate (haifa_insn_data_def, h_i_d, i, data); i++)
5481 if (data->reg_pressure != NULL)
5482 free (data->reg_pressure);
5483 for (use = data->reg_use_list; use != NULL; use = next)
5485 next = use->next_insn_use;
5486 free (use);
5489 VEC_free (haifa_insn_data_def, heap, h_i_d);
5492 /* Init data for the new insn INSN. */
5493 static void
5494 haifa_init_insn (rtx insn)
5496 gcc_assert (insn != NULL);
5498 sched_init_luids (NULL, NULL, NULL, insn);
5499 sched_extend_target ();
5500 sched_deps_init (false);
5501 haifa_init_h_i_d (NULL, NULL, NULL, insn);
5503 if (adding_bb_to_current_region_p)
5505 sd_init_insn (insn);
5507 /* Extend dependency caches by one element. */
5508 extend_dependency_caches (1, false);
5512 /* Init data for the new basic block BB which comes after AFTER. */
5513 static void
5514 haifa_init_only_bb (basic_block bb, basic_block after)
5516 gcc_assert (bb != NULL);
5518 sched_init_bbs ();
5520 if (common_sched_info->add_block)
5521 /* This changes only data structures of the front-end. */
5522 common_sched_info->add_block (bb, after);
5525 /* A generic version of sched_split_block (). */
5526 basic_block
5527 sched_split_block_1 (basic_block first_bb, rtx after)
5529 edge e;
5531 e = split_block (first_bb, after);
5532 gcc_assert (e->src == first_bb);
5534 /* sched_split_block emits note if *check == BB_END. Probably it
5535 is better to rip that note off. */
5537 return e->dest;
5540 /* A generic version of sched_create_empty_bb (). */
5541 basic_block
5542 sched_create_empty_bb_1 (basic_block after)
5544 return create_empty_bb (after);
5547 /* Insert PAT as an INSN into the schedule and update the necessary data
5548 structures to account for it. */
5550 sched_emit_insn (rtx pat)
5552 rtx insn = emit_insn_after (pat, last_scheduled_insn);
5553 last_scheduled_insn = insn;
5554 haifa_init_insn (insn);
5555 return insn;
5558 #endif /* INSN_SCHEDULING */