1/* Instruction scheduling pass.
2 Copyright (C) 1992-2024 Free Software Foundation, Inc.
3 Contributed by Michael Tiemann (tiemann@cygnus.com) Enhanced by,
4 and currently maintained by, Jim Wilson (wilson@cygnus.com)
5
6This file is part of GCC.
7
8GCC is free software; you can redistribute it and/or modify it under
9the terms of the GNU General Public License as published by the Free
10Software Foundation; either version 3, or (at your option) any later
11version.
12
13GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14WARRANTY; without even the implied warranty of MERCHANTABILITY or
15FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16for more details.
17
18You should have received a copy of the GNU General Public License
19along with GCC; see the file COPYING3. If not see
20<http://www.gnu.org/licenses/>. */
21
22/* Instruction scheduling pass. This file, along with sched-deps.cc,
23 contains the generic parts. The actual entry point for
24 the normal instruction scheduling pass is found in sched-rgn.cc.
25
26 We compute insn priorities based on data dependencies. Flow
27 analysis only creates a fraction of the data-dependencies we must
28 observe: namely, only those dependencies which the combiner can be
29 expected to use. For this pass, we must therefore create the
30 remaining dependencies we need to observe: register dependencies,
31 memory dependencies, dependencies to keep function calls in order,
32 and the dependence between a conditional branch and the setting of
33 condition codes are all dealt with here.
34
35 The scheduler first traverses the data flow graph, starting with
36 the last instruction, and proceeding to the first, assigning values
37 to insn_priority as it goes. This sorts the instructions
38 topologically by data dependence.
39
40 Once priorities have been established, we order the insns using
41 list scheduling. This works as follows: starting with a list of
42 all the ready insns, and sorted according to priority number, we
43 schedule the insn from the end of the list by placing its
44 predecessors in the list according to their priority order. We
45 consider this insn scheduled by setting the pointer to the "end" of
46 the list to point to the previous insn. When an insn has no
47 predecessors, we either queue it until sufficient time has elapsed
48 or add it to the ready list. As the instructions are scheduled or
49 when stalls are introduced, the queue advances and dumps insns into
50 the ready list. When all insns down to the lowest priority have
51 been scheduled, the critical path of the basic block has been made
52 as short as possible. The remaining insns are then scheduled in
53 remaining slots.
54
55 The following list shows the order in which we want to break ties
56 among insns in the ready list:
57
58 1. choose insn with the longest path to end of bb, ties
59 broken by
60 2. choose insn with least contribution to register pressure,
61 ties broken by
62 3. prefer in-block upon interblock motion, ties broken by
63 4. prefer useful upon speculative motion, ties broken by
64 5. choose insn with largest control flow probability, ties
65 broken by
66 6. choose insn with the least dependences upon the previously
67 scheduled insn, or finally
68 7 choose the insn which has the most insns dependent on it.
69 8. choose insn with lowest UID.
70
71 Memory references complicate matters. Only if we can be certain
72 that memory references are not part of the data dependency graph
73 (via true, anti, or output dependence), can we move operations past
74 memory references. To first approximation, reads can be done
75 independently, while writes introduce dependencies. Better
76 approximations will yield fewer dependencies.
77
78 Before reload, an extended analysis of interblock data dependences
79 is required for interblock scheduling. This is performed in
80 compute_block_dependences ().
81
82 Dependencies set up by memory references are treated in exactly the
83 same way as other dependencies, by using insn backward dependences
84 INSN_BACK_DEPS. INSN_BACK_DEPS are translated into forward dependences
85 INSN_FORW_DEPS for the purpose of forward list scheduling.
86
87 Having optimized the critical path, we may have also unduly
88 extended the lifetimes of some registers. If an operation requires
89 that constants be loaded into registers, it is certainly desirable
90 to load those constants as early as necessary, but no earlier.
91 I.e., it will not do to load up a bunch of registers at the
92 beginning of a basic block only to use them at the end, if they
93 could be loaded later, since this may result in excessive register
94 utilization.
95
96 Note that since branches are never in basic blocks, but only end
97 basic blocks, this pass will not move branches. But that is ok,
98 since we can use GNU's delayed branch scheduling pass to take care
99 of this case.
100
101 Also note that no further optimizations based on algebraic
102 identities are performed, so this pass would be a good one to
103 perform instruction splitting, such as breaking up a multiply
104 instruction into shifts and adds where that is profitable.
105
106 Given the memory aliasing analysis that this pass should perform,
107 it should be possible to remove redundant stores to memory, and to
108 load values from registers instead of hitting memory.
109
110 Before reload, speculative insns are moved only if a 'proof' exists
111 that no exception will be caused by this, and if no live registers
112 exist that inhibit the motion (live registers constraints are not
113 represented by data dependence edges).
114
115 This pass must update information that subsequent passes expect to
116 be correct. Namely: reg_n_refs, reg_n_sets, reg_n_deaths,
117 reg_n_calls_crossed, and reg_live_length. Also, BB_HEAD, BB_END.
118
119 The information in the line number notes is carefully retained by
120 this pass. Notes that refer to the starting and ending of
121 exception regions are also carefully retained by this pass. All
122 other NOTE insns are grouped in their same relative order at the
123 beginning of basic blocks and regions that have been scheduled. */
124
125#include "config.h"
126#include "system.h"
127#include "coretypes.h"
128#include "backend.h"
129#include "target.h"
130#include "rtl.h"
131#include "cfghooks.h"
132#include "df.h"
133#include "memmodel.h"
134#include "tm_p.h"
135#include "insn-config.h"
136#include "regs.h"
137#include "ira.h"
138#include "recog.h"
139#include "insn-attr.h"
140#include "cfgrtl.h"
141#include "cfgbuild.h"
142#include "sched-int.h"
143#include "common/common-target.h"
144#include "dbgcnt.h"
145#include "cfgloop.h"
146#include "dumpfile.h"
147#include "print-rtl.h"
148#include "function-abi.h"
149
150#ifdef INSN_SCHEDULING
151
152/* True if we do register pressure relief through live-range
153 shrinkage. */
154static bool live_range_shrinkage_p;
155
156/* Switch on live range shrinkage. */
157void
158initialize_live_range_shrinkage (void)
159{
160 live_range_shrinkage_p = true;
161}
162
163/* Switch off live range shrinkage. */
164void
165finish_live_range_shrinkage (void)
166{
167 live_range_shrinkage_p = false;
168}
169
170/* issue_rate is the number of insns that can be scheduled in the same
171 machine cycle. It can be defined in the config/mach/mach.h file,
172 otherwise we set it to 1. */
173
174int issue_rate;
175
176/* This can be set to true by a backend if the scheduler should not
177 enable a DCE pass. */
178bool sched_no_dce;
179
180/* The current initiation interval used when modulo scheduling. */
181static int modulo_ii;
182
183/* The maximum number of stages we are prepared to handle. */
184static int modulo_max_stages;
185
186/* The number of insns that exist in each iteration of the loop. We use this
187 to detect when we've scheduled all insns from the first iteration. */
188static int modulo_n_insns;
189
190/* The current count of insns in the first iteration of the loop that have
191 already been scheduled. */
192static int modulo_insns_scheduled;
193
194/* The maximum uid of insns from the first iteration of the loop. */
195static int modulo_iter0_max_uid;
196
197/* The number of times we should attempt to backtrack when modulo scheduling.
198 Decreased each time we have to backtrack. */
199static int modulo_backtracks_left;
200
201/* The stage in which the last insn from the original loop was
202 scheduled. */
203static int modulo_last_stage;
204
205/* sched-verbose controls the amount of debugging output the
206 scheduler prints. It is controlled by -fsched-verbose=N:
207 N=0: no debugging output.
208 N=1: default value.
209 N=2: bb's probabilities, detailed ready list info, unit/insn info.
210 N=3: rtl at abort point, control-flow, regions info.
211 N=5: dependences info. */
212int sched_verbose = 0;
213
214/* Debugging file. All printouts are sent to dump. */
215FILE *sched_dump = 0;
216
217/* This is a placeholder for the scheduler parameters common
218 to all schedulers. */
219struct common_sched_info_def *common_sched_info;
220
221#define INSN_TICK(INSN) (HID (INSN)->tick)
222#define INSN_EXACT_TICK(INSN) (HID (INSN)->exact_tick)
223#define INSN_TICK_ESTIMATE(INSN) (HID (INSN)->tick_estimate)
224#define INTER_TICK(INSN) (HID (INSN)->inter_tick)
225#define FEEDS_BACKTRACK_INSN(INSN) (HID (INSN)->feeds_backtrack_insn)
226#define SHADOW_P(INSN) (HID (INSN)->shadow_p)
227#define MUST_RECOMPUTE_SPEC_P(INSN) (HID (INSN)->must_recompute_spec)
228/* Cached cost of the instruction. Use insn_sched_cost to get cost of the
229 insn. -1 here means that the field is not initialized. */
230#define INSN_COST(INSN) (HID (INSN)->cost)
231
232/* If INSN_TICK of an instruction is equal to INVALID_TICK,
233 then it should be recalculated from scratch. */
234#define INVALID_TICK (-(max_insn_queue_index + 1))
235/* The minimal value of the INSN_TICK of an instruction. */
236#define MIN_TICK (-max_insn_queue_index)
237
238/* Original order of insns in the ready list.
239 Used to keep order of normal insns while separating DEBUG_INSNs. */
240#define INSN_RFS_DEBUG_ORIG_ORDER(INSN) (HID (INSN)->rfs_debug_orig_order)
241
242/* The deciding reason for INSN's place in the ready list. */
243#define INSN_LAST_RFS_WIN(INSN) (HID (INSN)->last_rfs_win)
244
245/* List of important notes we must keep around. This is a pointer to the
246 last element in the list. */
247rtx_insn *note_list;
248
249static struct spec_info_def spec_info_var;
250/* Description of the speculative part of the scheduling.
251 If NULL - no speculation. */
252spec_info_t spec_info = NULL;
253
254/* True, if recovery block was added during scheduling of current block.
255 Used to determine, if we need to fix INSN_TICKs. */
256static bool haifa_recovery_bb_recently_added_p;
257
258/* True, if recovery block was added during this scheduling pass.
259 Used to determine if we should have empty memory pools of dependencies
260 after finishing current region. */
261bool haifa_recovery_bb_ever_added_p;
262
263/* Counters of different types of speculative instructions. */
264static int nr_begin_data, nr_be_in_data, nr_begin_control, nr_be_in_control;
265
266/* Array used in {unlink, restore}_bb_notes. */
267static rtx_insn **bb_header = 0;
268
269/* Basic block after which recovery blocks will be created. */
270static basic_block before_recovery;
271
272/* Basic block just before the EXIT_BLOCK and after recovery, if we have
273 created it. */
274basic_block after_recovery;
275
276/* FALSE if we add bb to another region, so we don't need to initialize it. */
277bool adding_bb_to_current_region_p = true;
278
279/* Queues, etc. */
280
281/* An instruction is ready to be scheduled when all insns preceding it
282 have already been scheduled. It is important to ensure that all
283 insns which use its result will not be executed until its result
284 has been computed. An insn is maintained in one of four structures:
285
286 (P) the "Pending" set of insns which cannot be scheduled until
287 their dependencies have been satisfied.
288 (Q) the "Queued" set of insns that can be scheduled when sufficient
289 time has passed.
290 (R) the "Ready" list of unscheduled, uncommitted insns.
291 (S) the "Scheduled" list of insns.
292
293 Initially, all insns are either "Pending" or "Ready" depending on
294 whether their dependencies are satisfied.
295
296 Insns move from the "Ready" list to the "Scheduled" list as they
297 are committed to the schedule. As this occurs, the insns in the
298 "Pending" list have their dependencies satisfied and move to either
299 the "Ready" list or the "Queued" set depending on whether
300 sufficient time has passed to make them ready. As time passes,
301 insns move from the "Queued" set to the "Ready" list.
302
303 The "Pending" list (P) are the insns in the INSN_FORW_DEPS of the
304 unscheduled insns, i.e., those that are ready, queued, and pending.
305 The "Queued" set (Q) is implemented by the variable `insn_queue'.
306 The "Ready" list (R) is implemented by the variables `ready' and
307 `n_ready'.
308 The "Scheduled" list (S) is the new insn chain built by this pass.
309
310 The transition (R->S) is implemented in the scheduling loop in
311 `schedule_block' when the best insn to schedule is chosen.
312 The transitions (P->R and P->Q) are implemented in `schedule_insn' as
313 insns move from the ready list to the scheduled list.
314 The transition (Q->R) is implemented in 'queue_to_insn' as time
315 passes or stalls are introduced. */
316
317/* Implement a circular buffer to delay instructions until sufficient
318 time has passed. For the new pipeline description interface,
319 MAX_INSN_QUEUE_INDEX is a power of two minus one which is not less
320 than maximal time of instruction execution computed by genattr.cc on
321 the base maximal time of functional unit reservations and getting a
322 result. This is the longest time an insn may be queued. */
323
324static rtx_insn_list **insn_queue;
325static int q_ptr = 0;
326static int q_size = 0;
327#define NEXT_Q(X) (((X)+1) & max_insn_queue_index)
328#define NEXT_Q_AFTER(X, C) (((X)+C) & max_insn_queue_index)
329
330#define QUEUE_SCHEDULED (-3)
331#define QUEUE_NOWHERE (-2)
332#define QUEUE_READY (-1)
333/* QUEUE_SCHEDULED - INSN is scheduled.
334 QUEUE_NOWHERE - INSN isn't scheduled yet and is neither in
335 queue or ready list.
336 QUEUE_READY - INSN is in ready list.
337 N >= 0 - INSN queued for X [where NEXT_Q_AFTER (q_ptr, X) == N] cycles. */
338
339#define QUEUE_INDEX(INSN) (HID (INSN)->queue_index)
340
341/* The following variable value refers for all current and future
342 reservations of the processor units. */
343state_t curr_state;
344
345/* The following variable value is size of memory representing all
346 current and future reservations of the processor units. */
347size_t dfa_state_size;
348
349/* The following array is used to find the best insn from ready when
350 the automaton pipeline interface is used. */
351signed char *ready_try = NULL;
352
353/* The ready list. */
354struct ready_list ready = {NULL, .veclen: 0, .first: 0, .n_ready: 0, .n_debug: 0};
355
356/* The pointer to the ready list (to be removed). */
357static struct ready_list *readyp = &ready;
358
359/* Scheduling clock. */
360static int clock_var;
361
362/* Clock at which the previous instruction was issued. */
363static int last_clock_var;
364
365/* Set to true if, when queuing a shadow insn, we discover that it would be
366 scheduled too late. */
367static bool must_backtrack;
368
369/* The following variable value is number of essential insns issued on
370 the current cycle. An insn is essential one if it changes the
371 processors state. */
372int cycle_issued_insns;
373
374/* This records the actual schedule. It is built up during the main phase
375 of schedule_block, and afterwards used to reorder the insns in the RTL. */
376static vec<rtx_insn *> scheduled_insns;
377
378static int may_trap_exp (const_rtx, int);
379
380/* Nonzero iff the address is comprised from at most 1 register. */
381#define CONST_BASED_ADDRESS_P(x) \
382 (REG_P (x) \
383 || ((GET_CODE (x) == PLUS || GET_CODE (x) == MINUS \
384 || (GET_CODE (x) == LO_SUM)) \
385 && (CONSTANT_P (XEXP (x, 0)) \
386 || CONSTANT_P (XEXP (x, 1)))))
387
388/* Returns a class that insn with GET_DEST(insn)=x may belong to,
389 as found by analyzing insn's expression. */
390
391
392static int haifa_luid_for_non_insn (rtx x);
393
394/* Haifa version of sched_info hooks common to all headers. */
395const struct common_sched_info_def haifa_common_sched_info =
396 {
397 NULL, /* fix_recovery_cfg */
398 NULL, /* add_block */
399 NULL, /* estimate_number_of_insns */
400 .luid_for_non_insn: haifa_luid_for_non_insn, /* luid_for_non_insn */
401 .sched_pass_id: SCHED_PASS_UNKNOWN /* sched_pass_id */
402 };
403
404/* Mapping from instruction UID to its Logical UID. */
405vec<int> sched_luids;
406
407/* Next LUID to assign to an instruction. */
408int sched_max_luid = 1;
409
410/* Haifa Instruction Data. */
411vec<haifa_insn_data_def> h_i_d;
412
413void (* sched_init_only_bb) (basic_block, basic_block);
414
415/* Split block function. Different schedulers might use different functions
416 to handle their internal data consistent. */
417basic_block (* sched_split_block) (basic_block, rtx);
418
419/* Create empty basic block after the specified block. */
420basic_block (* sched_create_empty_bb) (basic_block);
421
422/* Return the number of cycles until INSN is expected to be ready.
423 Return zero if it already is. */
424static int
425insn_delay (rtx_insn *insn)
426{
427 return MAX (INSN_TICK (insn) - clock_var, 0);
428}
429
430static int
431may_trap_exp (const_rtx x, int is_store)
432{
433 enum rtx_code code;
434
435 if (x == 0)
436 return TRAP_FREE;
437 code = GET_CODE (x);
438 if (is_store)
439 {
440 if (code == MEM && may_trap_p (x))
441 return TRAP_RISKY;
442 else
443 return TRAP_FREE;
444 }
445 if (code == MEM)
446 {
447 /* The insn uses memory: a volatile load. */
448 if (MEM_VOLATILE_P (x))
449 return IRISKY;
450 /* An exception-free load. */
451 if (!may_trap_p (x))
452 return IFREE;
453 /* A load with 1 base register, to be further checked. */
454 if (CONST_BASED_ADDRESS_P (XEXP (x, 0)))
455 return PFREE_CANDIDATE;
456 /* No info on the load, to be further checked. */
457 return PRISKY_CANDIDATE;
458 }
459 else
460 {
461 const char *fmt;
462 int i, insn_class = TRAP_FREE;
463
464 /* Neither store nor load, check if it may cause a trap. */
465 if (may_trap_p (x))
466 return TRAP_RISKY;
467 /* Recursive step: walk the insn... */
468 fmt = GET_RTX_FORMAT (code);
469 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
470 {
471 if (fmt[i] == 'e')
472 {
473 int tmp_class = may_trap_exp (XEXP (x, i), is_store);
474 insn_class = WORST_CLASS (insn_class, tmp_class);
475 }
476 else if (fmt[i] == 'E')
477 {
478 int j;
479 for (j = 0; j < XVECLEN (x, i); j++)
480 {
481 int tmp_class = may_trap_exp (XVECEXP (x, i, j), is_store);
482 insn_class = WORST_CLASS (insn_class, tmp_class);
483 if (insn_class == TRAP_RISKY || insn_class == IRISKY)
484 break;
485 }
486 }
487 if (insn_class == TRAP_RISKY || insn_class == IRISKY)
488 break;
489 }
490 return insn_class;
491 }
492}
493
494/* Classifies rtx X of an insn for the purpose of verifying that X can be
495 executed speculatively (and consequently the insn can be moved
496 speculatively), by examining X, returning:
497 TRAP_RISKY: store, or risky non-load insn (e.g. division by variable).
498 TRAP_FREE: non-load insn.
499 IFREE: load from a globally safe location.
500 IRISKY: volatile load.
501 PFREE_CANDIDATE, PRISKY_CANDIDATE: load that need to be checked for
502 being either PFREE or PRISKY. */
503
504static int
505haifa_classify_rtx (const_rtx x)
506{
507 int tmp_class = TRAP_FREE;
508 int insn_class = TRAP_FREE;
509 enum rtx_code code;
510
511 if (GET_CODE (x) == PARALLEL)
512 {
513 int i, len = XVECLEN (x, 0);
514
515 for (i = len - 1; i >= 0; i--)
516 {
517 tmp_class = haifa_classify_rtx (XVECEXP (x, 0, i));
518 insn_class = WORST_CLASS (insn_class, tmp_class);
519 if (insn_class == TRAP_RISKY || insn_class == IRISKY)
520 break;
521 }
522 }
523 else
524 {
525 code = GET_CODE (x);
526 switch (code)
527 {
528 case CLOBBER:
529 /* Test if it is a 'store'. */
530 tmp_class = may_trap_exp (XEXP (x, 0), is_store: 1);
531 break;
532 case SET:
533 /* Test if it is a store. */
534 tmp_class = may_trap_exp (SET_DEST (x), is_store: 1);
535 if (tmp_class == TRAP_RISKY)
536 break;
537 /* Test if it is a load. */
538 tmp_class =
539 WORST_CLASS (tmp_class,
540 may_trap_exp (SET_SRC (x), 0));
541 break;
542 case COND_EXEC:
543 tmp_class = haifa_classify_rtx (COND_EXEC_CODE (x));
544 if (tmp_class == TRAP_RISKY)
545 break;
546 tmp_class = WORST_CLASS (tmp_class,
547 may_trap_exp (COND_EXEC_TEST (x), 0));
548 break;
549 case TRAP_IF:
550 tmp_class = TRAP_RISKY;
551 break;
552 default:;
553 }
554 insn_class = tmp_class;
555 }
556
557 return insn_class;
558}
559
560int
561haifa_classify_insn (const_rtx insn)
562{
563 return haifa_classify_rtx (x: PATTERN (insn));
564}
565
566/* After the scheduler initialization function has been called, this function
567 can be called to enable modulo scheduling. II is the initiation interval
568 we should use, it affects the delays for delay_pairs that were recorded as
569 separated by a given number of stages.
570
571 MAX_STAGES provides us with a limit
572 after which we give up scheduling; the caller must have unrolled at least
573 as many copies of the loop body and recorded delay_pairs for them.
574
575 INSNS is the number of real (non-debug) insns in one iteration of
576 the loop. MAX_UID can be used to test whether an insn belongs to
577 the first iteration of the loop; all of them have a uid lower than
578 MAX_UID. */
579void
580set_modulo_params (int ii, int max_stages, int insns, int max_uid)
581{
582 modulo_ii = ii;
583 modulo_max_stages = max_stages;
584 modulo_n_insns = insns;
585 modulo_iter0_max_uid = max_uid;
586 modulo_backtracks_left = param_max_modulo_backtrack_attempts;
587}
588
589/* A structure to record a pair of insns where the first one is a real
590 insn that has delay slots, and the second is its delayed shadow.
591 I1 is scheduled normally and will emit an assembly instruction,
592 while I2 describes the side effect that takes place at the
593 transition between cycles CYCLES and (CYCLES + 1) after I1. */
594struct delay_pair
595{
596 struct delay_pair *next_same_i1;
597 rtx_insn *i1, *i2;
598 int cycles;
599 /* When doing modulo scheduling, we a delay_pair can also be used to
600 show that I1 and I2 are the same insn in a different stage. If that
601 is the case, STAGES will be nonzero. */
602 int stages;
603};
604
605/* Helpers for delay hashing. */
606
607struct delay_i1_hasher : nofree_ptr_hash <delay_pair>
608{
609 typedef void *compare_type;
610 static inline hashval_t hash (const delay_pair *);
611 static inline bool equal (const delay_pair *, const void *);
612};
613
614/* Returns a hash value for X, based on hashing just I1. */
615
616inline hashval_t
617delay_i1_hasher::hash (const delay_pair *x)
618{
619 return htab_hash_pointer (x->i1);
620}
621
622/* Return true if I1 of pair X is the same as that of pair Y. */
623
624inline bool
625delay_i1_hasher::equal (const delay_pair *x, const void *y)
626{
627 return x->i1 == y;
628}
629
630struct delay_i2_hasher : free_ptr_hash <delay_pair>
631{
632 typedef void *compare_type;
633 static inline hashval_t hash (const delay_pair *);
634 static inline bool equal (const delay_pair *, const void *);
635};
636
637/* Returns a hash value for X, based on hashing just I2. */
638
639inline hashval_t
640delay_i2_hasher::hash (const delay_pair *x)
641{
642 return htab_hash_pointer (x->i2);
643}
644
645/* Return true if I2 of pair X is the same as that of pair Y. */
646
647inline bool
648delay_i2_hasher::equal (const delay_pair *x, const void *y)
649{
650 return x->i2 == y;
651}
652
653/* Two hash tables to record delay_pairs, one indexed by I1 and the other
654 indexed by I2. */
655static hash_table<delay_i1_hasher> *delay_htab;
656static hash_table<delay_i2_hasher> *delay_htab_i2;
657
658/* Called through htab_traverse. Walk the hashtable using I2 as
659 index, and delete all elements involving an UID higher than
660 that pointed to by *DATA. */
661int
662haifa_htab_i2_traverse (delay_pair **slot, int *data)
663{
664 int maxuid = *data;
665 struct delay_pair *p = *slot;
666 if (INSN_UID (insn: p->i2) >= maxuid || INSN_UID (insn: p->i1) >= maxuid)
667 {
668 delay_htab_i2->clear_slot (slot);
669 }
670 return 1;
671}
672
673/* Called through htab_traverse. Walk the hashtable using I2 as
674 index, and delete all elements involving an UID higher than
675 that pointed to by *DATA. */
676int
677haifa_htab_i1_traverse (delay_pair **pslot, int *data)
678{
679 int maxuid = *data;
680 struct delay_pair *p, *first, **pprev;
681
682 if (INSN_UID (insn: (*pslot)->i1) >= maxuid)
683 {
684 delay_htab->clear_slot (slot: pslot);
685 return 1;
686 }
687 pprev = &first;
688 for (p = *pslot; p; p = p->next_same_i1)
689 {
690 if (INSN_UID (insn: p->i2) < maxuid)
691 {
692 *pprev = p;
693 pprev = &p->next_same_i1;
694 }
695 }
696 *pprev = NULL;
697 if (first == NULL)
698 delay_htab->clear_slot (slot: pslot);
699 else
700 *pslot = first;
701 return 1;
702}
703
704/* Discard all delay pairs which involve an insn with an UID higher
705 than MAX_UID. */
706void
707discard_delay_pairs_above (int max_uid)
708{
709 delay_htab->traverse <int *, haifa_htab_i1_traverse> (argument: &max_uid);
710 delay_htab_i2->traverse <int *, haifa_htab_i2_traverse> (argument: &max_uid);
711}
712
713/* This function can be called by a port just before it starts the final
714 scheduling pass. It records the fact that an instruction with delay
715 slots has been split into two insns, I1 and I2. The first one will be
716 scheduled normally and initiates the operation. The second one is a
717 shadow which must follow a specific number of cycles after I1; its only
718 purpose is to show the side effect that occurs at that cycle in the RTL.
719 If a JUMP_INSN or a CALL_INSN has been split, I1 should be a normal INSN,
720 while I2 retains the original insn type.
721
722 There are two ways in which the number of cycles can be specified,
723 involving the CYCLES and STAGES arguments to this function. If STAGES
724 is zero, we just use the value of CYCLES. Otherwise, STAGES is a factor
725 which is multiplied by MODULO_II to give the number of cycles. This is
726 only useful if the caller also calls set_modulo_params to enable modulo
727 scheduling. */
728
729void
730record_delay_slot_pair (rtx_insn *i1, rtx_insn *i2, int cycles, int stages)
731{
732 struct delay_pair *p = XNEW (struct delay_pair);
733 struct delay_pair **slot;
734
735 p->i1 = i1;
736 p->i2 = i2;
737 p->cycles = cycles;
738 p->stages = stages;
739
740 if (!delay_htab)
741 {
742 delay_htab = new hash_table<delay_i1_hasher> (10);
743 delay_htab_i2 = new hash_table<delay_i2_hasher> (10);
744 }
745 slot = delay_htab->find_slot_with_hash (comparable: i1, hash: htab_hash_pointer (i1), insert: INSERT);
746 p->next_same_i1 = *slot;
747 *slot = p;
748 slot = delay_htab_i2->find_slot (value: p, insert: INSERT);
749 *slot = p;
750}
751
752/* Examine the delay pair hashtable to see if INSN is a shadow for another,
753 and return the other insn if so. Return NULL otherwise. */
754rtx_insn *
755real_insn_for_shadow (rtx_insn *insn)
756{
757 struct delay_pair *pair;
758
759 if (!delay_htab)
760 return NULL;
761
762 pair = delay_htab_i2->find_with_hash (comparable: insn, hash: htab_hash_pointer (insn));
763 if (!pair || pair->stages > 0)
764 return NULL;
765 return pair->i1;
766}
767
768/* For a pair P of insns, return the fixed distance in cycles from the first
769 insn after which the second must be scheduled. */
770static int
771pair_delay (struct delay_pair *p)
772{
773 if (p->stages == 0)
774 return p->cycles;
775 else
776 return p->stages * modulo_ii;
777}
778
779/* Given an insn INSN, add a dependence on its delayed shadow if it
780 has one. Also try to find situations where shadows depend on each other
781 and add dependencies to the real insns to limit the amount of backtracking
782 needed. */
783void
784add_delay_dependencies (rtx_insn *insn)
785{
786 struct delay_pair *pair;
787 sd_iterator_def sd_it;
788 dep_t dep;
789
790 if (!delay_htab)
791 return;
792
793 pair = delay_htab_i2->find_with_hash (comparable: insn, hash: htab_hash_pointer (insn));
794 if (!pair)
795 return;
796 add_dependence (insn, pair->i1, REG_DEP_ANTI);
797 if (pair->stages)
798 return;
799
800 FOR_EACH_DEP (pair->i2, SD_LIST_BACK, sd_it, dep)
801 {
802 rtx_insn *pro = DEP_PRO (dep);
803 struct delay_pair *other_pair
804 = delay_htab_i2->find_with_hash (comparable: pro, hash: htab_hash_pointer (pro));
805 if (!other_pair || other_pair->stages)
806 continue;
807 if (pair_delay (p: other_pair) >= pair_delay (p: pair))
808 {
809 if (sched_verbose >= 4)
810 {
811 fprintf (stream: sched_dump, format: ";;\tadding dependence %d <- %d\n",
812 INSN_UID (insn: other_pair->i1),
813 INSN_UID (insn: pair->i1));
814 fprintf (stream: sched_dump, format: ";;\tpair1 %d <- %d, cost %d\n",
815 INSN_UID (insn: pair->i1),
816 INSN_UID (insn: pair->i2),
817 pair_delay (p: pair));
818 fprintf (stream: sched_dump, format: ";;\tpair2 %d <- %d, cost %d\n",
819 INSN_UID (insn: other_pair->i1),
820 INSN_UID (insn: other_pair->i2),
821 pair_delay (p: other_pair));
822 }
823 add_dependence (pair->i1, other_pair->i1, REG_DEP_ANTI);
824 }
825 }
826}
827
828/* Forward declarations. */
829
830static int priority (rtx_insn *, bool force_recompute = false);
831static int autopref_rank_for_schedule (const rtx_insn *, const rtx_insn *);
832static int rank_for_schedule (const void *, const void *);
833static void swap_sort (rtx_insn **, int);
834static void queue_insn (rtx_insn *, int, const char *);
835static int schedule_insn (rtx_insn *);
836static void adjust_priority (rtx_insn *);
837static void advance_one_cycle (void);
838static void extend_h_i_d (void);
839
840
841/* Notes handling mechanism:
842 =========================
843 Generally, NOTES are saved before scheduling and restored after scheduling.
844 The scheduler distinguishes between two types of notes:
845
846 (1) LOOP_BEGIN, LOOP_END, SETJMP, EHREGION_BEG, EHREGION_END notes:
847 Before scheduling a region, a pointer to the note is added to the insn
848 that follows or precedes it. (This happens as part of the data dependence
849 computation). After scheduling an insn, the pointer contained in it is
850 used for regenerating the corresponding note (in reemit_notes).
851
852 (2) All other notes (e.g. INSN_DELETED): Before scheduling a block,
853 these notes are put in a list (in rm_other_notes() and
854 unlink_other_notes ()). After scheduling the block, these notes are
855 inserted at the beginning of the block (in schedule_block()). */
856
857static void ready_add (struct ready_list *, rtx_insn *, bool);
858static rtx_insn *ready_remove_first (struct ready_list *);
859static rtx_insn *ready_remove_first_dispatch (struct ready_list *ready);
860
861static void queue_to_ready (struct ready_list *);
862static int early_queue_to_ready (state_t, struct ready_list *);
863
864/* The following functions are used to implement multi-pass scheduling
865 on the first cycle. */
866static rtx_insn *ready_remove (struct ready_list *, int);
867static void ready_remove_insn (rtx_insn *);
868
869static void fix_inter_tick (rtx_insn *, rtx_insn *);
870static int fix_tick_ready (rtx_insn *);
871static void change_queue_index (rtx_insn *, int);
872
873/* The following functions are used to implement scheduling of data/control
874 speculative instructions. */
875
876static void extend_h_i_d (void);
877static void init_h_i_d (rtx_insn *);
878static int haifa_speculate_insn (rtx_insn *, ds_t, rtx *);
879static void generate_recovery_code (rtx_insn *);
880static void process_insn_forw_deps_be_in_spec (rtx_insn *, rtx_insn *, ds_t);
881static void begin_speculative_block (rtx_insn *);
882static void add_to_speculative_block (rtx_insn *);
883static void init_before_recovery (basic_block *);
884static void create_check_block_twin (rtx_insn *, bool);
885static void fix_recovery_deps (basic_block);
886static bool haifa_change_pattern (rtx_insn *, rtx);
887static void dump_new_block_header (int, basic_block, rtx_insn *, rtx_insn *);
888static void restore_bb_notes (basic_block);
889static void fix_jump_move (rtx_insn *);
890static void move_block_after_check (rtx_insn *);
891static void move_succs (vec<edge, va_gc> **, basic_block);
892static void sched_remove_insn (rtx_insn *);
893static void clear_priorities (rtx_insn *, rtx_vec_t *);
894static void calc_priorities (const rtx_vec_t &);
895static void add_jump_dependencies (rtx_insn *, rtx_insn *);
896
897#endif /* INSN_SCHEDULING */
898
899/* Point to state used for the current scheduling pass. */
900struct haifa_sched_info *current_sched_info;
901
902#ifndef INSN_SCHEDULING
903void
904schedule_insns (void)
905{
906}
907#else
908
909/* Do register pressure sensitive insn scheduling if the flag is set
910 up. */
911enum sched_pressure_algorithm sched_pressure;
912
913/* Map regno -> its pressure class. The map defined only when
914 SCHED_PRESSURE != SCHED_PRESSURE_NONE. */
915enum reg_class *sched_regno_pressure_class;
916
917/* The current register pressure. Only elements corresponding pressure
918 classes are defined. */
919static int curr_reg_pressure[N_REG_CLASSES];
920
921/* Saved value of the previous array. */
922static int saved_reg_pressure[N_REG_CLASSES];
923
924/* Register living at given scheduling point. */
925static bitmap curr_reg_live;
926
927/* Saved value of the previous array. */
928static bitmap saved_reg_live;
929
930/* Registers mentioned in the current region. */
931static bitmap region_ref_regs;
932
933/* Temporary bitmap used for SCHED_PRESSURE_MODEL. */
934static bitmap tmp_bitmap;
935
936/* Effective number of available registers of a given class (see comment
937 in sched_pressure_start_bb). */
938static int sched_class_regs_num[N_REG_CLASSES];
939/* The number of registers that the function would need to save before it
940 uses them, and the number of fixed_regs. Helpers for calculating of
941 sched_class_regs_num. */
942static int call_saved_regs_num[N_REG_CLASSES];
943static int fixed_regs_num[N_REG_CLASSES];
944
945/* Initiate register pressure relative info for scheduling the current
946 region. Currently it is only clearing register mentioned in the
947 current region. */
948void
949sched_init_region_reg_pressure_info (void)
950{
951 bitmap_clear (region_ref_regs);
952}
953
954/* PRESSURE[CL] describes the pressure on register class CL. Update it
955 for the birth (if BIRTH_P) or death (if !BIRTH_P) of register REGNO.
956 LIVE tracks the set of live registers; if it is null, assume that
957 every birth or death is genuine. */
958static inline void
959mark_regno_birth_or_death (bitmap live, int *pressure, int regno, bool birth_p)
960{
961 enum reg_class pressure_class;
962
963 pressure_class = sched_regno_pressure_class[regno];
964 if (regno >= FIRST_PSEUDO_REGISTER)
965 {
966 if (pressure_class != NO_REGS)
967 {
968 if (birth_p)
969 {
970 if (!live || bitmap_set_bit (live, regno))
971 pressure[pressure_class]
972 += (ira_reg_class_max_nregs
973 [pressure_class][PSEUDO_REGNO_MODE (regno)]);
974 }
975 else
976 {
977 if (!live || bitmap_clear_bit (live, regno))
978 pressure[pressure_class]
979 -= (ira_reg_class_max_nregs
980 [pressure_class][PSEUDO_REGNO_MODE (regno)]);
981 }
982 }
983 }
984 else if (pressure_class != NO_REGS
985 && ! TEST_HARD_REG_BIT (ira_no_alloc_regs, bit: regno))
986 {
987 if (birth_p)
988 {
989 if (!live || bitmap_set_bit (live, regno))
990 pressure[pressure_class]++;
991 }
992 else
993 {
994 if (!live || bitmap_clear_bit (live, regno))
995 pressure[pressure_class]--;
996 }
997 }
998}
999
1000/* Initiate current register pressure related info from living
1001 registers given by LIVE. */
1002static void
1003initiate_reg_pressure_info (bitmap live)
1004{
1005 int i;
1006 unsigned int j;
1007 bitmap_iterator bi;
1008
1009 for (i = 0; i < ira_pressure_classes_num; i++)
1010 curr_reg_pressure[ira_pressure_classes[i]] = 0;
1011 bitmap_clear (curr_reg_live);
1012 EXECUTE_IF_SET_IN_BITMAP (live, 0, j, bi)
1013 if (sched_pressure == SCHED_PRESSURE_MODEL
1014 || current_nr_blocks == 1
1015 || bitmap_bit_p (region_ref_regs, j))
1016 mark_regno_birth_or_death (live: curr_reg_live, pressure: curr_reg_pressure, regno: j, birth_p: true);
1017}
1018
1019/* Mark registers in X as mentioned in the current region. */
1020static void
1021setup_ref_regs (rtx x)
1022{
1023 int i, j;
1024 const RTX_CODE code = GET_CODE (x);
1025 const char *fmt;
1026
1027 if (REG_P (x))
1028 {
1029 bitmap_set_range (region_ref_regs, REGNO (x), REG_NREGS (x));
1030 return;
1031 }
1032 fmt = GET_RTX_FORMAT (code);
1033 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1034 if (fmt[i] == 'e')
1035 setup_ref_regs (XEXP (x, i));
1036 else if (fmt[i] == 'E')
1037 {
1038 for (j = 0; j < XVECLEN (x, i); j++)
1039 setup_ref_regs (XVECEXP (x, i, j));
1040 }
1041}
1042
1043/* Initiate current register pressure related info at the start of
1044 basic block BB. */
1045static void
1046initiate_bb_reg_pressure_info (basic_block bb)
1047{
1048 unsigned int i ATTRIBUTE_UNUSED;
1049 rtx_insn *insn;
1050
1051 if (current_nr_blocks > 1)
1052 FOR_BB_INSNS (bb, insn)
1053 if (NONDEBUG_INSN_P (insn))
1054 setup_ref_regs (PATTERN (insn));
1055 initiate_reg_pressure_info (live: df_get_live_in (bb));
1056 if (bb_has_eh_pred (bb))
1057 for (i = 0; ; ++i)
1058 {
1059 unsigned int regno = EH_RETURN_DATA_REGNO (i);
1060
1061 if (regno == INVALID_REGNUM)
1062 break;
1063 if (! bitmap_bit_p (df_get_live_in (bb), regno))
1064 mark_regno_birth_or_death (live: curr_reg_live, pressure: curr_reg_pressure,
1065 regno, birth_p: true);
1066 }
1067}
1068
1069/* Save current register pressure related info. */
1070static void
1071save_reg_pressure (void)
1072{
1073 int i;
1074
1075 for (i = 0; i < ira_pressure_classes_num; i++)
1076 saved_reg_pressure[ira_pressure_classes[i]]
1077 = curr_reg_pressure[ira_pressure_classes[i]];
1078 bitmap_copy (saved_reg_live, curr_reg_live);
1079}
1080
1081/* Restore saved register pressure related info. */
1082static void
1083restore_reg_pressure (void)
1084{
1085 int i;
1086
1087 for (i = 0; i < ira_pressure_classes_num; i++)
1088 curr_reg_pressure[ira_pressure_classes[i]]
1089 = saved_reg_pressure[ira_pressure_classes[i]];
1090 bitmap_copy (curr_reg_live, saved_reg_live);
1091}
1092
1093/* Return TRUE if the register is dying after its USE. */
1094static bool
1095dying_use_p (struct reg_use_data *use)
1096{
1097 struct reg_use_data *next;
1098
1099 for (next = use->next_regno_use; next != use; next = next->next_regno_use)
1100 if (NONDEBUG_INSN_P (next->insn)
1101 && QUEUE_INDEX (next->insn) != QUEUE_SCHEDULED)
1102 return false;
1103 return true;
1104}
1105
1106/* Print info about the current register pressure and its excess for
1107 each pressure class. */
1108static void
1109print_curr_reg_pressure (void)
1110{
1111 int i;
1112 enum reg_class cl;
1113
1114 fprintf (stream: sched_dump, format: ";;\t");
1115 for (i = 0; i < ira_pressure_classes_num; i++)
1116 {
1117 cl = ira_pressure_classes[i];
1118 gcc_assert (curr_reg_pressure[cl] >= 0);
1119 fprintf (stream: sched_dump, format: " %s:%d(%d)", reg_class_names[cl],
1120 curr_reg_pressure[cl],
1121 curr_reg_pressure[cl] - sched_class_regs_num[cl]);
1122 }
1123 fprintf (stream: sched_dump, format: "\n");
1124}
1125
1126/* Determine if INSN has a condition that is clobbered if a register
1127 in SET_REGS is modified. */
1128static bool
1129cond_clobbered_p (rtx_insn *insn, HARD_REG_SET set_regs)
1130{
1131 rtx pat = PATTERN (insn);
1132 gcc_assert (GET_CODE (pat) == COND_EXEC);
1133 if (TEST_HARD_REG_BIT (set: set_regs, REGNO (XEXP (COND_EXEC_TEST (pat), 0))))
1134 {
1135 sd_iterator_def sd_it;
1136 dep_t dep;
1137 haifa_change_pattern (insn, ORIG_PAT (insn));
1138 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
1139 DEP_STATUS (dep) &= ~DEP_CANCELLED;
1140 TODO_SPEC (insn) = HARD_DEP;
1141 if (sched_verbose >= 2)
1142 fprintf (stream: sched_dump,
1143 format: ";;\t\tdequeue insn %s because of clobbered condition\n",
1144 (*current_sched_info->print_insn) (insn, 0));
1145 return true;
1146 }
1147
1148 return false;
1149}
1150
1151/* This function should be called after modifying the pattern of INSN,
1152 to update scheduler data structures as needed. */
1153static void
1154update_insn_after_change (rtx_insn *insn)
1155{
1156 sd_iterator_def sd_it;
1157 dep_t dep;
1158
1159 dfa_clear_single_insn_cache (insn);
1160
1161 sd_it = sd_iterator_start (insn,
1162 SD_LIST_FORW | SD_LIST_BACK | SD_LIST_RES_BACK);
1163 while (sd_iterator_cond (it_ptr: &sd_it, dep_ptr: &dep))
1164 {
1165 DEP_COST (dep) = UNKNOWN_DEP_COST;
1166 sd_iterator_next (it_ptr: &sd_it);
1167 }
1168
1169 /* Invalidate INSN_COST, so it'll be recalculated. */
1170 INSN_COST (insn) = -1;
1171 /* Invalidate INSN_TICK, so it'll be recalculated. */
1172 INSN_TICK (insn) = INVALID_TICK;
1173
1174 /* Invalidate autoprefetch data entry. */
1175 INSN_AUTOPREF_MULTIPASS_DATA (insn)[0].status
1176 = AUTOPREF_MULTIPASS_DATA_UNINITIALIZED;
1177 INSN_AUTOPREF_MULTIPASS_DATA (insn)[1].status
1178 = AUTOPREF_MULTIPASS_DATA_UNINITIALIZED;
1179}
1180
1181
1182/* Two VECs, one to hold dependencies for which pattern replacements
1183 need to be applied or restored at the start of the next cycle, and
1184 another to hold an integer that is either one, to apply the
1185 corresponding replacement, or zero to restore it. */
1186static vec<dep_t> next_cycle_replace_deps;
1187static vec<int> next_cycle_apply;
1188
1189static void apply_replacement (dep_t, bool);
1190static void restore_pattern (dep_t, bool);
1191
1192/* Look at the remaining dependencies for insn NEXT, and compute and return
1193 the TODO_SPEC value we should use for it. This is called after one of
1194 NEXT's dependencies has been resolved.
1195 We also perform pattern replacements for predication, and for broken
1196 replacement dependencies. The latter is only done if FOR_BACKTRACK is
1197 false. */
1198
1199static ds_t
1200recompute_todo_spec (rtx_insn *next, bool for_backtrack)
1201{
1202 ds_t new_ds;
1203 sd_iterator_def sd_it;
1204 dep_t dep, modify_dep = NULL;
1205 int n_spec = 0;
1206 int n_control = 0;
1207 int n_replace = 0;
1208 bool first_p = true;
1209
1210 if (sd_lists_empty_p (next, SD_LIST_BACK))
1211 /* NEXT has all its dependencies resolved. */
1212 return 0;
1213
1214 if (!sd_lists_empty_p (next, SD_LIST_HARD_BACK))
1215 return HARD_DEP;
1216
1217 /* If NEXT is intended to sit adjacent to this instruction, we don't
1218 want to try to break any dependencies. Treat it as a HARD_DEP. */
1219 if (SCHED_GROUP_P (next))
1220 return HARD_DEP;
1221
1222 /* Now we've got NEXT with speculative deps only.
1223 1. Look at the deps to see what we have to do.
1224 2. Check if we can do 'todo'. */
1225 new_ds = 0;
1226
1227 FOR_EACH_DEP (next, SD_LIST_BACK, sd_it, dep)
1228 {
1229 rtx_insn *pro = DEP_PRO (dep);
1230 ds_t ds = DEP_STATUS (dep) & SPECULATIVE;
1231
1232 if (DEBUG_INSN_P (pro) && !DEBUG_INSN_P (next))
1233 continue;
1234
1235 if (ds)
1236 {
1237 n_spec++;
1238 if (first_p)
1239 {
1240 first_p = false;
1241
1242 new_ds = ds;
1243 }
1244 else
1245 new_ds = ds_merge (new_ds, ds);
1246 }
1247 else if (DEP_TYPE (dep) == REG_DEP_CONTROL)
1248 {
1249 if (QUEUE_INDEX (pro) != QUEUE_SCHEDULED)
1250 {
1251 n_control++;
1252 modify_dep = dep;
1253 }
1254 DEP_STATUS (dep) &= ~DEP_CANCELLED;
1255 }
1256 else if (DEP_REPLACE (dep) != NULL)
1257 {
1258 if (QUEUE_INDEX (pro) != QUEUE_SCHEDULED)
1259 {
1260 n_replace++;
1261 modify_dep = dep;
1262 }
1263 DEP_STATUS (dep) &= ~DEP_CANCELLED;
1264 }
1265 }
1266
1267 if (n_replace > 0 && n_control == 0 && n_spec == 0)
1268 {
1269 if (!dbg_cnt (index: sched_breakdep))
1270 return HARD_DEP;
1271 FOR_EACH_DEP (next, SD_LIST_BACK, sd_it, dep)
1272 {
1273 struct dep_replacement *desc = DEP_REPLACE (dep);
1274 if (desc != NULL)
1275 {
1276 if (desc->insn == next && !for_backtrack)
1277 {
1278 gcc_assert (n_replace == 1);
1279 apply_replacement (dep, true);
1280 }
1281 DEP_STATUS (dep) |= DEP_CANCELLED;
1282 }
1283 }
1284 return 0;
1285 }
1286
1287 else if (n_control == 1 && n_replace == 0 && n_spec == 0)
1288 {
1289 rtx_insn *pro, *other;
1290 rtx new_pat;
1291 rtx cond = NULL_RTX;
1292 bool success;
1293 rtx_insn *prev = NULL;
1294 int i;
1295 unsigned regno;
1296
1297 if ((current_sched_info->flags & DO_PREDICATION) == 0
1298 || (ORIG_PAT (next) != NULL_RTX
1299 && PREDICATED_PAT (next) == NULL_RTX))
1300 return HARD_DEP;
1301
1302 pro = DEP_PRO (modify_dep);
1303 other = real_insn_for_shadow (insn: pro);
1304 if (other != NULL_RTX)
1305 pro = other;
1306
1307 cond = sched_get_reverse_condition_uncached (pro);
1308 regno = REGNO (XEXP (cond, 0));
1309
1310 /* Find the last scheduled insn that modifies the condition register.
1311 We can stop looking once we find the insn we depend on through the
1312 REG_DEP_CONTROL; if the condition register isn't modified after it,
1313 we know that it still has the right value. */
1314 if (QUEUE_INDEX (pro) == QUEUE_SCHEDULED)
1315 FOR_EACH_VEC_ELT_REVERSE (scheduled_insns, i, prev)
1316 {
1317 HARD_REG_SET t;
1318
1319 find_all_hard_reg_sets (prev, &t, true);
1320 if (TEST_HARD_REG_BIT (set: t, bit: regno))
1321 return HARD_DEP;
1322 if (prev == pro)
1323 break;
1324 }
1325 if (ORIG_PAT (next) == NULL_RTX)
1326 {
1327 ORIG_PAT (next) = PATTERN (insn: next);
1328
1329 new_pat = gen_rtx_COND_EXEC (VOIDmode, cond, PATTERN (next));
1330 success = haifa_change_pattern (next, new_pat);
1331 if (!success)
1332 return HARD_DEP;
1333 PREDICATED_PAT (next) = new_pat;
1334 }
1335 else if (PATTERN (insn: next) != PREDICATED_PAT (next))
1336 {
1337 bool success = haifa_change_pattern (next,
1338 PREDICATED_PAT (next));
1339 gcc_assert (success);
1340 }
1341 DEP_STATUS (modify_dep) |= DEP_CANCELLED;
1342 return DEP_CONTROL;
1343 }
1344
1345 if (PREDICATED_PAT (next) != NULL_RTX)
1346 {
1347 int tick = INSN_TICK (next);
1348 bool success = haifa_change_pattern (next,
1349 ORIG_PAT (next));
1350 INSN_TICK (next) = tick;
1351 gcc_assert (success);
1352 }
1353
1354 /* We can't handle the case where there are both speculative and control
1355 dependencies, so we return HARD_DEP in such a case. Also fail if
1356 we have speculative dependencies with not enough points, or more than
1357 one control dependency. */
1358 if ((n_spec > 0 && (n_control > 0 || n_replace > 0))
1359 || (n_spec > 0
1360 /* Too few points? */
1361 && ds_weak (new_ds) < spec_info->data_weakness_cutoff)
1362 || n_control > 0
1363 || n_replace > 0)
1364 return HARD_DEP;
1365
1366 return new_ds;
1367}
1368
1369/* Pointer to the last instruction scheduled. */
1370static rtx_insn *last_scheduled_insn;
1371
1372/* Pointer to the last nondebug instruction scheduled within the
1373 block, or the prev_head of the scheduling block. Used by
1374 rank_for_schedule, so that insns independent of the last scheduled
1375 insn will be preferred over dependent instructions. */
1376static rtx_insn *last_nondebug_scheduled_insn;
1377
1378/* Pointer that iterates through the list of unscheduled insns if we
1379 have a dbg_cnt enabled. It always points at an insn prior to the
1380 first unscheduled one. */
1381static rtx_insn *nonscheduled_insns_begin;
1382
1383/* Compute cost of executing INSN.
1384 This is the number of cycles between instruction issue and
1385 instruction results. */
1386int
1387insn_sched_cost (rtx_insn *insn)
1388{
1389 int cost;
1390
1391 if (sched_fusion)
1392 return 0;
1393
1394 if (sel_sched_p ())
1395 {
1396 if (recog_memoized (insn) < 0)
1397 return 0;
1398
1399 cost = insn_default_latency (insn);
1400 if (cost < 0)
1401 cost = 0;
1402
1403 return cost;
1404 }
1405
1406 cost = INSN_COST (insn);
1407
1408 if (cost < 0)
1409 {
1410 /* A USE insn, or something else we don't need to
1411 understand. We can't pass these directly to
1412 result_ready_cost or insn_default_latency because it will
1413 trigger a fatal error for unrecognizable insns. */
1414 if (recog_memoized (insn) < 0)
1415 {
1416 INSN_COST (insn) = 0;
1417 return 0;
1418 }
1419 else
1420 {
1421 cost = insn_default_latency (insn);
1422 if (cost < 0)
1423 cost = 0;
1424
1425 INSN_COST (insn) = cost;
1426 }
1427 }
1428
1429 return cost;
1430}
1431
1432/* Compute cost of dependence LINK.
1433 This is the number of cycles between instruction issue and
1434 instruction results.
1435 ??? We also use this function to call recog_memoized on all insns. */
1436int
1437dep_cost_1 (dep_t link, dw_t dw)
1438{
1439 rtx_insn *insn = DEP_PRO (link);
1440 rtx_insn *used = DEP_CON (link);
1441 int cost;
1442
1443 if (DEP_COST (link) != UNKNOWN_DEP_COST)
1444 return DEP_COST (link);
1445
1446 if (delay_htab)
1447 {
1448 struct delay_pair *delay_entry;
1449 delay_entry
1450 = delay_htab_i2->find_with_hash (comparable: used, hash: htab_hash_pointer (used));
1451 if (delay_entry)
1452 {
1453 if (delay_entry->i1 == insn)
1454 {
1455 DEP_COST (link) = pair_delay (p: delay_entry);
1456 return DEP_COST (link);
1457 }
1458 }
1459 }
1460
1461 /* A USE insn should never require the value used to be computed.
1462 This allows the computation of a function's result and parameter
1463 values to overlap the return and call. We don't care about the
1464 dependence cost when only decreasing register pressure. */
1465 if (recog_memoized (insn: used) < 0)
1466 {
1467 cost = 0;
1468 recog_memoized (insn);
1469 }
1470 else
1471 {
1472 enum reg_note dep_type = DEP_TYPE (link);
1473
1474 cost = insn_sched_cost (insn);
1475
1476 if (INSN_CODE (insn) >= 0)
1477 {
1478 if (dep_type == REG_DEP_ANTI)
1479 cost = 0;
1480 else if (dep_type == REG_DEP_OUTPUT)
1481 {
1482 cost = (insn_default_latency (insn)
1483 - insn_default_latency (used));
1484 if (cost <= 0)
1485 cost = 1;
1486 }
1487 else if (bypass_p (insn))
1488 cost = insn_latency (insn, used);
1489 }
1490
1491
1492 if (targetm.sched.adjust_cost)
1493 cost = targetm.sched.adjust_cost (used, (int) dep_type, insn, cost,
1494 dw);
1495
1496 if (cost < 0)
1497 cost = 0;
1498 }
1499
1500 DEP_COST (link) = cost;
1501 return cost;
1502}
1503
1504/* Compute cost of dependence LINK.
1505 This is the number of cycles between instruction issue and
1506 instruction results. */
1507int
1508dep_cost (dep_t link)
1509{
1510 return dep_cost_1 (link, dw: 0);
1511}
1512
1513/* Use this sel-sched.cc friendly function in reorder2 instead of increasing
1514 INSN_PRIORITY explicitly. */
1515void
1516increase_insn_priority (rtx_insn *insn, int amount)
1517{
1518 if (!sel_sched_p ())
1519 {
1520 /* We're dealing with haifa-sched.cc INSN_PRIORITY. */
1521 if (INSN_PRIORITY_KNOWN (insn))
1522 INSN_PRIORITY (insn) += amount;
1523 }
1524 else
1525 {
1526 /* In sel-sched.cc INSN_PRIORITY is not kept up to date.
1527 Use EXPR_PRIORITY instead. */
1528 sel_add_to_insn_priority (insn, amount);
1529 }
1530}
1531
1532/* Return 'true' if DEP should be included in priority calculations. */
1533static bool
1534contributes_to_priority_p (dep_t dep)
1535{
1536 if (DEBUG_INSN_P (DEP_CON (dep))
1537 || DEBUG_INSN_P (DEP_PRO (dep)))
1538 return false;
1539
1540 /* Critical path is meaningful in block boundaries only. */
1541 if (!current_sched_info->contributes_to_priority (DEP_CON (dep),
1542 DEP_PRO (dep)))
1543 return false;
1544
1545 if (DEP_REPLACE (dep) != NULL)
1546 return false;
1547
1548 /* If flag COUNT_SPEC_IN_CRITICAL_PATH is set,
1549 then speculative instructions will less likely be
1550 scheduled. That is because the priority of
1551 their producers will increase, and, thus, the
1552 producers will more likely be scheduled, thus,
1553 resolving the dependence. */
1554 if (sched_deps_info->generate_spec_deps
1555 && !(spec_info->flags & COUNT_SPEC_IN_CRITICAL_PATH)
1556 && (DEP_STATUS (dep) & SPECULATIVE))
1557 return false;
1558
1559 return true;
1560}
1561
1562/* Compute the number of nondebug deps in list LIST for INSN. */
1563int
1564dep_list_size (rtx_insn *insn, sd_list_types_def list)
1565{
1566 sd_iterator_def sd_it;
1567 dep_t dep;
1568 int dbgcount = 0, nodbgcount = 0;
1569
1570 if (!MAY_HAVE_DEBUG_INSNS)
1571 return sd_lists_size (insn, list);
1572
1573 /* TODO: We should split normal and debug insns into separate SD_LIST_*
1574 sub-lists, and then we'll be able to use something like
1575 sd_lists_size(insn, list & SD_LIST_NON_DEBUG)
1576 instead of walking dependencies below. */
1577
1578 FOR_EACH_DEP (insn, list, sd_it, dep)
1579 {
1580 if (DEBUG_INSN_P (DEP_CON (dep)))
1581 dbgcount++;
1582 else if (!DEBUG_INSN_P (DEP_PRO (dep)))
1583 nodbgcount++;
1584 }
1585
1586 gcc_assert (dbgcount + nodbgcount == sd_lists_size (insn, list));
1587
1588 return nodbgcount;
1589}
1590
1591bool sched_fusion;
1592
1593/* Compute the priority number for INSN. */
1594static int
1595priority (rtx_insn *insn, bool force_recompute)
1596{
1597 if (! INSN_P (insn))
1598 return 0;
1599
1600 /* We should not be interested in priority of an already scheduled insn. */
1601 gcc_assert (QUEUE_INDEX (insn) != QUEUE_SCHEDULED);
1602
1603 if (force_recompute || !INSN_PRIORITY_KNOWN (insn))
1604 {
1605 int this_priority = -1;
1606
1607 if (sched_fusion)
1608 {
1609 int this_fusion_priority;
1610
1611 targetm.sched.fusion_priority (insn, FUSION_MAX_PRIORITY,
1612 &this_fusion_priority, &this_priority);
1613 INSN_FUSION_PRIORITY (insn) = this_fusion_priority;
1614 }
1615 else if (dep_list_size (insn, SD_LIST_FORW) == 0)
1616 /* ??? We should set INSN_PRIORITY to insn_sched_cost when and insn
1617 has some forward deps but all of them are ignored by
1618 contributes_to_priority hook. At the moment we set priority of
1619 such insn to 0. */
1620 this_priority = insn_sched_cost (insn);
1621 else
1622 {
1623 rtx_insn *prev_first, *twin;
1624 basic_block rec;
1625
1626 /* For recovery check instructions we calculate priority slightly
1627 different than that of normal instructions. Instead of walking
1628 through INSN_FORW_DEPS (check) list, we walk through
1629 INSN_FORW_DEPS list of each instruction in the corresponding
1630 recovery block. */
1631
1632 /* Selective scheduling does not define RECOVERY_BLOCK macro. */
1633 rec = sel_sched_p () ? NULL : RECOVERY_BLOCK (insn);
1634 if (!rec || rec == EXIT_BLOCK_PTR_FOR_FN (cfun))
1635 {
1636 prev_first = PREV_INSN (insn);
1637 twin = insn;
1638 }
1639 else
1640 {
1641 prev_first = NEXT_INSN (BB_HEAD (rec));
1642 twin = PREV_INSN (BB_END (rec));
1643 }
1644
1645 do
1646 {
1647 sd_iterator_def sd_it;
1648 dep_t dep;
1649
1650 FOR_EACH_DEP (twin, SD_LIST_FORW, sd_it, dep)
1651 {
1652 rtx_insn *next;
1653 int next_priority;
1654
1655 next = DEP_CON (dep);
1656
1657 if (BLOCK_FOR_INSN (insn: next) != rec)
1658 {
1659 int cost;
1660
1661 if (!contributes_to_priority_p (dep))
1662 continue;
1663
1664 if (twin == insn)
1665 cost = dep_cost (link: dep);
1666 else
1667 {
1668 struct _dep _dep1, *dep1 = &_dep1;
1669
1670 init_dep (dep1, insn, next, REG_DEP_ANTI);
1671
1672 cost = dep_cost (link: dep1);
1673 }
1674
1675 next_priority = cost + priority (insn: next);
1676
1677 if (next_priority > this_priority)
1678 this_priority = next_priority;
1679 }
1680 }
1681
1682 twin = PREV_INSN (insn: twin);
1683 }
1684 while (twin != prev_first);
1685 }
1686
1687 if (this_priority < 0)
1688 {
1689 gcc_assert (this_priority == -1);
1690
1691 this_priority = insn_sched_cost (insn);
1692 }
1693
1694 INSN_PRIORITY (insn) = this_priority;
1695 INSN_PRIORITY_STATUS (insn) = 1;
1696 }
1697
1698 return INSN_PRIORITY (insn);
1699}
1700
1701/* Macros and functions for keeping the priority queue sorted, and
1702 dealing with queuing and dequeuing of instructions. */
1703
1704/* For each pressure class CL, set DEATH[CL] to the number of registers
1705 in that class that die in INSN. */
1706
1707static void
1708calculate_reg_deaths (rtx_insn *insn, int *death)
1709{
1710 int i;
1711 struct reg_use_data *use;
1712
1713 for (i = 0; i < ira_pressure_classes_num; i++)
1714 death[ira_pressure_classes[i]] = 0;
1715 for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
1716 if (dying_use_p (use))
1717 mark_regno_birth_or_death (live: 0, pressure: death, regno: use->regno, birth_p: true);
1718}
1719
1720/* Setup info about the current register pressure impact of scheduling
1721 INSN at the current scheduling point. */
1722static void
1723setup_insn_reg_pressure_info (rtx_insn *insn)
1724{
1725 int i, change, before, after, hard_regno;
1726 int excess_cost_change;
1727 machine_mode mode;
1728 enum reg_class cl;
1729 struct reg_pressure_data *pressure_info;
1730 int *max_reg_pressure;
1731 static int death[N_REG_CLASSES];
1732
1733 gcc_checking_assert (!DEBUG_INSN_P (insn));
1734
1735 excess_cost_change = 0;
1736 calculate_reg_deaths (insn, death);
1737 pressure_info = INSN_REG_PRESSURE (insn);
1738 max_reg_pressure = INSN_MAX_REG_PRESSURE (insn);
1739 gcc_assert (pressure_info != NULL && max_reg_pressure != NULL);
1740 for (i = 0; i < ira_pressure_classes_num; i++)
1741 {
1742 cl = ira_pressure_classes[i];
1743 gcc_assert (curr_reg_pressure[cl] >= 0);
1744 change = (int) pressure_info[i].set_increase - death[cl];
1745 before = MAX (0, max_reg_pressure[i] - sched_class_regs_num[cl]);
1746 after = MAX (0, max_reg_pressure[i] + change
1747 - sched_class_regs_num[cl]);
1748 hard_regno = ira_class_hard_regs[cl][0];
1749 gcc_assert (hard_regno >= 0);
1750 mode = reg_raw_mode[hard_regno];
1751 excess_cost_change += ((after - before)
1752 * (ira_memory_move_cost[mode][cl][0]
1753 + ira_memory_move_cost[mode][cl][1]));
1754 }
1755 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insn) = excess_cost_change;
1756}
1757
1758/* This is the first page of code related to SCHED_PRESSURE_MODEL.
1759 It tries to make the scheduler take register pressure into account
1760 without introducing too many unnecessary stalls. It hooks into the
1761 main scheduling algorithm at several points:
1762
1763 - Before scheduling starts, model_start_schedule constructs a
1764 "model schedule" for the current block. This model schedule is
1765 chosen solely to keep register pressure down. It does not take the
1766 target's pipeline or the original instruction order into account,
1767 except as a tie-breaker. It also doesn't work to a particular
1768 pressure limit.
1769
1770 This model schedule gives us an idea of what pressure can be
1771 achieved for the block and gives us an example of a schedule that
1772 keeps to that pressure. It also makes the final schedule less
1773 dependent on the original instruction order. This is important
1774 because the original order can either be "wide" (many values live
1775 at once, such as in user-scheduled code) or "narrow" (few values
1776 live at once, such as after loop unrolling, where several
1777 iterations are executed sequentially).
1778
1779 We do not apply this model schedule to the rtx stream. We simply
1780 record it in model_schedule. We also compute the maximum pressure,
1781 MP, that was seen during this schedule.
1782
1783 - Instructions are added to the ready queue even if they require
1784 a stall. The length of the stall is instead computed as:
1785
1786 MAX (INSN_TICK (INSN) - clock_var, 0)
1787
1788 (= insn_delay). This allows rank_for_schedule to choose between
1789 introducing a deliberate stall or increasing pressure.
1790
1791 - Before sorting the ready queue, model_set_excess_costs assigns
1792 a pressure-based cost to each ready instruction in the queue.
1793 This is the instruction's INSN_REG_PRESSURE_EXCESS_COST_CHANGE
1794 (ECC for short) and is effectively measured in cycles.
1795
1796 - rank_for_schedule ranks instructions based on:
1797
1798 ECC (insn) + insn_delay (insn)
1799
1800 then as:
1801
1802 insn_delay (insn)
1803
1804 So, for example, an instruction X1 with an ECC of 1 that can issue
1805 now will win over an instruction X0 with an ECC of zero that would
1806 introduce a stall of one cycle. However, an instruction X2 with an
1807 ECC of 2 that can issue now will lose to both X0 and X1.
1808
1809 - When an instruction is scheduled, model_recompute updates the model
1810 schedule with the new pressures (some of which might now exceed the
1811 original maximum pressure MP). model_update_limit_points then searches
1812 for the new point of maximum pressure, if not already known. */
1813
1814/* Used to separate high-verbosity debug information for SCHED_PRESSURE_MODEL
1815 from surrounding debug information. */
1816#define MODEL_BAR \
1817 ";;\t\t+------------------------------------------------------\n"
1818
1819/* Information about the pressure on a particular register class at a
1820 particular point of the model schedule. */
1821struct model_pressure_data {
1822 /* The pressure at this point of the model schedule, or -1 if the
1823 point is associated with an instruction that has already been
1824 scheduled. */
1825 int ref_pressure;
1826
1827 /* The maximum pressure during or after this point of the model schedule. */
1828 int max_pressure;
1829};
1830
1831/* Per-instruction information that is used while building the model
1832 schedule. Here, "schedule" refers to the model schedule rather
1833 than the main schedule. */
1834struct model_insn_info {
1835 /* The instruction itself. */
1836 rtx_insn *insn;
1837
1838 /* If this instruction is in model_worklist, these fields link to the
1839 previous (higher-priority) and next (lower-priority) instructions
1840 in the list. */
1841 struct model_insn_info *prev;
1842 struct model_insn_info *next;
1843
1844 /* While constructing the schedule, QUEUE_INDEX describes whether an
1845 instruction has already been added to the schedule (QUEUE_SCHEDULED),
1846 is in model_worklist (QUEUE_READY), or neither (QUEUE_NOWHERE).
1847 old_queue records the value that QUEUE_INDEX had before scheduling
1848 started, so that we can restore it once the schedule is complete. */
1849 int old_queue;
1850
1851 /* The relative importance of an unscheduled instruction. Higher
1852 values indicate greater importance. */
1853 unsigned int model_priority;
1854
1855 /* The length of the longest path of satisfied true dependencies
1856 that leads to this instruction. */
1857 unsigned int depth;
1858
1859 /* The length of the longest path of dependencies of any kind
1860 that leads from this instruction. */
1861 unsigned int alap;
1862
1863 /* The number of predecessor nodes that must still be scheduled. */
1864 int unscheduled_preds;
1865};
1866
1867/* Information about the pressure limit for a particular register class.
1868 This structure is used when applying a model schedule to the main
1869 schedule. */
1870struct model_pressure_limit {
1871 /* The maximum register pressure seen in the original model schedule. */
1872 int orig_pressure;
1873
1874 /* The maximum register pressure seen in the current model schedule
1875 (which excludes instructions that have already been scheduled). */
1876 int pressure;
1877
1878 /* The point of the current model schedule at which PRESSURE is first
1879 reached. It is set to -1 if the value needs to be recomputed. */
1880 int point;
1881};
1882
1883/* Describes a particular way of measuring register pressure. */
1884struct model_pressure_group {
1885 /* Index PCI describes the maximum pressure on ira_pressure_classes[PCI]. */
1886 struct model_pressure_limit limits[N_REG_CLASSES];
1887
1888 /* Index (POINT * ira_num_pressure_classes + PCI) describes the pressure
1889 on register class ira_pressure_classes[PCI] at point POINT of the
1890 current model schedule. A POINT of model_num_insns describes the
1891 pressure at the end of the schedule. */
1892 struct model_pressure_data *model;
1893};
1894
1895/* Index POINT gives the instruction at point POINT of the model schedule.
1896 This array doesn't change during main scheduling. */
1897static vec<rtx_insn *> model_schedule;
1898
1899/* The list of instructions in the model worklist, sorted in order of
1900 decreasing priority. */
1901static struct model_insn_info *model_worklist;
1902
1903/* Index I describes the instruction with INSN_LUID I. */
1904static struct model_insn_info *model_insns;
1905
1906/* The number of instructions in the model schedule. */
1907static int model_num_insns;
1908
1909/* The index of the first instruction in model_schedule that hasn't yet been
1910 added to the main schedule, or model_num_insns if all of them have. */
1911static int model_curr_point;
1912
1913/* Describes the pressure before each instruction in the model schedule. */
1914static struct model_pressure_group model_before_pressure;
1915
1916/* The first unused model_priority value (as used in model_insn_info). */
1917static unsigned int model_next_priority;
1918
1919
1920/* The model_pressure_data for ira_pressure_classes[PCI] in GROUP
1921 at point POINT of the model schedule. */
1922#define MODEL_PRESSURE_DATA(GROUP, POINT, PCI) \
1923 (&(GROUP)->model[(POINT) * ira_pressure_classes_num + (PCI)])
1924
1925/* The maximum pressure on ira_pressure_classes[PCI] in GROUP at or
1926 after point POINT of the model schedule. */
1927#define MODEL_MAX_PRESSURE(GROUP, POINT, PCI) \
1928 (MODEL_PRESSURE_DATA (GROUP, POINT, PCI)->max_pressure)
1929
1930/* The pressure on ira_pressure_classes[PCI] in GROUP at point POINT
1931 of the model schedule. */
1932#define MODEL_REF_PRESSURE(GROUP, POINT, PCI) \
1933 (MODEL_PRESSURE_DATA (GROUP, POINT, PCI)->ref_pressure)
1934
1935/* Information about INSN that is used when creating the model schedule. */
1936#define MODEL_INSN_INFO(INSN) \
1937 (&model_insns[INSN_LUID (INSN)])
1938
1939/* The instruction at point POINT of the model schedule. */
1940#define MODEL_INSN(POINT) \
1941 (model_schedule[POINT])
1942
1943
1944/* Return INSN's index in the model schedule, or model_num_insns if it
1945 doesn't belong to that schedule. */
1946
1947static int
1948model_index (rtx_insn *insn)
1949{
1950 if (INSN_MODEL_INDEX (insn) == 0)
1951 return model_num_insns;
1952 return INSN_MODEL_INDEX (insn) - 1;
1953}
1954
1955/* Make sure that GROUP->limits is up-to-date for the current point
1956 of the model schedule. */
1957
1958static void
1959model_update_limit_points_in_group (struct model_pressure_group *group)
1960{
1961 int pci, max_pressure, point;
1962
1963 for (pci = 0; pci < ira_pressure_classes_num; pci++)
1964 {
1965 /* We may have passed the final point at which the pressure in
1966 group->limits[pci].pressure was reached. Update the limit if so. */
1967 max_pressure = MODEL_MAX_PRESSURE (group, model_curr_point, pci);
1968 group->limits[pci].pressure = max_pressure;
1969
1970 /* Find the point at which MAX_PRESSURE is first reached. We need
1971 to search in three cases:
1972
1973 - We've already moved past the previous pressure point.
1974 In this case we search forward from model_curr_point.
1975
1976 - We scheduled the previous point of maximum pressure ahead of
1977 its position in the model schedule, but doing so didn't bring
1978 the pressure point earlier. In this case we search forward
1979 from that previous pressure point.
1980
1981 - Scheduling an instruction early caused the maximum pressure
1982 to decrease. In this case we will have set the pressure
1983 point to -1, and we search forward from model_curr_point. */
1984 point = MAX (group->limits[pci].point, model_curr_point);
1985 while (point < model_num_insns
1986 && MODEL_REF_PRESSURE (group, point, pci) < max_pressure)
1987 point++;
1988 group->limits[pci].point = point;
1989
1990 gcc_assert (MODEL_REF_PRESSURE (group, point, pci) == max_pressure);
1991 gcc_assert (MODEL_MAX_PRESSURE (group, point, pci) == max_pressure);
1992 }
1993}
1994
1995/* Make sure that all register-pressure limits are up-to-date for the
1996 current position in the model schedule. */
1997
1998static void
1999model_update_limit_points (void)
2000{
2001 model_update_limit_points_in_group (group: &model_before_pressure);
2002}
2003
2004/* Return the model_index of the last unscheduled use in chain USE
2005 outside of USE's instruction. Return -1 if there are no other uses,
2006 or model_num_insns if the register is live at the end of the block. */
2007
2008static int
2009model_last_use_except (struct reg_use_data *use)
2010{
2011 struct reg_use_data *next;
2012 int last, index;
2013
2014 last = -1;
2015 for (next = use->next_regno_use; next != use; next = next->next_regno_use)
2016 if (NONDEBUG_INSN_P (next->insn)
2017 && QUEUE_INDEX (next->insn) != QUEUE_SCHEDULED)
2018 {
2019 index = model_index (insn: next->insn);
2020 if (index == model_num_insns)
2021 return model_num_insns;
2022 if (last < index)
2023 last = index;
2024 }
2025 return last;
2026}
2027
2028/* An instruction with model_index POINT has just been scheduled, and it
2029 adds DELTA to the pressure on ira_pressure_classes[PCI] after POINT - 1.
2030 Update MODEL_REF_PRESSURE (GROUP, POINT, PCI) and
2031 MODEL_MAX_PRESSURE (GROUP, POINT, PCI) accordingly. */
2032
2033static void
2034model_start_update_pressure (struct model_pressure_group *group,
2035 int point, int pci, int delta)
2036{
2037 int next_max_pressure;
2038
2039 if (point == model_num_insns)
2040 {
2041 /* The instruction wasn't part of the model schedule; it was moved
2042 from a different block. Update the pressure for the end of
2043 the model schedule. */
2044 MODEL_REF_PRESSURE (group, point, pci) += delta;
2045 MODEL_MAX_PRESSURE (group, point, pci) += delta;
2046 }
2047 else
2048 {
2049 /* Record that this instruction has been scheduled. Nothing now
2050 changes between POINT and POINT + 1, so get the maximum pressure
2051 from the latter. If the maximum pressure decreases, the new
2052 pressure point may be before POINT. */
2053 MODEL_REF_PRESSURE (group, point, pci) = -1;
2054 next_max_pressure = MODEL_MAX_PRESSURE (group, point + 1, pci);
2055 if (MODEL_MAX_PRESSURE (group, point, pci) > next_max_pressure)
2056 {
2057 MODEL_MAX_PRESSURE (group, point, pci) = next_max_pressure;
2058 if (group->limits[pci].point == point)
2059 group->limits[pci].point = -1;
2060 }
2061 }
2062}
2063
2064/* Record that scheduling a later instruction has changed the pressure
2065 at point POINT of the model schedule by DELTA (which might be 0).
2066 Update GROUP accordingly. Return nonzero if these changes might
2067 trigger changes to previous points as well. */
2068
2069static int
2070model_update_pressure (struct model_pressure_group *group,
2071 int point, int pci, int delta)
2072{
2073 int ref_pressure, max_pressure, next_max_pressure;
2074
2075 /* If POINT hasn't yet been scheduled, update its pressure. */
2076 ref_pressure = MODEL_REF_PRESSURE (group, point, pci);
2077 if (ref_pressure >= 0 && delta != 0)
2078 {
2079 ref_pressure += delta;
2080 MODEL_REF_PRESSURE (group, point, pci) = ref_pressure;
2081
2082 /* Check whether the maximum pressure in the overall schedule
2083 has increased. (This means that the MODEL_MAX_PRESSURE of
2084 every point <= POINT will need to increase too; see below.) */
2085 if (group->limits[pci].pressure < ref_pressure)
2086 group->limits[pci].pressure = ref_pressure;
2087
2088 /* If we are at maximum pressure, and the maximum pressure
2089 point was previously unknown or later than POINT,
2090 bring it forward. */
2091 if (group->limits[pci].pressure == ref_pressure
2092 && !IN_RANGE (group->limits[pci].point, 0, point))
2093 group->limits[pci].point = point;
2094
2095 /* If POINT used to be the point of maximum pressure, but isn't
2096 any longer, we need to recalculate it using a forward walk. */
2097 if (group->limits[pci].pressure > ref_pressure
2098 && group->limits[pci].point == point)
2099 group->limits[pci].point = -1;
2100 }
2101
2102 /* Update the maximum pressure at POINT. Changes here might also
2103 affect the maximum pressure at POINT - 1. */
2104 next_max_pressure = MODEL_MAX_PRESSURE (group, point + 1, pci);
2105 max_pressure = MAX (ref_pressure, next_max_pressure);
2106 if (MODEL_MAX_PRESSURE (group, point, pci) != max_pressure)
2107 {
2108 MODEL_MAX_PRESSURE (group, point, pci) = max_pressure;
2109 return 1;
2110 }
2111 return 0;
2112}
2113
2114/* INSN has just been scheduled. Update the model schedule accordingly. */
2115
2116static void
2117model_recompute (rtx_insn *insn)
2118{
2119 struct {
2120 int last_use;
2121 int regno;
2122 } uses[FIRST_PSEUDO_REGISTER + MAX_RECOG_OPERANDS];
2123 struct reg_use_data *use;
2124 struct reg_pressure_data *reg_pressure;
2125 int delta[N_REG_CLASSES];
2126 int pci, point, mix, new_last, cl, ref_pressure, queue;
2127 unsigned int i, num_uses, num_pending_births;
2128 bool print_p;
2129
2130 /* The destinations of INSN were previously live from POINT onwards, but are
2131 now live from model_curr_point onwards. Set up DELTA accordingly. */
2132 point = model_index (insn);
2133 reg_pressure = INSN_REG_PRESSURE (insn);
2134 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2135 {
2136 cl = ira_pressure_classes[pci];
2137 delta[cl] = reg_pressure[pci].set_increase;
2138 }
2139
2140 /* Record which registers previously died at POINT, but which now die
2141 before POINT. Adjust DELTA so that it represents the effect of
2142 this change after POINT - 1. Set NUM_PENDING_BIRTHS to the number of
2143 registers that will be born in the range [model_curr_point, POINT). */
2144 num_uses = 0;
2145 num_pending_births = 0;
2146 bitmap_clear (tmp_bitmap);
2147 for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
2148 {
2149 new_last = model_last_use_except (use);
2150 if (new_last < point && bitmap_set_bit (tmp_bitmap, use->regno))
2151 {
2152 gcc_assert (num_uses < ARRAY_SIZE (uses));
2153 uses[num_uses].last_use = new_last;
2154 uses[num_uses].regno = use->regno;
2155 /* This register is no longer live after POINT - 1. */
2156 mark_regno_birth_or_death (NULL, pressure: delta, regno: use->regno, birth_p: false);
2157 num_uses++;
2158 if (new_last >= 0)
2159 num_pending_births++;
2160 }
2161 }
2162
2163 /* Update the MODEL_REF_PRESSURE and MODEL_MAX_PRESSURE for POINT.
2164 Also set each group pressure limit for POINT. */
2165 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2166 {
2167 cl = ira_pressure_classes[pci];
2168 model_start_update_pressure (group: &model_before_pressure,
2169 point, pci, delta: delta[cl]);
2170 }
2171
2172 /* Walk the model schedule backwards, starting immediately before POINT. */
2173 print_p = false;
2174 if (point != model_curr_point)
2175 do
2176 {
2177 point--;
2178 insn = MODEL_INSN (point);
2179 queue = QUEUE_INDEX (insn);
2180
2181 if (queue != QUEUE_SCHEDULED)
2182 {
2183 /* DELTA describes the effect of the move on the register pressure
2184 after POINT. Make it describe the effect on the pressure
2185 before POINT. */
2186 i = 0;
2187 while (i < num_uses)
2188 {
2189 if (uses[i].last_use == point)
2190 {
2191 /* This register is now live again. */
2192 mark_regno_birth_or_death (NULL, pressure: delta,
2193 regno: uses[i].regno, birth_p: true);
2194
2195 /* Remove this use from the array. */
2196 uses[i] = uses[num_uses - 1];
2197 num_uses--;
2198 num_pending_births--;
2199 }
2200 else
2201 i++;
2202 }
2203
2204 if (sched_verbose >= 5)
2205 {
2206 if (!print_p)
2207 {
2208 fprintf (stream: sched_dump, MODEL_BAR);
2209 fprintf (stream: sched_dump, format: ";;\t\t| New pressure for model"
2210 " schedule\n");
2211 fprintf (stream: sched_dump, MODEL_BAR);
2212 print_p = true;
2213 }
2214
2215 fprintf (stream: sched_dump, format: ";;\t\t| %3d %4d %-30s ",
2216 point, INSN_UID (insn),
2217 str_pattern_slim (PATTERN (insn)));
2218 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2219 {
2220 cl = ira_pressure_classes[pci];
2221 ref_pressure = MODEL_REF_PRESSURE (&model_before_pressure,
2222 point, pci);
2223 fprintf (stream: sched_dump, format: " %s:[%d->%d]",
2224 reg_class_names[ira_pressure_classes[pci]],
2225 ref_pressure, ref_pressure + delta[cl]);
2226 }
2227 fprintf (stream: sched_dump, format: "\n");
2228 }
2229 }
2230
2231 /* Adjust the pressure at POINT. Set MIX to nonzero if POINT - 1
2232 might have changed as well. */
2233 mix = num_pending_births;
2234 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2235 {
2236 cl = ira_pressure_classes[pci];
2237 mix |= delta[cl];
2238 mix |= model_update_pressure (group: &model_before_pressure,
2239 point, pci, delta: delta[cl]);
2240 }
2241 }
2242 while (mix && point > model_curr_point);
2243
2244 if (print_p)
2245 fprintf (stream: sched_dump, MODEL_BAR);
2246}
2247
2248/* After DEP, which was cancelled, has been resolved for insn NEXT,
2249 check whether the insn's pattern needs restoring. */
2250static bool
2251must_restore_pattern_p (rtx_insn *next, dep_t dep)
2252{
2253 if (QUEUE_INDEX (next) == QUEUE_SCHEDULED)
2254 return false;
2255
2256 if (DEP_TYPE (dep) == REG_DEP_CONTROL)
2257 {
2258 gcc_assert (ORIG_PAT (next) != NULL_RTX);
2259 gcc_assert (next == DEP_CON (dep));
2260 }
2261 else
2262 {
2263 struct dep_replacement *desc = DEP_REPLACE (dep);
2264 if (desc->insn != next)
2265 {
2266 gcc_assert (*desc->loc == desc->orig);
2267 return false;
2268 }
2269 }
2270 return true;
2271}
2272
2273/* model_spill_cost (CL, P, P') returns the cost of increasing the
2274 pressure on CL from P to P'. We use this to calculate a "base ECC",
2275 baseECC (CL, X), for each pressure class CL and each instruction X.
2276 Supposing X changes the pressure on CL from P to P', and that the
2277 maximum pressure on CL in the current model schedule is MP', then:
2278
2279 * if X occurs before or at the next point of maximum pressure in
2280 the model schedule and P' > MP', then:
2281
2282 baseECC (CL, X) = model_spill_cost (CL, MP, P')
2283
2284 The idea is that the pressure after scheduling a fixed set of
2285 instructions -- in this case, the set up to and including the
2286 next maximum pressure point -- is going to be the same regardless
2287 of the order; we simply want to keep the intermediate pressure
2288 under control. Thus X has a cost of zero unless scheduling it
2289 now would exceed MP'.
2290
2291 If all increases in the set are by the same amount, no zero-cost
2292 instruction will ever cause the pressure to exceed MP'. However,
2293 if X is instead moved past an instruction X' with pressure in the
2294 range (MP' - (P' - P), MP'), the pressure at X' will increase
2295 beyond MP'. Since baseECC is very much a heuristic anyway,
2296 it doesn't seem worth the overhead of tracking cases like these.
2297
2298 The cost of exceeding MP' is always based on the original maximum
2299 pressure MP. This is so that going 2 registers over the original
2300 limit has the same cost regardless of whether it comes from two
2301 separate +1 deltas or from a single +2 delta.
2302
2303 * if X occurs after the next point of maximum pressure in the model
2304 schedule and P' > P, then:
2305
2306 baseECC (CL, X) = model_spill_cost (CL, MP, MP' + (P' - P))
2307
2308 That is, if we move X forward across a point of maximum pressure,
2309 and if X increases the pressure by P' - P, then we conservatively
2310 assume that scheduling X next would increase the maximum pressure
2311 by P' - P. Again, the cost of doing this is based on the original
2312 maximum pressure MP, for the same reason as above.
2313
2314 * if P' < P, P > MP, and X occurs at or after the next point of
2315 maximum pressure, then:
2316
2317 baseECC (CL, X) = -model_spill_cost (CL, MAX (MP, P'), P)
2318
2319 That is, if we have already exceeded the original maximum pressure MP,
2320 and if X might reduce the maximum pressure again -- or at least push
2321 it further back, and thus allow more scheduling freedom -- it is given
2322 a negative cost to reflect the improvement.
2323
2324 * otherwise,
2325
2326 baseECC (CL, X) = 0
2327
2328 In this case, X is not expected to affect the maximum pressure MP',
2329 so it has zero cost.
2330
2331 We then create a combined value baseECC (X) that is the sum of
2332 baseECC (CL, X) for each pressure class CL.
2333
2334 baseECC (X) could itself be used as the ECC value described above.
2335 However, this is often too conservative, in the sense that it
2336 tends to make high-priority instructions that increase pressure
2337 wait too long in cases where introducing a spill would be better.
2338 For this reason the final ECC is a priority-adjusted form of
2339 baseECC (X). Specifically, we calculate:
2340
2341 P (X) = INSN_PRIORITY (X) - insn_delay (X) - baseECC (X)
2342 baseP = MAX { P (X) | baseECC (X) <= 0 }
2343
2344 Then:
2345
2346 ECC (X) = MAX (MIN (baseP - P (X), baseECC (X)), 0)
2347
2348 Thus an instruction's effect on pressure is ignored if it has a high
2349 enough priority relative to the ones that don't increase pressure.
2350 Negative values of baseECC (X) do not increase the priority of X
2351 itself, but they do make it harder for other instructions to
2352 increase the pressure further.
2353
2354 This pressure cost is deliberately timid. The intention has been
2355 to choose a heuristic that rarely interferes with the normal list
2356 scheduler in cases where that scheduler would produce good code.
2357 We simply want to curb some of its worst excesses. */
2358
2359/* Return the cost of increasing the pressure in class CL from FROM to TO.
2360
2361 Here we use the very simplistic cost model that every register above
2362 sched_class_regs_num[CL] has a spill cost of 1. We could use other
2363 measures instead, such as one based on MEMORY_MOVE_COST. However:
2364
2365 (1) In order for an instruction to be scheduled, the higher cost
2366 would need to be justified in a single saving of that many stalls.
2367 This is overly pessimistic, because the benefit of spilling is
2368 often to avoid a sequence of several short stalls rather than
2369 a single long one.
2370
2371 (2) The cost is still arbitrary. Because we are not allocating
2372 registers during scheduling, we have no way of knowing for
2373 sure how many memory accesses will be required by each spill,
2374 where the spills will be placed within the block, or even
2375 which block(s) will contain the spills.
2376
2377 So a higher cost than 1 is often too conservative in practice,
2378 forcing blocks to contain unnecessary stalls instead of spill code.
2379 The simple cost below seems to be the best compromise. It reduces
2380 the interference with the normal list scheduler, which helps make
2381 it more suitable for a default-on option. */
2382
2383static int
2384model_spill_cost (int cl, int from, int to)
2385{
2386 from = MAX (from, sched_class_regs_num[cl]);
2387 return MAX (to, from) - from;
2388}
2389
2390/* Return baseECC (ira_pressure_classes[PCI], POINT), given that
2391 P = curr_reg_pressure[ira_pressure_classes[PCI]] and that
2392 P' = P + DELTA. */
2393
2394static int
2395model_excess_group_cost (struct model_pressure_group *group,
2396 int point, int pci, int delta)
2397{
2398 int pressure, cl;
2399
2400 cl = ira_pressure_classes[pci];
2401 if (delta < 0 && point >= group->limits[pci].point)
2402 {
2403 pressure = MAX (group->limits[pci].orig_pressure,
2404 curr_reg_pressure[cl] + delta);
2405 return -model_spill_cost (cl, from: pressure, to: curr_reg_pressure[cl]);
2406 }
2407
2408 if (delta > 0)
2409 {
2410 if (point > group->limits[pci].point)
2411 pressure = group->limits[pci].pressure + delta;
2412 else
2413 pressure = curr_reg_pressure[cl] + delta;
2414
2415 if (pressure > group->limits[pci].pressure)
2416 return model_spill_cost (cl, from: group->limits[pci].orig_pressure,
2417 to: pressure);
2418 }
2419
2420 return 0;
2421}
2422
2423/* Return baseECC (MODEL_INSN (INSN)). Dump the costs to sched_dump
2424 if PRINT_P. */
2425
2426static int
2427model_excess_cost (rtx_insn *insn, bool print_p)
2428{
2429 int point, pci, cl, cost, this_cost, delta;
2430 struct reg_pressure_data *insn_reg_pressure;
2431 int insn_death[N_REG_CLASSES];
2432
2433 calculate_reg_deaths (insn, death: insn_death);
2434 point = model_index (insn);
2435 insn_reg_pressure = INSN_REG_PRESSURE (insn);
2436 cost = 0;
2437
2438 if (print_p)
2439 fprintf (stream: sched_dump, format: ";;\t\t| %3d %4d | %4d %+3d |", point,
2440 INSN_UID (insn), INSN_PRIORITY (insn), insn_delay (insn));
2441
2442 /* Sum up the individual costs for each register class. */
2443 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2444 {
2445 cl = ira_pressure_classes[pci];
2446 delta = insn_reg_pressure[pci].set_increase - insn_death[cl];
2447 this_cost = model_excess_group_cost (group: &model_before_pressure,
2448 point, pci, delta);
2449 cost += this_cost;
2450 if (print_p)
2451 fprintf (stream: sched_dump, format: " %s:[%d base cost %d]",
2452 reg_class_names[cl], delta, this_cost);
2453 }
2454
2455 if (print_p)
2456 fprintf (stream: sched_dump, format: "\n");
2457
2458 return cost;
2459}
2460
2461/* Dump the next points of maximum pressure for GROUP. */
2462
2463static void
2464model_dump_pressure_points (struct model_pressure_group *group)
2465{
2466 int pci, cl;
2467
2468 fprintf (stream: sched_dump, format: ";;\t\t| pressure points");
2469 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2470 {
2471 cl = ira_pressure_classes[pci];
2472 fprintf (stream: sched_dump, format: " %s:[%d->%d at ", reg_class_names[cl],
2473 curr_reg_pressure[cl], group->limits[pci].pressure);
2474 if (group->limits[pci].point < model_num_insns)
2475 fprintf (stream: sched_dump, format: "%d:%d]", group->limits[pci].point,
2476 INSN_UID (MODEL_INSN (group->limits[pci].point)));
2477 else
2478 fprintf (stream: sched_dump, format: "end]");
2479 }
2480 fprintf (stream: sched_dump, format: "\n");
2481}
2482
2483/* Set INSN_REG_PRESSURE_EXCESS_COST_CHANGE for INSNS[0...COUNT-1]. */
2484
2485static void
2486model_set_excess_costs (rtx_insn **insns, int count)
2487{
2488 int i, cost, priority_base, priority;
2489 bool print_p;
2490
2491 /* Record the baseECC value for each instruction in the model schedule,
2492 except that negative costs are converted to zero ones now rather than
2493 later. Do not assign a cost to debug instructions, since they must
2494 not change code-generation decisions. Experiments suggest we also
2495 get better results by not assigning a cost to instructions from
2496 a different block.
2497
2498 Set PRIORITY_BASE to baseP in the block comment above. This is the
2499 maximum priority of the "cheap" instructions, which should always
2500 include the next model instruction. */
2501 priority_base = 0;
2502 print_p = false;
2503 for (i = 0; i < count; i++)
2504 if (INSN_MODEL_INDEX (insns[i]))
2505 {
2506 if (sched_verbose >= 6 && !print_p)
2507 {
2508 fprintf (stream: sched_dump, MODEL_BAR);
2509 fprintf (stream: sched_dump, format: ";;\t\t| Pressure costs for ready queue\n");
2510 model_dump_pressure_points (group: &model_before_pressure);
2511 fprintf (stream: sched_dump, MODEL_BAR);
2512 print_p = true;
2513 }
2514 cost = model_excess_cost (insn: insns[i], print_p);
2515 if (cost <= 0)
2516 {
2517 priority = INSN_PRIORITY (insns[i]) - insn_delay (insn: insns[i]) - cost;
2518 priority_base = MAX (priority_base, priority);
2519 cost = 0;
2520 }
2521 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i]) = cost;
2522 }
2523 if (print_p)
2524 fprintf (stream: sched_dump, MODEL_BAR);
2525
2526 /* Use MAX (baseECC, 0) and baseP to calculcate ECC for each
2527 instruction. */
2528 for (i = 0; i < count; i++)
2529 {
2530 cost = INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i]);
2531 priority = INSN_PRIORITY (insns[i]) - insn_delay (insn: insns[i]);
2532 if (cost > 0 && priority > priority_base)
2533 {
2534 cost += priority_base - priority;
2535 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i]) = MAX (cost, 0);
2536 }
2537 }
2538}
2539
2540
2541/* Enum of rank_for_schedule heuristic decisions. */
2542enum rfs_decision {
2543 RFS_LIVE_RANGE_SHRINK1, RFS_LIVE_RANGE_SHRINK2,
2544 RFS_SCHED_GROUP, RFS_PRESSURE_DELAY, RFS_PRESSURE_TICK,
2545 RFS_FEEDS_BACKTRACK_INSN, RFS_PRIORITY, RFS_AUTOPREF, RFS_SPECULATION,
2546 RFS_SCHED_RANK, RFS_LAST_INSN, RFS_PRESSURE_INDEX,
2547 RFS_DEP_COUNT, RFS_TIE, RFS_FUSION, RFS_COST, RFS_N };
2548
2549/* Corresponding strings for print outs. */
2550static const char *rfs_str[RFS_N] = {
2551 "RFS_LIVE_RANGE_SHRINK1", "RFS_LIVE_RANGE_SHRINK2",
2552 "RFS_SCHED_GROUP", "RFS_PRESSURE_DELAY", "RFS_PRESSURE_TICK",
2553 "RFS_FEEDS_BACKTRACK_INSN", "RFS_PRIORITY", "RFS_AUTOPREF", "RFS_SPECULATION",
2554 "RFS_SCHED_RANK", "RFS_LAST_INSN", "RFS_PRESSURE_INDEX",
2555 "RFS_DEP_COUNT", "RFS_TIE", "RFS_FUSION", "RFS_COST" };
2556
2557/* Statistical breakdown of rank_for_schedule decisions. */
2558struct rank_for_schedule_stats_t { unsigned stats[RFS_N]; };
2559static rank_for_schedule_stats_t rank_for_schedule_stats;
2560
2561/* Return the result of comparing insns TMP and TMP2 and update
2562 Rank_For_Schedule statistics. */
2563static int
2564rfs_result (enum rfs_decision decision, int result, rtx tmp, rtx tmp2)
2565{
2566 ++rank_for_schedule_stats.stats[decision];
2567 if (result < 0)
2568 INSN_LAST_RFS_WIN (tmp) = decision;
2569 else if (result > 0)
2570 INSN_LAST_RFS_WIN (tmp2) = decision;
2571 else
2572 gcc_unreachable ();
2573 return result;
2574}
2575
2576/* Sorting predicate to move DEBUG_INSNs to the top of ready list, while
2577 keeping normal insns in original order. */
2578
2579static int
2580rank_for_schedule_debug (const void *x, const void *y)
2581{
2582 rtx_insn *tmp = *(rtx_insn * const *) y;
2583 rtx_insn *tmp2 = *(rtx_insn * const *) x;
2584
2585 /* Schedule debug insns as early as possible. */
2586 if (DEBUG_INSN_P (tmp) && !DEBUG_INSN_P (tmp2))
2587 return -1;
2588 else if (!DEBUG_INSN_P (tmp) && DEBUG_INSN_P (tmp2))
2589 return 1;
2590 else if (DEBUG_INSN_P (tmp) && DEBUG_INSN_P (tmp2))
2591 return INSN_LUID (tmp) - INSN_LUID (tmp2);
2592 else
2593 return INSN_RFS_DEBUG_ORIG_ORDER (tmp2) - INSN_RFS_DEBUG_ORIG_ORDER (tmp);
2594}
2595
2596/* Returns a positive value if x is preferred; returns a negative value if
2597 y is preferred. Should never return 0, since that will make the sort
2598 unstable. */
2599
2600static int
2601rank_for_schedule (const void *x, const void *y)
2602{
2603 rtx_insn *tmp = *(rtx_insn * const *) y;
2604 rtx_insn *tmp2 = *(rtx_insn * const *) x;
2605 int tmp_class, tmp2_class;
2606 int val, priority_val, info_val, diff;
2607
2608 if (live_range_shrinkage_p)
2609 {
2610 /* Don't use SCHED_PRESSURE_MODEL -- it results in much worse
2611 code. */
2612 gcc_assert (sched_pressure == SCHED_PRESSURE_WEIGHTED);
2613 if ((INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp) < 0
2614 || INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2) < 0)
2615 && (diff = (INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp)
2616 - INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2))) != 0)
2617 return rfs_result (decision: RFS_LIVE_RANGE_SHRINK1, result: diff, tmp, tmp2);
2618 /* Sort by INSN_LUID (original insn order), so that we make the
2619 sort stable. This minimizes instruction movement, thus
2620 minimizing sched's effect on debugging and cross-jumping. */
2621 return rfs_result (decision: RFS_LIVE_RANGE_SHRINK2,
2622 INSN_LUID (tmp) - INSN_LUID (tmp2), tmp, tmp2);
2623 }
2624
2625 /* The insn in a schedule group should be issued the first. */
2626 if (flag_sched_group_heuristic &&
2627 SCHED_GROUP_P (tmp) != SCHED_GROUP_P (tmp2))
2628 return rfs_result (decision: RFS_SCHED_GROUP, SCHED_GROUP_P (tmp2) ? 1 : -1,
2629 tmp, tmp2);
2630
2631 /* Make sure that priority of TMP and TMP2 are initialized. */
2632 gcc_assert (INSN_PRIORITY_KNOWN (tmp) && INSN_PRIORITY_KNOWN (tmp2));
2633
2634 if (sched_fusion)
2635 {
2636 /* The instruction that has the same fusion priority as the last
2637 instruction is the instruction we picked next. If that is not
2638 the case, we sort ready list firstly by fusion priority, then
2639 by priority, and at last by INSN_LUID. */
2640 int a = INSN_FUSION_PRIORITY (tmp);
2641 int b = INSN_FUSION_PRIORITY (tmp2);
2642 int last = -1;
2643
2644 if (last_nondebug_scheduled_insn
2645 && !NOTE_P (last_nondebug_scheduled_insn)
2646 && BLOCK_FOR_INSN (insn: tmp)
2647 == BLOCK_FOR_INSN (insn: last_nondebug_scheduled_insn))
2648 last = INSN_FUSION_PRIORITY (last_nondebug_scheduled_insn);
2649
2650 if (a != last && b != last)
2651 {
2652 if (a == b)
2653 {
2654 a = INSN_PRIORITY (tmp);
2655 b = INSN_PRIORITY (tmp2);
2656 }
2657 if (a != b)
2658 return rfs_result (decision: RFS_FUSION, result: b - a, tmp, tmp2);
2659 else
2660 return rfs_result (decision: RFS_FUSION,
2661 INSN_LUID (tmp) - INSN_LUID (tmp2), tmp, tmp2);
2662 }
2663 else if (a == b)
2664 {
2665 gcc_assert (last_nondebug_scheduled_insn
2666 && !NOTE_P (last_nondebug_scheduled_insn));
2667 last = INSN_PRIORITY (last_nondebug_scheduled_insn);
2668
2669 a = abs (INSN_PRIORITY (tmp) - last);
2670 b = abs (INSN_PRIORITY (tmp2) - last);
2671 if (a != b)
2672 return rfs_result (decision: RFS_FUSION, result: a - b, tmp, tmp2);
2673 else
2674 return rfs_result (decision: RFS_FUSION,
2675 INSN_LUID (tmp) - INSN_LUID (tmp2), tmp, tmp2);
2676 }
2677 else if (a == last)
2678 return rfs_result (decision: RFS_FUSION, result: -1, tmp, tmp2);
2679 else
2680 return rfs_result (decision: RFS_FUSION, result: 1, tmp, tmp2);
2681 }
2682
2683 if (sched_pressure != SCHED_PRESSURE_NONE)
2684 {
2685 /* Prefer insn whose scheduling results in the smallest register
2686 pressure excess. */
2687 if ((diff = (INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp)
2688 + insn_delay (insn: tmp)
2689 - INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2)
2690 - insn_delay (insn: tmp2))))
2691 return rfs_result (decision: RFS_PRESSURE_DELAY, result: diff, tmp, tmp2);
2692 }
2693
2694 if (sched_pressure != SCHED_PRESSURE_NONE
2695 && (INSN_TICK (tmp2) > clock_var || INSN_TICK (tmp) > clock_var)
2696 && INSN_TICK (tmp2) != INSN_TICK (tmp))
2697 {
2698 diff = INSN_TICK (tmp) - INSN_TICK (tmp2);
2699 return rfs_result (decision: RFS_PRESSURE_TICK, result: diff, tmp, tmp2);
2700 }
2701
2702 /* If we are doing backtracking in this schedule, prefer insns that
2703 have forward dependencies with negative cost against an insn that
2704 was already scheduled. */
2705 if (current_sched_info->flags & DO_BACKTRACKING)
2706 {
2707 priority_val = FEEDS_BACKTRACK_INSN (tmp2) - FEEDS_BACKTRACK_INSN (tmp);
2708 if (priority_val)
2709 return rfs_result (decision: RFS_FEEDS_BACKTRACK_INSN, result: priority_val, tmp, tmp2);
2710 }
2711
2712 /* Prefer insn with higher priority. */
2713 priority_val = INSN_PRIORITY (tmp2) - INSN_PRIORITY (tmp);
2714
2715 if (flag_sched_critical_path_heuristic && priority_val)
2716 return rfs_result (decision: RFS_PRIORITY, result: priority_val, tmp, tmp2);
2717
2718 if (param_sched_autopref_queue_depth >= 0)
2719 {
2720 int autopref = autopref_rank_for_schedule (tmp, tmp2);
2721 if (autopref != 0)
2722 return rfs_result (decision: RFS_AUTOPREF, result: autopref, tmp, tmp2);
2723 }
2724
2725 /* Prefer speculative insn with greater dependencies weakness. */
2726 if (flag_sched_spec_insn_heuristic && spec_info)
2727 {
2728 ds_t ds1, ds2;
2729 dw_t dw1, dw2;
2730 int dw;
2731
2732 ds1 = TODO_SPEC (tmp) & SPECULATIVE;
2733 if (ds1)
2734 dw1 = ds_weak (ds1);
2735 else
2736 dw1 = NO_DEP_WEAK;
2737
2738 ds2 = TODO_SPEC (tmp2) & SPECULATIVE;
2739 if (ds2)
2740 dw2 = ds_weak (ds2);
2741 else
2742 dw2 = NO_DEP_WEAK;
2743
2744 dw = dw2 - dw1;
2745 if (dw > (NO_DEP_WEAK / 8) || dw < -(NO_DEP_WEAK / 8))
2746 return rfs_result (decision: RFS_SPECULATION, result: dw, tmp, tmp2);
2747 }
2748
2749 info_val = (*current_sched_info->rank) (tmp, tmp2);
2750 if (flag_sched_rank_heuristic && info_val)
2751 return rfs_result (decision: RFS_SCHED_RANK, result: info_val, tmp, tmp2);
2752
2753 /* Compare insns based on their relation to the last scheduled
2754 non-debug insn. */
2755 if (flag_sched_last_insn_heuristic && last_nondebug_scheduled_insn)
2756 {
2757 dep_t dep1;
2758 dep_t dep2;
2759 rtx_insn *last = last_nondebug_scheduled_insn;
2760
2761 /* Classify the instructions into three classes:
2762 1) Data dependent on last schedule insn.
2763 2) Anti/Output dependent on last scheduled insn.
2764 3) Independent of last scheduled insn, or has latency of one.
2765 Choose the insn from the highest numbered class if different. */
2766 dep1 = sd_find_dep_between (last, tmp, true);
2767
2768 if (dep1 == NULL || dep_cost (link: dep1) == 1)
2769 tmp_class = 3;
2770 else if (/* Data dependence. */
2771 DEP_TYPE (dep1) == REG_DEP_TRUE)
2772 tmp_class = 1;
2773 else
2774 tmp_class = 2;
2775
2776 dep2 = sd_find_dep_between (last, tmp2, true);
2777
2778 if (dep2 == NULL || dep_cost (link: dep2) == 1)
2779 tmp2_class = 3;
2780 else if (/* Data dependence. */
2781 DEP_TYPE (dep2) == REG_DEP_TRUE)
2782 tmp2_class = 1;
2783 else
2784 tmp2_class = 2;
2785
2786 if ((val = tmp2_class - tmp_class))
2787 return rfs_result (decision: RFS_LAST_INSN, result: val, tmp, tmp2);
2788 }
2789
2790 /* Prefer instructions that occur earlier in the model schedule. */
2791 if (sched_pressure == SCHED_PRESSURE_MODEL)
2792 {
2793 diff = model_index (insn: tmp) - model_index (insn: tmp2);
2794 if (diff != 0)
2795 return rfs_result (decision: RFS_PRESSURE_INDEX, result: diff, tmp, tmp2);
2796 }
2797
2798 /* Prefer the insn which has more later insns that depend on it.
2799 This gives the scheduler more freedom when scheduling later
2800 instructions at the expense of added register pressure. */
2801
2802 val = (dep_list_size (insn: tmp2, SD_LIST_FORW)
2803 - dep_list_size (insn: tmp, SD_LIST_FORW));
2804
2805 if (flag_sched_dep_count_heuristic && val != 0)
2806 return rfs_result (decision: RFS_DEP_COUNT, result: val, tmp, tmp2);
2807
2808 /* Sort by INSN_COST rather than INSN_LUID. This means that instructions
2809 which take longer to execute are prioritised and it leads to more
2810 dual-issue opportunities on in-order cores which have this feature. */
2811
2812 if (INSN_COST (tmp) != INSN_COST (tmp2))
2813 return rfs_result (decision: RFS_COST, INSN_COST (tmp2) - INSN_COST (tmp),
2814 tmp, tmp2);
2815
2816 /* If insns are equally good, sort by INSN_LUID (original insn order),
2817 so that we make the sort stable. This minimizes instruction movement,
2818 thus minimizing sched's effect on debugging and cross-jumping. */
2819 return rfs_result (decision: RFS_TIE, INSN_LUID (tmp) - INSN_LUID (tmp2), tmp, tmp2);
2820}
2821
2822/* Resort the array A in which only element at index N may be out of order. */
2823
2824HAIFA_INLINE static void
2825swap_sort (rtx_insn **a, int n)
2826{
2827 rtx_insn *insn = a[n - 1];
2828 int i = n - 2;
2829
2830 while (i >= 0 && rank_for_schedule (x: a + i, y: &insn) >= 0)
2831 {
2832 a[i + 1] = a[i];
2833 i -= 1;
2834 }
2835 a[i + 1] = insn;
2836}
2837
2838/* Add INSN to the insn queue so that it can be executed at least
2839 N_CYCLES after the currently executing insn. Preserve insns
2840 chain for debugging purposes. REASON will be printed in debugging
2841 output. */
2842
2843HAIFA_INLINE static void
2844queue_insn (rtx_insn *insn, int n_cycles, const char *reason)
2845{
2846 int next_q = NEXT_Q_AFTER (q_ptr, n_cycles);
2847 rtx_insn_list *link = alloc_INSN_LIST (insn, insn_queue[next_q]);
2848 int new_tick;
2849
2850 gcc_assert (n_cycles <= max_insn_queue_index);
2851 gcc_assert (!DEBUG_INSN_P (insn));
2852
2853 insn_queue[next_q] = link;
2854 q_size += 1;
2855
2856 if (sched_verbose >= 2)
2857 {
2858 fprintf (stream: sched_dump, format: ";;\t\tReady-->Q: insn %s: ",
2859 (*current_sched_info->print_insn) (insn, 0));
2860
2861 fprintf (stream: sched_dump, format: "queued for %d cycles (%s).\n", n_cycles, reason);
2862 }
2863
2864 QUEUE_INDEX (insn) = next_q;
2865
2866 if (current_sched_info->flags & DO_BACKTRACKING)
2867 {
2868 new_tick = clock_var + n_cycles;
2869 if (INSN_TICK (insn) == INVALID_TICK || INSN_TICK (insn) < new_tick)
2870 INSN_TICK (insn) = new_tick;
2871
2872 if (INSN_EXACT_TICK (insn) != INVALID_TICK
2873 && INSN_EXACT_TICK (insn) < clock_var + n_cycles)
2874 {
2875 must_backtrack = true;
2876 if (sched_verbose >= 2)
2877 fprintf (stream: sched_dump, format: ";;\t\tcausing a backtrack.\n");
2878 }
2879 }
2880}
2881
2882/* Remove INSN from queue. */
2883static void
2884queue_remove (rtx_insn *insn)
2885{
2886 gcc_assert (QUEUE_INDEX (insn) >= 0);
2887 remove_free_INSN_LIST_elem (insn, &insn_queue[QUEUE_INDEX (insn)]);
2888 q_size--;
2889 QUEUE_INDEX (insn) = QUEUE_NOWHERE;
2890}
2891
2892/* Return a pointer to the bottom of the ready list, i.e. the insn
2893 with the lowest priority. */
2894
2895rtx_insn **
2896ready_lastpos (struct ready_list *ready)
2897{
2898 gcc_assert (ready->n_ready >= 1);
2899 return ready->vec + ready->first - ready->n_ready + 1;
2900}
2901
2902/* Add an element INSN to the ready list so that it ends up with the
2903 lowest/highest priority depending on FIRST_P. */
2904
2905HAIFA_INLINE static void
2906ready_add (struct ready_list *ready, rtx_insn *insn, bool first_p)
2907{
2908 if (!first_p)
2909 {
2910 if (ready->first == ready->n_ready)
2911 {
2912 memmove (dest: ready->vec + ready->veclen - ready->n_ready,
2913 src: ready_lastpos (ready),
2914 n: ready->n_ready * sizeof (rtx));
2915 ready->first = ready->veclen - 1;
2916 }
2917 ready->vec[ready->first - ready->n_ready] = insn;
2918 }
2919 else
2920 {
2921 if (ready->first == ready->veclen - 1)
2922 {
2923 if (ready->n_ready)
2924 /* ready_lastpos() fails when called with (ready->n_ready == 0). */
2925 memmove (dest: ready->vec + ready->veclen - ready->n_ready - 1,
2926 src: ready_lastpos (ready),
2927 n: ready->n_ready * sizeof (rtx));
2928 ready->first = ready->veclen - 2;
2929 }
2930 ready->vec[++(ready->first)] = insn;
2931 }
2932
2933 ready->n_ready++;
2934 if (DEBUG_INSN_P (insn))
2935 ready->n_debug++;
2936
2937 gcc_assert (QUEUE_INDEX (insn) != QUEUE_READY);
2938 QUEUE_INDEX (insn) = QUEUE_READY;
2939
2940 if (INSN_EXACT_TICK (insn) != INVALID_TICK
2941 && INSN_EXACT_TICK (insn) < clock_var)
2942 {
2943 must_backtrack = true;
2944 }
2945}
2946
2947/* Remove the element with the highest priority from the ready list and
2948 return it. */
2949
2950HAIFA_INLINE static rtx_insn *
2951ready_remove_first (struct ready_list *ready)
2952{
2953 rtx_insn *t;
2954
2955 gcc_assert (ready->n_ready);
2956 t = ready->vec[ready->first--];
2957 ready->n_ready--;
2958 if (DEBUG_INSN_P (t))
2959 ready->n_debug--;
2960 /* If the queue becomes empty, reset it. */
2961 if (ready->n_ready == 0)
2962 ready->first = ready->veclen - 1;
2963
2964 gcc_assert (QUEUE_INDEX (t) == QUEUE_READY);
2965 QUEUE_INDEX (t) = QUEUE_NOWHERE;
2966
2967 return t;
2968}
2969
2970/* The following code implements multi-pass scheduling for the first
2971 cycle. In other words, we will try to choose ready insn which
2972 permits to start maximum number of insns on the same cycle. */
2973
2974/* Return a pointer to the element INDEX from the ready. INDEX for
2975 insn with the highest priority is 0, and the lowest priority has
2976 N_READY - 1. */
2977
2978rtx_insn *
2979ready_element (struct ready_list *ready, int index)
2980{
2981 gcc_assert (ready->n_ready && index < ready->n_ready);
2982
2983 return ready->vec[ready->first - index];
2984}
2985
2986/* Remove the element INDEX from the ready list and return it. INDEX
2987 for insn with the highest priority is 0, and the lowest priority
2988 has N_READY - 1. */
2989
2990HAIFA_INLINE static rtx_insn *
2991ready_remove (struct ready_list *ready, int index)
2992{
2993 rtx_insn *t;
2994 int i;
2995
2996 if (index == 0)
2997 return ready_remove_first (ready);
2998 gcc_assert (ready->n_ready && index < ready->n_ready);
2999 t = ready->vec[ready->first - index];
3000 ready->n_ready--;
3001 if (DEBUG_INSN_P (t))
3002 ready->n_debug--;
3003 for (i = index; i < ready->n_ready; i++)
3004 ready->vec[ready->first - i] = ready->vec[ready->first - i - 1];
3005 QUEUE_INDEX (t) = QUEUE_NOWHERE;
3006 return t;
3007}
3008
3009/* Remove INSN from the ready list. */
3010static void
3011ready_remove_insn (rtx_insn *insn)
3012{
3013 int i;
3014
3015 for (i = 0; i < readyp->n_ready; i++)
3016 if (ready_element (ready: readyp, index: i) == insn)
3017 {
3018 ready_remove (ready: readyp, index: i);
3019 return;
3020 }
3021 gcc_unreachable ();
3022}
3023
3024/* Calculate difference of two statistics set WAS and NOW.
3025 Result returned in WAS. */
3026static void
3027rank_for_schedule_stats_diff (rank_for_schedule_stats_t *was,
3028 const rank_for_schedule_stats_t *now)
3029{
3030 for (int i = 0; i < RFS_N; ++i)
3031 was->stats[i] = now->stats[i] - was->stats[i];
3032}
3033
3034/* Print rank_for_schedule statistics. */
3035static void
3036print_rank_for_schedule_stats (const char *prefix,
3037 const rank_for_schedule_stats_t *stats,
3038 struct ready_list *ready)
3039{
3040 for (int i = 0; i < RFS_N; ++i)
3041 if (stats->stats[i])
3042 {
3043 fprintf (stream: sched_dump, format: "%s%20s: %u", prefix, rfs_str[i], stats->stats[i]);
3044
3045 if (ready != NULL)
3046 /* Print out insns that won due to RFS_<I>. */
3047 {
3048 rtx_insn **p = ready_lastpos (ready);
3049
3050 fprintf (stream: sched_dump, format: ":");
3051 /* Start with 1 since least-priority insn didn't have any wins. */
3052 for (int j = 1; j < ready->n_ready; ++j)
3053 if (INSN_LAST_RFS_WIN (p[j]) == i)
3054 fprintf (stream: sched_dump, format: " %s",
3055 (*current_sched_info->print_insn) (p[j], 0));
3056 }
3057 fprintf (stream: sched_dump, format: "\n");
3058 }
3059}
3060
3061/* Separate DEBUG_INSNS from normal insns. DEBUG_INSNs go to the end
3062 of array. */
3063static void
3064ready_sort_debug (struct ready_list *ready)
3065{
3066 int i;
3067 rtx_insn **first = ready_lastpos (ready);
3068
3069 for (i = 0; i < ready->n_ready; ++i)
3070 if (!DEBUG_INSN_P (first[i]))
3071 INSN_RFS_DEBUG_ORIG_ORDER (first[i]) = i;
3072
3073 qsort (first, ready->n_ready, sizeof (rtx), rank_for_schedule_debug);
3074}
3075
3076/* Sort non-debug insns in the ready list READY by ascending priority.
3077 Assumes that all debug insns are separated from the real insns. */
3078static void
3079ready_sort_real (struct ready_list *ready)
3080{
3081 int i;
3082 rtx_insn **first = ready_lastpos (ready);
3083 int n_ready_real = ready->n_ready - ready->n_debug;
3084
3085 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
3086 for (i = 0; i < n_ready_real; ++i)
3087 setup_insn_reg_pressure_info (first[i]);
3088 else if (sched_pressure == SCHED_PRESSURE_MODEL
3089 && model_curr_point < model_num_insns)
3090 model_set_excess_costs (insns: first, count: n_ready_real);
3091
3092 rank_for_schedule_stats_t stats1;
3093 if (sched_verbose >= 4)
3094 stats1 = rank_for_schedule_stats;
3095
3096 if (n_ready_real == 2)
3097 swap_sort (a: first, n: n_ready_real);
3098 else if (n_ready_real > 2)
3099 qsort (first, n_ready_real, sizeof (rtx), rank_for_schedule);
3100
3101 if (sched_verbose >= 4)
3102 {
3103 rank_for_schedule_stats_diff (was: &stats1, now: &rank_for_schedule_stats);
3104 print_rank_for_schedule_stats (prefix: ";;\t\t", stats: &stats1, ready);
3105 }
3106}
3107
3108/* Sort the ready list READY by ascending priority. */
3109static void
3110ready_sort (struct ready_list *ready)
3111{
3112 if (ready->n_debug > 0)
3113 ready_sort_debug (ready);
3114 else
3115 ready_sort_real (ready);
3116}
3117
3118/* PREV is an insn that is ready to execute. Adjust its priority if that
3119 will help shorten or lengthen register lifetimes as appropriate. Also
3120 provide a hook for the target to tweak itself. */
3121
3122HAIFA_INLINE static void
3123adjust_priority (rtx_insn *prev)
3124{
3125 /* ??? There used to be code here to try and estimate how an insn
3126 affected register lifetimes, but it did it by looking at REG_DEAD
3127 notes, which we removed in schedule_region. Nor did it try to
3128 take into account register pressure or anything useful like that.
3129
3130 Revisit when we have a machine model to work with and not before. */
3131
3132 if (targetm.sched.adjust_priority)
3133 INSN_PRIORITY (prev) =
3134 targetm.sched.adjust_priority (prev, INSN_PRIORITY (prev));
3135}
3136
3137/* Advance DFA state STATE on one cycle. */
3138void
3139advance_state (state_t state)
3140{
3141 if (targetm.sched.dfa_pre_advance_cycle)
3142 targetm.sched.dfa_pre_advance_cycle ();
3143
3144 if (targetm.sched.dfa_pre_cycle_insn)
3145 state_transition (state,
3146 targetm.sched.dfa_pre_cycle_insn ());
3147
3148 state_transition (state, NULL);
3149
3150 if (targetm.sched.dfa_post_cycle_insn)
3151 state_transition (state,
3152 targetm.sched.dfa_post_cycle_insn ());
3153
3154 if (targetm.sched.dfa_post_advance_cycle)
3155 targetm.sched.dfa_post_advance_cycle ();
3156}
3157
3158/* Advance time on one cycle. */
3159HAIFA_INLINE static void
3160advance_one_cycle (void)
3161{
3162 int i;
3163
3164 advance_state (state: curr_state);
3165 for (i = 4; i <= sched_verbose; ++i)
3166 fprintf (stream: sched_dump, format: ";;\tAdvance the current state: %d.\n", clock_var);
3167}
3168
3169/* Update register pressure after scheduling INSN. */
3170static void
3171update_register_pressure (rtx_insn *insn)
3172{
3173 struct reg_use_data *use;
3174 struct reg_set_data *set;
3175
3176 gcc_checking_assert (!DEBUG_INSN_P (insn));
3177
3178 for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
3179 if (dying_use_p (use))
3180 mark_regno_birth_or_death (live: curr_reg_live, pressure: curr_reg_pressure,
3181 regno: use->regno, birth_p: false);
3182 for (set = INSN_REG_SET_LIST (insn); set != NULL; set = set->next_insn_set)
3183 mark_regno_birth_or_death (live: curr_reg_live, pressure: curr_reg_pressure,
3184 regno: set->regno, birth_p: true);
3185}
3186
3187/* Set up or update (if UPDATE_P) max register pressure (see its
3188 meaning in sched-int.h::_haifa_insn_data) for all current BB insns
3189 after insn AFTER. */
3190static void
3191setup_insn_max_reg_pressure (rtx_insn *after, bool update_p)
3192{
3193 int i, p;
3194 bool eq_p;
3195 rtx_insn *insn;
3196 static int max_reg_pressure[N_REG_CLASSES];
3197
3198 save_reg_pressure ();
3199 for (i = 0; i < ira_pressure_classes_num; i++)
3200 max_reg_pressure[ira_pressure_classes[i]]
3201 = curr_reg_pressure[ira_pressure_classes[i]];
3202 for (insn = NEXT_INSN (insn: after);
3203 insn != NULL_RTX && ! BARRIER_P (insn)
3204 && BLOCK_FOR_INSN (insn) == BLOCK_FOR_INSN (insn: after);
3205 insn = NEXT_INSN (insn))
3206 if (NONDEBUG_INSN_P (insn))
3207 {
3208 eq_p = true;
3209 for (i = 0; i < ira_pressure_classes_num; i++)
3210 {
3211 p = max_reg_pressure[ira_pressure_classes[i]];
3212 if (INSN_MAX_REG_PRESSURE (insn)[i] != p)
3213 {
3214 eq_p = false;
3215 INSN_MAX_REG_PRESSURE (insn)[i]
3216 = max_reg_pressure[ira_pressure_classes[i]];
3217 }
3218 }
3219 if (update_p && eq_p)
3220 break;
3221 update_register_pressure (insn);
3222 for (i = 0; i < ira_pressure_classes_num; i++)
3223 if (max_reg_pressure[ira_pressure_classes[i]]
3224 < curr_reg_pressure[ira_pressure_classes[i]])
3225 max_reg_pressure[ira_pressure_classes[i]]
3226 = curr_reg_pressure[ira_pressure_classes[i]];
3227 }
3228 restore_reg_pressure ();
3229}
3230
3231/* Update the current register pressure after scheduling INSN. Update
3232 also max register pressure for unscheduled insns of the current
3233 BB. */
3234static void
3235update_reg_and_insn_max_reg_pressure (rtx_insn *insn)
3236{
3237 int i;
3238 int before[N_REG_CLASSES];
3239
3240 for (i = 0; i < ira_pressure_classes_num; i++)
3241 before[i] = curr_reg_pressure[ira_pressure_classes[i]];
3242 update_register_pressure (insn);
3243 for (i = 0; i < ira_pressure_classes_num; i++)
3244 if (curr_reg_pressure[ira_pressure_classes[i]] != before[i])
3245 break;
3246 if (i < ira_pressure_classes_num)
3247 setup_insn_max_reg_pressure (after: insn, update_p: true);
3248}
3249
3250/* Set up register pressure at the beginning of basic block BB whose
3251 insns starting after insn AFTER. Set up also max register pressure
3252 for all insns of the basic block. */
3253void
3254sched_setup_bb_reg_pressure_info (basic_block bb, rtx_insn *after)
3255{
3256 gcc_assert (sched_pressure == SCHED_PRESSURE_WEIGHTED);
3257 initiate_bb_reg_pressure_info (bb);
3258 setup_insn_max_reg_pressure (after, update_p: false);
3259}
3260
3261/* If doing predication while scheduling, verify whether INSN, which
3262 has just been scheduled, clobbers the conditions of any
3263 instructions that must be predicated in order to break their
3264 dependencies. If so, remove them from the queues so that they will
3265 only be scheduled once their control dependency is resolved. */
3266
3267static void
3268check_clobbered_conditions (rtx_insn *insn)
3269{
3270 HARD_REG_SET t;
3271 int i;
3272
3273 if ((current_sched_info->flags & DO_PREDICATION) == 0)
3274 return;
3275
3276 find_all_hard_reg_sets (insn, &t, true);
3277
3278 restart:
3279 for (i = 0; i < ready.n_ready; i++)
3280 {
3281 rtx_insn *x = ready_element (ready: &ready, index: i);
3282 if (TODO_SPEC (x) == DEP_CONTROL && cond_clobbered_p (insn: x, set_regs: t))
3283 {
3284 ready_remove_insn (insn: x);
3285 goto restart;
3286 }
3287 }
3288 for (i = 0; i <= max_insn_queue_index; i++)
3289 {
3290 rtx_insn_list *link;
3291 int q = NEXT_Q_AFTER (q_ptr, i);
3292
3293 restart_queue:
3294 for (link = insn_queue[q]; link; link = link->next ())
3295 {
3296 rtx_insn *x = link->insn ();
3297 if (TODO_SPEC (x) == DEP_CONTROL && cond_clobbered_p (insn: x, set_regs: t))
3298 {
3299 queue_remove (insn: x);
3300 goto restart_queue;
3301 }
3302 }
3303 }
3304}
3305
3306/* Return (in order):
3307
3308 - positive if INSN adversely affects the pressure on one
3309 register class
3310
3311 - negative if INSN reduces the pressure on one register class
3312
3313 - 0 if INSN doesn't affect the pressure on any register class. */
3314
3315static int
3316model_classify_pressure (struct model_insn_info *insn)
3317{
3318 struct reg_pressure_data *reg_pressure;
3319 int death[N_REG_CLASSES];
3320 int pci, cl, sum;
3321
3322 calculate_reg_deaths (insn: insn->insn, death);
3323 reg_pressure = INSN_REG_PRESSURE (insn->insn);
3324 sum = 0;
3325 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3326 {
3327 cl = ira_pressure_classes[pci];
3328 if (death[cl] < reg_pressure[pci].set_increase)
3329 return 1;
3330 sum += reg_pressure[pci].set_increase - death[cl];
3331 }
3332 return sum;
3333}
3334
3335/* Return true if INSN1 should come before INSN2 in the model schedule. */
3336
3337static int
3338model_order_p (struct model_insn_info *insn1, struct model_insn_info *insn2)
3339{
3340 unsigned int height1, height2;
3341 unsigned int priority1, priority2;
3342
3343 /* Prefer instructions with a higher model priority. */
3344 if (insn1->model_priority != insn2->model_priority)
3345 return insn1->model_priority > insn2->model_priority;
3346
3347 /* Combine the length of the longest path of satisfied true dependencies
3348 that leads to each instruction (depth) with the length of the longest
3349 path of any dependencies that leads from the instruction (alap).
3350 Prefer instructions with the greatest combined length. If the combined
3351 lengths are equal, prefer instructions with the greatest depth.
3352
3353 The idea is that, if we have a set S of "equal" instructions that each
3354 have ALAP value X, and we pick one such instruction I, any true-dependent
3355 successors of I that have ALAP value X - 1 should be preferred over S.
3356 This encourages the schedule to be "narrow" rather than "wide".
3357 However, if I is a low-priority instruction that we decided to
3358 schedule because of its model_classify_pressure, and if there
3359 is a set of higher-priority instructions T, the aforementioned
3360 successors of I should not have the edge over T. */
3361 height1 = insn1->depth + insn1->alap;
3362 height2 = insn2->depth + insn2->alap;
3363 if (height1 != height2)
3364 return height1 > height2;
3365 if (insn1->depth != insn2->depth)
3366 return insn1->depth > insn2->depth;
3367
3368 /* We have no real preference between INSN1 an INSN2 as far as attempts
3369 to reduce pressure go. Prefer instructions with higher priorities. */
3370 priority1 = INSN_PRIORITY (insn1->insn);
3371 priority2 = INSN_PRIORITY (insn2->insn);
3372 if (priority1 != priority2)
3373 return priority1 > priority2;
3374
3375 /* Use the original rtl sequence as a tie-breaker. */
3376 return insn1 < insn2;
3377}
3378
3379/* Add INSN to the model worklist immediately after PREV. Add it to the
3380 beginning of the list if PREV is null. */
3381
3382static void
3383model_add_to_worklist_at (struct model_insn_info *insn,
3384 struct model_insn_info *prev)
3385{
3386 gcc_assert (QUEUE_INDEX (insn->insn) == QUEUE_NOWHERE);
3387 QUEUE_INDEX (insn->insn) = QUEUE_READY;
3388
3389 insn->prev = prev;
3390 if (prev)
3391 {
3392 insn->next = prev->next;
3393 prev->next = insn;
3394 }
3395 else
3396 {
3397 insn->next = model_worklist;
3398 model_worklist = insn;
3399 }
3400 if (insn->next)
3401 insn->next->prev = insn;
3402}
3403
3404/* Remove INSN from the model worklist. */
3405
3406static void
3407model_remove_from_worklist (struct model_insn_info *insn)
3408{
3409 gcc_assert (QUEUE_INDEX (insn->insn) == QUEUE_READY);
3410 QUEUE_INDEX (insn->insn) = QUEUE_NOWHERE;
3411
3412 if (insn->prev)
3413 insn->prev->next = insn->next;
3414 else
3415 model_worklist = insn->next;
3416 if (insn->next)
3417 insn->next->prev = insn->prev;
3418}
3419
3420/* Add INSN to the model worklist. Start looking for a suitable position
3421 between neighbors PREV and NEXT, testing at most param_max_sched_ready_insns
3422 insns either side. A null PREV indicates the beginning of the list and
3423 a null NEXT indicates the end. */
3424
3425static void
3426model_add_to_worklist (struct model_insn_info *insn,
3427 struct model_insn_info *prev,
3428 struct model_insn_info *next)
3429{
3430 int count;
3431
3432 count = param_max_sched_ready_insns;
3433 if (count > 0 && prev && model_order_p (insn1: insn, insn2: prev))
3434 do
3435 {
3436 count--;
3437 prev = prev->prev;
3438 }
3439 while (count > 0 && prev && model_order_p (insn1: insn, insn2: prev));
3440 else
3441 while (count > 0 && next && model_order_p (insn1: next, insn2: insn))
3442 {
3443 count--;
3444 prev = next;
3445 next = next->next;
3446 }
3447 model_add_to_worklist_at (insn, prev);
3448}
3449
3450/* INSN may now have a higher priority (in the model_order_p sense)
3451 than before. Move it up the worklist if necessary. */
3452
3453static void
3454model_promote_insn (struct model_insn_info *insn)
3455{
3456 struct model_insn_info *prev;
3457 int count;
3458
3459 prev = insn->prev;
3460 count = param_max_sched_ready_insns;
3461 while (count > 0 && prev && model_order_p (insn1: insn, insn2: prev))
3462 {
3463 count--;
3464 prev = prev->prev;
3465 }
3466 if (prev != insn->prev)
3467 {
3468 model_remove_from_worklist (insn);
3469 model_add_to_worklist_at (insn, prev);
3470 }
3471}
3472
3473/* Add INSN to the end of the model schedule. */
3474
3475static void
3476model_add_to_schedule (rtx_insn *insn)
3477{
3478 unsigned int point;
3479
3480 gcc_assert (QUEUE_INDEX (insn) == QUEUE_NOWHERE);
3481 QUEUE_INDEX (insn) = QUEUE_SCHEDULED;
3482
3483 point = model_schedule.length ();
3484 model_schedule.quick_push (obj: insn);
3485 INSN_MODEL_INDEX (insn) = point + 1;
3486}
3487
3488/* Analyze the instructions that are to be scheduled, setting up
3489 MODEL_INSN_INFO (...) and model_num_insns accordingly. Add ready
3490 instructions to model_worklist. */
3491
3492static void
3493model_analyze_insns (void)
3494{
3495 rtx_insn *start, *end, *iter;
3496 sd_iterator_def sd_it;
3497 dep_t dep;
3498 struct model_insn_info *insn, *con;
3499
3500 model_num_insns = 0;
3501 start = PREV_INSN (insn: current_sched_info->next_tail);
3502 end = current_sched_info->prev_head;
3503 for (iter = start; iter != end; iter = PREV_INSN (insn: iter))
3504 if (NONDEBUG_INSN_P (iter))
3505 {
3506 insn = MODEL_INSN_INFO (iter);
3507 insn->insn = iter;
3508 FOR_EACH_DEP (iter, SD_LIST_FORW, sd_it, dep)
3509 {
3510 con = MODEL_INSN_INFO (DEP_CON (dep));
3511 if (con->insn && insn->alap < con->alap + 1)
3512 insn->alap = con->alap + 1;
3513 }
3514
3515 insn->old_queue = QUEUE_INDEX (iter);
3516 QUEUE_INDEX (iter) = QUEUE_NOWHERE;
3517
3518 insn->unscheduled_preds = dep_list_size (insn: iter, SD_LIST_HARD_BACK);
3519 if (insn->unscheduled_preds == 0)
3520 model_add_to_worklist (insn, NULL, next: model_worklist);
3521
3522 model_num_insns++;
3523 }
3524}
3525
3526/* The global state describes the register pressure at the start of the
3527 model schedule. Initialize GROUP accordingly. */
3528
3529static void
3530model_init_pressure_group (struct model_pressure_group *group)
3531{
3532 int pci, cl;
3533
3534 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3535 {
3536 cl = ira_pressure_classes[pci];
3537 group->limits[pci].pressure = curr_reg_pressure[cl];
3538 group->limits[pci].point = 0;
3539 }
3540 /* Use index model_num_insns to record the state after the last
3541 instruction in the model schedule. */
3542 group->model = XNEWVEC (struct model_pressure_data,
3543 (model_num_insns + 1) * ira_pressure_classes_num);
3544}
3545
3546/* Record that MODEL_REF_PRESSURE (GROUP, POINT, PCI) is PRESSURE.
3547 Update the maximum pressure for the whole schedule. */
3548
3549static void
3550model_record_pressure (struct model_pressure_group *group,
3551 int point, int pci, int pressure)
3552{
3553 MODEL_REF_PRESSURE (group, point, pci) = pressure;
3554 if (group->limits[pci].pressure < pressure)
3555 {
3556 group->limits[pci].pressure = pressure;
3557 group->limits[pci].point = point;
3558 }
3559}
3560
3561/* INSN has just been added to the end of the model schedule. Record its
3562 register-pressure information. */
3563
3564static void
3565model_record_pressures (struct model_insn_info *insn)
3566{
3567 struct reg_pressure_data *reg_pressure;
3568 int point, pci, cl, delta;
3569 int death[N_REG_CLASSES];
3570
3571 point = model_index (insn: insn->insn);
3572 if (sched_verbose >= 2)
3573 {
3574 if (point == 0)
3575 {
3576 fprintf (stream: sched_dump, format: "\n;;\tModel schedule:\n;;\n");
3577 fprintf (stream: sched_dump, format: ";;\t| idx insn | mpri hght dpth prio |\n");
3578 }
3579 fprintf (stream: sched_dump, format: ";;\t| %3d %4d | %4d %4d %4d %4d | %-30s ",
3580 point, INSN_UID (insn: insn->insn), insn->model_priority,
3581 insn->depth + insn->alap, insn->depth,
3582 INSN_PRIORITY (insn->insn),
3583 str_pattern_slim (PATTERN (insn: insn->insn)));
3584 }
3585 calculate_reg_deaths (insn: insn->insn, death);
3586 reg_pressure = INSN_REG_PRESSURE (insn->insn);
3587 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3588 {
3589 cl = ira_pressure_classes[pci];
3590 delta = reg_pressure[pci].set_increase - death[cl];
3591 if (sched_verbose >= 2)
3592 fprintf (stream: sched_dump, format: " %s:[%d,%+d]", reg_class_names[cl],
3593 curr_reg_pressure[cl], delta);
3594 model_record_pressure (group: &model_before_pressure, point, pci,
3595 pressure: curr_reg_pressure[cl]);
3596 }
3597 if (sched_verbose >= 2)
3598 fprintf (stream: sched_dump, format: "\n");
3599}
3600
3601/* All instructions have been added to the model schedule. Record the
3602 final register pressure in GROUP and set up all MODEL_MAX_PRESSUREs. */
3603
3604static void
3605model_record_final_pressures (struct model_pressure_group *group)
3606{
3607 int point, pci, max_pressure, ref_pressure, cl;
3608
3609 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3610 {
3611 /* Record the final pressure for this class. */
3612 cl = ira_pressure_classes[pci];
3613 point = model_num_insns;
3614 ref_pressure = curr_reg_pressure[cl];
3615 model_record_pressure (group, point, pci, pressure: ref_pressure);
3616
3617 /* Record the original maximum pressure. */
3618 group->limits[pci].orig_pressure = group->limits[pci].pressure;
3619
3620 /* Update the MODEL_MAX_PRESSURE for every point of the schedule. */
3621 max_pressure = ref_pressure;
3622 MODEL_MAX_PRESSURE (group, point, pci) = max_pressure;
3623 while (point > 0)
3624 {
3625 point--;
3626 ref_pressure = MODEL_REF_PRESSURE (group, point, pci);
3627 max_pressure = MAX (max_pressure, ref_pressure);
3628 MODEL_MAX_PRESSURE (group, point, pci) = max_pressure;
3629 }
3630 }
3631}
3632
3633/* Update all successors of INSN, given that INSN has just been scheduled. */
3634
3635static void
3636model_add_successors_to_worklist (struct model_insn_info *insn)
3637{
3638 sd_iterator_def sd_it;
3639 struct model_insn_info *con;
3640 dep_t dep;
3641
3642 FOR_EACH_DEP (insn->insn, SD_LIST_FORW, sd_it, dep)
3643 {
3644 con = MODEL_INSN_INFO (DEP_CON (dep));
3645 /* Ignore debug instructions, and instructions from other blocks. */
3646 if (con->insn)
3647 {
3648 con->unscheduled_preds--;
3649
3650 /* Update the depth field of each true-dependent successor.
3651 Increasing the depth gives them a higher priority than
3652 before. */
3653 if (DEP_TYPE (dep) == REG_DEP_TRUE && con->depth < insn->depth + 1)
3654 {
3655 con->depth = insn->depth + 1;
3656 if (QUEUE_INDEX (con->insn) == QUEUE_READY)
3657 model_promote_insn (insn: con);
3658 }
3659
3660 /* If this is a true dependency, or if there are no remaining
3661 dependencies for CON (meaning that CON only had non-true
3662 dependencies), make sure that CON is on the worklist.
3663 We don't bother otherwise because it would tend to fill the
3664 worklist with a lot of low-priority instructions that are not
3665 yet ready to issue. */
3666 if ((con->depth > 0 || con->unscheduled_preds == 0)
3667 && QUEUE_INDEX (con->insn) == QUEUE_NOWHERE)
3668 model_add_to_worklist (insn: con, prev: insn, next: insn->next);
3669 }
3670 }
3671}
3672
3673/* Give INSN a higher priority than any current instruction, then give
3674 unscheduled predecessors of INSN a higher priority still. If any of
3675 those predecessors are not on the model worklist, do the same for its
3676 predecessors, and so on. */
3677
3678static void
3679model_promote_predecessors (struct model_insn_info *insn)
3680{
3681 struct model_insn_info *pro, *first;
3682 sd_iterator_def sd_it;
3683 dep_t dep;
3684
3685 if (sched_verbose >= 7)
3686 fprintf (stream: sched_dump, format: ";;\t+--- priority of %d = %d, priority of",
3687 INSN_UID (insn: insn->insn), model_next_priority);
3688 insn->model_priority = model_next_priority++;
3689 model_remove_from_worklist (insn);
3690 model_add_to_worklist_at (insn, NULL);
3691
3692 first = NULL;
3693 for (;;)
3694 {
3695 FOR_EACH_DEP (insn->insn, SD_LIST_HARD_BACK, sd_it, dep)
3696 {
3697 pro = MODEL_INSN_INFO (DEP_PRO (dep));
3698 /* The first test is to ignore debug instructions, and instructions
3699 from other blocks. */
3700 if (pro->insn
3701 && pro->model_priority != model_next_priority
3702 && QUEUE_INDEX (pro->insn) != QUEUE_SCHEDULED)
3703 {
3704 pro->model_priority = model_next_priority;
3705 if (sched_verbose >= 7)
3706 fprintf (stream: sched_dump, format: " %d", INSN_UID (insn: pro->insn));
3707 if (QUEUE_INDEX (pro->insn) == QUEUE_READY)
3708 {
3709 /* PRO is already in the worklist, but it now has
3710 a higher priority than before. Move it at the
3711 appropriate place. */
3712 model_remove_from_worklist (insn: pro);
3713 model_add_to_worklist (insn: pro, NULL, next: model_worklist);
3714 }
3715 else
3716 {
3717 /* PRO isn't in the worklist. Recursively process
3718 its predecessors until we find one that is. */
3719 pro->next = first;
3720 first = pro;
3721 }
3722 }
3723 }
3724 if (!first)
3725 break;
3726 insn = first;
3727 first = insn->next;
3728 }
3729 if (sched_verbose >= 7)
3730 fprintf (stream: sched_dump, format: " = %d\n", model_next_priority);
3731 model_next_priority++;
3732}
3733
3734/* Pick one instruction from model_worklist and process it. */
3735
3736static void
3737model_choose_insn (void)
3738{
3739 struct model_insn_info *insn, *fallback;
3740 int count;
3741
3742 if (sched_verbose >= 7)
3743 {
3744 fprintf (stream: sched_dump, format: ";;\t+--- worklist:\n");
3745 insn = model_worklist;
3746 count = param_max_sched_ready_insns;
3747 while (count > 0 && insn)
3748 {
3749 fprintf (stream: sched_dump, format: ";;\t+--- %d [%d, %d, %d, %d]\n",
3750 INSN_UID (insn: insn->insn), insn->model_priority,
3751 insn->depth + insn->alap, insn->depth,
3752 INSN_PRIORITY (insn->insn));
3753 count--;
3754 insn = insn->next;
3755 }
3756 }
3757
3758 /* Look for a ready instruction whose model_classify_priority is zero
3759 or negative, picking the highest-priority one. Adding such an
3760 instruction to the schedule now should do no harm, and may actually
3761 do some good.
3762
3763 Failing that, see whether there is an instruction with the highest
3764 extant model_priority that is not yet ready, but which would reduce
3765 pressure if it became ready. This is designed to catch cases like:
3766
3767 (set (mem (reg R1)) (reg R2))
3768
3769 where the instruction is the last remaining use of R1 and where the
3770 value of R2 is not yet available (or vice versa). The death of R1
3771 means that this instruction already reduces pressure. It is of
3772 course possible that the computation of R2 involves other registers
3773 that are hard to kill, but such cases are rare enough for this
3774 heuristic to be a win in general.
3775
3776 Failing that, just pick the highest-priority instruction in the
3777 worklist. */
3778 count = param_max_sched_ready_insns;
3779 insn = model_worklist;
3780 fallback = 0;
3781 for (;;)
3782 {
3783 if (count == 0 || !insn)
3784 {
3785 insn = fallback ? fallback : model_worklist;
3786 break;
3787 }
3788 if (insn->unscheduled_preds)
3789 {
3790 if (model_worklist->model_priority == insn->model_priority
3791 && !fallback
3792 && model_classify_pressure (insn) < 0)
3793 fallback = insn;
3794 }
3795 else
3796 {
3797 if (model_classify_pressure (insn) <= 0)
3798 break;
3799 }
3800 count--;
3801 insn = insn->next;
3802 }
3803
3804 if (sched_verbose >= 7 && insn != model_worklist)
3805 {
3806 if (insn->unscheduled_preds)
3807 fprintf (stream: sched_dump, format: ";;\t+--- promoting insn %d, with dependencies\n",
3808 INSN_UID (insn: insn->insn));
3809 else
3810 fprintf (stream: sched_dump, format: ";;\t+--- promoting insn %d, which is ready\n",
3811 INSN_UID (insn: insn->insn));
3812 }
3813 if (insn->unscheduled_preds)
3814 /* INSN isn't yet ready to issue. Give all its predecessors the
3815 highest priority. */
3816 model_promote_predecessors (insn);
3817 else
3818 {
3819 /* INSN is ready. Add it to the end of model_schedule and
3820 process its successors. */
3821 model_add_successors_to_worklist (insn);
3822 model_remove_from_worklist (insn);
3823 model_add_to_schedule (insn: insn->insn);
3824 model_record_pressures (insn);
3825 update_register_pressure (insn: insn->insn);
3826 }
3827}
3828
3829/* Restore all QUEUE_INDEXs to the values that they had before
3830 model_start_schedule was called. */
3831
3832static void
3833model_reset_queue_indices (void)
3834{
3835 unsigned int i;
3836 rtx_insn *insn;
3837
3838 FOR_EACH_VEC_ELT (model_schedule, i, insn)
3839 QUEUE_INDEX (insn) = MODEL_INSN_INFO (insn)->old_queue;
3840}
3841
3842/* We have calculated the model schedule and spill costs. Print a summary
3843 to sched_dump. */
3844
3845static void
3846model_dump_pressure_summary (void)
3847{
3848 int pci, cl;
3849
3850 fprintf (stream: sched_dump, format: ";; Pressure summary:");
3851 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3852 {
3853 cl = ira_pressure_classes[pci];
3854 fprintf (stream: sched_dump, format: " %s:%d", reg_class_names[cl],
3855 model_before_pressure.limits[pci].pressure);
3856 }
3857 fprintf (stream: sched_dump, format: "\n\n");
3858}
3859
3860/* Initialize the SCHED_PRESSURE_MODEL information for the current
3861 scheduling region. */
3862
3863static void
3864model_start_schedule (basic_block bb)
3865{
3866 model_next_priority = 1;
3867 model_schedule.create (nelems: sched_max_luid);
3868 model_insns = XCNEWVEC (struct model_insn_info, sched_max_luid);
3869
3870 gcc_assert (bb == BLOCK_FOR_INSN (NEXT_INSN (current_sched_info->prev_head)));
3871 initiate_reg_pressure_info (live: df_get_live_in (bb));
3872
3873 model_analyze_insns ();
3874 model_init_pressure_group (group: &model_before_pressure);
3875 while (model_worklist)
3876 model_choose_insn ();
3877 gcc_assert (model_num_insns == (int) model_schedule.length ());
3878 if (sched_verbose >= 2)
3879 fprintf (stream: sched_dump, format: "\n");
3880
3881 model_record_final_pressures (group: &model_before_pressure);
3882 model_reset_queue_indices ();
3883
3884 XDELETEVEC (model_insns);
3885
3886 model_curr_point = 0;
3887 initiate_reg_pressure_info (live: df_get_live_in (bb));
3888 if (sched_verbose >= 1)
3889 model_dump_pressure_summary ();
3890}
3891
3892/* Free the information associated with GROUP. */
3893
3894static void
3895model_finalize_pressure_group (struct model_pressure_group *group)
3896{
3897 XDELETEVEC (group->model);
3898}
3899
3900/* Free the information created by model_start_schedule. */
3901
3902static void
3903model_end_schedule (void)
3904{
3905 model_finalize_pressure_group (group: &model_before_pressure);
3906 model_schedule.release ();
3907}
3908
3909/* Prepare reg pressure scheduling for basic block BB. */
3910static void
3911sched_pressure_start_bb (basic_block bb)
3912{
3913 /* Set the number of available registers for each class taking into account
3914 relative probability of current basic block versus function prologue and
3915 epilogue.
3916 * If the basic block executes much more often than the prologue/epilogue
3917 (e.g., inside a hot loop), then cost of spill in the prologue is close to
3918 nil, so the effective number of available registers is
3919 (ira_class_hard_regs_num[cl] - fixed_regs_num[cl] - 0).
3920 * If the basic block executes as often as the prologue/epilogue,
3921 then spill in the block is as costly as in the prologue, so the effective
3922 number of available registers is
3923 (ira_class_hard_regs_num[cl] - fixed_regs_num[cl]
3924 - call_saved_regs_num[cl]).
3925 Note that all-else-equal, we prefer to spill in the prologue, since that
3926 allows "extra" registers for other basic blocks of the function.
3927 * If the basic block is on the cold path of the function and executes
3928 rarely, then we should always prefer to spill in the block, rather than
3929 in the prologue/epilogue. The effective number of available register is
3930 (ira_class_hard_regs_num[cl] - fixed_regs_num[cl]
3931 - call_saved_regs_num[cl]). */
3932 {
3933 int i;
3934 int entry_freq = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count.to_frequency (cfun);
3935 int bb_freq = bb->count.to_frequency (cfun);
3936
3937 if (bb_freq == 0)
3938 {
3939 if (entry_freq == 0)
3940 entry_freq = bb_freq = 1;
3941 }
3942 if (bb_freq < entry_freq)
3943 bb_freq = entry_freq;
3944
3945 for (i = 0; i < ira_pressure_classes_num; ++i)
3946 {
3947 enum reg_class cl = ira_pressure_classes[i];
3948 sched_class_regs_num[cl] = ira_class_hard_regs_num[cl]
3949 - fixed_regs_num[cl];
3950 sched_class_regs_num[cl]
3951 -= (call_saved_regs_num[cl] * entry_freq) / bb_freq;
3952 }
3953 }
3954
3955 if (sched_pressure == SCHED_PRESSURE_MODEL)
3956 model_start_schedule (bb);
3957}
3958
3959/* A structure that holds local state for the loop in schedule_block. */
3960struct sched_block_state
3961{
3962 /* True if no real insns have been scheduled in the current cycle. */
3963 bool first_cycle_insn_p;
3964 /* True if a shadow insn has been scheduled in the current cycle, which
3965 means that no more normal insns can be issued. */
3966 bool shadows_only_p;
3967 /* True if we're winding down a modulo schedule, which means that we only
3968 issue insns with INSN_EXACT_TICK set. */
3969 bool modulo_epilogue;
3970 /* Initialized with the machine's issue rate every cycle, and updated
3971 by calls to the variable_issue hook. */
3972 int can_issue_more;
3973};
3974
3975/* INSN is the "currently executing insn". Launch each insn which was
3976 waiting on INSN. READY is the ready list which contains the insns
3977 that are ready to fire. CLOCK is the current cycle. The function
3978 returns necessary cycle advance after issuing the insn (it is not
3979 zero for insns in a schedule group). */
3980
3981static int
3982schedule_insn (rtx_insn *insn)
3983{
3984 sd_iterator_def sd_it;
3985 dep_t dep;
3986 int i;
3987 int advance = 0;
3988
3989 if (sched_verbose >= 1)
3990 {
3991 struct reg_pressure_data *pressure_info;
3992 fprintf (stream: sched_dump, format: ";;\t%3i--> %s %-40s:",
3993 clock_var, (*current_sched_info->print_insn) (insn, 1),
3994 str_pattern_slim (PATTERN (insn)));
3995
3996 if (recog_memoized (insn) < 0)
3997 fprintf (stream: sched_dump, format: "nothing");
3998 else
3999 print_reservation (sched_dump, insn);
4000 pressure_info = INSN_REG_PRESSURE (insn);
4001 if (pressure_info != NULL)
4002 {
4003 fputc (c: ':', stream: sched_dump);
4004 for (i = 0; i < ira_pressure_classes_num; i++)
4005 fprintf (stream: sched_dump, format: "%s%s%+d(%d)",
4006 scheduled_insns.length () > 1
4007 && INSN_LUID (insn)
4008 < INSN_LUID (scheduled_insns[scheduled_insns.length () - 2]) ? "@" : "",
4009 reg_class_names[ira_pressure_classes[i]],
4010 pressure_info[i].set_increase, pressure_info[i].change);
4011 }
4012 if (sched_pressure == SCHED_PRESSURE_MODEL
4013 && model_curr_point < model_num_insns
4014 && model_index (insn) == model_curr_point)
4015 fprintf (stream: sched_dump, format: ":model %d", model_curr_point);
4016 fputc (c: '\n', stream: sched_dump);
4017 }
4018
4019 if (sched_pressure == SCHED_PRESSURE_WEIGHTED && !DEBUG_INSN_P (insn))
4020 update_reg_and_insn_max_reg_pressure (insn);
4021
4022 /* Scheduling instruction should have all its dependencies resolved and
4023 should have been removed from the ready list. */
4024 gcc_assert (sd_lists_empty_p (insn, SD_LIST_HARD_BACK));
4025
4026 /* Reset debug insns invalidated by moving this insn. */
4027 if (MAY_HAVE_DEBUG_BIND_INSNS && !DEBUG_INSN_P (insn))
4028 for (sd_it = sd_iterator_start (insn, SD_LIST_BACK);
4029 sd_iterator_cond (it_ptr: &sd_it, dep_ptr: &dep);)
4030 {
4031 rtx_insn *dbg = DEP_PRO (dep);
4032 struct reg_use_data *use, *next;
4033
4034 if (DEP_STATUS (dep) & DEP_CANCELLED)
4035 {
4036 sd_iterator_next (it_ptr: &sd_it);
4037 continue;
4038 }
4039
4040 gcc_assert (DEBUG_BIND_INSN_P (dbg));
4041
4042 if (sched_verbose >= 6)
4043 fprintf (stream: sched_dump, format: ";;\t\tresetting: debug insn %d\n",
4044 INSN_UID (insn: dbg));
4045
4046 /* ??? Rather than resetting the debug insn, we might be able
4047 to emit a debug temp before the just-scheduled insn, but
4048 this would involve checking that the expression at the
4049 point of the debug insn is equivalent to the expression
4050 before the just-scheduled insn. They might not be: the
4051 expression in the debug insn may depend on other insns not
4052 yet scheduled that set MEMs, REGs or even other debug
4053 insns. It's not clear that attempting to preserve debug
4054 information in these cases is worth the effort, given how
4055 uncommon these resets are and the likelihood that the debug
4056 temps introduced won't survive the schedule change. */
4057 INSN_VAR_LOCATION_LOC (dbg) = gen_rtx_UNKNOWN_VAR_LOC ();
4058 df_insn_rescan (dbg);
4059
4060 /* Unknown location doesn't use any registers. */
4061 for (use = INSN_REG_USE_LIST (dbg); use != NULL; use = next)
4062 {
4063 struct reg_use_data *prev = use;
4064
4065 /* Remove use from the cyclic next_regno_use chain first. */
4066 while (prev->next_regno_use != use)
4067 prev = prev->next_regno_use;
4068 prev->next_regno_use = use->next_regno_use;
4069 next = use->next_insn_use;
4070 free (ptr: use);
4071 }
4072 INSN_REG_USE_LIST (dbg) = NULL;
4073
4074 /* We delete rather than resolve these deps, otherwise we
4075 crash in sched_free_deps(), because forward deps are
4076 expected to be released before backward deps. */
4077 sd_delete_dep (sd_it);
4078 }
4079
4080 gcc_assert (QUEUE_INDEX (insn) == QUEUE_NOWHERE);
4081 QUEUE_INDEX (insn) = QUEUE_SCHEDULED;
4082
4083 if (sched_pressure == SCHED_PRESSURE_MODEL
4084 && model_curr_point < model_num_insns
4085 && NONDEBUG_INSN_P (insn))
4086 {
4087 if (model_index (insn) == model_curr_point)
4088 do
4089 model_curr_point++;
4090 while (model_curr_point < model_num_insns
4091 && (QUEUE_INDEX (MODEL_INSN (model_curr_point))
4092 == QUEUE_SCHEDULED));
4093 else
4094 model_recompute (insn);
4095 model_update_limit_points ();
4096 update_register_pressure (insn);
4097 if (sched_verbose >= 2)
4098 print_curr_reg_pressure ();
4099 }
4100
4101 gcc_assert (INSN_TICK (insn) >= MIN_TICK);
4102 if (INSN_TICK (insn) > clock_var)
4103 /* INSN has been prematurely moved from the queue to the ready list.
4104 This is possible only if following flags are set. */
4105 gcc_assert (flag_sched_stalled_insns || sched_fusion);
4106
4107 /* ??? Probably, if INSN is scheduled prematurely, we should leave
4108 INSN_TICK untouched. This is a machine-dependent issue, actually. */
4109 INSN_TICK (insn) = clock_var;
4110
4111 check_clobbered_conditions (insn);
4112
4113 /* Update dependent instructions. First, see if by scheduling this insn
4114 now we broke a dependence in a way that requires us to change another
4115 insn. */
4116 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
4117 sd_iterator_cond (it_ptr: &sd_it, dep_ptr: &dep); sd_iterator_next (it_ptr: &sd_it))
4118 {
4119 struct dep_replacement *desc = DEP_REPLACE (dep);
4120 rtx_insn *pro = DEP_PRO (dep);
4121 if (QUEUE_INDEX (pro) != QUEUE_SCHEDULED
4122 && desc != NULL && desc->insn == pro)
4123 apply_replacement (dep, false);
4124 }
4125
4126 /* Go through and resolve forward dependencies. */
4127 for (sd_it = sd_iterator_start (insn, SD_LIST_FORW);
4128 sd_iterator_cond (it_ptr: &sd_it, dep_ptr: &dep);)
4129 {
4130 rtx_insn *next = DEP_CON (dep);
4131 bool cancelled = (DEP_STATUS (dep) & DEP_CANCELLED) != 0;
4132
4133 /* Resolve the dependence between INSN and NEXT.
4134 sd_resolve_dep () moves current dep to another list thus
4135 advancing the iterator. */
4136 sd_resolve_dep (sd_it);
4137
4138 if (cancelled)
4139 {
4140 if (must_restore_pattern_p (next, dep))
4141 restore_pattern (dep, false);
4142 continue;
4143 }
4144
4145 /* Don't bother trying to mark next as ready if insn is a debug
4146 insn. If insn is the last hard dependency, it will have
4147 already been discounted. */
4148 if (DEBUG_INSN_P (insn) && !DEBUG_INSN_P (next))
4149 continue;
4150
4151 if (!IS_SPECULATION_BRANCHY_CHECK_P (insn))
4152 {
4153 int effective_cost;
4154
4155 effective_cost = try_ready (next);
4156
4157 if (effective_cost >= 0
4158 && SCHED_GROUP_P (next)
4159 && advance < effective_cost)
4160 advance = effective_cost;
4161 }
4162 else
4163 /* Check always has only one forward dependence (to the first insn in
4164 the recovery block), therefore, this will be executed only once. */
4165 {
4166 gcc_assert (sd_lists_empty_p (insn, SD_LIST_FORW));
4167 fix_recovery_deps (RECOVERY_BLOCK (insn));
4168 }
4169 }
4170
4171 /* Annotate the instruction with issue information -- TImode
4172 indicates that the instruction is expected not to be able
4173 to issue on the same cycle as the previous insn. A machine
4174 may use this information to decide how the instruction should
4175 be aligned. */
4176 if (issue_rate > 1
4177 && GET_CODE (PATTERN (insn)) != USE
4178 && GET_CODE (PATTERN (insn)) != CLOBBER
4179 && !DEBUG_INSN_P (insn))
4180 {
4181 if (reload_completed)
4182 PUT_MODE (x: insn, mode: clock_var > last_clock_var ? TImode : VOIDmode);
4183 last_clock_var = clock_var;
4184 }
4185
4186 if (nonscheduled_insns_begin != NULL_RTX)
4187 /* Indicate to debug counters that INSN is scheduled. */
4188 nonscheduled_insns_begin = insn;
4189
4190 return advance;
4191}
4192
4193/* Functions for handling of notes. */
4194
4195/* Add note list that ends on FROM_END to the end of TO_ENDP. */
4196void
4197concat_note_lists (rtx_insn *from_end, rtx_insn **to_endp)
4198{
4199 rtx_insn *from_start;
4200
4201 /* It's easy when have nothing to concat. */
4202 if (from_end == NULL)
4203 return;
4204
4205 /* It's also easy when destination is empty. */
4206 if (*to_endp == NULL)
4207 {
4208 *to_endp = from_end;
4209 return;
4210 }
4211
4212 from_start = from_end;
4213 while (PREV_INSN (insn: from_start) != NULL)
4214 from_start = PREV_INSN (insn: from_start);
4215
4216 SET_PREV_INSN (from_start) = *to_endp;
4217 SET_NEXT_INSN (*to_endp) = from_start;
4218 *to_endp = from_end;
4219}
4220
4221/* Delete notes between HEAD and TAIL and put them in the chain
4222 of notes ended by NOTE_LIST. */
4223void
4224remove_notes (rtx_insn *head, rtx_insn *tail)
4225{
4226 rtx_insn *next_tail, *insn, *next;
4227
4228 note_list = 0;
4229 if (head == tail && !INSN_P (head))
4230 return;
4231
4232 next_tail = NEXT_INSN (insn: tail);
4233 for (insn = head; insn != next_tail; insn = next)
4234 {
4235 next = NEXT_INSN (insn);
4236 if (!NOTE_P (insn))
4237 continue;
4238
4239 switch (NOTE_KIND (insn))
4240 {
4241 case NOTE_INSN_BASIC_BLOCK:
4242 continue;
4243
4244 case NOTE_INSN_EPILOGUE_BEG:
4245 if (insn != tail)
4246 {
4247 remove_insn (insn);
4248 /* If an insn was split just before the EPILOGUE_BEG note and
4249 that split created new basic blocks, we could have a
4250 BASIC_BLOCK note here. Safely advance over it in that case
4251 and assert that we land on a real insn. */
4252 if (NOTE_P (next)
4253 && NOTE_KIND (next) == NOTE_INSN_BASIC_BLOCK
4254 && next != next_tail)
4255 next = NEXT_INSN (insn: next);
4256 gcc_assert (INSN_P (next));
4257 add_reg_note (next, REG_SAVE_NOTE,
4258 GEN_INT (NOTE_INSN_EPILOGUE_BEG));
4259 break;
4260 }
4261 /* FALLTHRU */
4262
4263 default:
4264 remove_insn (insn);
4265
4266 /* Add the note to list that ends at NOTE_LIST. */
4267 SET_PREV_INSN (insn) = note_list;
4268 SET_NEXT_INSN (insn) = NULL_RTX;
4269 if (note_list)
4270 SET_NEXT_INSN (note_list) = insn;
4271 note_list = insn;
4272 break;
4273 }
4274
4275 gcc_assert ((sel_sched_p () || insn != tail) && insn != head);
4276 }
4277}
4278
4279/* A structure to record enough data to allow us to backtrack the scheduler to
4280 a previous state. */
4281struct haifa_saved_data
4282{
4283 /* Next entry on the list. */
4284 struct haifa_saved_data *next;
4285
4286 /* Backtracking is associated with scheduling insns that have delay slots.
4287 DELAY_PAIR points to the structure that contains the insns involved, and
4288 the number of cycles between them. */
4289 struct delay_pair *delay_pair;
4290
4291 /* Data used by the frontend (e.g. sched-ebb or sched-rgn). */
4292 void *fe_saved_data;
4293 /* Data used by the backend. */
4294 void *be_saved_data;
4295
4296 /* Copies of global state. */
4297 int clock_var, last_clock_var;
4298 struct ready_list ready;
4299 state_t curr_state;
4300
4301 rtx_insn *last_scheduled_insn;
4302 rtx_insn *last_nondebug_scheduled_insn;
4303 rtx_insn *nonscheduled_insns_begin;
4304 int cycle_issued_insns;
4305
4306 /* Copies of state used in the inner loop of schedule_block. */
4307 struct sched_block_state sched_block;
4308
4309 /* We don't need to save q_ptr, as its value is arbitrary and we can set it
4310 to 0 when restoring. */
4311 int q_size;
4312 rtx_insn_list **insn_queue;
4313
4314 /* Describe pattern replacements that occurred since this backtrack point
4315 was queued. */
4316 vec<dep_t> replacement_deps;
4317 vec<int> replace_apply;
4318
4319 /* A copy of the next-cycle replacement vectors at the time of the backtrack
4320 point. */
4321 vec<dep_t> next_cycle_deps;
4322 vec<int> next_cycle_apply;
4323};
4324
4325/* A record, in reverse order, of all scheduled insns which have delay slots
4326 and may require backtracking. */
4327static struct haifa_saved_data *backtrack_queue;
4328
4329/* For every dependency of INSN, set the FEEDS_BACKTRACK_INSN bit according
4330 to SET_P. */
4331static void
4332mark_backtrack_feeds (rtx_insn *insn, int set_p)
4333{
4334 sd_iterator_def sd_it;
4335 dep_t dep;
4336 FOR_EACH_DEP (insn, SD_LIST_HARD_BACK, sd_it, dep)
4337 {
4338 FEEDS_BACKTRACK_INSN (DEP_PRO (dep)) = set_p;
4339 }
4340}
4341
4342/* Save the current scheduler state so that we can backtrack to it
4343 later if necessary. PAIR gives the insns that make it necessary to
4344 save this point. SCHED_BLOCK is the local state of schedule_block
4345 that need to be saved. */
4346static void
4347save_backtrack_point (struct delay_pair *pair,
4348 struct sched_block_state sched_block)
4349{
4350 int i;
4351 struct haifa_saved_data *save = XNEW (struct haifa_saved_data);
4352
4353 save->curr_state = xmalloc (dfa_state_size);
4354 memcpy (dest: save->curr_state, src: curr_state, n: dfa_state_size);
4355
4356 save->ready.first = ready.first;
4357 save->ready.n_ready = ready.n_ready;
4358 save->ready.n_debug = ready.n_debug;
4359 save->ready.veclen = ready.veclen;
4360 save->ready.vec = XNEWVEC (rtx_insn *, ready.veclen);
4361 memcpy (dest: save->ready.vec, src: ready.vec, n: ready.veclen * sizeof (rtx));
4362
4363 save->insn_queue = XNEWVEC (rtx_insn_list *, max_insn_queue_index + 1);
4364 save->q_size = q_size;
4365 for (i = 0; i <= max_insn_queue_index; i++)
4366 {
4367 int q = NEXT_Q_AFTER (q_ptr, i);
4368 save->insn_queue[i] = copy_INSN_LIST (insn_queue[q]);
4369 }
4370
4371 save->clock_var = clock_var;
4372 save->last_clock_var = last_clock_var;
4373 save->cycle_issued_insns = cycle_issued_insns;
4374 save->last_scheduled_insn = last_scheduled_insn;
4375 save->last_nondebug_scheduled_insn = last_nondebug_scheduled_insn;
4376 save->nonscheduled_insns_begin = nonscheduled_insns_begin;
4377
4378 save->sched_block = sched_block;
4379
4380 save->replacement_deps.create (nelems: 0);
4381 save->replace_apply.create (nelems: 0);
4382 save->next_cycle_deps = next_cycle_replace_deps.copy ();
4383 save->next_cycle_apply = next_cycle_apply.copy ();
4384
4385 if (current_sched_info->save_state)
4386 save->fe_saved_data = (*current_sched_info->save_state) ();
4387
4388 if (targetm.sched.alloc_sched_context)
4389 {
4390 save->be_saved_data = targetm.sched.alloc_sched_context ();
4391 targetm.sched.init_sched_context (save->be_saved_data, false);
4392 }
4393 else
4394 save->be_saved_data = NULL;
4395
4396 save->delay_pair = pair;
4397
4398 save->next = backtrack_queue;
4399 backtrack_queue = save;
4400
4401 while (pair)
4402 {
4403 mark_backtrack_feeds (insn: pair->i2, set_p: 1);
4404 INSN_TICK (pair->i2) = INVALID_TICK;
4405 INSN_EXACT_TICK (pair->i2) = clock_var + pair_delay (p: pair);
4406 SHADOW_P (pair->i2) = pair->stages == 0;
4407 pair = pair->next_same_i1;
4408 }
4409}
4410
4411/* Walk the ready list and all queues. If any insns have unresolved backwards
4412 dependencies, these must be cancelled deps, broken by predication. Set or
4413 clear (depending on SET) the DEP_CANCELLED bit in DEP_STATUS. */
4414
4415static void
4416toggle_cancelled_flags (bool set)
4417{
4418 int i;
4419 sd_iterator_def sd_it;
4420 dep_t dep;
4421
4422 if (ready.n_ready > 0)
4423 {
4424 rtx_insn **first = ready_lastpos (ready: &ready);
4425 for (i = 0; i < ready.n_ready; i++)
4426 FOR_EACH_DEP (first[i], SD_LIST_BACK, sd_it, dep)
4427 if (!DEBUG_INSN_P (DEP_PRO (dep)))
4428 {
4429 if (set)
4430 DEP_STATUS (dep) |= DEP_CANCELLED;
4431 else
4432 DEP_STATUS (dep) &= ~DEP_CANCELLED;
4433 }
4434 }
4435 for (i = 0; i <= max_insn_queue_index; i++)
4436 {
4437 int q = NEXT_Q_AFTER (q_ptr, i);
4438 rtx_insn_list *link;
4439 for (link = insn_queue[q]; link; link = link->next ())
4440 {
4441 rtx_insn *insn = link->insn ();
4442 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
4443 if (!DEBUG_INSN_P (DEP_PRO (dep)))
4444 {
4445 if (set)
4446 DEP_STATUS (dep) |= DEP_CANCELLED;
4447 else
4448 DEP_STATUS (dep) &= ~DEP_CANCELLED;
4449 }
4450 }
4451 }
4452}
4453
4454/* Undo the replacements that have occurred after backtrack point SAVE
4455 was placed. */
4456static void
4457undo_replacements_for_backtrack (struct haifa_saved_data *save)
4458{
4459 while (!save->replacement_deps.is_empty ())
4460 {
4461 dep_t dep = save->replacement_deps.pop ();
4462 int apply_p = save->replace_apply.pop ();
4463
4464 if (apply_p)
4465 restore_pattern (dep, true);
4466 else
4467 apply_replacement (dep, true);
4468 }
4469 save->replacement_deps.release ();
4470 save->replace_apply.release ();
4471}
4472
4473/* Pop entries from the SCHEDULED_INSNS vector up to and including INSN.
4474 Restore their dependencies to an unresolved state, and mark them as
4475 queued nowhere. */
4476
4477static void
4478unschedule_insns_until (rtx_insn *insn)
4479{
4480 auto_vec<rtx_insn *> recompute_vec;
4481
4482 /* Make two passes over the insns to be unscheduled. First, we clear out
4483 dependencies and other trivial bookkeeping. */
4484 for (;;)
4485 {
4486 rtx_insn *last;
4487 sd_iterator_def sd_it;
4488 dep_t dep;
4489
4490 last = scheduled_insns.pop ();
4491
4492 /* This will be changed by restore_backtrack_point if the insn is in
4493 any queue. */
4494 QUEUE_INDEX (last) = QUEUE_NOWHERE;
4495 if (last != insn)
4496 INSN_TICK (last) = INVALID_TICK;
4497
4498 if (modulo_ii > 0 && INSN_UID (insn: last) < modulo_iter0_max_uid)
4499 modulo_insns_scheduled--;
4500
4501 for (sd_it = sd_iterator_start (insn: last, SD_LIST_RES_FORW);
4502 sd_iterator_cond (it_ptr: &sd_it, dep_ptr: &dep);)
4503 {
4504 rtx_insn *con = DEP_CON (dep);
4505 sd_unresolve_dep (sd_it);
4506 if (!MUST_RECOMPUTE_SPEC_P (con))
4507 {
4508 MUST_RECOMPUTE_SPEC_P (con) = 1;
4509 recompute_vec.safe_push (obj: con);
4510 }
4511 }
4512
4513 if (last == insn)
4514 break;
4515 }
4516
4517 /* A second pass, to update ready and speculation status for insns
4518 depending on the unscheduled ones. The first pass must have
4519 popped the scheduled_insns vector up to the point where we
4520 restart scheduling, as recompute_todo_spec requires it to be
4521 up-to-date. */
4522 while (!recompute_vec.is_empty ())
4523 {
4524 rtx_insn *con;
4525
4526 con = recompute_vec.pop ();
4527 MUST_RECOMPUTE_SPEC_P (con) = 0;
4528 if (!sd_lists_empty_p (con, SD_LIST_HARD_BACK))
4529 {
4530 TODO_SPEC (con) = HARD_DEP;
4531 INSN_TICK (con) = INVALID_TICK;
4532 if (PREDICATED_PAT (con) != NULL_RTX)
4533 haifa_change_pattern (con, ORIG_PAT (con));
4534 }
4535 else if (QUEUE_INDEX (con) != QUEUE_SCHEDULED)
4536 TODO_SPEC (con) = recompute_todo_spec (next: con, for_backtrack: true);
4537 }
4538}
4539
4540/* Restore scheduler state from the topmost entry on the backtracking queue.
4541 PSCHED_BLOCK_P points to the local data of schedule_block that we must
4542 overwrite with the saved data.
4543 The caller must already have called unschedule_insns_until. */
4544
4545static void
4546restore_last_backtrack_point (struct sched_block_state *psched_block)
4547{
4548 int i;
4549 struct haifa_saved_data *save = backtrack_queue;
4550
4551 backtrack_queue = save->next;
4552
4553 if (current_sched_info->restore_state)
4554 (*current_sched_info->restore_state) (save->fe_saved_data);
4555
4556 if (targetm.sched.alloc_sched_context)
4557 {
4558 targetm.sched.set_sched_context (save->be_saved_data);
4559 targetm.sched.free_sched_context (save->be_saved_data);
4560 }
4561
4562 /* Do this first since it clobbers INSN_TICK of the involved
4563 instructions. */
4564 undo_replacements_for_backtrack (save);
4565
4566 /* Clear the QUEUE_INDEX of everything in the ready list or one
4567 of the queues. */
4568 if (ready.n_ready > 0)
4569 {
4570 rtx_insn **first = ready_lastpos (ready: &ready);
4571 for (i = 0; i < ready.n_ready; i++)
4572 {
4573 rtx_insn *insn = first[i];
4574 QUEUE_INDEX (insn) = QUEUE_NOWHERE;
4575 INSN_TICK (insn) = INVALID_TICK;
4576 }
4577 }
4578 for (i = 0; i <= max_insn_queue_index; i++)
4579 {
4580 int q = NEXT_Q_AFTER (q_ptr, i);
4581
4582 for (rtx_insn_list *link = insn_queue[q]; link; link = link->next ())
4583 {
4584 rtx_insn *x = link->insn ();
4585 QUEUE_INDEX (x) = QUEUE_NOWHERE;
4586 INSN_TICK (x) = INVALID_TICK;
4587 }
4588 free_INSN_LIST_list (&insn_queue[q]);
4589 }
4590
4591 free (ptr: ready.vec);
4592 ready = save->ready;
4593
4594 if (ready.n_ready > 0)
4595 {
4596 rtx_insn **first = ready_lastpos (ready: &ready);
4597 for (i = 0; i < ready.n_ready; i++)
4598 {
4599 rtx_insn *insn = first[i];
4600 QUEUE_INDEX (insn) = QUEUE_READY;
4601 TODO_SPEC (insn) = recompute_todo_spec (next: insn, for_backtrack: true);
4602 INSN_TICK (insn) = save->clock_var;
4603 }
4604 }
4605
4606 q_ptr = 0;
4607 q_size = save->q_size;
4608 for (i = 0; i <= max_insn_queue_index; i++)
4609 {
4610 int q = NEXT_Q_AFTER (q_ptr, i);
4611
4612 insn_queue[q] = save->insn_queue[q];
4613
4614 for (rtx_insn_list *link = insn_queue[q]; link; link = link->next ())
4615 {
4616 rtx_insn *x = link->insn ();
4617 QUEUE_INDEX (x) = i;
4618 TODO_SPEC (x) = recompute_todo_spec (next: x, for_backtrack: true);
4619 INSN_TICK (x) = save->clock_var + i;
4620 }
4621 }
4622 free (ptr: save->insn_queue);
4623
4624 toggle_cancelled_flags (set: true);
4625
4626 clock_var = save->clock_var;
4627 last_clock_var = save->last_clock_var;
4628 cycle_issued_insns = save->cycle_issued_insns;
4629 last_scheduled_insn = save->last_scheduled_insn;
4630 last_nondebug_scheduled_insn = save->last_nondebug_scheduled_insn;
4631 nonscheduled_insns_begin = save->nonscheduled_insns_begin;
4632
4633 *psched_block = save->sched_block;
4634
4635 memcpy (dest: curr_state, src: save->curr_state, n: dfa_state_size);
4636 free (ptr: save->curr_state);
4637
4638 mark_backtrack_feeds (insn: save->delay_pair->i2, set_p: 0);
4639
4640 gcc_assert (next_cycle_replace_deps.is_empty ());
4641 next_cycle_replace_deps = save->next_cycle_deps.copy ();
4642 next_cycle_apply = save->next_cycle_apply.copy ();
4643
4644 free (ptr: save);
4645
4646 for (save = backtrack_queue; save; save = save->next)
4647 {
4648 mark_backtrack_feeds (insn: save->delay_pair->i2, set_p: 1);
4649 }
4650}
4651
4652/* Discard all data associated with the topmost entry in the backtrack
4653 queue. If RESET_TICK is false, we just want to free the data. If true,
4654 we are doing this because we discovered a reason to backtrack. In the
4655 latter case, also reset the INSN_TICK for the shadow insn. */
4656static void
4657free_topmost_backtrack_point (bool reset_tick)
4658{
4659 struct haifa_saved_data *save = backtrack_queue;
4660 int i;
4661
4662 backtrack_queue = save->next;
4663
4664 if (reset_tick)
4665 {
4666 struct delay_pair *pair = save->delay_pair;
4667 while (pair)
4668 {
4669 INSN_TICK (pair->i2) = INVALID_TICK;
4670 INSN_EXACT_TICK (pair->i2) = INVALID_TICK;
4671 pair = pair->next_same_i1;
4672 }
4673 undo_replacements_for_backtrack (save);
4674 }
4675 else
4676 {
4677 save->replacement_deps.release ();
4678 save->replace_apply.release ();
4679 }
4680
4681 if (targetm.sched.free_sched_context)
4682 targetm.sched.free_sched_context (save->be_saved_data);
4683 if (current_sched_info->restore_state)
4684 free (ptr: save->fe_saved_data);
4685 for (i = 0; i <= max_insn_queue_index; i++)
4686 free_INSN_LIST_list (&save->insn_queue[i]);
4687 free (ptr: save->insn_queue);
4688 free (ptr: save->curr_state);
4689 free (ptr: save->ready.vec);
4690 free (ptr: save);
4691}
4692
4693/* Free the entire backtrack queue. */
4694static void
4695free_backtrack_queue (void)
4696{
4697 while (backtrack_queue)
4698 free_topmost_backtrack_point (reset_tick: false);
4699}
4700
4701/* Apply a replacement described by DESC. If IMMEDIATELY is false, we
4702 may have to postpone the replacement until the start of the next cycle,
4703 at which point we will be called again with IMMEDIATELY true. This is
4704 only done for machines which have instruction packets with explicit
4705 parallelism however. */
4706static void
4707apply_replacement (dep_t dep, bool immediately)
4708{
4709 struct dep_replacement *desc = DEP_REPLACE (dep);
4710 if (!immediately && targetm.sched.exposed_pipeline && reload_completed)
4711 {
4712 next_cycle_replace_deps.safe_push (obj: dep);
4713 next_cycle_apply.safe_push (obj: 1);
4714 }
4715 else
4716 {
4717 bool success;
4718
4719 if (QUEUE_INDEX (desc->insn) == QUEUE_SCHEDULED)
4720 return;
4721
4722 if (sched_verbose >= 5)
4723 fprintf (stream: sched_dump, format: "applying replacement for insn %d\n",
4724 INSN_UID (insn: desc->insn));
4725
4726 success = validate_change (desc->insn, desc->loc, desc->newval, 0);
4727 gcc_assert (success);
4728
4729 rtx_insn *insn = DEP_PRO (dep);
4730
4731 /* Recompute priority since dependent priorities may have changed. */
4732 priority (insn, force_recompute: true);
4733 update_insn_after_change (insn: desc->insn);
4734
4735 if ((TODO_SPEC (desc->insn) & (HARD_DEP | DEP_POSTPONED)) == 0)
4736 fix_tick_ready (desc->insn);
4737
4738 if (backtrack_queue != NULL)
4739 {
4740 backtrack_queue->replacement_deps.safe_push (obj: dep);
4741 backtrack_queue->replace_apply.safe_push (obj: 1);
4742 }
4743 }
4744}
4745
4746/* We have determined that a pattern involved in DEP must be restored.
4747 If IMMEDIATELY is false, we may have to postpone the replacement
4748 until the start of the next cycle, at which point we will be called
4749 again with IMMEDIATELY true. */
4750static void
4751restore_pattern (dep_t dep, bool immediately)
4752{
4753 rtx_insn *next = DEP_CON (dep);
4754 int tick = INSN_TICK (next);
4755
4756 /* If we already scheduled the insn, the modified version is
4757 correct. */
4758 if (QUEUE_INDEX (next) == QUEUE_SCHEDULED)
4759 return;
4760
4761 if (!immediately && targetm.sched.exposed_pipeline && reload_completed)
4762 {
4763 next_cycle_replace_deps.safe_push (obj: dep);
4764 next_cycle_apply.safe_push (obj: 0);
4765 return;
4766 }
4767
4768
4769 if (DEP_TYPE (dep) == REG_DEP_CONTROL)
4770 {
4771 if (sched_verbose >= 5)
4772 fprintf (stream: sched_dump, format: "restoring pattern for insn %d\n",
4773 INSN_UID (insn: next));
4774 haifa_change_pattern (next, ORIG_PAT (next));
4775 }
4776 else
4777 {
4778 struct dep_replacement *desc = DEP_REPLACE (dep);
4779 bool success;
4780
4781 if (sched_verbose >= 5)
4782 fprintf (stream: sched_dump, format: "restoring pattern for insn %d\n",
4783 INSN_UID (insn: desc->insn));
4784 tick = INSN_TICK (desc->insn);
4785
4786 success = validate_change (desc->insn, desc->loc, desc->orig, 0);
4787 gcc_assert (success);
4788
4789 rtx_insn *insn = DEP_PRO (dep);
4790
4791 if (QUEUE_INDEX (insn) != QUEUE_SCHEDULED)
4792 {
4793 /* Recompute priority since dependent priorities may have changed. */
4794 priority (insn, force_recompute: true);
4795 }
4796
4797 update_insn_after_change (insn: desc->insn);
4798
4799 if (backtrack_queue != NULL)
4800 {
4801 backtrack_queue->replacement_deps.safe_push (obj: dep);
4802 backtrack_queue->replace_apply.safe_push (obj: 0);
4803 }
4804 }
4805 INSN_TICK (next) = tick;
4806 if (TODO_SPEC (next) == DEP_POSTPONED)
4807 return;
4808
4809 if (sd_lists_empty_p (next, SD_LIST_BACK))
4810 TODO_SPEC (next) = 0;
4811 else if (!sd_lists_empty_p (next, SD_LIST_HARD_BACK))
4812 TODO_SPEC (next) = HARD_DEP;
4813}
4814
4815/* Perform pattern replacements that were queued up until the next
4816 cycle. */
4817static void
4818perform_replacements_new_cycle (void)
4819{
4820 int i;
4821 dep_t dep;
4822 FOR_EACH_VEC_ELT (next_cycle_replace_deps, i, dep)
4823 {
4824 int apply_p = next_cycle_apply[i];
4825 if (apply_p)
4826 apply_replacement (dep, immediately: true);
4827 else
4828 restore_pattern (dep, immediately: true);
4829 }
4830 next_cycle_replace_deps.truncate (size: 0);
4831 next_cycle_apply.truncate (size: 0);
4832}
4833
4834/* Compute INSN_TICK_ESTIMATE for INSN. PROCESSED is a bitmap of
4835 instructions we've previously encountered, a set bit prevents
4836 recursion. BUDGET is a limit on how far ahead we look, it is
4837 reduced on recursive calls. Return true if we produced a good
4838 estimate, or false if we exceeded the budget. */
4839static bool
4840estimate_insn_tick (bitmap processed, rtx_insn *insn, int budget)
4841{
4842 sd_iterator_def sd_it;
4843 dep_t dep;
4844 int earliest = INSN_TICK (insn);
4845
4846 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
4847 {
4848 rtx_insn *pro = DEP_PRO (dep);
4849 int t;
4850
4851 if (DEP_STATUS (dep) & DEP_CANCELLED)
4852 continue;
4853
4854 if (QUEUE_INDEX (pro) == QUEUE_SCHEDULED)
4855 gcc_assert (INSN_TICK (pro) + dep_cost (dep) <= INSN_TICK (insn));
4856 else
4857 {
4858 int cost = dep_cost (link: dep);
4859 if (cost >= budget)
4860 return false;
4861 if (!bitmap_bit_p (processed, INSN_LUID (pro)))
4862 {
4863 if (!estimate_insn_tick (processed, insn: pro, budget: budget - cost))
4864 return false;
4865 }
4866 gcc_assert (INSN_TICK_ESTIMATE (pro) != INVALID_TICK);
4867 t = INSN_TICK_ESTIMATE (pro) + cost;
4868 if (earliest == INVALID_TICK || t > earliest)
4869 earliest = t;
4870 }
4871 }
4872 bitmap_set_bit (processed, INSN_LUID (insn));
4873 INSN_TICK_ESTIMATE (insn) = earliest;
4874 return true;
4875}
4876
4877/* Examine the pair of insns in P, and estimate (optimistically, assuming
4878 infinite resources) the cycle in which the delayed shadow can be issued.
4879 Return the number of cycles that must pass before the real insn can be
4880 issued in order to meet this constraint. */
4881static int
4882estimate_shadow_tick (struct delay_pair *p)
4883{
4884 auto_bitmap processed;
4885 int t;
4886 bool cutoff;
4887
4888 cutoff = !estimate_insn_tick (processed, insn: p->i2,
4889 budget: max_insn_queue_index + pair_delay (p));
4890 if (cutoff)
4891 return max_insn_queue_index;
4892 t = INSN_TICK_ESTIMATE (p->i2) - (clock_var + pair_delay (p) + 1);
4893 if (t > 0)
4894 return t;
4895 return 0;
4896}
4897
4898/* If INSN has no unresolved backwards dependencies, add it to the schedule and
4899 recursively resolve all its forward dependencies. */
4900static void
4901resolve_dependencies (rtx_insn *insn)
4902{
4903 sd_iterator_def sd_it;
4904 dep_t dep;
4905
4906 /* Don't use sd_lists_empty_p; it ignores debug insns. */
4907 if (DEPS_LIST_FIRST (INSN_HARD_BACK_DEPS (insn)) != NULL
4908 || DEPS_LIST_FIRST (INSN_SPEC_BACK_DEPS (insn)) != NULL)
4909 return;
4910
4911 if (sched_verbose >= 4)
4912 fprintf (stream: sched_dump, format: ";;\tquickly resolving %d\n", INSN_UID (insn));
4913
4914 if (QUEUE_INDEX (insn) >= 0)
4915 queue_remove (insn);
4916
4917 scheduled_insns.safe_push (obj: insn);
4918
4919 /* Update dependent instructions. */
4920 for (sd_it = sd_iterator_start (insn, SD_LIST_FORW);
4921 sd_iterator_cond (it_ptr: &sd_it, dep_ptr: &dep);)
4922 {
4923 rtx_insn *next = DEP_CON (dep);
4924
4925 if (sched_verbose >= 4)
4926 fprintf (stream: sched_dump, format: ";;\t\tdep %d against %d\n", INSN_UID (insn),
4927 INSN_UID (insn: next));
4928
4929 /* Resolve the dependence between INSN and NEXT.
4930 sd_resolve_dep () moves current dep to another list thus
4931 advancing the iterator. */
4932 sd_resolve_dep (sd_it);
4933
4934 if (!IS_SPECULATION_BRANCHY_CHECK_P (insn))
4935 {
4936 resolve_dependencies (insn: next);
4937 }
4938 else
4939 /* Check always has only one forward dependence (to the first insn in
4940 the recovery block), therefore, this will be executed only once. */
4941 {
4942 gcc_assert (sd_lists_empty_p (insn, SD_LIST_FORW));
4943 }
4944 }
4945}
4946
4947
4948/* Return the head and tail pointers of ebb starting at BEG and ending
4949 at END. */
4950void
4951get_ebb_head_tail (basic_block beg, basic_block end,
4952 rtx_insn **headp, rtx_insn **tailp)
4953{
4954 rtx_insn *beg_head = BB_HEAD (beg);
4955 rtx_insn * beg_tail = BB_END (beg);
4956 rtx_insn * end_head = BB_HEAD (end);
4957 rtx_insn * end_tail = BB_END (end);
4958
4959 /* Don't include any notes or labels at the beginning of the BEG
4960 basic block, or notes at the end of the END basic blocks. */
4961
4962 if (LABEL_P (beg_head))
4963 beg_head = NEXT_INSN (insn: beg_head);
4964
4965 while (beg_head != beg_tail)
4966 if (NOTE_P (beg_head))
4967 beg_head = NEXT_INSN (insn: beg_head);
4968 else if (DEBUG_INSN_P (beg_head))
4969 {
4970 rtx_insn * note, *next;
4971
4972 for (note = NEXT_INSN (insn: beg_head);
4973 note != beg_tail;
4974 note = next)
4975 {
4976 next = NEXT_INSN (insn: note);
4977 if (NOTE_P (note))
4978 {
4979 if (sched_verbose >= 9)
4980 fprintf (stream: sched_dump, format: "reorder %i\n", INSN_UID (insn: note));
4981
4982 reorder_insns_nobb (note, note, PREV_INSN (insn: beg_head));
4983
4984 if (BLOCK_FOR_INSN (insn: note) != beg)
4985 df_insn_change_bb (note, beg);
4986 }
4987 else if (!DEBUG_INSN_P (note))
4988 break;
4989 }
4990
4991 break;
4992 }
4993 else
4994 break;
4995
4996 *headp = beg_head;
4997
4998 if (beg == end)
4999 end_head = beg_head;
5000 else if (LABEL_P (end_head))
5001 end_head = NEXT_INSN (insn: end_head);
5002
5003 while (end_head != end_tail)
5004 if (NOTE_P (end_tail))
5005 end_tail = PREV_INSN (insn: end_tail);
5006 else if (DEBUG_INSN_P (end_tail))
5007 {
5008 rtx_insn * note, *prev;
5009
5010 for (note = PREV_INSN (insn: end_tail);
5011 note != end_head;
5012 note = prev)
5013 {
5014 prev = PREV_INSN (insn: note);
5015 if (NOTE_P (note))
5016 {
5017 if (sched_verbose >= 9)
5018 fprintf (stream: sched_dump, format: "reorder %i\n", INSN_UID (insn: note));
5019
5020 reorder_insns_nobb (note, note, end_tail);
5021
5022 if (end_tail == BB_END (end))
5023 BB_END (end) = note;
5024
5025 if (BLOCK_FOR_INSN (insn: note) != end)
5026 df_insn_change_bb (note, end);
5027 }
5028 else if (!DEBUG_INSN_P (note))
5029 break;
5030 }
5031
5032 break;
5033 }
5034 else
5035 break;
5036
5037 *tailp = end_tail;
5038}
5039
5040/* Return true if there are no real insns in the range [ HEAD, TAIL ]. */
5041
5042bool
5043no_real_insns_p (const rtx_insn *head, const rtx_insn *tail)
5044{
5045 while (head != NEXT_INSN (insn: tail))
5046 {
5047 if (!NOTE_P (head) && !LABEL_P (head))
5048 return false;
5049 head = NEXT_INSN (insn: head);
5050 }
5051 return true;
5052}
5053
5054/* Restore-other-notes: NOTE_LIST is the end of a chain of notes
5055 previously found among the insns. Insert them just before HEAD. */
5056rtx_insn *
5057restore_other_notes (rtx_insn *head, basic_block head_bb)
5058{
5059 if (note_list != 0)
5060 {
5061 rtx_insn *note_head = note_list;
5062
5063 if (head)
5064 head_bb = BLOCK_FOR_INSN (insn: head);
5065 else
5066 head = NEXT_INSN (insn: bb_note (head_bb));
5067
5068 while (PREV_INSN (insn: note_head))
5069 {
5070 set_block_for_insn (insn: note_head, bb: head_bb);
5071 note_head = PREV_INSN (insn: note_head);
5072 }
5073 /* In the above cycle we've missed this note. */
5074 set_block_for_insn (insn: note_head, bb: head_bb);
5075
5076 SET_PREV_INSN (note_head) = PREV_INSN (insn: head);
5077 SET_NEXT_INSN (PREV_INSN (insn: head)) = note_head;
5078 SET_PREV_INSN (head) = note_list;
5079 SET_NEXT_INSN (note_list) = head;
5080
5081 if (BLOCK_FOR_INSN (insn: head) != head_bb)
5082 BB_END (head_bb) = note_list;
5083
5084 head = note_head;
5085 }
5086
5087 return head;
5088}
5089
5090/* When we know we are going to discard the schedule due to a failed attempt
5091 at modulo scheduling, undo all replacements. */
5092static void
5093undo_all_replacements (void)
5094{
5095 rtx_insn *insn;
5096 int i;
5097
5098 FOR_EACH_VEC_ELT (scheduled_insns, i, insn)
5099 {
5100 sd_iterator_def sd_it;
5101 dep_t dep;
5102
5103 /* See if we must undo a replacement. */
5104 for (sd_it = sd_iterator_start (insn, SD_LIST_RES_FORW);
5105 sd_iterator_cond (it_ptr: &sd_it, dep_ptr: &dep); sd_iterator_next (it_ptr: &sd_it))
5106 {
5107 struct dep_replacement *desc = DEP_REPLACE (dep);
5108 if (desc != NULL)
5109 validate_change (desc->insn, desc->loc, desc->orig, 0);
5110 }
5111 }
5112}
5113
5114/* Return first non-scheduled insn in the current scheduling block.
5115 This is mostly used for debug-counter purposes. */
5116static rtx_insn *
5117first_nonscheduled_insn (void)
5118{
5119 rtx_insn *insn = (nonscheduled_insns_begin != NULL_RTX
5120 ? nonscheduled_insns_begin
5121 : current_sched_info->prev_head);
5122
5123 do
5124 {
5125 insn = next_nonnote_nondebug_insn (insn);
5126 }
5127 while (QUEUE_INDEX (insn) == QUEUE_SCHEDULED);
5128
5129 return insn;
5130}
5131
5132/* Move insns that became ready to fire from queue to ready list. */
5133
5134static void
5135queue_to_ready (struct ready_list *ready)
5136{
5137 rtx_insn *insn;
5138 rtx_insn_list *link;
5139 rtx_insn *skip_insn;
5140
5141 q_ptr = NEXT_Q (q_ptr);
5142
5143 if (dbg_cnt (index: sched_insn) == false)
5144 /* If debug counter is activated do not requeue the first
5145 nonscheduled insn. */
5146 skip_insn = first_nonscheduled_insn ();
5147 else
5148 skip_insn = NULL;
5149
5150 /* Add all pending insns that can be scheduled without stalls to the
5151 ready list. */
5152 for (link = insn_queue[q_ptr]; link; link = link->next ())
5153 {
5154 insn = link->insn ();
5155 q_size -= 1;
5156
5157 if (sched_verbose >= 2)
5158 fprintf (stream: sched_dump, format: ";;\t\tQ-->Ready: insn %s: ",
5159 (*current_sched_info->print_insn) (insn, 0));
5160
5161 /* If the ready list is full, delay the insn for 1 cycle.
5162 See the comment in schedule_block for the rationale. */
5163 if (!reload_completed
5164 && (ready->n_ready - ready->n_debug > param_max_sched_ready_insns
5165 || (sched_pressure == SCHED_PRESSURE_MODEL
5166 /* Limit pressure recalculations to
5167 param_max_sched_ready_insns instructions too. */
5168 && model_index (insn) > (model_curr_point
5169 + param_max_sched_ready_insns)))
5170 && !(sched_pressure == SCHED_PRESSURE_MODEL
5171 && model_curr_point < model_num_insns
5172 /* Always allow the next model instruction to issue. */
5173 && model_index (insn) == model_curr_point)
5174 && !SCHED_GROUP_P (insn)
5175 && insn != skip_insn)
5176 {
5177 if (sched_verbose >= 2)
5178 fprintf (stream: sched_dump, format: "keeping in queue, ready full\n");
5179 queue_insn (insn, n_cycles: 1, reason: "ready full");
5180 }
5181 else
5182 {
5183 ready_add (ready, insn, first_p: false);
5184 if (sched_verbose >= 2)
5185 fprintf (stream: sched_dump, format: "moving to ready without stalls\n");
5186 }
5187 }
5188 free_INSN_LIST_list (&insn_queue[q_ptr]);
5189
5190 /* If there are no ready insns, stall until one is ready and add all
5191 of the pending insns at that point to the ready list. */
5192 if (ready->n_ready == 0)
5193 {
5194 int stalls;
5195
5196 for (stalls = 1; stalls <= max_insn_queue_index; stalls++)
5197 {
5198 if ((link = insn_queue[NEXT_Q_AFTER (q_ptr, stalls)]))
5199 {
5200 for (; link; link = link->next ())
5201 {
5202 insn = link->insn ();
5203 q_size -= 1;
5204
5205 if (sched_verbose >= 2)
5206 fprintf (stream: sched_dump, format: ";;\t\tQ-->Ready: insn %s: ",
5207 (*current_sched_info->print_insn) (insn, 0));
5208
5209 ready_add (ready, insn, first_p: false);
5210 if (sched_verbose >= 2)
5211 fprintf (stream: sched_dump, format: "moving to ready with %d stalls\n", stalls);
5212 }
5213 free_INSN_LIST_list (&insn_queue[NEXT_Q_AFTER (q_ptr, stalls)]);
5214
5215 advance_one_cycle ();
5216
5217 break;
5218 }
5219
5220 advance_one_cycle ();
5221 }
5222
5223 q_ptr = NEXT_Q_AFTER (q_ptr, stalls);
5224 clock_var += stalls;
5225 if (sched_verbose >= 2)
5226 fprintf (stream: sched_dump, format: ";;\tAdvancing clock by %d cycle[s] to %d\n",
5227 stalls, clock_var);
5228 }
5229}
5230
5231/* Used by early_queue_to_ready. Determines whether it is "ok" to
5232 prematurely move INSN from the queue to the ready list. Currently,
5233 if a target defines the hook 'is_costly_dependence', this function
5234 uses the hook to check whether there exist any dependences which are
5235 considered costly by the target, between INSN and other insns that
5236 have already been scheduled. Dependences are checked up to Y cycles
5237 back, with default Y=1; The flag -fsched-stalled-insns-dep=Y allows
5238 controlling this value.
5239 (Other considerations could be taken into account instead (or in
5240 addition) depending on user flags and target hooks. */
5241
5242static bool
5243ok_for_early_queue_removal (rtx_insn *insn)
5244{
5245 if (targetm.sched.is_costly_dependence)
5246 {
5247 int n_cycles;
5248 int i = scheduled_insns.length ();
5249 for (n_cycles = flag_sched_stalled_insns_dep; n_cycles; n_cycles--)
5250 {
5251 while (i-- > 0)
5252 {
5253 int cost;
5254
5255 rtx_insn *prev_insn = scheduled_insns[i];
5256
5257 if (!NOTE_P (prev_insn))
5258 {
5259 dep_t dep;
5260
5261 dep = sd_find_dep_between (prev_insn, insn, true);
5262
5263 if (dep != NULL)
5264 {
5265 cost = dep_cost (link: dep);
5266
5267 if (targetm.sched.is_costly_dependence (dep, cost,
5268 flag_sched_stalled_insns_dep - n_cycles))
5269 return false;
5270 }
5271 }
5272
5273 if (GET_MODE (prev_insn) == TImode) /* end of dispatch group */
5274 break;
5275 }
5276
5277 if (i == 0)
5278 break;
5279 }
5280 }
5281
5282 return true;
5283}
5284
5285
5286/* Remove insns from the queue, before they become "ready" with respect
5287 to FU latency considerations. */
5288
5289static int
5290early_queue_to_ready (state_t state, struct ready_list *ready)
5291{
5292 rtx_insn *insn;
5293 rtx_insn_list *link;
5294 rtx_insn_list *next_link;
5295 rtx_insn_list *prev_link;
5296 bool move_to_ready;
5297 int cost;
5298 state_t temp_state = alloca (dfa_state_size);
5299 int stalls;
5300 int insns_removed = 0;
5301
5302 /*
5303 Flag '-fsched-stalled-insns=X' determines the aggressiveness of this
5304 function:
5305
5306 X == 0: There is no limit on how many queued insns can be removed
5307 prematurely. (flag_sched_stalled_insns = -1).
5308
5309 X >= 1: Only X queued insns can be removed prematurely in each
5310 invocation. (flag_sched_stalled_insns = X).
5311
5312 Otherwise: Early queue removal is disabled.
5313 (flag_sched_stalled_insns = 0)
5314 */
5315
5316 if (! flag_sched_stalled_insns)
5317 return 0;
5318
5319 for (stalls = 0; stalls <= max_insn_queue_index; stalls++)
5320 {
5321 if ((link = insn_queue[NEXT_Q_AFTER (q_ptr, stalls)]))
5322 {
5323 if (sched_verbose > 6)
5324 fprintf (stream: sched_dump, format: ";; look at index %d + %d\n", q_ptr, stalls);
5325
5326 prev_link = 0;
5327 while (link)
5328 {
5329 next_link = link->next ();
5330 insn = link->insn ();
5331 if (insn && sched_verbose > 6)
5332 print_rtl_single (sched_dump, insn);
5333
5334 memcpy (dest: temp_state, src: state, n: dfa_state_size);
5335 if (recog_memoized (insn) < 0)
5336 /* non-negative to indicate that it's not ready
5337 to avoid infinite Q->R->Q->R... */
5338 cost = 0;
5339 else
5340 cost = state_transition (temp_state, insn);
5341
5342 if (sched_verbose >= 6)
5343 fprintf (stream: sched_dump, format: "transition cost = %d\n", cost);
5344
5345 move_to_ready = false;
5346 if (cost < 0)
5347 {
5348 move_to_ready = ok_for_early_queue_removal (insn);
5349 if (move_to_ready == true)
5350 {
5351 /* move from Q to R */
5352 q_size -= 1;
5353 ready_add (ready, insn, first_p: false);
5354
5355 if (prev_link)
5356 XEXP (prev_link, 1) = next_link;
5357 else
5358 insn_queue[NEXT_Q_AFTER (q_ptr, stalls)] = next_link;
5359
5360 free_INSN_LIST_node (link);
5361
5362 if (sched_verbose >= 2)
5363 fprintf (stream: sched_dump, format: ";;\t\tEarly Q-->Ready: insn %s\n",
5364 (*current_sched_info->print_insn) (insn, 0));
5365
5366 insns_removed++;
5367 if (insns_removed == flag_sched_stalled_insns)
5368 /* Remove no more than flag_sched_stalled_insns insns
5369 from Q at a time. */
5370 return insns_removed;
5371 }
5372 }
5373
5374 if (move_to_ready == false)
5375 prev_link = link;
5376
5377 link = next_link;
5378 } /* while link */
5379 } /* if link */
5380
5381 } /* for stalls.. */
5382
5383 return insns_removed;
5384}
5385
5386
5387/* Print the ready list for debugging purposes.
5388 If READY_TRY is non-zero then only print insns that max_issue
5389 will consider. */
5390static void
5391debug_ready_list_1 (struct ready_list *ready, signed char *ready_try)
5392{
5393 rtx_insn **p;
5394 int i;
5395
5396 if (ready->n_ready == 0)
5397 {
5398 fprintf (stream: sched_dump, format: "\n");
5399 return;
5400 }
5401
5402 p = ready_lastpos (ready);
5403 for (i = 0; i < ready->n_ready; i++)
5404 {
5405 if (ready_try != NULL && ready_try[ready->n_ready - i - 1])
5406 continue;
5407
5408 fprintf (stream: sched_dump, format: " %s:%d",
5409 (*current_sched_info->print_insn) (p[i], 0),
5410 INSN_LUID (p[i]));
5411 if (sched_pressure != SCHED_PRESSURE_NONE)
5412 fprintf (stream: sched_dump, format: "(cost=%d",
5413 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (p[i]));
5414 fprintf (stream: sched_dump, format: ":prio=%d", INSN_PRIORITY (p[i]));
5415 if (INSN_TICK (p[i]) > clock_var)
5416 fprintf (stream: sched_dump, format: ":delay=%d", INSN_TICK (p[i]) - clock_var);
5417 if (sched_pressure == SCHED_PRESSURE_MODEL)
5418 fprintf (stream: sched_dump, format: ":idx=%d",
5419 model_index (insn: p[i]));
5420 if (sched_pressure != SCHED_PRESSURE_NONE)
5421 fprintf (stream: sched_dump, format: ")");
5422 }
5423 fprintf (stream: sched_dump, format: "\n");
5424}
5425
5426/* Print the ready list. Callable from debugger. */
5427static void
5428debug_ready_list (struct ready_list *ready)
5429{
5430 debug_ready_list_1 (ready, NULL);
5431}
5432
5433/* Search INSN for REG_SAVE_NOTE notes and convert them back into insn
5434 NOTEs. This is used for NOTE_INSN_EPILOGUE_BEG, so that sched-ebb
5435 replaces the epilogue note in the correct basic block. */
5436void
5437reemit_notes (rtx_insn *insn)
5438{
5439 rtx note;
5440 rtx_insn *last = insn;
5441
5442 for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
5443 {
5444 if (REG_NOTE_KIND (note) == REG_SAVE_NOTE)
5445 {
5446 enum insn_note note_type = (enum insn_note) INTVAL (XEXP (note, 0));
5447
5448 last = emit_note_before (note_type, last);
5449 remove_note (insn, note);
5450 df_insn_create_insn_record (last);
5451 }
5452 }
5453}
5454
5455/* Move INSN. Reemit notes if needed. Update CFG, if needed. */
5456static void
5457move_insn (rtx_insn *insn, rtx_insn *last, rtx nt)
5458{
5459 if (PREV_INSN (insn) != last)
5460 {
5461 basic_block bb;
5462 rtx_insn *note;
5463 int jump_p = 0;
5464
5465 bb = BLOCK_FOR_INSN (insn);
5466
5467 /* BB_HEAD is either LABEL or NOTE. */
5468 gcc_assert (BB_HEAD (bb) != insn);
5469
5470 if (BB_END (bb) == insn)
5471 /* If this is last instruction in BB, move end marker one
5472 instruction up. */
5473 {
5474 /* Jumps are always placed at the end of basic block. */
5475 jump_p = control_flow_insn_p (insn);
5476
5477 gcc_assert (!jump_p
5478 || ((common_sched_info->sched_pass_id == SCHED_RGN_PASS)
5479 && IS_SPECULATION_BRANCHY_CHECK_P (insn))
5480 || (common_sched_info->sched_pass_id
5481 == SCHED_EBB_PASS));
5482
5483 gcc_assert (BLOCK_FOR_INSN (PREV_INSN (insn)) == bb);
5484
5485 BB_END (bb) = PREV_INSN (insn);
5486 }
5487
5488 gcc_assert (BB_END (bb) != last);
5489
5490 if (jump_p)
5491 /* We move the block note along with jump. */
5492 {
5493 gcc_assert (nt);
5494
5495 note = NEXT_INSN (insn);
5496 while (NOTE_NOT_BB_P (note) && note != nt)
5497 note = NEXT_INSN (insn: note);
5498
5499 if (note != nt
5500 && (LABEL_P (note)
5501 || BARRIER_P (note)))
5502 note = NEXT_INSN (insn: note);
5503
5504 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
5505 }
5506 else
5507 note = insn;
5508
5509 SET_NEXT_INSN (PREV_INSN (insn)) = NEXT_INSN (insn: note);
5510 SET_PREV_INSN (NEXT_INSN (insn: note)) = PREV_INSN (insn);
5511
5512 SET_NEXT_INSN (note) = NEXT_INSN (insn: last);
5513 SET_PREV_INSN (NEXT_INSN (insn: last)) = note;
5514
5515 SET_NEXT_INSN (last) = insn;
5516 SET_PREV_INSN (insn) = last;
5517
5518 bb = BLOCK_FOR_INSN (insn: last);
5519
5520 if (jump_p)
5521 {
5522 fix_jump_move (insn);
5523
5524 if (BLOCK_FOR_INSN (insn) != bb)
5525 move_block_after_check (insn);
5526
5527 gcc_assert (BB_END (bb) == last);
5528 }
5529
5530 df_insn_change_bb (insn, bb);
5531
5532 /* Update BB_END, if needed. */
5533 if (BB_END (bb) == last)
5534 BB_END (bb) = insn;
5535 }
5536
5537 SCHED_GROUP_P (insn) = 0;
5538}
5539
5540/* Return true if scheduling INSN will finish current clock cycle. */
5541static bool
5542insn_finishes_cycle_p (rtx_insn *insn)
5543{
5544 if (SCHED_GROUP_P (insn))
5545 /* After issuing INSN, rest of the sched_group will be forced to issue
5546 in order. Don't make any plans for the rest of cycle. */
5547 return true;
5548
5549 /* Finishing the block will, apparently, finish the cycle. */
5550 if (current_sched_info->insn_finishes_block_p
5551 && current_sched_info->insn_finishes_block_p (insn))
5552 return true;
5553
5554 return false;
5555}
5556
5557/* Helper for autopref_multipass_init. Given a SET in PAT and whether
5558 we're expecting a memory WRITE or not, check that the insn is relevant to
5559 the autoprefetcher modelling code. Return true iff that is the case.
5560 If it is relevant, record the base register of the memory op in BASE and
5561 the offset in OFFSET. */
5562
5563static bool
5564analyze_set_insn_for_autopref (rtx pat, bool write, rtx *base, int *offset)
5565{
5566 if (GET_CODE (pat) != SET)
5567 return false;
5568
5569 rtx mem = write ? SET_DEST (pat) : SET_SRC (pat);
5570 if (!MEM_P (mem))
5571 return false;
5572
5573 struct address_info info;
5574 decompose_mem_address (&info, mem);
5575
5576 /* TODO: Currently only (base+const) addressing is supported. */
5577 if (info.base == NULL || !REG_P (*info.base)
5578 || (info.disp != NULL && !CONST_INT_P (*info.disp)))
5579 return false;
5580
5581 *base = *info.base;
5582 *offset = info.disp ? INTVAL (*info.disp) : 0;
5583 return true;
5584}
5585
5586/* Functions to model cache auto-prefetcher.
5587
5588 Some of the CPUs have cache auto-prefetcher, which /seems/ to initiate
5589 memory prefetches if it sees instructions with consequitive memory accesses
5590 in the instruction stream. Details of such hardware units are not published,
5591 so we can only guess what exactly is going on there.
5592 In the scheduler, we model abstract auto-prefetcher. If there are memory
5593 insns in the ready list (or the queue) that have same memory base, but
5594 different offsets, then we delay the insns with larger offsets until insns
5595 with smaller offsets get scheduled. If PARAM_SCHED_AUTOPREF_QUEUE_DEPTH
5596 is "1", then we look at the ready list; if it is N>1, then we also look
5597 through N-1 queue entries.
5598 If the param is N>=0, then rank_for_schedule will consider auto-prefetching
5599 among its heuristics.
5600 Param value of "-1" disables modelling of the auto-prefetcher. */
5601
5602/* Initialize autoprefetcher model data for INSN. */
5603static void
5604autopref_multipass_init (const rtx_insn *insn, int write)
5605{
5606 autopref_multipass_data_t data = &INSN_AUTOPREF_MULTIPASS_DATA (insn)[write];
5607
5608 gcc_assert (data->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED);
5609 data->base = NULL_RTX;
5610 data->offset = 0;
5611 /* Set insn entry initialized, but not relevant for auto-prefetcher. */
5612 data->status = AUTOPREF_MULTIPASS_DATA_IRRELEVANT;
5613
5614 rtx pat = PATTERN (insn);
5615
5616 /* We have a multi-set insn like a load-multiple or store-multiple.
5617 We care about these as long as all the memory ops inside the PARALLEL
5618 have the same base register. We care about the minimum and maximum
5619 offsets from that base but don't check for the order of those offsets
5620 within the PARALLEL insn itself. */
5621 if (GET_CODE (pat) == PARALLEL)
5622 {
5623 int n_elems = XVECLEN (pat, 0);
5624
5625 int i, offset;
5626 rtx base, prev_base = NULL_RTX;
5627 int min_offset = INT_MAX;
5628
5629 for (i = 0; i < n_elems; i++)
5630 {
5631 rtx set = XVECEXP (pat, 0, i);
5632 if (GET_CODE (set) != SET)
5633 return;
5634
5635 if (!analyze_set_insn_for_autopref (pat: set, write, base: &base, offset: &offset))
5636 return;
5637
5638 /* Ensure that all memory operations in the PARALLEL use the same
5639 base register. */
5640 if (i > 0 && REGNO (base) != REGNO (prev_base))
5641 return;
5642 prev_base = base;
5643 min_offset = MIN (min_offset, offset);
5644 }
5645
5646 /* If we reached here then we have a valid PARALLEL of multiple memory ops
5647 with prev_base as the base and min_offset containing the offset. */
5648 gcc_assert (prev_base);
5649 data->base = prev_base;
5650 data->offset = min_offset;
5651 data->status = AUTOPREF_MULTIPASS_DATA_NORMAL;
5652 return;
5653 }
5654
5655 /* Otherwise this is a single set memory operation. */
5656 rtx set = single_set (insn);
5657 if (set == NULL_RTX)
5658 return;
5659
5660 if (!analyze_set_insn_for_autopref (pat: set, write, base: &data->base,
5661 offset: &data->offset))
5662 return;
5663
5664 /* This insn is relevant for the auto-prefetcher.
5665 The base and offset fields will have been filled in the
5666 analyze_set_insn_for_autopref call above. */
5667 data->status = AUTOPREF_MULTIPASS_DATA_NORMAL;
5668}
5669
5670/* Helper function for rank_for_schedule sorting. */
5671static int
5672autopref_rank_for_schedule (const rtx_insn *insn1, const rtx_insn *insn2)
5673{
5674 int r = 0;
5675 for (int write = 0; write < 2 && !r; ++write)
5676 {
5677 autopref_multipass_data_t data1
5678 = &INSN_AUTOPREF_MULTIPASS_DATA (insn1)[write];
5679 autopref_multipass_data_t data2
5680 = &INSN_AUTOPREF_MULTIPASS_DATA (insn2)[write];
5681
5682 if (data1->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED)
5683 autopref_multipass_init (insn: insn1, write);
5684
5685 if (data2->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED)
5686 autopref_multipass_init (insn: insn2, write);
5687
5688 int irrel1 = data1->status == AUTOPREF_MULTIPASS_DATA_IRRELEVANT;
5689 int irrel2 = data2->status == AUTOPREF_MULTIPASS_DATA_IRRELEVANT;
5690
5691 if (!irrel1 && !irrel2)
5692 /* Sort memory references from lowest offset to the largest. */
5693 r = (data1->offset > data2->offset) - (data1->offset < data2->offset);
5694 else if (write)
5695 /* Schedule "irrelevant" insns before memory stores to resolve
5696 as many producer dependencies of stores as possible. */
5697 r = irrel2 - irrel1;
5698 else
5699 /* Schedule "irrelevant" insns after memory reads to avoid breaking
5700 memory read sequences. */
5701 r = irrel1 - irrel2;
5702 }
5703
5704 return r;
5705}
5706
5707/* True if header of debug dump was printed. */
5708static bool autopref_multipass_dfa_lookahead_guard_started_dump_p;
5709
5710/* Helper for autopref_multipass_dfa_lookahead_guard.
5711 Return "1" if INSN1 should be delayed in favor of INSN2. */
5712static int
5713autopref_multipass_dfa_lookahead_guard_1 (const rtx_insn *insn1,
5714 const rtx_insn *insn2, int write)
5715{
5716 autopref_multipass_data_t data1
5717 = &INSN_AUTOPREF_MULTIPASS_DATA (insn1)[write];
5718 autopref_multipass_data_t data2
5719 = &INSN_AUTOPREF_MULTIPASS_DATA (insn2)[write];
5720
5721 if (data2->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED)
5722 autopref_multipass_init (insn: insn2, write);
5723 if (data2->status == AUTOPREF_MULTIPASS_DATA_IRRELEVANT)
5724 return 0;
5725
5726 if (rtx_equal_p (data1->base, data2->base)
5727 && data1->offset > data2->offset)
5728 {
5729 if (sched_verbose >= 2)
5730 {
5731 if (!autopref_multipass_dfa_lookahead_guard_started_dump_p)
5732 {
5733 fprintf (stream: sched_dump,
5734 format: ";;\t\tnot trying in max_issue due to autoprefetch "
5735 "model: ");
5736 autopref_multipass_dfa_lookahead_guard_started_dump_p = true;
5737 }
5738
5739 fprintf (stream: sched_dump, format: " %d(%d)", INSN_UID (insn: insn1), INSN_UID (insn: insn2));
5740 }
5741
5742 return 1;
5743 }
5744
5745 return 0;
5746}
5747
5748/* General note:
5749
5750 We could have also hooked autoprefetcher model into
5751 first_cycle_multipass_backtrack / first_cycle_multipass_issue hooks
5752 to enable intelligent selection of "[r1+0]=r2; [r1+4]=r3" on the same cycle
5753 (e.g., once "[r1+0]=r2" is issued in max_issue(), "[r1+4]=r3" gets
5754 unblocked). We don't bother about this yet because target of interest
5755 (ARM Cortex-A15) can issue only 1 memory operation per cycle. */
5756
5757/* Implementation of first_cycle_multipass_dfa_lookahead_guard hook.
5758 Return "1" if INSN1 should not be considered in max_issue due to
5759 auto-prefetcher considerations. */
5760int
5761autopref_multipass_dfa_lookahead_guard (rtx_insn *insn1, int ready_index)
5762{
5763 int r = 0;
5764
5765 /* Exit early if the param forbids this or if we're not entering here through
5766 normal haifa scheduling. This can happen if selective scheduling is
5767 explicitly enabled. */
5768 if (!insn_queue || param_sched_autopref_queue_depth <= 0)
5769 return 0;
5770
5771 if (sched_verbose >= 2 && ready_index == 0)
5772 autopref_multipass_dfa_lookahead_guard_started_dump_p = false;
5773
5774 for (int write = 0; write < 2; ++write)
5775 {
5776 autopref_multipass_data_t data1
5777 = &INSN_AUTOPREF_MULTIPASS_DATA (insn1)[write];
5778
5779 if (data1->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED)
5780 autopref_multipass_init (insn: insn1, write);
5781 if (data1->status == AUTOPREF_MULTIPASS_DATA_IRRELEVANT)
5782 continue;
5783
5784 if (ready_index == 0
5785 && data1->status == AUTOPREF_MULTIPASS_DATA_DONT_DELAY)
5786 /* We allow only a single delay on priviledged instructions.
5787 Doing otherwise would cause infinite loop. */
5788 {
5789 if (sched_verbose >= 2)
5790 {
5791 if (!autopref_multipass_dfa_lookahead_guard_started_dump_p)
5792 {
5793 fprintf (stream: sched_dump,
5794 format: ";;\t\tnot trying in max_issue due to autoprefetch "
5795 "model: ");
5796 autopref_multipass_dfa_lookahead_guard_started_dump_p = true;
5797 }
5798
5799 fprintf (stream: sched_dump, format: " *%d*", INSN_UID (insn: insn1));
5800 }
5801 continue;
5802 }
5803
5804 for (int i2 = 0; i2 < ready.n_ready; ++i2)
5805 {
5806 rtx_insn *insn2 = get_ready_element (i2);
5807 if (insn1 == insn2)
5808 continue;
5809 r = autopref_multipass_dfa_lookahead_guard_1 (insn1, insn2, write);
5810 if (r)
5811 {
5812 if (ready_index == 0)
5813 {
5814 r = -1;
5815 data1->status = AUTOPREF_MULTIPASS_DATA_DONT_DELAY;
5816 }
5817 goto finish;
5818 }
5819 }
5820
5821 if (param_sched_autopref_queue_depth == 1)
5822 continue;
5823
5824 /* Everything from the current queue slot should have been moved to
5825 the ready list. */
5826 gcc_assert (insn_queue[NEXT_Q_AFTER (q_ptr, 0)] == NULL_RTX);
5827
5828 int n_stalls = param_sched_autopref_queue_depth - 1;
5829 if (n_stalls > max_insn_queue_index)
5830 n_stalls = max_insn_queue_index;
5831
5832 for (int stalls = 1; stalls <= n_stalls; ++stalls)
5833 {
5834 for (rtx_insn_list *link = insn_queue[NEXT_Q_AFTER (q_ptr, stalls)];
5835 link != NULL_RTX;
5836 link = link->next ())
5837 {
5838 rtx_insn *insn2 = link->insn ();
5839 r = autopref_multipass_dfa_lookahead_guard_1 (insn1, insn2,
5840 write);
5841 if (r)
5842 {
5843 /* Queue INSN1 until INSN2 can issue. */
5844 r = -stalls;
5845 if (ready_index == 0)
5846 data1->status = AUTOPREF_MULTIPASS_DATA_DONT_DELAY;
5847 goto finish;
5848 }
5849 }
5850 }
5851 }
5852
5853 finish:
5854 if (sched_verbose >= 2
5855 && autopref_multipass_dfa_lookahead_guard_started_dump_p
5856 && (ready_index == ready.n_ready - 1 || r < 0))
5857 /* This does not /always/ trigger. We don't output EOL if the last
5858 insn is not recognized (INSN_CODE < 0) and lookahead_guard is not
5859 called. We can live with this. */
5860 fprintf (stream: sched_dump, format: "\n");
5861
5862 return r;
5863}
5864
5865/* Define type for target data used in multipass scheduling. */
5866#ifndef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DATA_T
5867# define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DATA_T int
5868#endif
5869typedef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DATA_T first_cycle_multipass_data_t;
5870
5871/* The following structure describe an entry of the stack of choices. */
5872struct choice_entry
5873{
5874 /* Ordinal number of the issued insn in the ready queue. */
5875 int index;
5876 /* The number of the rest insns whose issues we should try. */
5877 int rest;
5878 /* The number of issued essential insns. */
5879 int n;
5880 /* State after issuing the insn. */
5881 state_t state;
5882 /* Target-specific data. */
5883 first_cycle_multipass_data_t target_data;
5884};
5885
5886/* The following array is used to implement a stack of choices used in
5887 function max_issue. */
5888static struct choice_entry *choice_stack;
5889
5890/* This holds the value of the target dfa_lookahead hook. */
5891int dfa_lookahead;
5892
5893/* The following variable value is maximal number of tries of issuing
5894 insns for the first cycle multipass insn scheduling. We define
5895 this value as constant*(DFA_LOOKAHEAD**ISSUE_RATE). We would not
5896 need this constraint if all real insns (with non-negative codes)
5897 had reservations because in this case the algorithm complexity is
5898 O(DFA_LOOKAHEAD**ISSUE_RATE). Unfortunately, the dfa descriptions
5899 might be incomplete and such insn might occur. For such
5900 descriptions, the complexity of algorithm (without the constraint)
5901 could achieve DFA_LOOKAHEAD ** N , where N is the queue length. */
5902static int max_lookahead_tries;
5903
5904/* The following function returns maximal (or close to maximal) number
5905 of insns which can be issued on the same cycle and one of which
5906 insns is insns with the best rank (the first insn in READY). To
5907 make this function tries different samples of ready insns. READY
5908 is current queue `ready'. Global array READY_TRY reflects what
5909 insns are already issued in this try. The function stops immediately,
5910 if it reached the such a solution, that all instruction can be issued.
5911 INDEX will contain index of the best insn in READY. The following
5912 function is used only for first cycle multipass scheduling.
5913
5914 PRIVILEGED_N >= 0
5915
5916 This function expects recognized insns only. All USEs,
5917 CLOBBERs, etc must be filtered elsewhere. */
5918int
5919max_issue (struct ready_list *ready, int privileged_n, state_t state,
5920 bool first_cycle_insn_p, int *index)
5921{
5922 int n, i, all, n_ready, best, delay, tries_num;
5923 int more_issue;
5924 struct choice_entry *top;
5925 rtx_insn *insn;
5926
5927 if (sched_fusion)
5928 return 0;
5929
5930 n_ready = ready->n_ready;
5931 gcc_assert (dfa_lookahead >= 1 && privileged_n >= 0
5932 && privileged_n <= n_ready);
5933
5934 /* Init MAX_LOOKAHEAD_TRIES. */
5935 if (max_lookahead_tries == 0)
5936 {
5937 max_lookahead_tries = 100;
5938 for (i = 0; i < issue_rate; i++)
5939 max_lookahead_tries *= dfa_lookahead;
5940 }
5941
5942 /* Init max_points. */
5943 more_issue = issue_rate - cycle_issued_insns;
5944 gcc_assert (more_issue >= 0);
5945
5946 /* The number of the issued insns in the best solution. */
5947 best = 0;
5948
5949 top = choice_stack;
5950
5951 /* Set initial state of the search. */
5952 memcpy (dest: top->state, src: state, n: dfa_state_size);
5953 top->rest = dfa_lookahead;
5954 top->n = 0;
5955 if (targetm.sched.first_cycle_multipass_begin)
5956 targetm.sched.first_cycle_multipass_begin (&top->target_data,
5957 ready_try, n_ready,
5958 first_cycle_insn_p);
5959
5960 /* Count the number of the insns to search among. */
5961 for (all = i = 0; i < n_ready; i++)
5962 if (!ready_try [i])
5963 all++;
5964
5965 if (sched_verbose >= 2)
5966 {
5967 fprintf (stream: sched_dump, format: ";;\t\tmax_issue among %d insns:", all);
5968 debug_ready_list_1 (ready, ready_try);
5969 }
5970
5971 /* I is the index of the insn to try next. */
5972 i = 0;
5973 tries_num = 0;
5974 for (;;)
5975 {
5976 if (/* If we've reached a dead end or searched enough of what we have
5977 been asked... */
5978 top->rest == 0
5979 /* or have nothing else to try... */
5980 || i >= n_ready
5981 /* or should not issue more. */
5982 || top->n >= more_issue)
5983 {
5984 /* ??? (... || i == n_ready). */
5985 gcc_assert (i <= n_ready);
5986
5987 /* We should not issue more than issue_rate instructions. */
5988 gcc_assert (top->n <= more_issue);
5989
5990 if (top == choice_stack)
5991 break;
5992
5993 if (best < top - choice_stack)
5994 {
5995 if (privileged_n)
5996 {
5997 n = privileged_n;
5998 /* Try to find issued privileged insn. */
5999 while (n && !ready_try[--n])
6000 ;
6001 }
6002
6003 if (/* If all insns are equally good... */
6004 privileged_n == 0
6005 /* Or a privileged insn will be issued. */
6006 || ready_try[n])
6007 /* Then we have a solution. */
6008 {
6009 best = top - choice_stack;
6010 /* This is the index of the insn issued first in this
6011 solution. */
6012 *index = choice_stack [1].index;
6013 if (top->n == more_issue || best == all)
6014 break;
6015 }
6016 }
6017
6018 /* Set ready-list index to point to the last insn
6019 ('i++' below will advance it to the next insn). */
6020 i = top->index;
6021
6022 /* Backtrack. */
6023 ready_try [i] = 0;
6024
6025 if (targetm.sched.first_cycle_multipass_backtrack)
6026 targetm.sched.first_cycle_multipass_backtrack (&top->target_data,
6027 ready_try, n_ready);
6028
6029 top--;
6030 memcpy (dest: state, src: top->state, n: dfa_state_size);
6031 }
6032 else if (!ready_try [i])
6033 {
6034 tries_num++;
6035 if (tries_num > max_lookahead_tries)
6036 break;
6037 insn = ready_element (ready, index: i);
6038 delay = state_transition (state, insn);
6039 if (delay < 0)
6040 {
6041 if (state_dead_lock_p (state)
6042 || insn_finishes_cycle_p (insn))
6043 /* We won't issue any more instructions in the next
6044 choice_state. */
6045 top->rest = 0;
6046 else
6047 top->rest--;
6048
6049 n = top->n;
6050 if (memcmp (s1: top->state, s2: state, n: dfa_state_size) != 0)
6051 n++;
6052
6053 /* Advance to the next choice_entry. */
6054 top++;
6055 /* Initialize it. */
6056 top->rest = dfa_lookahead;
6057 top->index = i;
6058 top->n = n;
6059 memcpy (dest: top->state, src: state, n: dfa_state_size);
6060 ready_try [i] = 1;
6061
6062 if (targetm.sched.first_cycle_multipass_issue)
6063 targetm.sched.first_cycle_multipass_issue (&top->target_data,
6064 ready_try, n_ready,
6065 insn,
6066 &((top - 1)
6067 ->target_data));
6068
6069 i = -1;
6070 }
6071 }
6072
6073 /* Increase ready-list index. */
6074 i++;
6075 }
6076
6077 if (targetm.sched.first_cycle_multipass_end)
6078 targetm.sched.first_cycle_multipass_end (best != 0
6079 ? &choice_stack[1].target_data
6080 : NULL);
6081
6082 /* Restore the original state of the DFA. */
6083 memcpy (dest: state, src: choice_stack->state, n: dfa_state_size);
6084
6085 return best;
6086}
6087
6088/* The following function chooses insn from READY and modifies
6089 READY. The following function is used only for first
6090 cycle multipass scheduling.
6091 Return:
6092 -1 if cycle should be advanced,
6093 0 if INSN_PTR is set to point to the desirable insn,
6094 1 if choose_ready () should be restarted without advancing the cycle. */
6095static int
6096choose_ready (struct ready_list *ready, bool first_cycle_insn_p,
6097 rtx_insn **insn_ptr)
6098{
6099 if (dbg_cnt (index: sched_insn) == false)
6100 {
6101 if (nonscheduled_insns_begin == NULL_RTX)
6102 nonscheduled_insns_begin = current_sched_info->prev_head;
6103
6104 rtx_insn *insn = first_nonscheduled_insn ();
6105
6106 if (QUEUE_INDEX (insn) == QUEUE_READY)
6107 /* INSN is in the ready_list. */
6108 {
6109 ready_remove_insn (insn);
6110 *insn_ptr = insn;
6111 return 0;
6112 }
6113
6114 /* INSN is in the queue. Advance cycle to move it to the ready list. */
6115 gcc_assert (QUEUE_INDEX (insn) >= 0);
6116 return -1;
6117 }
6118
6119 if (dfa_lookahead <= 0 || SCHED_GROUP_P (ready_element (ready, 0))
6120 || DEBUG_INSN_P (ready_element (ready, 0)))
6121 {
6122 if (targetm.sched.dispatch (NULL, IS_DISPATCH_ON))
6123 *insn_ptr = ready_remove_first_dispatch (ready);
6124 else
6125 *insn_ptr = ready_remove_first (ready);
6126
6127 return 0;
6128 }
6129 else
6130 {
6131 /* Try to choose the best insn. */
6132 int index = 0, i;
6133 rtx_insn *insn;
6134
6135 insn = ready_element (ready, index: 0);
6136 if (INSN_CODE (insn) < 0)
6137 {
6138 *insn_ptr = ready_remove_first (ready);
6139 return 0;
6140 }
6141
6142 /* Filter the search space. */
6143 for (i = 0; i < ready->n_ready; i++)
6144 {
6145 ready_try[i] = 0;
6146
6147 insn = ready_element (ready, index: i);
6148
6149 /* If this insn is recognizable we should have already
6150 recognized it earlier.
6151 ??? Not very clear where this is supposed to be done.
6152 See dep_cost_1. */
6153 gcc_checking_assert (INSN_CODE (insn) >= 0
6154 || recog_memoized (insn) < 0);
6155 if (INSN_CODE (insn) < 0)
6156 {
6157 /* Non-recognized insns at position 0 are handled above. */
6158 gcc_assert (i > 0);
6159 ready_try[i] = 1;
6160 continue;
6161 }
6162
6163 if (targetm.sched.first_cycle_multipass_dfa_lookahead_guard)
6164 {
6165 ready_try[i]
6166 = (targetm.sched.first_cycle_multipass_dfa_lookahead_guard
6167 (insn, i));
6168
6169 if (ready_try[i] < 0)
6170 /* Queue instruction for several cycles.
6171 We need to restart choose_ready as we have changed
6172 the ready list. */
6173 {
6174 change_queue_index (insn, -ready_try[i]);
6175 return 1;
6176 }
6177
6178 /* Make sure that we didn't end up with 0'th insn filtered out.
6179 Don't be tempted to make life easier for backends and just
6180 requeue 0'th insn if (ready_try[0] == 0) and restart
6181 choose_ready. Backends should be very considerate about
6182 requeueing instructions -- especially the highest priority
6183 one at position 0. */
6184 gcc_assert (ready_try[i] == 0 || i > 0);
6185 if (ready_try[i])
6186 continue;
6187 }
6188
6189 gcc_assert (ready_try[i] == 0);
6190 /* INSN made it through the scrutiny of filters! */
6191 }
6192
6193 if (max_issue (ready, privileged_n: 1, state: curr_state, first_cycle_insn_p, index: &index) == 0)
6194 {
6195 *insn_ptr = ready_remove_first (ready);
6196 if (sched_verbose >= 4)
6197 fprintf (stream: sched_dump, format: ";;\t\tChosen insn (but can't issue) : %s \n",
6198 (*current_sched_info->print_insn) (*insn_ptr, 0));
6199 return 0;
6200 }
6201 else
6202 {
6203 if (sched_verbose >= 4)
6204 fprintf (stream: sched_dump, format: ";;\t\tChosen insn : %s\n",
6205 (*current_sched_info->print_insn)
6206 (ready_element (ready, index), 0));
6207
6208 *insn_ptr = ready_remove (ready, index);
6209 return 0;
6210 }
6211 }
6212}
6213
6214/* This function is called when we have successfully scheduled a
6215 block. It uses the schedule stored in the scheduled_insns vector
6216 to rearrange the RTL. PREV_HEAD is used as the anchor to which we
6217 append the scheduled insns; TAIL is the insn after the scheduled
6218 block. TARGET_BB is the argument passed to schedule_block. */
6219
6220static void
6221commit_schedule (rtx_insn *prev_head, rtx_insn *tail, basic_block *target_bb)
6222{
6223 unsigned int i;
6224 rtx_insn *insn;
6225
6226 last_scheduled_insn = prev_head;
6227 for (i = 0;
6228 scheduled_insns.iterate (ix: i, ptr: &insn);
6229 i++)
6230 {
6231 if (control_flow_insn_p (last_scheduled_insn)
6232 || current_sched_info->advance_target_bb (*target_bb, insn))
6233 {
6234 *target_bb = current_sched_info->advance_target_bb (*target_bb, 0);
6235
6236 if (sched_verbose)
6237 {
6238 rtx_insn *x;
6239
6240 x = next_real_insn (last_scheduled_insn);
6241 gcc_assert (x);
6242 dump_new_block_header (1, *target_bb, x, tail);
6243 }
6244
6245 last_scheduled_insn = bb_note (*target_bb);
6246 }
6247
6248 if (current_sched_info->begin_move_insn)
6249 (*current_sched_info->begin_move_insn) (insn, last_scheduled_insn);
6250 move_insn (insn, last: last_scheduled_insn,
6251 nt: current_sched_info->next_tail);
6252 if (!DEBUG_INSN_P (insn))
6253 reemit_notes (insn);
6254 last_scheduled_insn = insn;
6255 }
6256
6257 scheduled_insns.truncate (size: 0);
6258}
6259
6260/* Examine all insns on the ready list and queue those which can't be
6261 issued in this cycle. TEMP_STATE is temporary scheduler state we
6262 can use as scratch space. If FIRST_CYCLE_INSN_P is true, no insns
6263 have been issued for the current cycle, which means it is valid to
6264 issue an asm statement.
6265
6266 If SHADOWS_ONLY_P is true, we eliminate all real insns and only
6267 leave those for which SHADOW_P is true. If MODULO_EPILOGUE is true,
6268 we only leave insns which have an INSN_EXACT_TICK. */
6269
6270static void
6271prune_ready_list (state_t temp_state, bool first_cycle_insn_p,
6272 bool shadows_only_p, bool modulo_epilogue_p)
6273{
6274 int i, pass;
6275 bool sched_group_found = false;
6276 int min_cost_group = 0;
6277
6278 if (sched_fusion)
6279 return;
6280
6281 for (i = 0; i < ready.n_ready; i++)
6282 {
6283 rtx_insn *insn = ready_element (ready: &ready, index: i);
6284 if (SCHED_GROUP_P (insn))
6285 {
6286 sched_group_found = true;
6287 break;
6288 }
6289 }
6290
6291 /* Make two passes if there's a SCHED_GROUP_P insn; make sure to handle
6292 such an insn first and note its cost. If at least one SCHED_GROUP_P insn
6293 gets queued, then all other insns get queued for one cycle later. */
6294 for (pass = sched_group_found ? 0 : 1; pass < 2; )
6295 {
6296 int n = ready.n_ready;
6297 for (i = 0; i < n; i++)
6298 {
6299 rtx_insn *insn = ready_element (ready: &ready, index: i);
6300 int cost = 0;
6301 const char *reason = "resource conflict";
6302
6303 if (DEBUG_INSN_P (insn))
6304 continue;
6305
6306 if (sched_group_found && !SCHED_GROUP_P (insn)
6307 && ((pass == 0) || (min_cost_group >= 1)))
6308 {
6309 if (pass == 0)
6310 continue;
6311 cost = min_cost_group;
6312 reason = "not in sched group";
6313 }
6314 else if (modulo_epilogue_p
6315 && INSN_EXACT_TICK (insn) == INVALID_TICK)
6316 {
6317 cost = max_insn_queue_index;
6318 reason = "not an epilogue insn";
6319 }
6320 else if (shadows_only_p && !SHADOW_P (insn))
6321 {
6322 cost = 1;
6323 reason = "not a shadow";
6324 }
6325 else if (recog_memoized (insn) < 0)
6326 {
6327 if (!first_cycle_insn_p
6328 && (GET_CODE (PATTERN (insn)) == ASM_INPUT
6329 || asm_noperands (PATTERN (insn)) >= 0))
6330 cost = 1;
6331 reason = "asm";
6332 }
6333 else if (sched_pressure != SCHED_PRESSURE_NONE)
6334 {
6335 if (sched_pressure == SCHED_PRESSURE_MODEL
6336 && INSN_TICK (insn) <= clock_var)
6337 {
6338 memcpy (dest: temp_state, src: curr_state, n: dfa_state_size);
6339 if (state_transition (temp_state, insn) >= 0)
6340 INSN_TICK (insn) = clock_var + 1;
6341 }
6342 cost = 0;
6343 }
6344 else
6345 {
6346 int delay_cost = 0;
6347
6348 if (delay_htab)
6349 {
6350 struct delay_pair *delay_entry;
6351 delay_entry
6352 = delay_htab->find_with_hash (comparable: insn,
6353 hash: htab_hash_pointer (insn));
6354 while (delay_entry && delay_cost == 0)
6355 {
6356 delay_cost = estimate_shadow_tick (p: delay_entry);
6357 if (delay_cost > max_insn_queue_index)
6358 delay_cost = max_insn_queue_index;
6359 delay_entry = delay_entry->next_same_i1;
6360 }
6361 }
6362
6363 memcpy (dest: temp_state, src: curr_state, n: dfa_state_size);
6364 cost = state_transition (temp_state, insn);
6365 if (cost < 0)
6366 cost = 0;
6367 else if (cost == 0)
6368 cost = 1;
6369 if (cost < delay_cost)
6370 {
6371 cost = delay_cost;
6372 reason = "shadow tick";
6373 }
6374 }
6375 if (cost >= 1)
6376 {
6377 if (SCHED_GROUP_P (insn) && cost > min_cost_group)
6378 min_cost_group = cost;
6379 ready_remove (ready: &ready, index: i);
6380 /* Normally we'd want to queue INSN for COST cycles. However,
6381 if SCHED_GROUP_P is set, then we must ensure that nothing
6382 else comes between INSN and its predecessor. If there is
6383 some other insn ready to fire on the next cycle, then that
6384 invariant would be broken.
6385
6386 So when SCHED_GROUP_P is set, just queue this insn for a
6387 single cycle. */
6388 queue_insn (insn, SCHED_GROUP_P (insn) ? 1 : cost, reason);
6389 if (i + 1 < n)
6390 break;
6391 }
6392 }
6393 if (i == n)
6394 pass++;
6395 }
6396}
6397
6398/* Called when we detect that the schedule is impossible. We examine the
6399 backtrack queue to find the earliest insn that caused this condition. */
6400
6401static struct haifa_saved_data *
6402verify_shadows (void)
6403{
6404 struct haifa_saved_data *save, *earliest_fail = NULL;
6405 for (save = backtrack_queue; save; save = save->next)
6406 {
6407 int t;
6408 struct delay_pair *pair = save->delay_pair;
6409 rtx_insn *i1 = pair->i1;
6410
6411 for (; pair; pair = pair->next_same_i1)
6412 {
6413 rtx_insn *i2 = pair->i2;
6414
6415 if (QUEUE_INDEX (i2) == QUEUE_SCHEDULED)
6416 continue;
6417
6418 t = INSN_TICK (i1) + pair_delay (p: pair);
6419 if (t < clock_var)
6420 {
6421 if (sched_verbose >= 2)
6422 fprintf (stream: sched_dump,
6423 format: ";;\t\tfailed delay requirements for %d/%d (%d->%d)"
6424 ", not ready\n",
6425 INSN_UID (insn: pair->i1), INSN_UID (insn: pair->i2),
6426 INSN_TICK (pair->i1), INSN_EXACT_TICK (pair->i2));
6427 earliest_fail = save;
6428 break;
6429 }
6430 if (QUEUE_INDEX (i2) >= 0)
6431 {
6432 int queued_for = INSN_TICK (i2);
6433
6434 if (t < queued_for)
6435 {
6436 if (sched_verbose >= 2)
6437 fprintf (stream: sched_dump,
6438 format: ";;\t\tfailed delay requirements for %d/%d"
6439 " (%d->%d), queued too late\n",
6440 INSN_UID (insn: pair->i1), INSN_UID (insn: pair->i2),
6441 INSN_TICK (pair->i1), INSN_EXACT_TICK (pair->i2));
6442 earliest_fail = save;
6443 break;
6444 }
6445 }
6446 }
6447 }
6448
6449 return earliest_fail;
6450}
6451
6452/* Print instructions together with useful scheduling information between
6453 HEAD and TAIL (inclusive). */
6454static void
6455dump_insn_stream (rtx_insn *head, rtx_insn *tail)
6456{
6457 fprintf (stream: sched_dump, format: ";;\t| insn | prio |\n");
6458
6459 rtx_insn *next_tail = NEXT_INSN (insn: tail);
6460 for (rtx_insn *insn = head; insn != next_tail; insn = NEXT_INSN (insn))
6461 {
6462 int priority = NOTE_P (insn) ? 0 : INSN_PRIORITY (insn);
6463 const char *pattern = (NOTE_P (insn)
6464 ? "note"
6465 : str_pattern_slim (PATTERN (insn)));
6466
6467 fprintf (stream: sched_dump, format: ";;\t| %4d | %4d | %-30s ",
6468 INSN_UID (insn), priority, pattern);
6469
6470 if (sched_verbose >= 4)
6471 {
6472 if (NOTE_P (insn) || LABEL_P (insn) || recog_memoized (insn) < 0)
6473 fprintf (stream: sched_dump, format: "nothing");
6474 else
6475 print_reservation (sched_dump, insn);
6476 }
6477 fprintf (stream: sched_dump, format: "\n");
6478 }
6479}
6480
6481/* Use forward list scheduling to rearrange insns of block pointed to by
6482 TARGET_BB, possibly bringing insns from subsequent blocks in the same
6483 region. */
6484
6485bool
6486schedule_block (basic_block *target_bb, state_t init_state)
6487{
6488 int i;
6489 bool success = modulo_ii == 0;
6490 struct sched_block_state ls;
6491 state_t temp_state = NULL; /* It is used for multipass scheduling. */
6492 int sort_p, advance, start_clock_var;
6493
6494 /* Head/tail info for this block. */
6495 rtx_insn *prev_head = current_sched_info->prev_head;
6496 rtx_insn *next_tail = current_sched_info->next_tail;
6497 rtx_insn *head = NEXT_INSN (insn: prev_head);
6498 rtx_insn *tail = PREV_INSN (insn: next_tail);
6499
6500 if ((current_sched_info->flags & DONT_BREAK_DEPENDENCIES) == 0
6501 && sched_pressure != SCHED_PRESSURE_MODEL && !sched_fusion)
6502 find_modifiable_mems (head, tail);
6503
6504 /* We used to have code to avoid getting parameters moved from hard
6505 argument registers into pseudos.
6506
6507 However, it was removed when it proved to be of marginal benefit
6508 and caused problems because schedule_block and compute_forward_dependences
6509 had different notions of what the "head" insn was. */
6510
6511 gcc_assert (head != tail || INSN_P (head));
6512
6513 haifa_recovery_bb_recently_added_p = false;
6514
6515 backtrack_queue = NULL;
6516
6517 /* Debug info. */
6518 if (sched_verbose)
6519 {
6520 dump_new_block_header (0, *target_bb, head, tail);
6521
6522 if (sched_verbose >= 2)
6523 {
6524 dump_insn_stream (head, tail);
6525 memset (s: &rank_for_schedule_stats, c: 0,
6526 n: sizeof (rank_for_schedule_stats));
6527 }
6528 }
6529
6530 if (init_state == NULL)
6531 state_reset (curr_state);
6532 else
6533 memcpy (dest: curr_state, src: init_state, n: dfa_state_size);
6534
6535 /* Clear the ready list. */
6536 ready.first = ready.veclen - 1;
6537 ready.n_ready = 0;
6538 ready.n_debug = 0;
6539
6540 /* It is used for first cycle multipass scheduling. */
6541 temp_state = alloca (dfa_state_size);
6542
6543 if (targetm.sched.init)
6544 targetm.sched.init (sched_dump, sched_verbose, ready.veclen);
6545
6546 /* We start inserting insns after PREV_HEAD. */
6547 last_scheduled_insn = prev_head;
6548 last_nondebug_scheduled_insn = NULL;
6549 nonscheduled_insns_begin = NULL;
6550
6551 gcc_assert ((NOTE_P (last_scheduled_insn)
6552 || DEBUG_INSN_P (last_scheduled_insn))
6553 && BLOCK_FOR_INSN (last_scheduled_insn) == *target_bb);
6554
6555 /* Initialize INSN_QUEUE. Q_SIZE is the total number of insns in the
6556 queue. */
6557 q_ptr = 0;
6558 q_size = 0;
6559
6560 insn_queue = XALLOCAVEC (rtx_insn_list *, max_insn_queue_index + 1);
6561 memset (s: insn_queue, c: 0, n: (max_insn_queue_index + 1) * sizeof (rtx));
6562
6563 /* Start just before the beginning of time. */
6564 clock_var = -1;
6565
6566 /* We need queue and ready lists and clock_var be initialized
6567 in try_ready () (which is called through init_ready_list ()). */
6568 (*current_sched_info->init_ready_list) ();
6569
6570 if (sched_pressure)
6571 sched_pressure_start_bb (bb: *target_bb);
6572
6573 /* The algorithm is O(n^2) in the number of ready insns at any given
6574 time in the worst case. Before reload we are more likely to have
6575 big lists so truncate them to a reasonable size. */
6576 if (!reload_completed
6577 && ready.n_ready - ready.n_debug > param_max_sched_ready_insns)
6578 {
6579 ready_sort_debug (ready: &ready);
6580 ready_sort_real (ready: &ready);
6581
6582 /* Find first free-standing insn past param_max_sched_ready_insns.
6583 If there are debug insns, we know they're first. */
6584 for (i = param_max_sched_ready_insns + ready.n_debug; i < ready.n_ready;
6585 i++)
6586 if (!SCHED_GROUP_P (ready_element (&ready, i)))
6587 break;
6588
6589 if (sched_verbose >= 2)
6590 {
6591 fprintf (stream: sched_dump,
6592 format: ";;\t\tReady list on entry: %d insns: ", ready.n_ready);
6593 debug_ready_list (ready: &ready);
6594 fprintf (stream: sched_dump,
6595 format: ";;\t\t before reload => truncated to %d insns\n", i);
6596 }
6597
6598 /* Delay all insns past it for 1 cycle. If debug counter is
6599 activated make an exception for the insn right after
6600 nonscheduled_insns_begin. */
6601 {
6602 rtx_insn *skip_insn;
6603
6604 if (dbg_cnt (index: sched_insn) == false)
6605 skip_insn = first_nonscheduled_insn ();
6606 else
6607 skip_insn = NULL;
6608
6609 while (i < ready.n_ready)
6610 {
6611 rtx_insn *insn;
6612
6613 insn = ready_remove (ready: &ready, index: i);
6614
6615 if (insn != skip_insn)
6616 queue_insn (insn, n_cycles: 1, reason: "list truncated");
6617 }
6618 if (skip_insn)
6619 ready_add (ready: &ready, insn: skip_insn, first_p: true);
6620 }
6621 }
6622
6623 /* Now we can restore basic block notes and maintain precise cfg. */
6624 restore_bb_notes (*target_bb);
6625
6626 last_clock_var = -1;
6627
6628 advance = 0;
6629
6630 gcc_assert (scheduled_insns.length () == 0);
6631 sort_p = true;
6632 must_backtrack = false;
6633 modulo_insns_scheduled = 0;
6634
6635 ls.modulo_epilogue = false;
6636 ls.first_cycle_insn_p = true;
6637
6638 /* Loop until all the insns in BB are scheduled. */
6639 while ((*current_sched_info->schedule_more_p) ())
6640 {
6641 perform_replacements_new_cycle ();
6642 do
6643 {
6644 start_clock_var = clock_var;
6645
6646 clock_var++;
6647
6648 advance_one_cycle ();
6649
6650 /* Add to the ready list all pending insns that can be issued now.
6651 If there are no ready insns, increment clock until one
6652 is ready and add all pending insns at that point to the ready
6653 list. */
6654 queue_to_ready (ready: &ready);
6655
6656 gcc_assert (ready.n_ready);
6657
6658 if (sched_verbose >= 2)
6659 {
6660 fprintf (stream: sched_dump, format: ";;\t\tReady list after queue_to_ready:");
6661 debug_ready_list (ready: &ready);
6662 }
6663 advance -= clock_var - start_clock_var;
6664 }
6665 while (advance > 0);
6666
6667 if (ls.modulo_epilogue)
6668 {
6669 int stage = clock_var / modulo_ii;
6670 if (stage > modulo_last_stage * 2 + 2)
6671 {
6672 if (sched_verbose >= 2)
6673 fprintf (stream: sched_dump,
6674 format: ";;\t\tmodulo scheduled succeeded at II %d\n",
6675 modulo_ii);
6676 success = true;
6677 goto end_schedule;
6678 }
6679 }
6680 else if (modulo_ii > 0)
6681 {
6682 int stage = clock_var / modulo_ii;
6683 if (stage > modulo_max_stages)
6684 {
6685 if (sched_verbose >= 2)
6686 fprintf (stream: sched_dump,
6687 format: ";;\t\tfailing schedule due to excessive stages\n");
6688 goto end_schedule;
6689 }
6690 if (modulo_n_insns == modulo_insns_scheduled
6691 && stage > modulo_last_stage)
6692 {
6693 if (sched_verbose >= 2)
6694 fprintf (stream: sched_dump,
6695 format: ";;\t\tfound kernel after %d stages, II %d\n",
6696 stage, modulo_ii);
6697 ls.modulo_epilogue = true;
6698 }
6699 }
6700
6701 prune_ready_list (temp_state, first_cycle_insn_p: true, shadows_only_p: false, modulo_epilogue_p: ls.modulo_epilogue);
6702 if (ready.n_ready == 0)
6703 continue;
6704 if (must_backtrack)
6705 goto do_backtrack;
6706
6707 ls.shadows_only_p = false;
6708 cycle_issued_insns = 0;
6709 ls.can_issue_more = issue_rate;
6710 for (;;)
6711 {
6712 rtx_insn *insn;
6713 int cost;
6714 bool asm_p;
6715
6716 if (sort_p && ready.n_ready > 0)
6717 {
6718 /* Sort the ready list based on priority. This must be
6719 done every iteration through the loop, as schedule_insn
6720 may have readied additional insns that will not be
6721 sorted correctly. */
6722 ready_sort (ready: &ready);
6723
6724 if (sched_verbose >= 2)
6725 {
6726 fprintf (stream: sched_dump,
6727 format: ";;\t\tReady list after ready_sort: ");
6728 debug_ready_list (ready: &ready);
6729 }
6730 }
6731
6732 /* We don't want md sched reorder to even see debug isns, so put
6733 them out right away. */
6734 if (ready.n_ready && DEBUG_INSN_P (ready_element (&ready, 0))
6735 && (*current_sched_info->schedule_more_p) ())
6736 {
6737 while (ready.n_ready && DEBUG_INSN_P (ready_element (&ready, 0)))
6738 {
6739 rtx_insn *insn = ready_remove_first (ready: &ready);
6740 gcc_assert (DEBUG_INSN_P (insn));
6741 (*current_sched_info->begin_schedule_ready) (insn);
6742 scheduled_insns.safe_push (obj: insn);
6743 last_scheduled_insn = insn;
6744 advance = schedule_insn (insn);
6745 gcc_assert (advance == 0);
6746 if (ready.n_ready > 0)
6747 ready_sort (ready: &ready);
6748 }
6749 }
6750
6751 if (ls.first_cycle_insn_p && !ready.n_ready)
6752 break;
6753
6754 resume_after_backtrack:
6755 /* Allow the target to reorder the list, typically for
6756 better instruction bundling. */
6757 if (sort_p
6758 && (ready.n_ready == 0
6759 || !SCHED_GROUP_P (ready_element (&ready, 0))))
6760 {
6761 if (ls.first_cycle_insn_p && targetm.sched.reorder)
6762 ls.can_issue_more
6763 = targetm.sched.reorder (sched_dump, sched_verbose,
6764 ready_lastpos (ready: &ready),
6765 &ready.n_ready, clock_var);
6766 else if (!ls.first_cycle_insn_p && targetm.sched.reorder2)
6767 ls.can_issue_more
6768 = targetm.sched.reorder2 (sched_dump, sched_verbose,
6769 ready.n_ready
6770 ? ready_lastpos (ready: &ready) : NULL,
6771 &ready.n_ready, clock_var);
6772 }
6773
6774 restart_choose_ready:
6775 if (sched_verbose >= 2)
6776 {
6777 fprintf (stream: sched_dump, format: ";;\tReady list (t = %3d): ",
6778 clock_var);
6779 debug_ready_list (ready: &ready);
6780 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
6781 print_curr_reg_pressure ();
6782 }
6783
6784 if (ready.n_ready == 0
6785 && ls.can_issue_more
6786 && reload_completed)
6787 {
6788 /* Allow scheduling insns directly from the queue in case
6789 there's nothing better to do (ready list is empty) but
6790 there are still vacant dispatch slots in the current cycle. */
6791 if (sched_verbose >= 6)
6792 fprintf (stream: sched_dump,format: ";;\t\tSecond chance\n");
6793 memcpy (dest: temp_state, src: curr_state, n: dfa_state_size);
6794 if (early_queue_to_ready (state: temp_state, ready: &ready))
6795 ready_sort (ready: &ready);
6796 }
6797
6798 if (ready.n_ready == 0
6799 || !ls.can_issue_more
6800 || state_dead_lock_p (curr_state)
6801 || !(*current_sched_info->schedule_more_p) ())
6802 break;
6803
6804 /* Select and remove the insn from the ready list. */
6805 if (sort_p)
6806 {
6807 int res;
6808
6809 insn = NULL;
6810 res = choose_ready (ready: &ready, first_cycle_insn_p: ls.first_cycle_insn_p, insn_ptr: &insn);
6811
6812 if (res < 0)
6813 /* Finish cycle. */
6814 break;
6815 if (res > 0)
6816 goto restart_choose_ready;
6817
6818 gcc_assert (insn != NULL_RTX);
6819 }
6820 else
6821 insn = ready_remove_first (ready: &ready);
6822
6823 if (sched_pressure != SCHED_PRESSURE_NONE
6824 && INSN_TICK (insn) > clock_var)
6825 {
6826 ready_add (ready: &ready, insn, first_p: true);
6827 advance = 1;
6828 break;
6829 }
6830
6831 if (targetm.sched.dfa_new_cycle
6832 && targetm.sched.dfa_new_cycle (sched_dump, sched_verbose,
6833 insn, last_clock_var,
6834 clock_var, &sort_p))
6835 /* SORT_P is used by the target to override sorting
6836 of the ready list. This is needed when the target
6837 has modified its internal structures expecting that
6838 the insn will be issued next. As we need the insn
6839 to have the highest priority (so it will be returned by
6840 the ready_remove_first call above), we invoke
6841 ready_add (&ready, insn, true).
6842 But, still, there is one issue: INSN can be later
6843 discarded by scheduler's front end through
6844 current_sched_info->can_schedule_ready_p, hence, won't
6845 be issued next. */
6846 {
6847 ready_add (ready: &ready, insn, first_p: true);
6848 break;
6849 }
6850
6851 sort_p = true;
6852
6853 if (current_sched_info->can_schedule_ready_p
6854 && ! (*current_sched_info->can_schedule_ready_p) (insn))
6855 /* We normally get here only if we don't want to move
6856 insn from the split block. */
6857 {
6858 TODO_SPEC (insn) = DEP_POSTPONED;
6859 goto restart_choose_ready;
6860 }
6861
6862 if (delay_htab)
6863 {
6864 /* If this insn is the first part of a delay-slot pair, record a
6865 backtrack point. */
6866 struct delay_pair *delay_entry;
6867 delay_entry
6868 = delay_htab->find_with_hash (comparable: insn, hash: htab_hash_pointer (insn));
6869 if (delay_entry)
6870 {
6871 save_backtrack_point (pair: delay_entry, sched_block: ls);
6872 if (sched_verbose >= 2)
6873 fprintf (stream: sched_dump, format: ";;\t\tsaving backtrack point\n");
6874 }
6875 }
6876
6877 /* DECISION is made. */
6878
6879 if (modulo_ii > 0 && INSN_UID (insn) < modulo_iter0_max_uid)
6880 {
6881 modulo_insns_scheduled++;
6882 modulo_last_stage = clock_var / modulo_ii;
6883 }
6884 if (TODO_SPEC (insn) & SPECULATIVE)
6885 generate_recovery_code (insn);
6886
6887 if (targetm.sched.dispatch (NULL, IS_DISPATCH_ON))
6888 targetm.sched.dispatch_do (insn, ADD_TO_DISPATCH_WINDOW);
6889
6890 /* Update counters, etc in the scheduler's front end. */
6891 (*current_sched_info->begin_schedule_ready) (insn);
6892 scheduled_insns.safe_push (obj: insn);
6893 gcc_assert (NONDEBUG_INSN_P (insn));
6894 last_nondebug_scheduled_insn = last_scheduled_insn = insn;
6895
6896 if (recog_memoized (insn) >= 0)
6897 {
6898 memcpy (dest: temp_state, src: curr_state, n: dfa_state_size);
6899 cost = state_transition (curr_state, insn);
6900 if (sched_pressure != SCHED_PRESSURE_WEIGHTED && !sched_fusion)
6901 gcc_assert (cost < 0);
6902 if (memcmp (s1: temp_state, s2: curr_state, n: dfa_state_size) != 0)
6903 cycle_issued_insns++;
6904 asm_p = false;
6905 }
6906 else
6907 asm_p = (GET_CODE (PATTERN (insn)) == ASM_INPUT
6908 || asm_noperands (PATTERN (insn)) >= 0);
6909
6910 if (targetm.sched.variable_issue)
6911 ls.can_issue_more =
6912 targetm.sched.variable_issue (sched_dump, sched_verbose,
6913 insn, ls.can_issue_more);
6914 /* A naked CLOBBER or USE generates no instruction, so do
6915 not count them against the issue rate. */
6916 else if (GET_CODE (PATTERN (insn)) != USE
6917 && GET_CODE (PATTERN (insn)) != CLOBBER)
6918 ls.can_issue_more--;
6919 advance = schedule_insn (insn);
6920
6921 if (SHADOW_P (insn))
6922 ls.shadows_only_p = true;
6923
6924 /* After issuing an asm insn we should start a new cycle. */
6925 if (advance == 0 && asm_p)
6926 advance = 1;
6927
6928 if (must_backtrack)
6929 break;
6930
6931 if (advance != 0)
6932 break;
6933
6934 ls.first_cycle_insn_p = false;
6935 if (ready.n_ready > 0)
6936 prune_ready_list (temp_state, first_cycle_insn_p: false, shadows_only_p: ls.shadows_only_p,
6937 modulo_epilogue_p: ls.modulo_epilogue);
6938 }
6939
6940 do_backtrack:
6941 if (!must_backtrack)
6942 for (i = 0; i < ready.n_ready; i++)
6943 {
6944 rtx_insn *insn = ready_element (ready: &ready, index: i);
6945 if (INSN_EXACT_TICK (insn) == clock_var)
6946 {
6947 must_backtrack = true;
6948 clock_var++;
6949 break;
6950 }
6951 }
6952 if (must_backtrack && modulo_ii > 0)
6953 {
6954 if (modulo_backtracks_left == 0)
6955 goto end_schedule;
6956 modulo_backtracks_left--;
6957 }
6958 while (must_backtrack)
6959 {
6960 struct haifa_saved_data *failed;
6961 rtx_insn *failed_insn;
6962
6963 must_backtrack = false;
6964 failed = verify_shadows ();
6965 gcc_assert (failed);
6966
6967 failed_insn = failed->delay_pair->i1;
6968 /* Clear these queues. */
6969 perform_replacements_new_cycle ();
6970 toggle_cancelled_flags (set: false);
6971 unschedule_insns_until (insn: failed_insn);
6972 while (failed != backtrack_queue)
6973 free_topmost_backtrack_point (reset_tick: true);
6974 restore_last_backtrack_point (psched_block: &ls);
6975 if (sched_verbose >= 2)
6976 fprintf (stream: sched_dump, format: ";;\t\trewind to cycle %d\n", clock_var);
6977 /* Delay by at least a cycle. This could cause additional
6978 backtracking. */
6979 queue_insn (insn: failed_insn, n_cycles: 1, reason: "backtracked");
6980 advance = 0;
6981 if (must_backtrack)
6982 continue;
6983 if (ready.n_ready > 0)
6984 goto resume_after_backtrack;
6985 else
6986 {
6987 if (clock_var == 0 && ls.first_cycle_insn_p)
6988 goto end_schedule;
6989 advance = 1;
6990 break;
6991 }
6992 }
6993 ls.first_cycle_insn_p = true;
6994 }
6995 if (ls.modulo_epilogue)
6996 success = true;
6997 end_schedule:
6998 if (!ls.first_cycle_insn_p || advance)
6999 advance_one_cycle ();
7000 perform_replacements_new_cycle ();
7001 if (modulo_ii > 0)
7002 {
7003 /* Once again, debug insn suckiness: they can be on the ready list
7004 even if they have unresolved dependencies. To make our view
7005 of the world consistent, remove such "ready" insns. */
7006 restart_debug_insn_loop:
7007 for (i = ready.n_ready - 1; i >= 0; i--)
7008 {
7009 rtx_insn *x;
7010
7011 x = ready_element (ready: &ready, index: i);
7012 if (DEPS_LIST_FIRST (INSN_HARD_BACK_DEPS (x)) != NULL
7013 || DEPS_LIST_FIRST (INSN_SPEC_BACK_DEPS (x)) != NULL)
7014 {
7015 ready_remove (ready: &ready, index: i);
7016 goto restart_debug_insn_loop;
7017 }
7018 }
7019 for (i = ready.n_ready - 1; i >= 0; i--)
7020 {
7021 rtx_insn *x;
7022
7023 x = ready_element (ready: &ready, index: i);
7024 resolve_dependencies (insn: x);
7025 }
7026 for (i = 0; i <= max_insn_queue_index; i++)
7027 {
7028 rtx_insn_list *link;
7029 while ((link = insn_queue[i]) != NULL)
7030 {
7031 rtx_insn *x = link->insn ();
7032 insn_queue[i] = link->next ();
7033 QUEUE_INDEX (x) = QUEUE_NOWHERE;
7034 free_INSN_LIST_node (link);
7035 resolve_dependencies (insn: x);
7036 }
7037 }
7038 }
7039
7040 if (!success)
7041 undo_all_replacements ();
7042
7043 /* Debug info. */
7044 if (sched_verbose)
7045 {
7046 fprintf (stream: sched_dump, format: ";;\tReady list (final): ");
7047 debug_ready_list (ready: &ready);
7048 }
7049
7050 if (modulo_ii == 0 && current_sched_info->queue_must_finish_empty)
7051 /* Sanity check -- queue must be empty now. Meaningless if region has
7052 multiple bbs. */
7053 gcc_assert (!q_size && !ready.n_ready && !ready.n_debug);
7054 else if (modulo_ii == 0)
7055 {
7056 /* We must maintain QUEUE_INDEX between blocks in region. */
7057 for (i = ready.n_ready - 1; i >= 0; i--)
7058 {
7059 rtx_insn *x;
7060
7061 x = ready_element (ready: &ready, index: i);
7062 QUEUE_INDEX (x) = QUEUE_NOWHERE;
7063 TODO_SPEC (x) = HARD_DEP;
7064 }
7065
7066 if (q_size)
7067 for (i = 0; i <= max_insn_queue_index; i++)
7068 {
7069 rtx_insn_list *link;
7070 for (link = insn_queue[i]; link; link = link->next ())
7071 {
7072 rtx_insn *x;
7073
7074 x = link->insn ();
7075 QUEUE_INDEX (x) = QUEUE_NOWHERE;
7076 TODO_SPEC (x) = HARD_DEP;
7077 }
7078 free_INSN_LIST_list (&insn_queue[i]);
7079 }
7080 }
7081
7082 if (sched_pressure == SCHED_PRESSURE_MODEL)
7083 model_end_schedule ();
7084
7085 if (success)
7086 {
7087 commit_schedule (prev_head, tail, target_bb);
7088 if (sched_verbose)
7089 fprintf (stream: sched_dump, format: ";; total time = %d\n", clock_var);
7090 }
7091 else
7092 last_scheduled_insn = tail;
7093
7094 scheduled_insns.truncate (size: 0);
7095
7096 if (!current_sched_info->queue_must_finish_empty
7097 || haifa_recovery_bb_recently_added_p)
7098 {
7099 /* INSN_TICK (minimum clock tick at which the insn becomes
7100 ready) may be not correct for the insn in the subsequent
7101 blocks of the region. We should use a correct value of
7102 `clock_var' or modify INSN_TICK. It is better to keep
7103 clock_var value equal to 0 at the start of a basic block.
7104 Therefore we modify INSN_TICK here. */
7105 fix_inter_tick (NEXT_INSN (insn: prev_head), last_scheduled_insn);
7106 }
7107
7108 if (targetm.sched.finish)
7109 {
7110 targetm.sched.finish (sched_dump, sched_verbose);
7111 /* Target might have added some instructions to the scheduled block
7112 in its md_finish () hook. These new insns don't have any data
7113 initialized and to identify them we extend h_i_d so that they'll
7114 get zero luids. */
7115 sched_extend_luids ();
7116 }
7117
7118 /* Update head/tail boundaries. */
7119 head = NEXT_INSN (insn: prev_head);
7120 tail = last_scheduled_insn;
7121
7122 if (sched_verbose)
7123 {
7124 fprintf (stream: sched_dump, format: ";; new head = %d\n;; new tail = %d\n",
7125 INSN_UID (insn: head), INSN_UID (insn: tail));
7126
7127 if (sched_verbose >= 2)
7128 {
7129 dump_insn_stream (head, tail);
7130 print_rank_for_schedule_stats (prefix: ";; TOTAL ", stats: &rank_for_schedule_stats,
7131 NULL);
7132 }
7133
7134 fprintf (stream: sched_dump, format: "\n");
7135 }
7136
7137 head = restore_other_notes (head, NULL);
7138
7139 current_sched_info->head = head;
7140 current_sched_info->tail = tail;
7141
7142 free_backtrack_queue ();
7143
7144 return success;
7145}
7146
7147/* Set_priorities: compute priority of each insn in the block. */
7148
7149int
7150set_priorities (rtx_insn *head, rtx_insn *tail)
7151{
7152 rtx_insn *insn;
7153 int n_insn;
7154 int sched_max_insns_priority =
7155 current_sched_info->sched_max_insns_priority;
7156 rtx_insn *prev_head;
7157
7158 if (head == tail && ! INSN_P (head))
7159 gcc_unreachable ();
7160
7161 n_insn = 0;
7162
7163 prev_head = PREV_INSN (insn: head);
7164 for (insn = tail; insn != prev_head; insn = PREV_INSN (insn))
7165 {
7166 if (!INSN_P (insn))
7167 continue;
7168
7169 n_insn++;
7170 (void) priority (insn);
7171
7172 gcc_assert (INSN_PRIORITY_KNOWN (insn));
7173
7174 sched_max_insns_priority = MAX (sched_max_insns_priority,
7175 INSN_PRIORITY (insn));
7176 }
7177
7178 current_sched_info->sched_max_insns_priority = sched_max_insns_priority;
7179
7180 return n_insn;
7181}
7182
7183/* Set sched_dump and sched_verbose for the desired debugging output. */
7184void
7185setup_sched_dump (void)
7186{
7187 sched_verbose = sched_verbose_param;
7188 sched_dump = dump_file;
7189 if (!dump_file)
7190 sched_verbose = 0;
7191}
7192
7193/* Allocate data for register pressure sensitive scheduling. */
7194static void
7195alloc_global_sched_pressure_data (void)
7196{
7197 if (sched_pressure != SCHED_PRESSURE_NONE)
7198 {
7199 int i, max_regno = max_reg_num ();
7200
7201 if (sched_dump != NULL)
7202 /* We need info about pseudos for rtl dumps about pseudo
7203 classes and costs. */
7204 regstat_init_n_sets_and_refs ();
7205 ira_set_pseudo_classes (true, sched_verbose ? sched_dump : NULL);
7206 sched_regno_pressure_class
7207 = (enum reg_class *) xmalloc (max_regno * sizeof (enum reg_class));
7208 for (i = 0; i < max_regno; i++)
7209 sched_regno_pressure_class[i]
7210 = (i < FIRST_PSEUDO_REGISTER
7211 ? ira_pressure_class_translate[REGNO_REG_CLASS (i)]
7212 : ira_pressure_class_translate[reg_allocno_class (i)]);
7213 curr_reg_live = BITMAP_ALLOC (NULL);
7214 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
7215 {
7216 saved_reg_live = BITMAP_ALLOC (NULL);
7217 region_ref_regs = BITMAP_ALLOC (NULL);
7218 }
7219 if (sched_pressure == SCHED_PRESSURE_MODEL)
7220 tmp_bitmap = BITMAP_ALLOC (NULL);
7221
7222 /* Calculate number of CALL_SAVED_REGS and FIXED_REGS in register classes
7223 that we calculate register pressure for. */
7224 for (int c = 0; c < ira_pressure_classes_num; ++c)
7225 {
7226 enum reg_class cl = ira_pressure_classes[c];
7227
7228 call_saved_regs_num[cl] = 0;
7229 fixed_regs_num[cl] = 0;
7230
7231 for (int i = 0; i < ira_class_hard_regs_num[cl]; ++i)
7232 {
7233 unsigned int regno = ira_class_hard_regs[cl][i];
7234 if (fixed_regs[regno])
7235 ++fixed_regs_num[cl];
7236 else if (!crtl->abi->clobbers_full_reg_p (regno))
7237 ++call_saved_regs_num[cl];
7238 }
7239 }
7240 }
7241}
7242
7243/* Free data for register pressure sensitive scheduling. Also called
7244 from schedule_region when stopping sched-pressure early. */
7245void
7246free_global_sched_pressure_data (void)
7247{
7248 if (sched_pressure != SCHED_PRESSURE_NONE)
7249 {
7250 if (regstat_n_sets_and_refs != NULL)
7251 regstat_free_n_sets_and_refs ();
7252 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
7253 {
7254 BITMAP_FREE (region_ref_regs);
7255 BITMAP_FREE (saved_reg_live);
7256 }
7257 if (sched_pressure == SCHED_PRESSURE_MODEL)
7258 BITMAP_FREE (tmp_bitmap);
7259 BITMAP_FREE (curr_reg_live);
7260 free (ptr: sched_regno_pressure_class);
7261 }
7262}
7263
7264/* Initialize some global state for the scheduler. This function works
7265 with the common data shared between all the schedulers. It is called
7266 from the scheduler specific initialization routine. */
7267
7268void
7269sched_init (void)
7270{
7271 if (targetm.sched.dispatch (NULL, IS_DISPATCH_ON))
7272 targetm.sched.dispatch_do (NULL, DISPATCH_INIT);
7273
7274 if (live_range_shrinkage_p)
7275 sched_pressure = SCHED_PRESSURE_WEIGHTED;
7276 else if (flag_sched_pressure
7277 && !reload_completed
7278 && common_sched_info->sched_pass_id == SCHED_RGN_PASS)
7279 sched_pressure = ((enum sched_pressure_algorithm)
7280 param_sched_pressure_algorithm);
7281 else
7282 sched_pressure = SCHED_PRESSURE_NONE;
7283
7284 if (sched_pressure != SCHED_PRESSURE_NONE)
7285 ira_setup_eliminable_regset ();
7286
7287 /* Initialize SPEC_INFO. */
7288 if (targetm.sched.set_sched_flags)
7289 {
7290 spec_info = &spec_info_var;
7291 targetm.sched.set_sched_flags (spec_info);
7292
7293 if (spec_info->mask != 0)
7294 {
7295 spec_info->data_weakness_cutoff
7296 = (param_sched_spec_prob_cutoff * MAX_DEP_WEAK) / 100;
7297 spec_info->control_weakness_cutoff
7298 = (param_sched_spec_prob_cutoff * REG_BR_PROB_BASE) / 100;
7299 }
7300 else
7301 /* So we won't read anything accidentally. */
7302 spec_info = NULL;
7303
7304 }
7305 else
7306 /* So we won't read anything accidentally. */
7307 spec_info = 0;
7308
7309 /* Initialize issue_rate. */
7310 if (targetm.sched.issue_rate)
7311 issue_rate = targetm.sched.issue_rate ();
7312 else
7313 issue_rate = 1;
7314
7315 if (targetm.sched.first_cycle_multipass_dfa_lookahead
7316 /* Don't use max_issue with reg_pressure scheduling. Multipass
7317 scheduling and reg_pressure scheduling undo each other's decisions. */
7318 && sched_pressure == SCHED_PRESSURE_NONE)
7319 dfa_lookahead = targetm.sched.first_cycle_multipass_dfa_lookahead ();
7320 else
7321 dfa_lookahead = 0;
7322
7323 /* Set to "0" so that we recalculate. */
7324 max_lookahead_tries = 0;
7325
7326 if (targetm.sched.init_dfa_pre_cycle_insn)
7327 targetm.sched.init_dfa_pre_cycle_insn ();
7328
7329 if (targetm.sched.init_dfa_post_cycle_insn)
7330 targetm.sched.init_dfa_post_cycle_insn ();
7331
7332 dfa_start ();
7333 dfa_state_size = state_size ();
7334
7335 init_alias_analysis ();
7336
7337 if (!sched_no_dce)
7338 df_set_flags (DF_LR_RUN_DCE);
7339 df_note_add_problem ();
7340
7341 /* More problems needed for interloop dep calculation in SMS. */
7342 if (common_sched_info->sched_pass_id == SCHED_SMS_PASS)
7343 {
7344 df_rd_add_problem ();
7345 df_chain_add_problem (DF_DU_CHAIN + DF_UD_CHAIN);
7346 }
7347
7348 df_analyze ();
7349
7350 /* Do not run DCE after reload, as this can kill nops inserted
7351 by bundling. */
7352 if (reload_completed)
7353 df_clear_flags (DF_LR_RUN_DCE);
7354
7355 regstat_compute_calls_crossed ();
7356
7357 if (targetm.sched.init_global)
7358 targetm.sched.init_global (sched_dump, sched_verbose, get_max_uid () + 1);
7359
7360 alloc_global_sched_pressure_data ();
7361
7362 curr_state = xmalloc (dfa_state_size);
7363}
7364
7365static void haifa_init_only_bb (basic_block, basic_block);
7366
7367/* Initialize data structures specific to the Haifa scheduler. */
7368void
7369haifa_sched_init (void)
7370{
7371 setup_sched_dump ();
7372 sched_init ();
7373
7374 scheduled_insns.create (nelems: 0);
7375
7376 if (spec_info != NULL)
7377 {
7378 sched_deps_info->use_deps_list = 1;
7379 sched_deps_info->generate_spec_deps = 1;
7380 }
7381
7382 /* Initialize luids, dependency caches, target and h_i_d for the
7383 whole function. */
7384 {
7385 sched_init_bbs ();
7386
7387 auto_vec<basic_block> bbs (n_basic_blocks_for_fn (cfun));
7388 basic_block bb;
7389 FOR_EACH_BB_FN (bb, cfun)
7390 bbs.quick_push (obj: bb);
7391 sched_init_luids (bbs);
7392 sched_deps_init (true);
7393 sched_extend_target ();
7394 haifa_init_h_i_d (bbs);
7395 }
7396
7397 sched_init_only_bb = haifa_init_only_bb;
7398 sched_split_block = sched_split_block_1;
7399 sched_create_empty_bb = sched_create_empty_bb_1;
7400 haifa_recovery_bb_ever_added_p = false;
7401
7402 nr_begin_data = nr_begin_control = nr_be_in_data = nr_be_in_control = 0;
7403 before_recovery = 0;
7404 after_recovery = 0;
7405
7406 modulo_ii = 0;
7407}
7408
7409/* Finish work with the data specific to the Haifa scheduler. */
7410void
7411haifa_sched_finish (void)
7412{
7413 sched_create_empty_bb = NULL;
7414 sched_split_block = NULL;
7415 sched_init_only_bb = NULL;
7416
7417 if (spec_info && spec_info->dump)
7418 {
7419 char c = reload_completed ? 'a' : 'b';
7420
7421 fprintf (stream: spec_info->dump,
7422 format: ";; %s:\n", current_function_name ());
7423
7424 fprintf (stream: spec_info->dump,
7425 format: ";; Procedure %cr-begin-data-spec motions == %d\n",
7426 c, nr_begin_data);
7427 fprintf (stream: spec_info->dump,
7428 format: ";; Procedure %cr-be-in-data-spec motions == %d\n",
7429 c, nr_be_in_data);
7430 fprintf (stream: spec_info->dump,
7431 format: ";; Procedure %cr-begin-control-spec motions == %d\n",
7432 c, nr_begin_control);
7433 fprintf (stream: spec_info->dump,
7434 format: ";; Procedure %cr-be-in-control-spec motions == %d\n",
7435 c, nr_be_in_control);
7436 }
7437
7438 scheduled_insns.release ();
7439
7440 /* Finalize h_i_d, dependency caches, and luids for the whole
7441 function. Target will be finalized in md_global_finish (). */
7442 sched_deps_finish ();
7443 sched_finish_luids ();
7444 current_sched_info = NULL;
7445 insn_queue = NULL;
7446 sched_finish ();
7447}
7448
7449/* Free global data used during insn scheduling. This function works with
7450 the common data shared between the schedulers. */
7451
7452void
7453sched_finish (void)
7454{
7455 haifa_finish_h_i_d ();
7456 free_global_sched_pressure_data ();
7457 free (ptr: curr_state);
7458
7459 if (targetm.sched.finish_global)
7460 targetm.sched.finish_global (sched_dump, sched_verbose);
7461
7462 end_alias_analysis ();
7463
7464 regstat_free_calls_crossed ();
7465
7466 dfa_finish ();
7467}
7468
7469/* Free all delay_pair structures that were recorded. */
7470void
7471free_delay_pairs (void)
7472{
7473 if (delay_htab)
7474 {
7475 delay_htab->empty ();
7476 delay_htab_i2->empty ();
7477 }
7478}
7479
7480/* Fix INSN_TICKs of the instructions in the current block as well as
7481 INSN_TICKs of their dependents.
7482 HEAD and TAIL are the begin and the end of the current scheduled block. */
7483static void
7484fix_inter_tick (rtx_insn *head, rtx_insn *tail)
7485{
7486 /* Set of instructions with corrected INSN_TICK. */
7487 auto_bitmap processed;
7488 /* ??? It is doubtful if we should assume that cycle advance happens on
7489 basic block boundaries. Basically insns that are unconditionally ready
7490 on the start of the block are more preferable then those which have
7491 a one cycle dependency over insn from the previous block. */
7492 int next_clock = clock_var + 1;
7493
7494 /* Iterates over scheduled instructions and fix their INSN_TICKs and
7495 INSN_TICKs of dependent instructions, so that INSN_TICKs are consistent
7496 across different blocks. */
7497 for (tail = NEXT_INSN (insn: tail); head != tail; head = NEXT_INSN (insn: head))
7498 {
7499 if (INSN_P (head))
7500 {
7501 int tick;
7502 sd_iterator_def sd_it;
7503 dep_t dep;
7504
7505 tick = INSN_TICK (head);
7506 gcc_assert (tick >= MIN_TICK);
7507
7508 /* Fix INSN_TICK of instruction from just scheduled block. */
7509 if (bitmap_set_bit (processed, INSN_LUID (head)))
7510 {
7511 tick -= next_clock;
7512
7513 if (tick < MIN_TICK)
7514 tick = MIN_TICK;
7515
7516 INSN_TICK (head) = tick;
7517 }
7518
7519 if (DEBUG_INSN_P (head))
7520 continue;
7521
7522 FOR_EACH_DEP (head, SD_LIST_RES_FORW, sd_it, dep)
7523 {
7524 rtx_insn *next;
7525
7526 next = DEP_CON (dep);
7527 tick = INSN_TICK (next);
7528
7529 if (tick != INVALID_TICK
7530 /* If NEXT has its INSN_TICK calculated, fix it.
7531 If not - it will be properly calculated from
7532 scratch later in fix_tick_ready. */
7533 && bitmap_set_bit (processed, INSN_LUID (next)))
7534 {
7535 tick -= next_clock;
7536
7537 if (tick < MIN_TICK)
7538 tick = MIN_TICK;
7539
7540 if (tick > INTER_TICK (next))
7541 INTER_TICK (next) = tick;
7542 else
7543 tick = INTER_TICK (next);
7544
7545 INSN_TICK (next) = tick;
7546 }
7547 }
7548 }
7549 }
7550}
7551
7552/* Check if NEXT is ready to be added to the ready or queue list.
7553 If "yes", add it to the proper list.
7554 Returns:
7555 -1 - is not ready yet,
7556 0 - added to the ready list,
7557 0 < N - queued for N cycles. */
7558int
7559try_ready (rtx_insn *next)
7560{
7561 ds_t old_ts, new_ts;
7562
7563 old_ts = TODO_SPEC (next);
7564
7565 gcc_assert (!(old_ts & ~(SPECULATIVE | HARD_DEP | DEP_CONTROL | DEP_POSTPONED))
7566 && (old_ts == HARD_DEP
7567 || old_ts == DEP_POSTPONED
7568 || (old_ts & SPECULATIVE)
7569 || old_ts == DEP_CONTROL));
7570
7571 new_ts = recompute_todo_spec (next, for_backtrack: false);
7572
7573 if (new_ts & (HARD_DEP | DEP_POSTPONED))
7574 gcc_assert (new_ts == old_ts
7575 && QUEUE_INDEX (next) == QUEUE_NOWHERE);
7576 else if (current_sched_info->new_ready)
7577 new_ts = current_sched_info->new_ready (next, new_ts);
7578
7579 /* * if !(old_ts & SPECULATIVE) (e.g. HARD_DEP or 0), then insn might
7580 have its original pattern or changed (speculative) one. This is due
7581 to changing ebb in region scheduling.
7582 * But if (old_ts & SPECULATIVE), then we are pretty sure that insn
7583 has speculative pattern.
7584
7585 We can't assert (!(new_ts & HARD_DEP) || new_ts == old_ts) here because
7586 control-speculative NEXT could have been discarded by sched-rgn.cc
7587 (the same case as when discarded by can_schedule_ready_p ()). */
7588
7589 if ((new_ts & SPECULATIVE)
7590 /* If (old_ts == new_ts), then (old_ts & SPECULATIVE) and we don't
7591 need to change anything. */
7592 && new_ts != old_ts)
7593 {
7594 int res;
7595 rtx new_pat;
7596
7597 gcc_assert ((new_ts & SPECULATIVE) && !(new_ts & ~SPECULATIVE));
7598
7599 res = haifa_speculate_insn (next, new_ts, &new_pat);
7600
7601 switch (res)
7602 {
7603 case -1:
7604 /* It would be nice to change DEP_STATUS of all dependences,
7605 which have ((DEP_STATUS & SPECULATIVE) == new_ts) to HARD_DEP,
7606 so we won't reanalyze anything. */
7607 new_ts = HARD_DEP;
7608 break;
7609
7610 case 0:
7611 /* We follow the rule, that every speculative insn
7612 has non-null ORIG_PAT. */
7613 if (!ORIG_PAT (next))
7614 ORIG_PAT (next) = PATTERN (insn: next);
7615 break;
7616
7617 case 1:
7618 if (!ORIG_PAT (next))
7619 /* If we gonna to overwrite the original pattern of insn,
7620 save it. */
7621 ORIG_PAT (next) = PATTERN (insn: next);
7622
7623 res = haifa_change_pattern (next, new_pat);
7624 gcc_assert (res);
7625 break;
7626
7627 default:
7628 gcc_unreachable ();
7629 }
7630 }
7631
7632 /* We need to restore pattern only if (new_ts == 0), because otherwise it is
7633 either correct (new_ts & SPECULATIVE),
7634 or we simply don't care (new_ts & HARD_DEP). */
7635
7636 gcc_assert (!ORIG_PAT (next)
7637 || !IS_SPECULATION_BRANCHY_CHECK_P (next));
7638
7639 TODO_SPEC (next) = new_ts;
7640
7641 if (new_ts & (HARD_DEP | DEP_POSTPONED))
7642 {
7643 /* We can't assert (QUEUE_INDEX (next) == QUEUE_NOWHERE) here because
7644 control-speculative NEXT could have been discarded by sched-rgn.cc
7645 (the same case as when discarded by can_schedule_ready_p ()). */
7646 /*gcc_assert (QUEUE_INDEX (next) == QUEUE_NOWHERE);*/
7647
7648 change_queue_index (next, QUEUE_NOWHERE);
7649
7650 return -1;
7651 }
7652 else if (!(new_ts & BEGIN_SPEC)
7653 && ORIG_PAT (next) && PREDICATED_PAT (next) == NULL_RTX
7654 && !IS_SPECULATION_CHECK_P (next))
7655 /* We should change pattern of every previously speculative
7656 instruction - and we determine if NEXT was speculative by using
7657 ORIG_PAT field. Except one case - speculation checks have ORIG_PAT
7658 pat too, so skip them. */
7659 {
7660 bool success = haifa_change_pattern (next, ORIG_PAT (next));
7661 gcc_assert (success);
7662 ORIG_PAT (next) = 0;
7663 }
7664
7665 if (sched_verbose >= 2)
7666 {
7667 fprintf (stream: sched_dump, format: ";;\t\tdependencies resolved: insn %s",
7668 (*current_sched_info->print_insn) (next, 0));
7669
7670 if (spec_info && spec_info->dump)
7671 {
7672 if (new_ts & BEGIN_DATA)
7673 fprintf (stream: spec_info->dump, format: "; data-spec;");
7674 if (new_ts & BEGIN_CONTROL)
7675 fprintf (stream: spec_info->dump, format: "; control-spec;");
7676 if (new_ts & BE_IN_CONTROL)
7677 fprintf (stream: spec_info->dump, format: "; in-control-spec;");
7678 }
7679 if (TODO_SPEC (next) & DEP_CONTROL)
7680 fprintf (stream: sched_dump, format: " predicated");
7681 fprintf (stream: sched_dump, format: "\n");
7682 }
7683
7684 adjust_priority (prev: next);
7685
7686 return fix_tick_ready (next);
7687}
7688
7689/* Calculate INSN_TICK of NEXT and add it to either ready or queue list. */
7690static int
7691fix_tick_ready (rtx_insn *next)
7692{
7693 int tick, delay;
7694
7695 if (!DEBUG_INSN_P (next) && !sd_lists_empty_p (next, SD_LIST_RES_BACK))
7696 {
7697 int full_p;
7698 sd_iterator_def sd_it;
7699 dep_t dep;
7700
7701 tick = INSN_TICK (next);
7702 /* if tick is not equal to INVALID_TICK, then update
7703 INSN_TICK of NEXT with the most recent resolved dependence
7704 cost. Otherwise, recalculate from scratch. */
7705 full_p = (tick == INVALID_TICK);
7706
7707 FOR_EACH_DEP (next, SD_LIST_RES_BACK, sd_it, dep)
7708 {
7709 rtx_insn *pro = DEP_PRO (dep);
7710 int tick1;
7711
7712 gcc_assert (INSN_TICK (pro) >= MIN_TICK);
7713
7714 tick1 = INSN_TICK (pro) + dep_cost (link: dep);
7715 if (tick1 > tick)
7716 tick = tick1;
7717
7718 if (!full_p)
7719 break;
7720 }
7721 }
7722 else
7723 tick = -1;
7724
7725 INSN_TICK (next) = tick;
7726
7727 delay = tick - clock_var;
7728 if (delay <= 0 || sched_pressure != SCHED_PRESSURE_NONE || sched_fusion)
7729 delay = QUEUE_READY;
7730
7731 change_queue_index (next, delay);
7732
7733 return delay;
7734}
7735
7736/* Move NEXT to the proper queue list with (DELAY >= 1),
7737 or add it to the ready list (DELAY == QUEUE_READY),
7738 or remove it from ready and queue lists at all (DELAY == QUEUE_NOWHERE). */
7739static void
7740change_queue_index (rtx_insn *next, int delay)
7741{
7742 int i = QUEUE_INDEX (next);
7743
7744 gcc_assert (QUEUE_NOWHERE <= delay && delay <= max_insn_queue_index
7745 && delay != 0);
7746 gcc_assert (i != QUEUE_SCHEDULED);
7747
7748 if ((delay > 0 && NEXT_Q_AFTER (q_ptr, delay) == i)
7749 || (delay < 0 && delay == i))
7750 /* We have nothing to do. */
7751 return;
7752
7753 /* Remove NEXT from wherever it is now. */
7754 if (i == QUEUE_READY)
7755 ready_remove_insn (insn: next);
7756 else if (i >= 0)
7757 queue_remove (insn: next);
7758
7759 /* Add it to the proper place. */
7760 if (delay == QUEUE_READY)
7761 ready_add (ready: readyp, insn: next, first_p: false);
7762 else if (delay >= 1)
7763 queue_insn (insn: next, n_cycles: delay, reason: "change queue index");
7764
7765 if (sched_verbose >= 2)
7766 {
7767 fprintf (stream: sched_dump, format: ";;\t\ttick updated: insn %s",
7768 (*current_sched_info->print_insn) (next, 0));
7769
7770 if (delay == QUEUE_READY)
7771 fprintf (stream: sched_dump, format: " into ready\n");
7772 else if (delay >= 1)
7773 fprintf (stream: sched_dump, format: " into queue with cost=%d\n", delay);
7774 else
7775 fprintf (stream: sched_dump, format: " removed from ready or queue lists\n");
7776 }
7777}
7778
7779static int sched_ready_n_insns = -1;
7780
7781/* Initialize per region data structures. */
7782void
7783sched_extend_ready_list (int new_sched_ready_n_insns)
7784{
7785 int i;
7786
7787 if (sched_ready_n_insns == -1)
7788 /* At the first call we need to initialize one more choice_stack
7789 entry. */
7790 {
7791 i = 0;
7792 sched_ready_n_insns = 0;
7793 scheduled_insns.reserve (nelems: new_sched_ready_n_insns);
7794 }
7795 else
7796 i = sched_ready_n_insns + 1;
7797
7798 ready.veclen = new_sched_ready_n_insns + issue_rate;
7799 ready.vec = XRESIZEVEC (rtx_insn *, ready.vec, ready.veclen);
7800
7801 gcc_assert (new_sched_ready_n_insns >= sched_ready_n_insns);
7802
7803 ready_try = (signed char *) xrecalloc (ready_try, new_sched_ready_n_insns,
7804 sched_ready_n_insns,
7805 sizeof (*ready_try));
7806
7807 /* We allocate +1 element to save initial state in the choice_stack[0]
7808 entry. */
7809 choice_stack = XRESIZEVEC (struct choice_entry, choice_stack,
7810 new_sched_ready_n_insns + 1);
7811
7812 for (; i <= new_sched_ready_n_insns; i++)
7813 {
7814 choice_stack[i].state = xmalloc (dfa_state_size);
7815
7816 if (targetm.sched.first_cycle_multipass_init)
7817 targetm.sched.first_cycle_multipass_init (&(choice_stack[i]
7818 .target_data));
7819 }
7820
7821 sched_ready_n_insns = new_sched_ready_n_insns;
7822}
7823
7824/* Free per region data structures. */
7825void
7826sched_finish_ready_list (void)
7827{
7828 int i;
7829
7830 free (ptr: ready.vec);
7831 ready.vec = NULL;
7832 ready.veclen = 0;
7833
7834 free (ptr: ready_try);
7835 ready_try = NULL;
7836
7837 for (i = 0; i <= sched_ready_n_insns; i++)
7838 {
7839 if (targetm.sched.first_cycle_multipass_fini)
7840 targetm.sched.first_cycle_multipass_fini (&(choice_stack[i]
7841 .target_data));
7842
7843 free (ptr: choice_stack [i].state);
7844 }
7845 free (ptr: choice_stack);
7846 choice_stack = NULL;
7847
7848 sched_ready_n_insns = -1;
7849}
7850
7851static int
7852haifa_luid_for_non_insn (rtx x)
7853{
7854 gcc_assert (NOTE_P (x) || LABEL_P (x));
7855
7856 return 0;
7857}
7858
7859/* Generates recovery code for INSN. */
7860static void
7861generate_recovery_code (rtx_insn *insn)
7862{
7863 if (TODO_SPEC (insn) & BEGIN_SPEC)
7864 begin_speculative_block (insn);
7865
7866 /* Here we have insn with no dependencies to
7867 instructions other then CHECK_SPEC ones. */
7868
7869 if (TODO_SPEC (insn) & BE_IN_SPEC)
7870 add_to_speculative_block (insn);
7871}
7872
7873/* Helper function.
7874 Tries to add speculative dependencies of type FS between instructions
7875 in deps_list L and TWIN. */
7876static void
7877process_insn_forw_deps_be_in_spec (rtx_insn *insn, rtx_insn *twin, ds_t fs)
7878{
7879 sd_iterator_def sd_it;
7880 dep_t dep;
7881
7882 FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep)
7883 {
7884 ds_t ds;
7885 rtx_insn *consumer;
7886
7887 consumer = DEP_CON (dep);
7888
7889 ds = DEP_STATUS (dep);
7890
7891 if (/* If we want to create speculative dep. */
7892 fs
7893 /* And we can do that because this is a true dep. */
7894 && (ds & DEP_TYPES) == DEP_TRUE)
7895 {
7896 gcc_assert (!(ds & BE_IN_SPEC));
7897
7898 if (/* If this dep can be overcome with 'begin speculation'. */
7899 ds & BEGIN_SPEC)
7900 /* Then we have a choice: keep the dep 'begin speculative'
7901 or transform it into 'be in speculative'. */
7902 {
7903 if (/* In try_ready we assert that if insn once became ready
7904 it can be removed from the ready (or queue) list only
7905 due to backend decision. Hence we can't let the
7906 probability of the speculative dep to decrease. */
7907 ds_weak (ds) <= ds_weak (fs))
7908 {
7909 ds_t new_ds;
7910
7911 new_ds = (ds & ~BEGIN_SPEC) | fs;
7912
7913 if (/* consumer can 'be in speculative'. */
7914 sched_insn_is_legitimate_for_speculation_p (consumer,
7915 new_ds))
7916 /* Transform it to be in speculative. */
7917 ds = new_ds;
7918 }
7919 }
7920 else
7921 /* Mark the dep as 'be in speculative'. */
7922 ds |= fs;
7923 }
7924
7925 {
7926 dep_def _new_dep, *new_dep = &_new_dep;
7927
7928 init_dep_1 (new_dep, twin, consumer, DEP_TYPE (dep), ds);
7929 sd_add_dep (new_dep, false);
7930 }
7931 }
7932}
7933
7934/* Generates recovery code for BEGIN speculative INSN. */
7935static void
7936begin_speculative_block (rtx_insn *insn)
7937{
7938 if (TODO_SPEC (insn) & BEGIN_DATA)
7939 nr_begin_data++;
7940 if (TODO_SPEC (insn) & BEGIN_CONTROL)
7941 nr_begin_control++;
7942
7943 create_check_block_twin (insn, false);
7944
7945 TODO_SPEC (insn) &= ~BEGIN_SPEC;
7946}
7947
7948static void haifa_init_insn (rtx_insn *);
7949
7950/* Generates recovery code for BE_IN speculative INSN. */
7951static void
7952add_to_speculative_block (rtx_insn *insn)
7953{
7954 ds_t ts;
7955 sd_iterator_def sd_it;
7956 dep_t dep;
7957 auto_vec<rtx_insn *, 10> twins;
7958
7959 ts = TODO_SPEC (insn);
7960 gcc_assert (!(ts & ~BE_IN_SPEC));
7961
7962 if (ts & BE_IN_DATA)
7963 nr_be_in_data++;
7964 if (ts & BE_IN_CONTROL)
7965 nr_be_in_control++;
7966
7967 TODO_SPEC (insn) &= ~BE_IN_SPEC;
7968 gcc_assert (!TODO_SPEC (insn));
7969
7970 DONE_SPEC (insn) |= ts;
7971
7972 /* First we convert all simple checks to branchy. */
7973 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
7974 sd_iterator_cond (it_ptr: &sd_it, dep_ptr: &dep);)
7975 {
7976 rtx_insn *check = DEP_PRO (dep);
7977
7978 if (IS_SPECULATION_SIMPLE_CHECK_P (check))
7979 {
7980 create_check_block_twin (check, true);
7981
7982 /* Restart search. */
7983 sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
7984 }
7985 else
7986 /* Continue search. */
7987 sd_iterator_next (it_ptr: &sd_it);
7988 }
7989
7990 auto_vec<rtx_insn *> priorities_roots;
7991 clear_priorities (insn, &priorities_roots);
7992
7993 while (1)
7994 {
7995 rtx_insn *check, *twin;
7996 basic_block rec;
7997
7998 /* Get the first backward dependency of INSN. */
7999 sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
8000 if (!sd_iterator_cond (it_ptr: &sd_it, dep_ptr: &dep))
8001 /* INSN has no backward dependencies left. */
8002 break;
8003
8004 gcc_assert ((DEP_STATUS (dep) & BEGIN_SPEC) == 0
8005 && (DEP_STATUS (dep) & BE_IN_SPEC) != 0
8006 && (DEP_STATUS (dep) & DEP_TYPES) == DEP_TRUE);
8007
8008 check = DEP_PRO (dep);
8009
8010 gcc_assert (!IS_SPECULATION_CHECK_P (check) && !ORIG_PAT (check)
8011 && QUEUE_INDEX (check) == QUEUE_NOWHERE);
8012
8013 rec = BLOCK_FOR_INSN (insn: check);
8014
8015 twin = emit_insn_before (copy_insn (PATTERN (insn)), BB_END (rec));
8016 haifa_init_insn (twin);
8017
8018 sd_copy_back_deps (twin, insn, true);
8019
8020 if (sched_verbose && spec_info->dump)
8021 /* INSN_BB (insn) isn't determined for twin insns yet.
8022 So we can't use current_sched_info->print_insn. */
8023 fprintf (stream: spec_info->dump, format: ";;\t\tGenerated twin insn : %d/rec%d\n",
8024 INSN_UID (insn: twin), rec->index);
8025
8026 twins.safe_push (obj: twin);
8027
8028 /* Add dependences between TWIN and all appropriate
8029 instructions from REC. */
8030 FOR_EACH_DEP (insn, SD_LIST_SPEC_BACK, sd_it, dep)
8031 {
8032 rtx_insn *pro = DEP_PRO (dep);
8033
8034 gcc_assert (DEP_TYPE (dep) == REG_DEP_TRUE);
8035
8036 /* INSN might have dependencies from the instructions from
8037 several recovery blocks. At this iteration we process those
8038 producers that reside in REC. */
8039 if (BLOCK_FOR_INSN (insn: pro) == rec)
8040 {
8041 dep_def _new_dep, *new_dep = &_new_dep;
8042
8043 init_dep (new_dep, pro, twin, REG_DEP_TRUE);
8044 sd_add_dep (new_dep, false);
8045 }
8046 }
8047
8048 process_insn_forw_deps_be_in_spec (insn, twin, fs: ts);
8049
8050 /* Remove all dependencies between INSN and insns in REC. */
8051 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
8052 sd_iterator_cond (it_ptr: &sd_it, dep_ptr: &dep);)
8053 {
8054 rtx_insn *pro = DEP_PRO (dep);
8055
8056 if (BLOCK_FOR_INSN (insn: pro) == rec)
8057 sd_delete_dep (sd_it);
8058 else
8059 sd_iterator_next (it_ptr: &sd_it);
8060 }
8061 }
8062
8063 /* We couldn't have added the dependencies between INSN and TWINS earlier
8064 because that would make TWINS appear in the INSN_BACK_DEPS (INSN). */
8065 unsigned int i;
8066 rtx_insn *twin;
8067 FOR_EACH_VEC_ELT_REVERSE (twins, i, twin)
8068 {
8069 dep_def _new_dep, *new_dep = &_new_dep;
8070
8071 init_dep (new_dep, insn, twin, REG_DEP_OUTPUT);
8072 sd_add_dep (new_dep, false);
8073 }
8074
8075 calc_priorities (priorities_roots);
8076}
8077
8078/* Extends and fills with zeros (only the new part) array pointed to by P. */
8079void *
8080xrecalloc (void *p, size_t new_nmemb, size_t old_nmemb, size_t size)
8081{
8082 gcc_assert (new_nmemb >= old_nmemb);
8083 p = XRESIZEVAR (void, p, new_nmemb * size);
8084 memset (s: ((char *) p) + old_nmemb * size, c: 0, n: (new_nmemb - old_nmemb) * size);
8085 return p;
8086}
8087
8088/* Helper function.
8089 Find fallthru edge from PRED. */
8090edge
8091find_fallthru_edge_from (basic_block pred)
8092{
8093 edge e;
8094 basic_block succ;
8095
8096 succ = pred->next_bb;
8097 gcc_assert (succ->prev_bb == pred);
8098
8099 if (EDGE_COUNT (pred->succs) <= EDGE_COUNT (succ->preds))
8100 {
8101 e = find_fallthru_edge (edges: pred->succs);
8102
8103 if (e)
8104 {
8105 gcc_assert (e->dest == succ || e->dest->index == EXIT_BLOCK);
8106 return e;
8107 }
8108 }
8109 else
8110 {
8111 e = find_fallthru_edge (edges: succ->preds);
8112
8113 if (e)
8114 {
8115 gcc_assert (e->src == pred);
8116 return e;
8117 }
8118 }
8119
8120 return NULL;
8121}
8122
8123/* Extend per basic block data structures. */
8124static void
8125sched_extend_bb (void)
8126{
8127 /* The following is done to keep current_sched_info->next_tail non null. */
8128 rtx_insn *end = BB_END (EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb);
8129 rtx_insn *insn = DEBUG_INSN_P (end) ? prev_nondebug_insn (end) : end;
8130 if (NEXT_INSN (insn: end) == 0
8131 || (!NOTE_P (insn)
8132 && !LABEL_P (insn)
8133 /* Don't emit a NOTE if it would end up before a BARRIER. */
8134 && !BARRIER_P (next_nondebug_insn (end))))
8135 {
8136 rtx_note *note = emit_note_after (NOTE_INSN_DELETED, end);
8137 /* Make note appear outside BB. */
8138 set_block_for_insn (insn: note, NULL);
8139 BB_END (EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb) = end;
8140 }
8141}
8142
8143/* Init per basic block data structures. */
8144void
8145sched_init_bbs (void)
8146{
8147 sched_extend_bb ();
8148}
8149
8150/* Initialize BEFORE_RECOVERY variable. */
8151static void
8152init_before_recovery (basic_block *before_recovery_ptr)
8153{
8154 basic_block last;
8155 edge e;
8156
8157 last = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
8158 e = find_fallthru_edge_from (pred: last);
8159
8160 if (e)
8161 {
8162 /* We create two basic blocks:
8163 1. Single instruction block is inserted right after E->SRC
8164 and has jump to
8165 2. Empty block right before EXIT_BLOCK.
8166 Between these two blocks recovery blocks will be emitted. */
8167
8168 basic_block single, empty;
8169
8170 /* If the fallthrough edge to exit we've found is from the block we've
8171 created before, don't do anything more. */
8172 if (last == after_recovery)
8173 return;
8174
8175 adding_bb_to_current_region_p = false;
8176
8177 single = sched_create_empty_bb (last);
8178 empty = sched_create_empty_bb (single);
8179
8180 /* Add new blocks to the root loop. */
8181 if (current_loops != NULL)
8182 {
8183 add_bb_to_loop (single, (*current_loops->larray)[0]);
8184 add_bb_to_loop (empty, (*current_loops->larray)[0]);
8185 }
8186
8187 single->count = last->count;
8188 empty->count = last->count;
8189 BB_COPY_PARTITION (single, last);
8190 BB_COPY_PARTITION (empty, last);
8191
8192 redirect_edge_succ (e, single);
8193 make_single_succ_edge (single, empty, 0);
8194 make_single_succ_edge (empty, EXIT_BLOCK_PTR_FOR_FN (cfun),
8195 EDGE_FALLTHRU);
8196
8197 rtx_code_label *label = block_label (empty);
8198 rtx_jump_insn *x = emit_jump_insn_after (targetm.gen_jump (label),
8199 BB_END (single));
8200 JUMP_LABEL (x) = label;
8201 LABEL_NUSES (label)++;
8202 haifa_init_insn (x);
8203
8204 emit_barrier_after (x);
8205
8206 sched_init_only_bb (empty, NULL);
8207 sched_init_only_bb (single, NULL);
8208 sched_extend_bb ();
8209
8210 adding_bb_to_current_region_p = true;
8211 before_recovery = single;
8212 after_recovery = empty;
8213
8214 if (before_recovery_ptr)
8215 *before_recovery_ptr = before_recovery;
8216
8217 if (sched_verbose >= 2 && spec_info->dump)
8218 fprintf (stream: spec_info->dump,
8219 format: ";;\t\tFixed fallthru to EXIT : %d->>%d->%d->>EXIT\n",
8220 last->index, single->index, empty->index);
8221 }
8222 else
8223 before_recovery = last;
8224}
8225
8226/* Returns new recovery block. */
8227basic_block
8228sched_create_recovery_block (basic_block *before_recovery_ptr)
8229{
8230 rtx_insn *barrier;
8231 basic_block rec;
8232
8233 haifa_recovery_bb_recently_added_p = true;
8234 haifa_recovery_bb_ever_added_p = true;
8235
8236 init_before_recovery (before_recovery_ptr);
8237
8238 barrier = get_last_bb_insn (before_recovery);
8239 gcc_assert (BARRIER_P (barrier));
8240
8241 rtx_insn *label = emit_label_after (gen_label_rtx (), barrier);
8242
8243 rec = create_basic_block (label, label, before_recovery);
8244
8245 /* A recovery block always ends with an unconditional jump. */
8246 emit_barrier_after (BB_END (rec));
8247
8248 if (BB_PARTITION (before_recovery) != BB_UNPARTITIONED)
8249 BB_SET_PARTITION (rec, BB_COLD_PARTITION);
8250
8251 if (sched_verbose && spec_info->dump)
8252 fprintf (stream: spec_info->dump, format: ";;\t\tGenerated recovery block rec%d\n",
8253 rec->index);
8254
8255 return rec;
8256}
8257
8258/* Create edges: FIRST_BB -> REC; FIRST_BB -> SECOND_BB; REC -> SECOND_BB
8259 and emit necessary jumps. */
8260void
8261sched_create_recovery_edges (basic_block first_bb, basic_block rec,
8262 basic_block second_bb)
8263{
8264 int edge_flags;
8265
8266 /* This is fixing of incoming edge. */
8267 /* ??? Which other flags should be specified? */
8268 if (BB_PARTITION (first_bb) != BB_PARTITION (rec))
8269 /* Partition type is the same, if it is "unpartitioned". */
8270 edge_flags = EDGE_CROSSING;
8271 else
8272 edge_flags = 0;
8273
8274 edge e2 = single_succ_edge (bb: first_bb);
8275 edge e = make_edge (first_bb, rec, edge_flags);
8276
8277 /* TODO: The actual probability can be determined and is computed as
8278 'todo_spec' variable in create_check_block_twin and
8279 in sel-sched.cc `check_ds' in create_speculation_check. */
8280 e->probability = profile_probability::very_unlikely ();
8281 rec->count = e->count ();
8282 e2->probability = e->probability.invert ();
8283
8284 rtx_code_label *label = block_label (second_bb);
8285 rtx_jump_insn *jump = emit_jump_insn_after (targetm.gen_jump (label),
8286 BB_END (rec));
8287 JUMP_LABEL (jump) = label;
8288 LABEL_NUSES (label)++;
8289
8290 if (BB_PARTITION (second_bb) != BB_PARTITION (rec))
8291 /* Partition type is the same, if it is "unpartitioned". */
8292 {
8293 /* Rewritten from cfgrtl.cc. */
8294 if (crtl->has_bb_partition && targetm_common.have_named_sections)
8295 {
8296 /* We don't need the same note for the check because
8297 any_condjump_p (check) == true. */
8298 CROSSING_JUMP_P (jump) = 1;
8299 }
8300 edge_flags = EDGE_CROSSING;
8301 }
8302 else
8303 edge_flags = 0;
8304
8305 make_single_succ_edge (rec, second_bb, edge_flags);
8306 if (dom_info_available_p (CDI_DOMINATORS))
8307 set_immediate_dominator (CDI_DOMINATORS, rec, first_bb);
8308}
8309
8310/* This function creates recovery code for INSN. If MUTATE_P is nonzero,
8311 INSN is a simple check, that should be converted to branchy one. */
8312static void
8313create_check_block_twin (rtx_insn *insn, bool mutate_p)
8314{
8315 basic_block rec;
8316 rtx_insn *label, *check, *twin;
8317 rtx check_pat;
8318 ds_t fs;
8319 sd_iterator_def sd_it;
8320 dep_t dep;
8321 dep_def _new_dep, *new_dep = &_new_dep;
8322 ds_t todo_spec;
8323
8324 gcc_assert (ORIG_PAT (insn) != NULL_RTX);
8325
8326 if (!mutate_p)
8327 todo_spec = TODO_SPEC (insn);
8328 else
8329 {
8330 gcc_assert (IS_SPECULATION_SIMPLE_CHECK_P (insn)
8331 && (TODO_SPEC (insn) & SPECULATIVE) == 0);
8332
8333 todo_spec = CHECK_SPEC (insn);
8334 }
8335
8336 todo_spec &= SPECULATIVE;
8337
8338 /* Create recovery block. */
8339 if (mutate_p || targetm.sched.needs_block_p (todo_spec))
8340 {
8341 rec = sched_create_recovery_block (NULL);
8342 label = BB_HEAD (rec);
8343 }
8344 else
8345 {
8346 rec = EXIT_BLOCK_PTR_FOR_FN (cfun);
8347 label = NULL;
8348 }
8349
8350 /* Emit CHECK. */
8351 check_pat = targetm.sched.gen_spec_check (insn, label, todo_spec);
8352
8353 if (rec != EXIT_BLOCK_PTR_FOR_FN (cfun))
8354 {
8355 /* To have mem_reg alive at the beginning of second_bb,
8356 we emit check BEFORE insn, so insn after splitting
8357 insn will be at the beginning of second_bb, which will
8358 provide us with the correct life information. */
8359 check = emit_jump_insn_before (check_pat, insn);
8360 JUMP_LABEL (check) = label;
8361 LABEL_NUSES (label)++;
8362 }
8363 else
8364 check = emit_insn_before (check_pat, insn);
8365
8366 /* Extend data structures. */
8367 haifa_init_insn (check);
8368
8369 /* CHECK is being added to current region. Extend ready list. */
8370 gcc_assert (sched_ready_n_insns != -1);
8371 sched_extend_ready_list (new_sched_ready_n_insns: sched_ready_n_insns + 1);
8372
8373 if (current_sched_info->add_remove_insn)
8374 current_sched_info->add_remove_insn (insn, 0);
8375
8376 RECOVERY_BLOCK (check) = rec;
8377
8378 if (sched_verbose && spec_info->dump)
8379 fprintf (stream: spec_info->dump, format: ";;\t\tGenerated check insn : %s\n",
8380 (*current_sched_info->print_insn) (check, 0));
8381
8382 gcc_assert (ORIG_PAT (insn));
8383
8384 /* Initialize TWIN (twin is a duplicate of original instruction
8385 in the recovery block). */
8386 if (rec != EXIT_BLOCK_PTR_FOR_FN (cfun))
8387 {
8388 sd_iterator_def sd_it;
8389 dep_t dep;
8390
8391 FOR_EACH_DEP (insn, SD_LIST_RES_BACK, sd_it, dep)
8392 if ((DEP_STATUS (dep) & DEP_OUTPUT) != 0)
8393 {
8394 struct _dep _dep2, *dep2 = &_dep2;
8395
8396 init_dep (dep2, DEP_PRO (dep), check, REG_DEP_TRUE);
8397
8398 sd_add_dep (dep2, true);
8399 }
8400
8401 twin = emit_insn_after (ORIG_PAT (insn), BB_END (rec));
8402 haifa_init_insn (twin);
8403
8404 if (sched_verbose && spec_info->dump)
8405 /* INSN_BB (insn) isn't determined for twin insns yet.
8406 So we can't use current_sched_info->print_insn. */
8407 fprintf (stream: spec_info->dump, format: ";;\t\tGenerated twin insn : %d/rec%d\n",
8408 INSN_UID (insn: twin), rec->index);
8409 }
8410 else
8411 {
8412 ORIG_PAT (check) = ORIG_PAT (insn);
8413 HAS_INTERNAL_DEP (check) = 1;
8414 twin = check;
8415 /* ??? We probably should change all OUTPUT dependencies to
8416 (TRUE | OUTPUT). */
8417 }
8418
8419 /* Copy all resolved back dependencies of INSN to TWIN. This will
8420 provide correct value for INSN_TICK (TWIN). */
8421 sd_copy_back_deps (twin, insn, true);
8422
8423 if (rec != EXIT_BLOCK_PTR_FOR_FN (cfun))
8424 /* In case of branchy check, fix CFG. */
8425 {
8426 basic_block first_bb, second_bb;
8427 rtx_insn *jump;
8428
8429 first_bb = BLOCK_FOR_INSN (insn: check);
8430 second_bb = sched_split_block (first_bb, check);
8431
8432 sched_create_recovery_edges (first_bb, rec, second_bb);
8433
8434 sched_init_only_bb (second_bb, first_bb);
8435 sched_init_only_bb (rec, EXIT_BLOCK_PTR_FOR_FN (cfun));
8436
8437 jump = BB_END (rec);
8438 haifa_init_insn (jump);
8439 }
8440
8441 /* Move backward dependences from INSN to CHECK and
8442 move forward dependences from INSN to TWIN. */
8443
8444 /* First, create dependencies between INSN's producers and CHECK & TWIN. */
8445 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
8446 {
8447 rtx_insn *pro = DEP_PRO (dep);
8448 ds_t ds;
8449
8450 /* If BEGIN_DATA: [insn ~~TRUE~~> producer]:
8451 check --TRUE--> producer ??? or ANTI ???
8452 twin --TRUE--> producer
8453 twin --ANTI--> check
8454
8455 If BEGIN_CONTROL: [insn ~~ANTI~~> producer]:
8456 check --ANTI--> producer
8457 twin --ANTI--> producer
8458 twin --ANTI--> check
8459
8460 If BE_IN_SPEC: [insn ~~TRUE~~> producer]:
8461 check ~~TRUE~~> producer
8462 twin ~~TRUE~~> producer
8463 twin --ANTI--> check */
8464
8465 ds = DEP_STATUS (dep);
8466
8467 if (ds & BEGIN_SPEC)
8468 {
8469 gcc_assert (!mutate_p);
8470 ds &= ~BEGIN_SPEC;
8471 }
8472
8473 init_dep_1 (new_dep, pro, check, DEP_TYPE (dep), ds);
8474 sd_add_dep (new_dep, false);
8475
8476 if (rec != EXIT_BLOCK_PTR_FOR_FN (cfun))
8477 {
8478 DEP_CON (new_dep) = twin;
8479 sd_add_dep (new_dep, false);
8480 }
8481 }
8482
8483 /* Second, remove backward dependencies of INSN. */
8484 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
8485 sd_iterator_cond (it_ptr: &sd_it, dep_ptr: &dep);)
8486 {
8487 if ((DEP_STATUS (dep) & BEGIN_SPEC)
8488 || mutate_p)
8489 /* We can delete this dep because we overcome it with
8490 BEGIN_SPECULATION. */
8491 sd_delete_dep (sd_it);
8492 else
8493 sd_iterator_next (it_ptr: &sd_it);
8494 }
8495
8496 /* Future Speculations. Determine what BE_IN speculations will be like. */
8497 fs = 0;
8498
8499 /* Fields (DONE_SPEC (x) & BEGIN_SPEC) and CHECK_SPEC (x) are set only
8500 here. */
8501
8502 gcc_assert (!DONE_SPEC (insn));
8503
8504 if (!mutate_p)
8505 {
8506 ds_t ts = TODO_SPEC (insn);
8507
8508 DONE_SPEC (insn) = ts & BEGIN_SPEC;
8509 CHECK_SPEC (check) = ts & BEGIN_SPEC;
8510
8511 /* Luckiness of future speculations solely depends upon initial
8512 BEGIN speculation. */
8513 if (ts & BEGIN_DATA)
8514 fs = set_dep_weak (fs, BE_IN_DATA, get_dep_weak (ts, BEGIN_DATA));
8515 if (ts & BEGIN_CONTROL)
8516 fs = set_dep_weak (fs, BE_IN_CONTROL,
8517 get_dep_weak (ts, BEGIN_CONTROL));
8518 }
8519 else
8520 CHECK_SPEC (check) = CHECK_SPEC (insn);
8521
8522 /* Future speculations: call the helper. */
8523 process_insn_forw_deps_be_in_spec (insn, twin, fs);
8524
8525 if (rec != EXIT_BLOCK_PTR_FOR_FN (cfun))
8526 {
8527 /* Which types of dependencies should we use here is,
8528 generally, machine-dependent question... But, for now,
8529 it is not. */
8530
8531 if (!mutate_p)
8532 {
8533 init_dep (new_dep, insn, check, REG_DEP_TRUE);
8534 sd_add_dep (new_dep, false);
8535
8536 init_dep (new_dep, insn, twin, REG_DEP_OUTPUT);
8537 sd_add_dep (new_dep, false);
8538 }
8539 else
8540 {
8541 if (spec_info->dump)
8542 fprintf (stream: spec_info->dump, format: ";;\t\tRemoved simple check : %s\n",
8543 (*current_sched_info->print_insn) (insn, 0));
8544
8545 /* Remove all dependencies of the INSN. */
8546 {
8547 sd_it = sd_iterator_start (insn, types: (SD_LIST_FORW
8548 | SD_LIST_BACK
8549 | SD_LIST_RES_BACK));
8550 while (sd_iterator_cond (it_ptr: &sd_it, dep_ptr: &dep))
8551 sd_delete_dep (sd_it);
8552 }
8553
8554 /* If former check (INSN) already was moved to the ready (or queue)
8555 list, add new check (CHECK) there too. */
8556 if (QUEUE_INDEX (insn) != QUEUE_NOWHERE)
8557 try_ready (next: check);
8558
8559 /* Remove old check from instruction stream and free its
8560 data. */
8561 sched_remove_insn (insn);
8562 }
8563
8564 init_dep (new_dep, check, twin, REG_DEP_ANTI);
8565 sd_add_dep (new_dep, false);
8566 }
8567 else
8568 {
8569 init_dep_1 (new_dep, insn, check, REG_DEP_TRUE, DEP_TRUE | DEP_OUTPUT);
8570 sd_add_dep (new_dep, false);
8571 }
8572
8573 if (!mutate_p)
8574 /* Fix priorities. If MUTATE_P is nonzero, this is not necessary,
8575 because it'll be done later in add_to_speculative_block. */
8576 {
8577 auto_vec<rtx_insn *> priorities_roots;
8578
8579 clear_priorities (twin, &priorities_roots);
8580 calc_priorities (priorities_roots);
8581 }
8582}
8583
8584/* Removes dependency between instructions in the recovery block REC
8585 and usual region instructions. It keeps inner dependences so it
8586 won't be necessary to recompute them. */
8587static void
8588fix_recovery_deps (basic_block rec)
8589{
8590 rtx_insn *note, *insn, *jump;
8591 auto_vec<rtx_insn *, 10> ready_list;
8592 auto_bitmap in_ready;
8593
8594 /* NOTE - a basic block note. */
8595 note = NEXT_INSN (BB_HEAD (rec));
8596 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
8597 insn = BB_END (rec);
8598 gcc_assert (JUMP_P (insn));
8599 insn = PREV_INSN (insn);
8600
8601 do
8602 {
8603 sd_iterator_def sd_it;
8604 dep_t dep;
8605
8606 for (sd_it = sd_iterator_start (insn, SD_LIST_FORW);
8607 sd_iterator_cond (it_ptr: &sd_it, dep_ptr: &dep);)
8608 {
8609 rtx_insn *consumer = DEP_CON (dep);
8610
8611 if (BLOCK_FOR_INSN (insn: consumer) != rec)
8612 {
8613 sd_delete_dep (sd_it);
8614
8615 if (bitmap_set_bit (in_ready, INSN_LUID (consumer)))
8616 ready_list.safe_push (obj: consumer);
8617 }
8618 else
8619 {
8620 gcc_assert ((DEP_STATUS (dep) & DEP_TYPES) == DEP_TRUE);
8621
8622 sd_iterator_next (it_ptr: &sd_it);
8623 }
8624 }
8625
8626 insn = PREV_INSN (insn);
8627 }
8628 while (insn != note);
8629
8630 /* Try to add instructions to the ready or queue list. */
8631 unsigned int i;
8632 rtx_insn *temp;
8633 FOR_EACH_VEC_ELT_REVERSE (ready_list, i, temp)
8634 try_ready (next: temp);
8635
8636 /* Fixing jump's dependences. */
8637 insn = BB_HEAD (rec);
8638 jump = BB_END (rec);
8639
8640 gcc_assert (LABEL_P (insn));
8641 insn = NEXT_INSN (insn);
8642
8643 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (insn));
8644 add_jump_dependencies (insn, jump);
8645}
8646
8647/* Change pattern of INSN to NEW_PAT. Invalidate cached haifa
8648 instruction data. */
8649static bool
8650haifa_change_pattern (rtx_insn *insn, rtx new_pat)
8651{
8652 int t;
8653
8654 t = validate_change (insn, &PATTERN (insn), new_pat, 0);
8655 if (!t)
8656 return false;
8657
8658 update_insn_after_change (insn);
8659 return true;
8660}
8661
8662/* -1 - can't speculate,
8663 0 - for speculation with REQUEST mode it is OK to use
8664 current instruction pattern,
8665 1 - need to change pattern for *NEW_PAT to be speculative. */
8666int
8667sched_speculate_insn (rtx_insn *insn, ds_t request, rtx *new_pat)
8668{
8669 gcc_assert (current_sched_info->flags & DO_SPECULATION
8670 && (request & SPECULATIVE)
8671 && sched_insn_is_legitimate_for_speculation_p (insn, request));
8672
8673 if ((request & spec_info->mask) != request)
8674 return -1;
8675
8676 if (request & BE_IN_SPEC
8677 && !(request & BEGIN_SPEC))
8678 return 0;
8679
8680 return targetm.sched.speculate_insn (insn, request, new_pat);
8681}
8682
8683static int
8684haifa_speculate_insn (rtx_insn *insn, ds_t request, rtx *new_pat)
8685{
8686 gcc_assert (sched_deps_info->generate_spec_deps
8687 && !IS_SPECULATION_CHECK_P (insn));
8688
8689 if (HAS_INTERNAL_DEP (insn)
8690 || SCHED_GROUP_P (insn))
8691 return -1;
8692
8693 return sched_speculate_insn (insn, request, new_pat);
8694}
8695
8696/* Print some information about block BB, which starts with HEAD and
8697 ends with TAIL, before scheduling it.
8698 I is zero, if scheduler is about to start with the fresh ebb. */
8699static void
8700dump_new_block_header (int i, basic_block bb, rtx_insn *head, rtx_insn *tail)
8701{
8702 if (!i)
8703 fprintf (stream: sched_dump,
8704 format: ";; ======================================================\n");
8705 else
8706 fprintf (stream: sched_dump,
8707 format: ";; =====================ADVANCING TO=====================\n");
8708 fprintf (stream: sched_dump,
8709 format: ";; -- basic block %d from %d to %d -- %s reload\n",
8710 bb->index, INSN_UID (insn: head), INSN_UID (insn: tail),
8711 (reload_completed ? "after" : "before"));
8712 fprintf (stream: sched_dump,
8713 format: ";; ======================================================\n");
8714 fprintf (stream: sched_dump, format: "\n");
8715}
8716
8717/* Unlink basic block notes and labels and saves them, so they
8718 can be easily restored. We unlink basic block notes in EBB to
8719 provide back-compatibility with the previous code, as target backends
8720 assume, that there'll be only instructions between
8721 current_sched_info->{head and tail}. We restore these notes as soon
8722 as we can.
8723 FIRST (LAST) is the first (last) basic block in the ebb.
8724 NB: In usual case (FIRST == LAST) nothing is really done. */
8725void
8726unlink_bb_notes (basic_block first, basic_block last)
8727{
8728 /* We DON'T unlink basic block notes of the first block in the ebb. */
8729 if (first == last)
8730 return;
8731
8732 bb_header = XNEWVEC (rtx_insn *, last_basic_block_for_fn (cfun));
8733
8734 /* Make a sentinel. */
8735 if (last->next_bb != EXIT_BLOCK_PTR_FOR_FN (cfun))
8736 bb_header[last->next_bb->index] = 0;
8737
8738 first = first->next_bb;
8739 do
8740 {
8741 rtx_insn *prev, *label, *note, *next;
8742
8743 label = BB_HEAD (last);
8744 if (LABEL_P (label))
8745 note = NEXT_INSN (insn: label);
8746 else
8747 note = label;
8748 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
8749
8750 prev = PREV_INSN (insn: label);
8751 next = NEXT_INSN (insn: note);
8752 gcc_assert (prev && next);
8753
8754 SET_NEXT_INSN (prev) = next;
8755 SET_PREV_INSN (next) = prev;
8756
8757 bb_header[last->index] = label;
8758
8759 if (last == first)
8760 break;
8761
8762 last = last->prev_bb;
8763 }
8764 while (1);
8765}
8766
8767/* Restore basic block notes.
8768 FIRST is the first basic block in the ebb. */
8769static void
8770restore_bb_notes (basic_block first)
8771{
8772 if (!bb_header)
8773 return;
8774
8775 /* We DON'T unlink basic block notes of the first block in the ebb. */
8776 first = first->next_bb;
8777 /* Remember: FIRST is actually a second basic block in the ebb. */
8778
8779 while (first != EXIT_BLOCK_PTR_FOR_FN (cfun)
8780 && bb_header[first->index])
8781 {
8782 rtx_insn *prev, *label, *note, *next;
8783
8784 label = bb_header[first->index];
8785 prev = PREV_INSN (insn: label);
8786 next = NEXT_INSN (insn: prev);
8787
8788 if (LABEL_P (label))
8789 note = NEXT_INSN (insn: label);
8790 else
8791 note = label;
8792 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
8793
8794 bb_header[first->index] = 0;
8795
8796 SET_NEXT_INSN (prev) = label;
8797 SET_NEXT_INSN (note) = next;
8798 SET_PREV_INSN (next) = note;
8799
8800 first = first->next_bb;
8801 }
8802
8803 free (ptr: bb_header);
8804 bb_header = 0;
8805}
8806
8807/* Helper function.
8808 Fix CFG after both in- and inter-block movement of
8809 control_flow_insn_p JUMP. */
8810static void
8811fix_jump_move (rtx_insn *jump)
8812{
8813 basic_block bb, jump_bb, jump_bb_next;
8814
8815 bb = BLOCK_FOR_INSN (insn: PREV_INSN (insn: jump));
8816 jump_bb = BLOCK_FOR_INSN (insn: jump);
8817 jump_bb_next = jump_bb->next_bb;
8818
8819 gcc_assert (common_sched_info->sched_pass_id == SCHED_EBB_PASS
8820 || IS_SPECULATION_BRANCHY_CHECK_P (jump));
8821
8822 if (!NOTE_INSN_BASIC_BLOCK_P (BB_END (jump_bb_next)))
8823 /* if jump_bb_next is not empty. */
8824 BB_END (jump_bb) = BB_END (jump_bb_next);
8825
8826 if (BB_END (bb) != PREV_INSN (insn: jump))
8827 /* Then there are instruction after jump that should be placed
8828 to jump_bb_next. */
8829 BB_END (jump_bb_next) = BB_END (bb);
8830 else
8831 /* Otherwise jump_bb_next is empty. */
8832 BB_END (jump_bb_next) = NEXT_INSN (BB_HEAD (jump_bb_next));
8833
8834 /* To make assertion in move_insn happy. */
8835 BB_END (bb) = PREV_INSN (insn: jump);
8836
8837 update_bb_for_insn (jump_bb_next);
8838}
8839
8840/* Fix CFG after interblock movement of control_flow_insn_p JUMP. */
8841static void
8842move_block_after_check (rtx_insn *jump)
8843{
8844 basic_block bb, jump_bb, jump_bb_next;
8845 vec<edge, va_gc> *t;
8846
8847 bb = BLOCK_FOR_INSN (insn: PREV_INSN (insn: jump));
8848 jump_bb = BLOCK_FOR_INSN (insn: jump);
8849 jump_bb_next = jump_bb->next_bb;
8850
8851 update_bb_for_insn (jump_bb);
8852
8853 gcc_assert (IS_SPECULATION_CHECK_P (jump)
8854 || IS_SPECULATION_CHECK_P (BB_END (jump_bb_next)));
8855
8856 unlink_block (jump_bb_next);
8857 link_block (jump_bb_next, bb);
8858
8859 t = bb->succs;
8860 bb->succs = 0;
8861 move_succs (&(jump_bb->succs), bb);
8862 move_succs (&(jump_bb_next->succs), jump_bb);
8863 move_succs (&t, jump_bb_next);
8864
8865 df_mark_solutions_dirty ();
8866
8867 common_sched_info->fix_recovery_cfg
8868 (bb->index, jump_bb->index, jump_bb_next->index);
8869}
8870
8871/* Helper function for move_block_after_check.
8872 This functions attaches edge vector pointed to by SUCCSP to
8873 block TO. */
8874static void
8875move_succs (vec<edge, va_gc> **succsp, basic_block to)
8876{
8877 edge e;
8878 edge_iterator ei;
8879
8880 gcc_assert (to->succs == 0);
8881
8882 to->succs = *succsp;
8883
8884 FOR_EACH_EDGE (e, ei, to->succs)
8885 e->src = to;
8886
8887 *succsp = 0;
8888}
8889
8890/* Remove INSN from the instruction stream.
8891 INSN should have any dependencies. */
8892static void
8893sched_remove_insn (rtx_insn *insn)
8894{
8895 sd_finish_insn (insn);
8896
8897 change_queue_index (next: insn, QUEUE_NOWHERE);
8898 current_sched_info->add_remove_insn (insn, 1);
8899 delete_insn (insn);
8900}
8901
8902/* Clear priorities of all instructions, that are forward dependent on INSN.
8903 Store in vector pointed to by ROOTS_PTR insns on which priority () should
8904 be invoked to initialize all cleared priorities. */
8905static void
8906clear_priorities (rtx_insn *insn, rtx_vec_t *roots_ptr)
8907{
8908 sd_iterator_def sd_it;
8909 dep_t dep;
8910 bool insn_is_root_p = true;
8911
8912 gcc_assert (QUEUE_INDEX (insn) != QUEUE_SCHEDULED);
8913
8914 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
8915 {
8916 rtx_insn *pro = DEP_PRO (dep);
8917
8918 if (INSN_PRIORITY_STATUS (pro) >= 0
8919 && QUEUE_INDEX (insn) != QUEUE_SCHEDULED)
8920 {
8921 /* If DEP doesn't contribute to priority then INSN itself should
8922 be added to priority roots. */
8923 if (contributes_to_priority_p (dep))
8924 insn_is_root_p = false;
8925
8926 INSN_PRIORITY_STATUS (pro) = -1;
8927 clear_priorities (insn: pro, roots_ptr);
8928 }
8929 }
8930
8931 if (insn_is_root_p)
8932 roots_ptr->safe_push (obj: insn);
8933}
8934
8935/* Recompute priorities of instructions, whose priorities might have been
8936 changed. ROOTS is a vector of instructions whose priority computation will
8937 trigger initialization of all cleared priorities. */
8938static void
8939calc_priorities (const rtx_vec_t &roots)
8940{
8941 int i;
8942 rtx_insn *insn;
8943
8944 FOR_EACH_VEC_ELT (roots, i, insn)
8945 priority (insn);
8946}
8947
8948
8949/* Add dependences between JUMP and other instructions in the recovery
8950 block. INSN is the first insn the recovery block. */
8951static void
8952add_jump_dependencies (rtx_insn *insn, rtx_insn *jump)
8953{
8954 do
8955 {
8956 insn = NEXT_INSN (insn);
8957 if (insn == jump)
8958 break;
8959
8960 if (dep_list_size (insn, SD_LIST_FORW) == 0)
8961 {
8962 dep_def _new_dep, *new_dep = &_new_dep;
8963
8964 init_dep (new_dep, insn, jump, REG_DEP_ANTI);
8965 sd_add_dep (new_dep, false);
8966 }
8967 }
8968 while (1);
8969
8970 gcc_assert (!sd_lists_empty_p (jump, SD_LIST_BACK));
8971}
8972
8973/* Extend data structures for logical insn UID. */
8974void
8975sched_extend_luids (void)
8976{
8977 int new_luids_max_uid = get_max_uid () + 1;
8978
8979 sched_luids.safe_grow_cleared (len: new_luids_max_uid, exact: true);
8980}
8981
8982/* Initialize LUID for INSN. */
8983void
8984sched_init_insn_luid (rtx_insn *insn)
8985{
8986 int i = INSN_P (insn) ? 1 : common_sched_info->luid_for_non_insn (insn);
8987 int luid;
8988
8989 if (i >= 0)
8990 {
8991 luid = sched_max_luid;
8992 sched_max_luid += i;
8993 }
8994 else
8995 luid = -1;
8996
8997 SET_INSN_LUID (insn, luid);
8998}
8999
9000/* Initialize luids for BBS.
9001 The hook common_sched_info->luid_for_non_insn () is used to determine
9002 if notes, labels, etc. need luids. */
9003void
9004sched_init_luids (const bb_vec_t &bbs)
9005{
9006 int i;
9007 basic_block bb;
9008
9009 sched_extend_luids ();
9010 FOR_EACH_VEC_ELT (bbs, i, bb)
9011 {
9012 rtx_insn *insn;
9013
9014 FOR_BB_INSNS (bb, insn)
9015 sched_init_insn_luid (insn);
9016 }
9017}
9018
9019/* Free LUIDs. */
9020void
9021sched_finish_luids (void)
9022{
9023 sched_luids.release ();
9024 sched_max_luid = 1;
9025}
9026
9027/* Return logical uid of INSN. Helpful while debugging. */
9028int
9029insn_luid (rtx_insn *insn)
9030{
9031 return INSN_LUID (insn);
9032}
9033
9034/* Extend per insn data in the target. */
9035void
9036sched_extend_target (void)
9037{
9038 if (targetm.sched.h_i_d_extended)
9039 targetm.sched.h_i_d_extended ();
9040}
9041
9042/* Extend global scheduler structures (those, that live across calls to
9043 schedule_block) to include information about just emitted INSN. */
9044static void
9045extend_h_i_d (void)
9046{
9047 int reserve = (get_max_uid () + 1 - h_i_d.length ());
9048 if (reserve > 0
9049 && ! h_i_d.space (nelems: reserve))
9050 {
9051 h_i_d.safe_grow_cleared (len: 3U * get_max_uid () / 2, exact: true);
9052 sched_extend_target ();
9053 }
9054}
9055
9056/* Initialize h_i_d entry of the INSN with default values.
9057 Values, that are not explicitly initialized here, hold zero. */
9058static void
9059init_h_i_d (rtx_insn *insn)
9060{
9061 if (INSN_LUID (insn) > 0)
9062 {
9063 INSN_COST (insn) = -1;
9064 QUEUE_INDEX (insn) = QUEUE_NOWHERE;
9065 INSN_TICK (insn) = INVALID_TICK;
9066 INSN_EXACT_TICK (insn) = INVALID_TICK;
9067 INTER_TICK (insn) = INVALID_TICK;
9068 TODO_SPEC (insn) = HARD_DEP;
9069 INSN_AUTOPREF_MULTIPASS_DATA (insn)[0].status
9070 = AUTOPREF_MULTIPASS_DATA_UNINITIALIZED;
9071 INSN_AUTOPREF_MULTIPASS_DATA (insn)[1].status
9072 = AUTOPREF_MULTIPASS_DATA_UNINITIALIZED;
9073 }
9074}
9075
9076/* Initialize haifa_insn_data for BBS. */
9077void
9078haifa_init_h_i_d (const bb_vec_t &bbs)
9079{
9080 int i;
9081 basic_block bb;
9082
9083 extend_h_i_d ();
9084 FOR_EACH_VEC_ELT (bbs, i, bb)
9085 {
9086 rtx_insn *insn;
9087
9088 FOR_BB_INSNS (bb, insn)
9089 init_h_i_d (insn);
9090 }
9091}
9092
9093/* Finalize haifa_insn_data. */
9094void
9095haifa_finish_h_i_d (void)
9096{
9097 int i;
9098 haifa_insn_data_t data;
9099 reg_use_data *use, *next_use;
9100 reg_set_data *set, *next_set;
9101
9102 FOR_EACH_VEC_ELT (h_i_d, i, data)
9103 {
9104 free (ptr: data->max_reg_pressure);
9105 free (ptr: data->reg_pressure);
9106 for (use = data->reg_use_list; use != NULL; use = next_use)
9107 {
9108 next_use = use->next_insn_use;
9109 free (ptr: use);
9110 }
9111 for (set = data->reg_set_list; set != NULL; set = next_set)
9112 {
9113 next_set = set->next_insn_set;
9114 free (ptr: set);
9115 }
9116
9117 }
9118 h_i_d.release ();
9119}
9120
9121/* Init data for the new insn INSN. */
9122static void
9123haifa_init_insn (rtx_insn *insn)
9124{
9125 gcc_assert (insn != NULL);
9126
9127 sched_extend_luids ();
9128 sched_init_insn_luid (insn);
9129 sched_extend_target ();
9130 sched_deps_init (false);
9131 extend_h_i_d ();
9132 init_h_i_d (insn);
9133
9134 if (adding_bb_to_current_region_p)
9135 {
9136 sd_init_insn (insn);
9137
9138 /* Extend dependency caches by one element. */
9139 extend_dependency_caches (1, false);
9140 }
9141 if (sched_pressure != SCHED_PRESSURE_NONE)
9142 init_insn_reg_pressure_info (insn);
9143}
9144
9145/* Init data for the new basic block BB which comes after AFTER. */
9146static void
9147haifa_init_only_bb (basic_block bb, basic_block after)
9148{
9149 gcc_assert (bb != NULL);
9150
9151 sched_init_bbs ();
9152
9153 if (common_sched_info->add_block)
9154 /* This changes only data structures of the front-end. */
9155 common_sched_info->add_block (bb, after);
9156}
9157
9158/* A generic version of sched_split_block (). */
9159basic_block
9160sched_split_block_1 (basic_block first_bb, rtx after)
9161{
9162 edge e;
9163
9164 e = split_block (first_bb, after);
9165 gcc_assert (e->src == first_bb);
9166
9167 /* sched_split_block emits note if *check == BB_END. Probably it
9168 is better to rip that note off. */
9169
9170 return e->dest;
9171}
9172
9173/* A generic version of sched_create_empty_bb (). */
9174basic_block
9175sched_create_empty_bb_1 (basic_block after)
9176{
9177 return create_empty_bb (after);
9178}
9179
9180/* Insert PAT as an INSN into the schedule and update the necessary data
9181 structures to account for it. */
9182rtx_insn *
9183sched_emit_insn (rtx pat)
9184{
9185 rtx_insn *insn = emit_insn_before (pat, first_nonscheduled_insn ());
9186 haifa_init_insn (insn);
9187
9188 if (current_sched_info->add_remove_insn)
9189 current_sched_info->add_remove_insn (insn, 0);
9190
9191 (*current_sched_info->begin_schedule_ready) (insn);
9192 scheduled_insns.safe_push (obj: insn);
9193
9194 last_scheduled_insn = insn;
9195 return insn;
9196}
9197
9198/* This function returns a candidate satisfying dispatch constraints from
9199 the ready list. */
9200
9201static rtx_insn *
9202ready_remove_first_dispatch (struct ready_list *ready)
9203{
9204 int i;
9205 rtx_insn *insn = ready_element (ready, index: 0);
9206
9207 if (ready->n_ready == 1
9208 || !INSN_P (insn)
9209 || INSN_CODE (insn) < 0
9210 || !active_insn_p (insn)
9211 || targetm.sched.dispatch (insn, FITS_DISPATCH_WINDOW))
9212 return ready_remove_first (ready);
9213
9214 for (i = 1; i < ready->n_ready; i++)
9215 {
9216 insn = ready_element (ready, index: i);
9217
9218 if (!INSN_P (insn)
9219 || INSN_CODE (insn) < 0
9220 || !active_insn_p (insn))
9221 continue;
9222
9223 if (targetm.sched.dispatch (insn, FITS_DISPATCH_WINDOW))
9224 {
9225 /* Return ith element of ready. */
9226 insn = ready_remove (ready, index: i);
9227 return insn;
9228 }
9229 }
9230
9231 if (targetm.sched.dispatch (NULL, DISPATCH_VIOLATION))
9232 return ready_remove_first (ready);
9233
9234 for (i = 1; i < ready->n_ready; i++)
9235 {
9236 insn = ready_element (ready, index: i);
9237
9238 if (!INSN_P (insn)
9239 || INSN_CODE (insn) < 0
9240 || !active_insn_p (insn))
9241 continue;
9242
9243 /* Return i-th element of ready. */
9244 if (targetm.sched.dispatch (insn, IS_CMP))
9245 return ready_remove (ready, index: i);
9246 }
9247
9248 return ready_remove_first (ready);
9249}
9250
9251/* Get number of ready insn in the ready list. */
9252
9253int
9254number_in_ready (void)
9255{
9256 return ready.n_ready;
9257}
9258
9259/* Get number of ready's in the ready list. */
9260
9261rtx_insn *
9262get_ready_element (int i)
9263{
9264 return ready_element (ready: &ready, index: i);
9265}
9266
9267#endif /* INSN_SCHEDULING */
9268

source code of gcc/haifa-sched.cc