re PR target/57293 (not needed frame pointers on IA-32 (performance regression?))
[gcc.git] / gcc / haifa-sched.c
1 /* Instruction scheduling pass.
2 Copyright (C) 1992-2013 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
6 This file is part of GCC.
7
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
12
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
17
18 You should have received a copy of the GNU General Public License
19 along 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.c,
23 contains the generic parts. The actual entry point is found for
24 the normal instruction scheduling pass is found in sched-rgn.c.
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_backward_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 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 \f
125 #include "config.h"
126 #include "system.h"
127 #include "coretypes.h"
128 #include "tm.h"
129 #include "diagnostic-core.h"
130 #include "hard-reg-set.h"
131 #include "rtl.h"
132 #include "tm_p.h"
133 #include "regs.h"
134 #include "function.h"
135 #include "flags.h"
136 #include "insn-config.h"
137 #include "insn-attr.h"
138 #include "except.h"
139 #include "recog.h"
140 #include "sched-int.h"
141 #include "target.h"
142 #include "common/common-target.h"
143 #include "params.h"
144 #include "dbgcnt.h"
145 #include "cfgloop.h"
146 #include "ira.h"
147 #include "emit-rtl.h" /* FIXME: Can go away once crtl is moved to rtl.h. */
148 #include "hash-table.h"
149 #include "dumpfile.h"
150
151 #ifdef INSN_SCHEDULING
152
153 /* True if we do register pressure relief through live-range
154 shrinkage. */
155 static bool live_range_shrinkage_p;
156
157 /* Switch on live range shrinkage. */
158 void
159 initialize_live_range_shrinkage (void)
160 {
161 live_range_shrinkage_p = true;
162 }
163
164 /* Switch off live range shrinkage. */
165 void
166 finish_live_range_shrinkage (void)
167 {
168 live_range_shrinkage_p = false;
169 }
170
171 /* issue_rate is the number of insns that can be scheduled in the same
172 machine cycle. It can be defined in the config/mach/mach.h file,
173 otherwise we set it to 1. */
174
175 int issue_rate;
176
177 /* This can be set to true by a backend if the scheduler should not
178 enable a DCE pass. */
179 bool sched_no_dce;
180
181 /* The current initiation interval used when modulo scheduling. */
182 static int modulo_ii;
183
184 /* The maximum number of stages we are prepared to handle. */
185 static int modulo_max_stages;
186
187 /* The number of insns that exist in each iteration of the loop. We use this
188 to detect when we've scheduled all insns from the first iteration. */
189 static int modulo_n_insns;
190
191 /* The current count of insns in the first iteration of the loop that have
192 already been scheduled. */
193 static int modulo_insns_scheduled;
194
195 /* The maximum uid of insns from the first iteration of the loop. */
196 static int modulo_iter0_max_uid;
197
198 /* The number of times we should attempt to backtrack when modulo scheduling.
199 Decreased each time we have to backtrack. */
200 static int modulo_backtracks_left;
201
202 /* The stage in which the last insn from the original loop was
203 scheduled. */
204 static int modulo_last_stage;
205
206 /* sched-verbose controls the amount of debugging output the
207 scheduler prints. It is controlled by -fsched-verbose=N:
208 N>0 and no -DSR : the output is directed to stderr.
209 N>=10 will direct the printouts to stderr (regardless of -dSR).
210 N=1: same as -dSR.
211 N=2: bb's probabilities, detailed ready list info, unit/insn info.
212 N=3: rtl at abort point, control-flow, regions info.
213 N=5: dependences info. */
214
215 int sched_verbose = 0;
216
217 /* Debugging file. All printouts are sent to dump, which is always set,
218 either to stderr, or to the dump listing file (-dRS). */
219 FILE *sched_dump = 0;
220
221 /* This is a placeholder for the scheduler parameters common
222 to all schedulers. */
223 struct common_sched_info_def *common_sched_info;
224
225 #define INSN_TICK(INSN) (HID (INSN)->tick)
226 #define INSN_EXACT_TICK(INSN) (HID (INSN)->exact_tick)
227 #define INSN_TICK_ESTIMATE(INSN) (HID (INSN)->tick_estimate)
228 #define INTER_TICK(INSN) (HID (INSN)->inter_tick)
229 #define FEEDS_BACKTRACK_INSN(INSN) (HID (INSN)->feeds_backtrack_insn)
230 #define SHADOW_P(INSN) (HID (INSN)->shadow_p)
231 #define MUST_RECOMPUTE_SPEC_P(INSN) (HID (INSN)->must_recompute_spec)
232 /* Cached cost of the instruction. Use insn_cost to get cost of the
233 insn. -1 here means that the field is not initialized. */
234 #define INSN_COST(INSN) (HID (INSN)->cost)
235
236 /* If INSN_TICK of an instruction is equal to INVALID_TICK,
237 then it should be recalculated from scratch. */
238 #define INVALID_TICK (-(max_insn_queue_index + 1))
239 /* The minimal value of the INSN_TICK of an instruction. */
240 #define MIN_TICK (-max_insn_queue_index)
241
242 /* List of important notes we must keep around. This is a pointer to the
243 last element in the list. */
244 rtx note_list;
245
246 static struct spec_info_def spec_info_var;
247 /* Description of the speculative part of the scheduling.
248 If NULL - no speculation. */
249 spec_info_t spec_info = NULL;
250
251 /* True, if recovery block was added during scheduling of current block.
252 Used to determine, if we need to fix INSN_TICKs. */
253 static bool haifa_recovery_bb_recently_added_p;
254
255 /* True, if recovery block was added during this scheduling pass.
256 Used to determine if we should have empty memory pools of dependencies
257 after finishing current region. */
258 bool haifa_recovery_bb_ever_added_p;
259
260 /* Counters of different types of speculative instructions. */
261 static int nr_begin_data, nr_be_in_data, nr_begin_control, nr_be_in_control;
262
263 /* Array used in {unlink, restore}_bb_notes. */
264 static rtx *bb_header = 0;
265
266 /* Basic block after which recovery blocks will be created. */
267 static basic_block before_recovery;
268
269 /* Basic block just before the EXIT_BLOCK and after recovery, if we have
270 created it. */
271 basic_block after_recovery;
272
273 /* FALSE if we add bb to another region, so we don't need to initialize it. */
274 bool adding_bb_to_current_region_p = true;
275
276 /* Queues, etc. */
277
278 /* An instruction is ready to be scheduled when all insns preceding it
279 have already been scheduled. It is important to ensure that all
280 insns which use its result will not be executed until its result
281 has been computed. An insn is maintained in one of four structures:
282
283 (P) the "Pending" set of insns which cannot be scheduled until
284 their dependencies have been satisfied.
285 (Q) the "Queued" set of insns that can be scheduled when sufficient
286 time has passed.
287 (R) the "Ready" list of unscheduled, uncommitted insns.
288 (S) the "Scheduled" list of insns.
289
290 Initially, all insns are either "Pending" or "Ready" depending on
291 whether their dependencies are satisfied.
292
293 Insns move from the "Ready" list to the "Scheduled" list as they
294 are committed to the schedule. As this occurs, the insns in the
295 "Pending" list have their dependencies satisfied and move to either
296 the "Ready" list or the "Queued" set depending on whether
297 sufficient time has passed to make them ready. As time passes,
298 insns move from the "Queued" set to the "Ready" list.
299
300 The "Pending" list (P) are the insns in the INSN_FORW_DEPS of the
301 unscheduled insns, i.e., those that are ready, queued, and pending.
302 The "Queued" set (Q) is implemented by the variable `insn_queue'.
303 The "Ready" list (R) is implemented by the variables `ready' and
304 `n_ready'.
305 The "Scheduled" list (S) is the new insn chain built by this pass.
306
307 The transition (R->S) is implemented in the scheduling loop in
308 `schedule_block' when the best insn to schedule is chosen.
309 The transitions (P->R and P->Q) are implemented in `schedule_insn' as
310 insns move from the ready list to the scheduled list.
311 The transition (Q->R) is implemented in 'queue_to_insn' as time
312 passes or stalls are introduced. */
313
314 /* Implement a circular buffer to delay instructions until sufficient
315 time has passed. For the new pipeline description interface,
316 MAX_INSN_QUEUE_INDEX is a power of two minus one which is not less
317 than maximal time of instruction execution computed by genattr.c on
318 the base maximal time of functional unit reservations and getting a
319 result. This is the longest time an insn may be queued. */
320
321 static rtx *insn_queue;
322 static int q_ptr = 0;
323 static int q_size = 0;
324 #define NEXT_Q(X) (((X)+1) & max_insn_queue_index)
325 #define NEXT_Q_AFTER(X, C) (((X)+C) & max_insn_queue_index)
326
327 #define QUEUE_SCHEDULED (-3)
328 #define QUEUE_NOWHERE (-2)
329 #define QUEUE_READY (-1)
330 /* QUEUE_SCHEDULED - INSN is scheduled.
331 QUEUE_NOWHERE - INSN isn't scheduled yet and is neither in
332 queue or ready list.
333 QUEUE_READY - INSN is in ready list.
334 N >= 0 - INSN queued for X [where NEXT_Q_AFTER (q_ptr, X) == N] cycles. */
335
336 #define QUEUE_INDEX(INSN) (HID (INSN)->queue_index)
337
338 /* The following variable value refers for all current and future
339 reservations of the processor units. */
340 state_t curr_state;
341
342 /* The following variable value is size of memory representing all
343 current and future reservations of the processor units. */
344 size_t dfa_state_size;
345
346 /* The following array is used to find the best insn from ready when
347 the automaton pipeline interface is used. */
348 char *ready_try = NULL;
349
350 /* The ready list. */
351 struct ready_list ready = {NULL, 0, 0, 0, 0};
352
353 /* The pointer to the ready list (to be removed). */
354 static struct ready_list *readyp = &ready;
355
356 /* Scheduling clock. */
357 static int clock_var;
358
359 /* Clock at which the previous instruction was issued. */
360 static int last_clock_var;
361
362 /* Set to true if, when queuing a shadow insn, we discover that it would be
363 scheduled too late. */
364 static bool must_backtrack;
365
366 /* The following variable value is number of essential insns issued on
367 the current cycle. An insn is essential one if it changes the
368 processors state. */
369 int cycle_issued_insns;
370
371 /* This records the actual schedule. It is built up during the main phase
372 of schedule_block, and afterwards used to reorder the insns in the RTL. */
373 static vec<rtx> scheduled_insns;
374
375 static int may_trap_exp (const_rtx, int);
376
377 /* Nonzero iff the address is comprised from at most 1 register. */
378 #define CONST_BASED_ADDRESS_P(x) \
379 (REG_P (x) \
380 || ((GET_CODE (x) == PLUS || GET_CODE (x) == MINUS \
381 || (GET_CODE (x) == LO_SUM)) \
382 && (CONSTANT_P (XEXP (x, 0)) \
383 || CONSTANT_P (XEXP (x, 1)))))
384
385 /* Returns a class that insn with GET_DEST(insn)=x may belong to,
386 as found by analyzing insn's expression. */
387
388 \f
389 static int haifa_luid_for_non_insn (rtx x);
390
391 /* Haifa version of sched_info hooks common to all headers. */
392 const struct common_sched_info_def haifa_common_sched_info =
393 {
394 NULL, /* fix_recovery_cfg */
395 NULL, /* add_block */
396 NULL, /* estimate_number_of_insns */
397 haifa_luid_for_non_insn, /* luid_for_non_insn */
398 SCHED_PASS_UNKNOWN /* sched_pass_id */
399 };
400
401 /* Mapping from instruction UID to its Logical UID. */
402 vec<int> sched_luids = vNULL;
403
404 /* Next LUID to assign to an instruction. */
405 int sched_max_luid = 1;
406
407 /* Haifa Instruction Data. */
408 vec<haifa_insn_data_def> h_i_d = vNULL;
409
410 void (* sched_init_only_bb) (basic_block, basic_block);
411
412 /* Split block function. Different schedulers might use different functions
413 to handle their internal data consistent. */
414 basic_block (* sched_split_block) (basic_block, rtx);
415
416 /* Create empty basic block after the specified block. */
417 basic_block (* sched_create_empty_bb) (basic_block);
418
419 /* Return the number of cycles until INSN is expected to be ready.
420 Return zero if it already is. */
421 static int
422 insn_delay (rtx insn)
423 {
424 return MAX (INSN_TICK (insn) - clock_var, 0);
425 }
426
427 static int
428 may_trap_exp (const_rtx x, int is_store)
429 {
430 enum rtx_code code;
431
432 if (x == 0)
433 return TRAP_FREE;
434 code = GET_CODE (x);
435 if (is_store)
436 {
437 if (code == MEM && may_trap_p (x))
438 return TRAP_RISKY;
439 else
440 return TRAP_FREE;
441 }
442 if (code == MEM)
443 {
444 /* The insn uses memory: a volatile load. */
445 if (MEM_VOLATILE_P (x))
446 return IRISKY;
447 /* An exception-free load. */
448 if (!may_trap_p (x))
449 return IFREE;
450 /* A load with 1 base register, to be further checked. */
451 if (CONST_BASED_ADDRESS_P (XEXP (x, 0)))
452 return PFREE_CANDIDATE;
453 /* No info on the load, to be further checked. */
454 return PRISKY_CANDIDATE;
455 }
456 else
457 {
458 const char *fmt;
459 int i, insn_class = TRAP_FREE;
460
461 /* Neither store nor load, check if it may cause a trap. */
462 if (may_trap_p (x))
463 return TRAP_RISKY;
464 /* Recursive step: walk the insn... */
465 fmt = GET_RTX_FORMAT (code);
466 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
467 {
468 if (fmt[i] == 'e')
469 {
470 int tmp_class = may_trap_exp (XEXP (x, i), is_store);
471 insn_class = WORST_CLASS (insn_class, tmp_class);
472 }
473 else if (fmt[i] == 'E')
474 {
475 int j;
476 for (j = 0; j < XVECLEN (x, i); j++)
477 {
478 int tmp_class = may_trap_exp (XVECEXP (x, i, j), is_store);
479 insn_class = WORST_CLASS (insn_class, tmp_class);
480 if (insn_class == TRAP_RISKY || insn_class == IRISKY)
481 break;
482 }
483 }
484 if (insn_class == TRAP_RISKY || insn_class == IRISKY)
485 break;
486 }
487 return insn_class;
488 }
489 }
490
491 /* Classifies rtx X of an insn for the purpose of verifying that X can be
492 executed speculatively (and consequently the insn can be moved
493 speculatively), by examining X, returning:
494 TRAP_RISKY: store, or risky non-load insn (e.g. division by variable).
495 TRAP_FREE: non-load insn.
496 IFREE: load from a globally safe location.
497 IRISKY: volatile load.
498 PFREE_CANDIDATE, PRISKY_CANDIDATE: load that need to be checked for
499 being either PFREE or PRISKY. */
500
501 static int
502 haifa_classify_rtx (const_rtx x)
503 {
504 int tmp_class = TRAP_FREE;
505 int insn_class = TRAP_FREE;
506 enum rtx_code code;
507
508 if (GET_CODE (x) == PARALLEL)
509 {
510 int i, len = XVECLEN (x, 0);
511
512 for (i = len - 1; i >= 0; i--)
513 {
514 tmp_class = haifa_classify_rtx (XVECEXP (x, 0, i));
515 insn_class = WORST_CLASS (insn_class, tmp_class);
516 if (insn_class == TRAP_RISKY || insn_class == IRISKY)
517 break;
518 }
519 }
520 else
521 {
522 code = GET_CODE (x);
523 switch (code)
524 {
525 case CLOBBER:
526 /* Test if it is a 'store'. */
527 tmp_class = may_trap_exp (XEXP (x, 0), 1);
528 break;
529 case SET:
530 /* Test if it is a store. */
531 tmp_class = may_trap_exp (SET_DEST (x), 1);
532 if (tmp_class == TRAP_RISKY)
533 break;
534 /* Test if it is a load. */
535 tmp_class =
536 WORST_CLASS (tmp_class,
537 may_trap_exp (SET_SRC (x), 0));
538 break;
539 case COND_EXEC:
540 tmp_class = haifa_classify_rtx (COND_EXEC_CODE (x));
541 if (tmp_class == TRAP_RISKY)
542 break;
543 tmp_class = WORST_CLASS (tmp_class,
544 may_trap_exp (COND_EXEC_TEST (x), 0));
545 break;
546 case TRAP_IF:
547 tmp_class = TRAP_RISKY;
548 break;
549 default:;
550 }
551 insn_class = tmp_class;
552 }
553
554 return insn_class;
555 }
556
557 int
558 haifa_classify_insn (const_rtx insn)
559 {
560 return haifa_classify_rtx (PATTERN (insn));
561 }
562 \f
563 /* After the scheduler initialization function has been called, this function
564 can be called to enable modulo scheduling. II is the initiation interval
565 we should use, it affects the delays for delay_pairs that were recorded as
566 separated by a given number of stages.
567
568 MAX_STAGES provides us with a limit
569 after which we give up scheduling; the caller must have unrolled at least
570 as many copies of the loop body and recorded delay_pairs for them.
571
572 INSNS is the number of real (non-debug) insns in one iteration of
573 the loop. MAX_UID can be used to test whether an insn belongs to
574 the first iteration of the loop; all of them have a uid lower than
575 MAX_UID. */
576 void
577 set_modulo_params (int ii, int max_stages, int insns, int max_uid)
578 {
579 modulo_ii = ii;
580 modulo_max_stages = max_stages;
581 modulo_n_insns = insns;
582 modulo_iter0_max_uid = max_uid;
583 modulo_backtracks_left = PARAM_VALUE (PARAM_MAX_MODULO_BACKTRACK_ATTEMPTS);
584 }
585
586 /* A structure to record a pair of insns where the first one is a real
587 insn that has delay slots, and the second is its delayed shadow.
588 I1 is scheduled normally and will emit an assembly instruction,
589 while I2 describes the side effect that takes place at the
590 transition between cycles CYCLES and (CYCLES + 1) after I1. */
591 struct delay_pair
592 {
593 struct delay_pair *next_same_i1;
594 rtx i1, i2;
595 int cycles;
596 /* When doing modulo scheduling, we a delay_pair can also be used to
597 show that I1 and I2 are the same insn in a different stage. If that
598 is the case, STAGES will be nonzero. */
599 int stages;
600 };
601
602 /* Helpers for delay hashing. */
603
604 struct delay_i1_hasher : typed_noop_remove <delay_pair>
605 {
606 typedef delay_pair value_type;
607 typedef void compare_type;
608 static inline hashval_t hash (const value_type *);
609 static inline bool equal (const value_type *, const compare_type *);
610 };
611
612 /* Returns a hash value for X, based on hashing just I1. */
613
614 inline hashval_t
615 delay_i1_hasher::hash (const value_type *x)
616 {
617 return htab_hash_pointer (x->i1);
618 }
619
620 /* Return true if I1 of pair X is the same as that of pair Y. */
621
622 inline bool
623 delay_i1_hasher::equal (const value_type *x, const compare_type *y)
624 {
625 return x->i1 == y;
626 }
627
628 struct delay_i2_hasher : typed_free_remove <delay_pair>
629 {
630 typedef delay_pair value_type;
631 typedef void compare_type;
632 static inline hashval_t hash (const value_type *);
633 static inline bool equal (const value_type *, const compare_type *);
634 };
635
636 /* Returns a hash value for X, based on hashing just I2. */
637
638 inline hashval_t
639 delay_i2_hasher::hash (const value_type *x)
640 {
641 return htab_hash_pointer (x->i2);
642 }
643
644 /* Return true if I2 of pair X is the same as that of pair Y. */
645
646 inline bool
647 delay_i2_hasher::equal (const value_type *x, const compare_type *y)
648 {
649 return x->i2 == y;
650 }
651
652 /* Two hash tables to record delay_pairs, one indexed by I1 and the other
653 indexed by I2. */
654 static hash_table <delay_i1_hasher> delay_htab;
655 static hash_table <delay_i2_hasher> delay_htab_i2;
656
657 /* Called through htab_traverse. Walk the hashtable using I2 as
658 index, and delete all elements involving an UID higher than
659 that pointed to by *DATA. */
660 int
661 haifa_htab_i2_traverse (delay_pair **slot, int *data)
662 {
663 int maxuid = *data;
664 struct delay_pair *p = *slot;
665 if (INSN_UID (p->i2) >= maxuid || INSN_UID (p->i1) >= maxuid)
666 {
667 delay_htab_i2.clear_slot (slot);
668 }
669 return 1;
670 }
671
672 /* Called through htab_traverse. Walk the hashtable using I2 as
673 index, and delete all elements involving an UID higher than
674 that pointed to by *DATA. */
675 int
676 haifa_htab_i1_traverse (delay_pair **pslot, int *data)
677 {
678 int maxuid = *data;
679 struct delay_pair *p, *first, **pprev;
680
681 if (INSN_UID ((*pslot)->i1) >= maxuid)
682 {
683 delay_htab.clear_slot (pslot);
684 return 1;
685 }
686 pprev = &first;
687 for (p = *pslot; p; p = p->next_same_i1)
688 {
689 if (INSN_UID (p->i2) < maxuid)
690 {
691 *pprev = p;
692 pprev = &p->next_same_i1;
693 }
694 }
695 *pprev = NULL;
696 if (first == NULL)
697 delay_htab.clear_slot (pslot);
698 else
699 *pslot = first;
700 return 1;
701 }
702
703 /* Discard all delay pairs which involve an insn with an UID higher
704 than MAX_UID. */
705 void
706 discard_delay_pairs_above (int max_uid)
707 {
708 delay_htab.traverse <int *, haifa_htab_i1_traverse> (&max_uid);
709 delay_htab_i2.traverse <int *, haifa_htab_i2_traverse> (&max_uid);
710 }
711
712 /* This function can be called by a port just before it starts the final
713 scheduling pass. It records the fact that an instruction with delay
714 slots has been split into two insns, I1 and I2. The first one will be
715 scheduled normally and initiates the operation. The second one is a
716 shadow which must follow a specific number of cycles after I1; its only
717 purpose is to show the side effect that occurs at that cycle in the RTL.
718 If a JUMP_INSN or a CALL_INSN has been split, I1 should be a normal INSN,
719 while I2 retains the original insn type.
720
721 There are two ways in which the number of cycles can be specified,
722 involving the CYCLES and STAGES arguments to this function. If STAGES
723 is zero, we just use the value of CYCLES. Otherwise, STAGES is a factor
724 which is multiplied by MODULO_II to give the number of cycles. This is
725 only useful if the caller also calls set_modulo_params to enable modulo
726 scheduling. */
727
728 void
729 record_delay_slot_pair (rtx i1, rtx i2, int cycles, int stages)
730 {
731 struct delay_pair *p = XNEW (struct delay_pair);
732 struct delay_pair **slot;
733
734 p->i1 = i1;
735 p->i2 = i2;
736 p->cycles = cycles;
737 p->stages = stages;
738
739 if (!delay_htab.is_created ())
740 {
741 delay_htab.create (10);
742 delay_htab_i2.create (10);
743 }
744 slot = delay_htab.find_slot_with_hash (i1, htab_hash_pointer (i1), INSERT);
745 p->next_same_i1 = *slot;
746 *slot = p;
747 slot = delay_htab_i2.find_slot_with_hash (i2, htab_hash_pointer (i2), INSERT);
748 *slot = p;
749 }
750
751 /* Examine the delay pair hashtable to see if INSN is a shadow for another,
752 and return the other insn if so. Return NULL otherwise. */
753 rtx
754 real_insn_for_shadow (rtx insn)
755 {
756 struct delay_pair *pair;
757
758 if (!delay_htab.is_created ())
759 return NULL_RTX;
760
761 pair = delay_htab_i2.find_with_hash (insn, htab_hash_pointer (insn));
762 if (!pair || pair->stages > 0)
763 return NULL_RTX;
764 return pair->i1;
765 }
766
767 /* For a pair P of insns, return the fixed distance in cycles from the first
768 insn after which the second must be scheduled. */
769 static int
770 pair_delay (struct delay_pair *p)
771 {
772 if (p->stages == 0)
773 return p->cycles;
774 else
775 return p->stages * modulo_ii;
776 }
777
778 /* Given an insn INSN, add a dependence on its delayed shadow if it
779 has one. Also try to find situations where shadows depend on each other
780 and add dependencies to the real insns to limit the amount of backtracking
781 needed. */
782 void
783 add_delay_dependencies (rtx insn)
784 {
785 struct delay_pair *pair;
786 sd_iterator_def sd_it;
787 dep_t dep;
788
789 if (!delay_htab.is_created ())
790 return;
791
792 pair = delay_htab_i2.find_with_hash (insn, htab_hash_pointer (insn));
793 if (!pair)
794 return;
795 add_dependence (insn, pair->i1, REG_DEP_ANTI);
796 if (pair->stages)
797 return;
798
799 FOR_EACH_DEP (pair->i2, SD_LIST_BACK, sd_it, dep)
800 {
801 rtx pro = DEP_PRO (dep);
802 struct delay_pair *other_pair
803 = delay_htab_i2.find_with_hash (pro, htab_hash_pointer (pro));
804 if (!other_pair || other_pair->stages)
805 continue;
806 if (pair_delay (other_pair) >= pair_delay (pair))
807 {
808 if (sched_verbose >= 4)
809 {
810 fprintf (sched_dump, ";;\tadding dependence %d <- %d\n",
811 INSN_UID (other_pair->i1),
812 INSN_UID (pair->i1));
813 fprintf (sched_dump, ";;\tpair1 %d <- %d, cost %d\n",
814 INSN_UID (pair->i1),
815 INSN_UID (pair->i2),
816 pair_delay (pair));
817 fprintf (sched_dump, ";;\tpair2 %d <- %d, cost %d\n",
818 INSN_UID (other_pair->i1),
819 INSN_UID (other_pair->i2),
820 pair_delay (other_pair));
821 }
822 add_dependence (pair->i1, other_pair->i1, REG_DEP_ANTI);
823 }
824 }
825 }
826 \f
827 /* Forward declarations. */
828
829 static int priority (rtx);
830 static int rank_for_schedule (const void *, const void *);
831 static void swap_sort (rtx *, int);
832 static void queue_insn (rtx, int, const char *);
833 static int schedule_insn (rtx);
834 static void adjust_priority (rtx);
835 static void advance_one_cycle (void);
836 static void extend_h_i_d (void);
837
838
839 /* Notes handling mechanism:
840 =========================
841 Generally, NOTES are saved before scheduling and restored after scheduling.
842 The scheduler distinguishes between two types of notes:
843
844 (1) LOOP_BEGIN, LOOP_END, SETJMP, EHREGION_BEG, EHREGION_END notes:
845 Before scheduling a region, a pointer to the note is added to the insn
846 that follows or precedes it. (This happens as part of the data dependence
847 computation). After scheduling an insn, the pointer contained in it is
848 used for regenerating the corresponding note (in reemit_notes).
849
850 (2) All other notes (e.g. INSN_DELETED): Before scheduling a block,
851 these notes are put in a list (in rm_other_notes() and
852 unlink_other_notes ()). After scheduling the block, these notes are
853 inserted at the beginning of the block (in schedule_block()). */
854
855 static void ready_add (struct ready_list *, rtx, bool);
856 static rtx ready_remove_first (struct ready_list *);
857 static rtx ready_remove_first_dispatch (struct ready_list *ready);
858
859 static void queue_to_ready (struct ready_list *);
860 static int early_queue_to_ready (state_t, struct ready_list *);
861
862 static void debug_ready_list (struct ready_list *);
863
864 /* The following functions are used to implement multi-pass scheduling
865 on the first cycle. */
866 static rtx ready_remove (struct ready_list *, int);
867 static void ready_remove_insn (rtx);
868
869 static void fix_inter_tick (rtx, rtx);
870 static int fix_tick_ready (rtx);
871 static void change_queue_index (rtx, int);
872
873 /* The following functions are used to implement scheduling of data/control
874 speculative instructions. */
875
876 static void extend_h_i_d (void);
877 static void init_h_i_d (rtx);
878 static int haifa_speculate_insn (rtx, ds_t, rtx *);
879 static void generate_recovery_code (rtx);
880 static void process_insn_forw_deps_be_in_spec (rtx, rtx, ds_t);
881 static void begin_speculative_block (rtx);
882 static void add_to_speculative_block (rtx);
883 static void init_before_recovery (basic_block *);
884 static void create_check_block_twin (rtx, bool);
885 static void fix_recovery_deps (basic_block);
886 static bool haifa_change_pattern (rtx, rtx);
887 static void dump_new_block_header (int, basic_block, rtx, rtx);
888 static void restore_bb_notes (basic_block);
889 static void fix_jump_move (rtx);
890 static void move_block_after_check (rtx);
891 static void move_succs (vec<edge, va_gc> **, basic_block);
892 static void sched_remove_insn (rtx);
893 static void clear_priorities (rtx, rtx_vec_t *);
894 static void calc_priorities (rtx_vec_t);
895 static void add_jump_dependencies (rtx, rtx);
896
897 #endif /* INSN_SCHEDULING */
898 \f
899 /* Point to state used for the current scheduling pass. */
900 struct haifa_sched_info *current_sched_info;
901 \f
902 #ifndef INSN_SCHEDULING
903 void
904 schedule_insns (void)
905 {
906 }
907 #else
908
909 /* Do register pressure sensitive insn scheduling if the flag is set
910 up. */
911 enum sched_pressure_algorithm sched_pressure;
912
913 /* Map regno -> its pressure class. The map defined only when
914 SCHED_PRESSURE != SCHED_PRESSURE_NONE. */
915 enum reg_class *sched_regno_pressure_class;
916
917 /* The current register pressure. Only elements corresponding pressure
918 classes are defined. */
919 static int curr_reg_pressure[N_REG_CLASSES];
920
921 /* Saved value of the previous array. */
922 static int saved_reg_pressure[N_REG_CLASSES];
923
924 /* Register living at given scheduling point. */
925 static bitmap curr_reg_live;
926
927 /* Saved value of the previous array. */
928 static bitmap saved_reg_live;
929
930 /* Registers mentioned in the current region. */
931 static bitmap region_ref_regs;
932
933 /* Initiate register pressure relative info for scheduling the current
934 region. Currently it is only clearing register mentioned in the
935 current region. */
936 void
937 sched_init_region_reg_pressure_info (void)
938 {
939 bitmap_clear (region_ref_regs);
940 }
941
942 /* PRESSURE[CL] describes the pressure on register class CL. Update it
943 for the birth (if BIRTH_P) or death (if !BIRTH_P) of register REGNO.
944 LIVE tracks the set of live registers; if it is null, assume that
945 every birth or death is genuine. */
946 static inline void
947 mark_regno_birth_or_death (bitmap live, int *pressure, int regno, bool birth_p)
948 {
949 enum reg_class pressure_class;
950
951 pressure_class = sched_regno_pressure_class[regno];
952 if (regno >= FIRST_PSEUDO_REGISTER)
953 {
954 if (pressure_class != NO_REGS)
955 {
956 if (birth_p)
957 {
958 if (!live || bitmap_set_bit (live, regno))
959 pressure[pressure_class]
960 += (ira_reg_class_max_nregs
961 [pressure_class][PSEUDO_REGNO_MODE (regno)]);
962 }
963 else
964 {
965 if (!live || bitmap_clear_bit (live, regno))
966 pressure[pressure_class]
967 -= (ira_reg_class_max_nregs
968 [pressure_class][PSEUDO_REGNO_MODE (regno)]);
969 }
970 }
971 }
972 else if (pressure_class != NO_REGS
973 && ! TEST_HARD_REG_BIT (ira_no_alloc_regs, regno))
974 {
975 if (birth_p)
976 {
977 if (!live || bitmap_set_bit (live, regno))
978 pressure[pressure_class]++;
979 }
980 else
981 {
982 if (!live || bitmap_clear_bit (live, regno))
983 pressure[pressure_class]--;
984 }
985 }
986 }
987
988 /* Initiate current register pressure related info from living
989 registers given by LIVE. */
990 static void
991 initiate_reg_pressure_info (bitmap live)
992 {
993 int i;
994 unsigned int j;
995 bitmap_iterator bi;
996
997 for (i = 0; i < ira_pressure_classes_num; i++)
998 curr_reg_pressure[ira_pressure_classes[i]] = 0;
999 bitmap_clear (curr_reg_live);
1000 EXECUTE_IF_SET_IN_BITMAP (live, 0, j, bi)
1001 if (sched_pressure == SCHED_PRESSURE_MODEL
1002 || current_nr_blocks == 1
1003 || bitmap_bit_p (region_ref_regs, j))
1004 mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure, j, true);
1005 }
1006
1007 /* Mark registers in X as mentioned in the current region. */
1008 static void
1009 setup_ref_regs (rtx x)
1010 {
1011 int i, j, regno;
1012 const RTX_CODE code = GET_CODE (x);
1013 const char *fmt;
1014
1015 if (REG_P (x))
1016 {
1017 regno = REGNO (x);
1018 if (HARD_REGISTER_NUM_P (regno))
1019 bitmap_set_range (region_ref_regs, regno,
1020 hard_regno_nregs[regno][GET_MODE (x)]);
1021 else
1022 bitmap_set_bit (region_ref_regs, REGNO (x));
1023 return;
1024 }
1025 fmt = GET_RTX_FORMAT (code);
1026 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1027 if (fmt[i] == 'e')
1028 setup_ref_regs (XEXP (x, i));
1029 else if (fmt[i] == 'E')
1030 {
1031 for (j = 0; j < XVECLEN (x, i); j++)
1032 setup_ref_regs (XVECEXP (x, i, j));
1033 }
1034 }
1035
1036 /* Initiate current register pressure related info at the start of
1037 basic block BB. */
1038 static void
1039 initiate_bb_reg_pressure_info (basic_block bb)
1040 {
1041 unsigned int i ATTRIBUTE_UNUSED;
1042 rtx insn;
1043
1044 if (current_nr_blocks > 1)
1045 FOR_BB_INSNS (bb, insn)
1046 if (NONDEBUG_INSN_P (insn))
1047 setup_ref_regs (PATTERN (insn));
1048 initiate_reg_pressure_info (df_get_live_in (bb));
1049 #ifdef EH_RETURN_DATA_REGNO
1050 if (bb_has_eh_pred (bb))
1051 for (i = 0; ; ++i)
1052 {
1053 unsigned int regno = EH_RETURN_DATA_REGNO (i);
1054
1055 if (regno == INVALID_REGNUM)
1056 break;
1057 if (! bitmap_bit_p (df_get_live_in (bb), regno))
1058 mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure,
1059 regno, true);
1060 }
1061 #endif
1062 }
1063
1064 /* Save current register pressure related info. */
1065 static void
1066 save_reg_pressure (void)
1067 {
1068 int i;
1069
1070 for (i = 0; i < ira_pressure_classes_num; i++)
1071 saved_reg_pressure[ira_pressure_classes[i]]
1072 = curr_reg_pressure[ira_pressure_classes[i]];
1073 bitmap_copy (saved_reg_live, curr_reg_live);
1074 }
1075
1076 /* Restore saved register pressure related info. */
1077 static void
1078 restore_reg_pressure (void)
1079 {
1080 int i;
1081
1082 for (i = 0; i < ira_pressure_classes_num; i++)
1083 curr_reg_pressure[ira_pressure_classes[i]]
1084 = saved_reg_pressure[ira_pressure_classes[i]];
1085 bitmap_copy (curr_reg_live, saved_reg_live);
1086 }
1087
1088 /* Return TRUE if the register is dying after its USE. */
1089 static bool
1090 dying_use_p (struct reg_use_data *use)
1091 {
1092 struct reg_use_data *next;
1093
1094 for (next = use->next_regno_use; next != use; next = next->next_regno_use)
1095 if (NONDEBUG_INSN_P (next->insn)
1096 && QUEUE_INDEX (next->insn) != QUEUE_SCHEDULED)
1097 return false;
1098 return true;
1099 }
1100
1101 /* Print info about the current register pressure and its excess for
1102 each pressure class. */
1103 static void
1104 print_curr_reg_pressure (void)
1105 {
1106 int i;
1107 enum reg_class cl;
1108
1109 fprintf (sched_dump, ";;\t");
1110 for (i = 0; i < ira_pressure_classes_num; i++)
1111 {
1112 cl = ira_pressure_classes[i];
1113 gcc_assert (curr_reg_pressure[cl] >= 0);
1114 fprintf (sched_dump, " %s:%d(%d)", reg_class_names[cl],
1115 curr_reg_pressure[cl],
1116 curr_reg_pressure[cl] - ira_class_hard_regs_num[cl]);
1117 }
1118 fprintf (sched_dump, "\n");
1119 }
1120 \f
1121 /* Determine if INSN has a condition that is clobbered if a register
1122 in SET_REGS is modified. */
1123 static bool
1124 cond_clobbered_p (rtx insn, HARD_REG_SET set_regs)
1125 {
1126 rtx pat = PATTERN (insn);
1127 gcc_assert (GET_CODE (pat) == COND_EXEC);
1128 if (TEST_HARD_REG_BIT (set_regs, REGNO (XEXP (COND_EXEC_TEST (pat), 0))))
1129 {
1130 sd_iterator_def sd_it;
1131 dep_t dep;
1132 haifa_change_pattern (insn, ORIG_PAT (insn));
1133 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
1134 DEP_STATUS (dep) &= ~DEP_CANCELLED;
1135 TODO_SPEC (insn) = HARD_DEP;
1136 if (sched_verbose >= 2)
1137 fprintf (sched_dump,
1138 ";;\t\tdequeue insn %s because of clobbered condition\n",
1139 (*current_sched_info->print_insn) (insn, 0));
1140 return true;
1141 }
1142
1143 return false;
1144 }
1145
1146 /* This function should be called after modifying the pattern of INSN,
1147 to update scheduler data structures as needed. */
1148 static void
1149 update_insn_after_change (rtx insn)
1150 {
1151 sd_iterator_def sd_it;
1152 dep_t dep;
1153
1154 dfa_clear_single_insn_cache (insn);
1155
1156 sd_it = sd_iterator_start (insn,
1157 SD_LIST_FORW | SD_LIST_BACK | SD_LIST_RES_BACK);
1158 while (sd_iterator_cond (&sd_it, &dep))
1159 {
1160 DEP_COST (dep) = UNKNOWN_DEP_COST;
1161 sd_iterator_next (&sd_it);
1162 }
1163
1164 /* Invalidate INSN_COST, so it'll be recalculated. */
1165 INSN_COST (insn) = -1;
1166 /* Invalidate INSN_TICK, so it'll be recalculated. */
1167 INSN_TICK (insn) = INVALID_TICK;
1168 }
1169
1170
1171 /* Two VECs, one to hold dependencies for which pattern replacements
1172 need to be applied or restored at the start of the next cycle, and
1173 another to hold an integer that is either one, to apply the
1174 corresponding replacement, or zero to restore it. */
1175 static vec<dep_t> next_cycle_replace_deps;
1176 static vec<int> next_cycle_apply;
1177
1178 static void apply_replacement (dep_t, bool);
1179 static void restore_pattern (dep_t, bool);
1180
1181 /* Look at the remaining dependencies for insn NEXT, and compute and return
1182 the TODO_SPEC value we should use for it. This is called after one of
1183 NEXT's dependencies has been resolved.
1184 We also perform pattern replacements for predication, and for broken
1185 replacement dependencies. The latter is only done if FOR_BACKTRACK is
1186 false. */
1187
1188 static ds_t
1189 recompute_todo_spec (rtx next, bool for_backtrack)
1190 {
1191 ds_t new_ds;
1192 sd_iterator_def sd_it;
1193 dep_t dep, modify_dep = NULL;
1194 int n_spec = 0;
1195 int n_control = 0;
1196 int n_replace = 0;
1197 bool first_p = true;
1198
1199 if (sd_lists_empty_p (next, SD_LIST_BACK))
1200 /* NEXT has all its dependencies resolved. */
1201 return 0;
1202
1203 if (!sd_lists_empty_p (next, SD_LIST_HARD_BACK))
1204 return HARD_DEP;
1205
1206 /* Now we've got NEXT with speculative deps only.
1207 1. Look at the deps to see what we have to do.
1208 2. Check if we can do 'todo'. */
1209 new_ds = 0;
1210
1211 FOR_EACH_DEP (next, SD_LIST_BACK, sd_it, dep)
1212 {
1213 rtx pro = DEP_PRO (dep);
1214 ds_t ds = DEP_STATUS (dep) & SPECULATIVE;
1215
1216 if (DEBUG_INSN_P (pro) && !DEBUG_INSN_P (next))
1217 continue;
1218
1219 if (ds)
1220 {
1221 n_spec++;
1222 if (first_p)
1223 {
1224 first_p = false;
1225
1226 new_ds = ds;
1227 }
1228 else
1229 new_ds = ds_merge (new_ds, ds);
1230 }
1231 else if (DEP_TYPE (dep) == REG_DEP_CONTROL)
1232 {
1233 if (QUEUE_INDEX (pro) != QUEUE_SCHEDULED)
1234 {
1235 n_control++;
1236 modify_dep = dep;
1237 }
1238 DEP_STATUS (dep) &= ~DEP_CANCELLED;
1239 }
1240 else if (DEP_REPLACE (dep) != NULL)
1241 {
1242 if (QUEUE_INDEX (pro) != QUEUE_SCHEDULED)
1243 {
1244 n_replace++;
1245 modify_dep = dep;
1246 }
1247 DEP_STATUS (dep) &= ~DEP_CANCELLED;
1248 }
1249 }
1250
1251 if (n_replace > 0 && n_control == 0 && n_spec == 0)
1252 {
1253 if (!dbg_cnt (sched_breakdep))
1254 return HARD_DEP;
1255 FOR_EACH_DEP (next, SD_LIST_BACK, sd_it, dep)
1256 {
1257 struct dep_replacement *desc = DEP_REPLACE (dep);
1258 if (desc != NULL)
1259 {
1260 if (desc->insn == next && !for_backtrack)
1261 {
1262 gcc_assert (n_replace == 1);
1263 apply_replacement (dep, true);
1264 }
1265 DEP_STATUS (dep) |= DEP_CANCELLED;
1266 }
1267 }
1268 return 0;
1269 }
1270
1271 else if (n_control == 1 && n_replace == 0 && n_spec == 0)
1272 {
1273 rtx pro, other, new_pat;
1274 rtx cond = NULL_RTX;
1275 bool success;
1276 rtx prev = NULL_RTX;
1277 int i;
1278 unsigned regno;
1279
1280 if ((current_sched_info->flags & DO_PREDICATION) == 0
1281 || (ORIG_PAT (next) != NULL_RTX
1282 && PREDICATED_PAT (next) == NULL_RTX))
1283 return HARD_DEP;
1284
1285 pro = DEP_PRO (modify_dep);
1286 other = real_insn_for_shadow (pro);
1287 if (other != NULL_RTX)
1288 pro = other;
1289
1290 cond = sched_get_reverse_condition_uncached (pro);
1291 regno = REGNO (XEXP (cond, 0));
1292
1293 /* Find the last scheduled insn that modifies the condition register.
1294 We can stop looking once we find the insn we depend on through the
1295 REG_DEP_CONTROL; if the condition register isn't modified after it,
1296 we know that it still has the right value. */
1297 if (QUEUE_INDEX (pro) == QUEUE_SCHEDULED)
1298 FOR_EACH_VEC_ELT_REVERSE (scheduled_insns, i, prev)
1299 {
1300 HARD_REG_SET t;
1301
1302 find_all_hard_reg_sets (prev, &t);
1303 if (TEST_HARD_REG_BIT (t, regno))
1304 return HARD_DEP;
1305 if (prev == pro)
1306 break;
1307 }
1308 if (ORIG_PAT (next) == NULL_RTX)
1309 {
1310 ORIG_PAT (next) = PATTERN (next);
1311
1312 new_pat = gen_rtx_COND_EXEC (VOIDmode, cond, PATTERN (next));
1313 success = haifa_change_pattern (next, new_pat);
1314 if (!success)
1315 return HARD_DEP;
1316 PREDICATED_PAT (next) = new_pat;
1317 }
1318 else if (PATTERN (next) != PREDICATED_PAT (next))
1319 {
1320 bool success = haifa_change_pattern (next,
1321 PREDICATED_PAT (next));
1322 gcc_assert (success);
1323 }
1324 DEP_STATUS (modify_dep) |= DEP_CANCELLED;
1325 return DEP_CONTROL;
1326 }
1327
1328 if (PREDICATED_PAT (next) != NULL_RTX)
1329 {
1330 int tick = INSN_TICK (next);
1331 bool success = haifa_change_pattern (next,
1332 ORIG_PAT (next));
1333 INSN_TICK (next) = tick;
1334 gcc_assert (success);
1335 }
1336
1337 /* We can't handle the case where there are both speculative and control
1338 dependencies, so we return HARD_DEP in such a case. Also fail if
1339 we have speculative dependencies with not enough points, or more than
1340 one control dependency. */
1341 if ((n_spec > 0 && (n_control > 0 || n_replace > 0))
1342 || (n_spec > 0
1343 /* Too few points? */
1344 && ds_weak (new_ds) < spec_info->data_weakness_cutoff)
1345 || n_control > 0
1346 || n_replace > 0)
1347 return HARD_DEP;
1348
1349 return new_ds;
1350 }
1351 \f
1352 /* Pointer to the last instruction scheduled. */
1353 static rtx last_scheduled_insn;
1354
1355 /* Pointer to the last nondebug instruction scheduled within the
1356 block, or the prev_head of the scheduling block. Used by
1357 rank_for_schedule, so that insns independent of the last scheduled
1358 insn will be preferred over dependent instructions. */
1359 static rtx last_nondebug_scheduled_insn;
1360
1361 /* Pointer that iterates through the list of unscheduled insns if we
1362 have a dbg_cnt enabled. It always points at an insn prior to the
1363 first unscheduled one. */
1364 static rtx nonscheduled_insns_begin;
1365
1366 /* Compute cost of executing INSN.
1367 This is the number of cycles between instruction issue and
1368 instruction results. */
1369 int
1370 insn_cost (rtx insn)
1371 {
1372 int cost;
1373
1374 if (sel_sched_p ())
1375 {
1376 if (recog_memoized (insn) < 0)
1377 return 0;
1378
1379 cost = insn_default_latency (insn);
1380 if (cost < 0)
1381 cost = 0;
1382
1383 return cost;
1384 }
1385
1386 cost = INSN_COST (insn);
1387
1388 if (cost < 0)
1389 {
1390 /* A USE insn, or something else we don't need to
1391 understand. We can't pass these directly to
1392 result_ready_cost or insn_default_latency because it will
1393 trigger a fatal error for unrecognizable insns. */
1394 if (recog_memoized (insn) < 0)
1395 {
1396 INSN_COST (insn) = 0;
1397 return 0;
1398 }
1399 else
1400 {
1401 cost = insn_default_latency (insn);
1402 if (cost < 0)
1403 cost = 0;
1404
1405 INSN_COST (insn) = cost;
1406 }
1407 }
1408
1409 return cost;
1410 }
1411
1412 /* Compute cost of dependence LINK.
1413 This is the number of cycles between instruction issue and
1414 instruction results.
1415 ??? We also use this function to call recog_memoized on all insns. */
1416 int
1417 dep_cost_1 (dep_t link, dw_t dw)
1418 {
1419 rtx insn = DEP_PRO (link);
1420 rtx used = DEP_CON (link);
1421 int cost;
1422
1423 if (DEP_COST (link) != UNKNOWN_DEP_COST)
1424 return DEP_COST (link);
1425
1426 if (delay_htab.is_created ())
1427 {
1428 struct delay_pair *delay_entry;
1429 delay_entry
1430 = delay_htab_i2.find_with_hash (used, htab_hash_pointer (used));
1431 if (delay_entry)
1432 {
1433 if (delay_entry->i1 == insn)
1434 {
1435 DEP_COST (link) = pair_delay (delay_entry);
1436 return DEP_COST (link);
1437 }
1438 }
1439 }
1440
1441 /* A USE insn should never require the value used to be computed.
1442 This allows the computation of a function's result and parameter
1443 values to overlap the return and call. We don't care about the
1444 dependence cost when only decreasing register pressure. */
1445 if (recog_memoized (used) < 0)
1446 {
1447 cost = 0;
1448 recog_memoized (insn);
1449 }
1450 else
1451 {
1452 enum reg_note dep_type = DEP_TYPE (link);
1453
1454 cost = insn_cost (insn);
1455
1456 if (INSN_CODE (insn) >= 0)
1457 {
1458 if (dep_type == REG_DEP_ANTI)
1459 cost = 0;
1460 else if (dep_type == REG_DEP_OUTPUT)
1461 {
1462 cost = (insn_default_latency (insn)
1463 - insn_default_latency (used));
1464 if (cost <= 0)
1465 cost = 1;
1466 }
1467 else if (bypass_p (insn))
1468 cost = insn_latency (insn, used);
1469 }
1470
1471
1472 if (targetm.sched.adjust_cost_2)
1473 cost = targetm.sched.adjust_cost_2 (used, (int) dep_type, insn, cost,
1474 dw);
1475 else if (targetm.sched.adjust_cost != NULL)
1476 {
1477 /* This variable is used for backward compatibility with the
1478 targets. */
1479 rtx dep_cost_rtx_link = alloc_INSN_LIST (NULL_RTX, NULL_RTX);
1480
1481 /* Make it self-cycled, so that if some tries to walk over this
1482 incomplete list he/she will be caught in an endless loop. */
1483 XEXP (dep_cost_rtx_link, 1) = dep_cost_rtx_link;
1484
1485 /* Targets use only REG_NOTE_KIND of the link. */
1486 PUT_REG_NOTE_KIND (dep_cost_rtx_link, DEP_TYPE (link));
1487
1488 cost = targetm.sched.adjust_cost (used, dep_cost_rtx_link,
1489 insn, cost);
1490
1491 free_INSN_LIST_node (dep_cost_rtx_link);
1492 }
1493
1494 if (cost < 0)
1495 cost = 0;
1496 }
1497
1498 DEP_COST (link) = cost;
1499 return cost;
1500 }
1501
1502 /* Compute cost of dependence LINK.
1503 This is the number of cycles between instruction issue and
1504 instruction results. */
1505 int
1506 dep_cost (dep_t link)
1507 {
1508 return dep_cost_1 (link, 0);
1509 }
1510
1511 /* Use this sel-sched.c friendly function in reorder2 instead of increasing
1512 INSN_PRIORITY explicitly. */
1513 void
1514 increase_insn_priority (rtx insn, int amount)
1515 {
1516 if (!sel_sched_p ())
1517 {
1518 /* We're dealing with haifa-sched.c INSN_PRIORITY. */
1519 if (INSN_PRIORITY_KNOWN (insn))
1520 INSN_PRIORITY (insn) += amount;
1521 }
1522 else
1523 {
1524 /* In sel-sched.c INSN_PRIORITY is not kept up to date.
1525 Use EXPR_PRIORITY instead. */
1526 sel_add_to_insn_priority (insn, amount);
1527 }
1528 }
1529
1530 /* Return 'true' if DEP should be included in priority calculations. */
1531 static bool
1532 contributes_to_priority_p (dep_t dep)
1533 {
1534 if (DEBUG_INSN_P (DEP_CON (dep))
1535 || DEBUG_INSN_P (DEP_PRO (dep)))
1536 return false;
1537
1538 /* Critical path is meaningful in block boundaries only. */
1539 if (!current_sched_info->contributes_to_priority (DEP_CON (dep),
1540 DEP_PRO (dep)))
1541 return false;
1542
1543 if (DEP_REPLACE (dep) != NULL)
1544 return false;
1545
1546 /* If flag COUNT_SPEC_IN_CRITICAL_PATH is set,
1547 then speculative instructions will less likely be
1548 scheduled. That is because the priority of
1549 their producers will increase, and, thus, the
1550 producers will more likely be scheduled, thus,
1551 resolving the dependence. */
1552 if (sched_deps_info->generate_spec_deps
1553 && !(spec_info->flags & COUNT_SPEC_IN_CRITICAL_PATH)
1554 && (DEP_STATUS (dep) & SPECULATIVE))
1555 return false;
1556
1557 return true;
1558 }
1559
1560 /* Compute the number of nondebug deps in list LIST for INSN. */
1561
1562 static int
1563 dep_list_size (rtx insn, sd_list_types_def list)
1564 {
1565 sd_iterator_def sd_it;
1566 dep_t dep;
1567 int dbgcount = 0, nodbgcount = 0;
1568
1569 if (!MAY_HAVE_DEBUG_INSNS)
1570 return sd_lists_size (insn, list);
1571
1572 FOR_EACH_DEP (insn, list, sd_it, dep)
1573 {
1574 if (DEBUG_INSN_P (DEP_CON (dep)))
1575 dbgcount++;
1576 else if (!DEBUG_INSN_P (DEP_PRO (dep)))
1577 nodbgcount++;
1578 }
1579
1580 gcc_assert (dbgcount + nodbgcount == sd_lists_size (insn, list));
1581
1582 return nodbgcount;
1583 }
1584
1585 /* Compute the priority number for INSN. */
1586 static int
1587 priority (rtx insn)
1588 {
1589 if (! INSN_P (insn))
1590 return 0;
1591
1592 /* We should not be interested in priority of an already scheduled insn. */
1593 gcc_assert (QUEUE_INDEX (insn) != QUEUE_SCHEDULED);
1594
1595 if (!INSN_PRIORITY_KNOWN (insn))
1596 {
1597 int this_priority = -1;
1598
1599 if (dep_list_size (insn, SD_LIST_FORW) == 0)
1600 /* ??? We should set INSN_PRIORITY to insn_cost when and insn has
1601 some forward deps but all of them are ignored by
1602 contributes_to_priority hook. At the moment we set priority of
1603 such insn to 0. */
1604 this_priority = insn_cost (insn);
1605 else
1606 {
1607 rtx prev_first, twin;
1608 basic_block rec;
1609
1610 /* For recovery check instructions we calculate priority slightly
1611 different than that of normal instructions. Instead of walking
1612 through INSN_FORW_DEPS (check) list, we walk through
1613 INSN_FORW_DEPS list of each instruction in the corresponding
1614 recovery block. */
1615
1616 /* Selective scheduling does not define RECOVERY_BLOCK macro. */
1617 rec = sel_sched_p () ? NULL : RECOVERY_BLOCK (insn);
1618 if (!rec || rec == EXIT_BLOCK_PTR_FOR_FN (cfun))
1619 {
1620 prev_first = PREV_INSN (insn);
1621 twin = insn;
1622 }
1623 else
1624 {
1625 prev_first = NEXT_INSN (BB_HEAD (rec));
1626 twin = PREV_INSN (BB_END (rec));
1627 }
1628
1629 do
1630 {
1631 sd_iterator_def sd_it;
1632 dep_t dep;
1633
1634 FOR_EACH_DEP (twin, SD_LIST_FORW, sd_it, dep)
1635 {
1636 rtx next;
1637 int next_priority;
1638
1639 next = DEP_CON (dep);
1640
1641 if (BLOCK_FOR_INSN (next) != rec)
1642 {
1643 int cost;
1644
1645 if (!contributes_to_priority_p (dep))
1646 continue;
1647
1648 if (twin == insn)
1649 cost = dep_cost (dep);
1650 else
1651 {
1652 struct _dep _dep1, *dep1 = &_dep1;
1653
1654 init_dep (dep1, insn, next, REG_DEP_ANTI);
1655
1656 cost = dep_cost (dep1);
1657 }
1658
1659 next_priority = cost + priority (next);
1660
1661 if (next_priority > this_priority)
1662 this_priority = next_priority;
1663 }
1664 }
1665
1666 twin = PREV_INSN (twin);
1667 }
1668 while (twin != prev_first);
1669 }
1670
1671 if (this_priority < 0)
1672 {
1673 gcc_assert (this_priority == -1);
1674
1675 this_priority = insn_cost (insn);
1676 }
1677
1678 INSN_PRIORITY (insn) = this_priority;
1679 INSN_PRIORITY_STATUS (insn) = 1;
1680 }
1681
1682 return INSN_PRIORITY (insn);
1683 }
1684 \f
1685 /* Macros and functions for keeping the priority queue sorted, and
1686 dealing with queuing and dequeuing of instructions. */
1687
1688 #define SCHED_SORT(READY, N_READY) \
1689 do { if ((N_READY) == 2) \
1690 swap_sort (READY, N_READY); \
1691 else if ((N_READY) > 2) \
1692 qsort (READY, N_READY, sizeof (rtx), rank_for_schedule); } \
1693 while (0)
1694
1695 /* For each pressure class CL, set DEATH[CL] to the number of registers
1696 in that class that die in INSN. */
1697
1698 static void
1699 calculate_reg_deaths (rtx insn, int *death)
1700 {
1701 int i;
1702 struct reg_use_data *use;
1703
1704 for (i = 0; i < ira_pressure_classes_num; i++)
1705 death[ira_pressure_classes[i]] = 0;
1706 for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
1707 if (dying_use_p (use))
1708 mark_regno_birth_or_death (0, death, use->regno, true);
1709 }
1710
1711 /* Setup info about the current register pressure impact of scheduling
1712 INSN at the current scheduling point. */
1713 static void
1714 setup_insn_reg_pressure_info (rtx insn)
1715 {
1716 int i, change, before, after, hard_regno;
1717 int excess_cost_change;
1718 enum machine_mode mode;
1719 enum reg_class cl;
1720 struct reg_pressure_data *pressure_info;
1721 int *max_reg_pressure;
1722 static int death[N_REG_CLASSES];
1723
1724 gcc_checking_assert (!DEBUG_INSN_P (insn));
1725
1726 excess_cost_change = 0;
1727 calculate_reg_deaths (insn, death);
1728 pressure_info = INSN_REG_PRESSURE (insn);
1729 max_reg_pressure = INSN_MAX_REG_PRESSURE (insn);
1730 gcc_assert (pressure_info != NULL && max_reg_pressure != NULL);
1731 for (i = 0; i < ira_pressure_classes_num; i++)
1732 {
1733 cl = ira_pressure_classes[i];
1734 gcc_assert (curr_reg_pressure[cl] >= 0);
1735 change = (int) pressure_info[i].set_increase - death[cl];
1736 before = MAX (0, max_reg_pressure[i] - ira_class_hard_regs_num[cl]);
1737 after = MAX (0, max_reg_pressure[i] + change
1738 - ira_class_hard_regs_num[cl]);
1739 hard_regno = ira_class_hard_regs[cl][0];
1740 gcc_assert (hard_regno >= 0);
1741 mode = reg_raw_mode[hard_regno];
1742 excess_cost_change += ((after - before)
1743 * (ira_memory_move_cost[mode][cl][0]
1744 + ira_memory_move_cost[mode][cl][1]));
1745 }
1746 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insn) = excess_cost_change;
1747 }
1748 \f
1749 /* This is the first page of code related to SCHED_PRESSURE_MODEL.
1750 It tries to make the scheduler take register pressure into account
1751 without introducing too many unnecessary stalls. It hooks into the
1752 main scheduling algorithm at several points:
1753
1754 - Before scheduling starts, model_start_schedule constructs a
1755 "model schedule" for the current block. This model schedule is
1756 chosen solely to keep register pressure down. It does not take the
1757 target's pipeline or the original instruction order into account,
1758 except as a tie-breaker. It also doesn't work to a particular
1759 pressure limit.
1760
1761 This model schedule gives us an idea of what pressure can be
1762 achieved for the block and gives us an example of a schedule that
1763 keeps to that pressure. It also makes the final schedule less
1764 dependent on the original instruction order. This is important
1765 because the original order can either be "wide" (many values live
1766 at once, such as in user-scheduled code) or "narrow" (few values
1767 live at once, such as after loop unrolling, where several
1768 iterations are executed sequentially).
1769
1770 We do not apply this model schedule to the rtx stream. We simply
1771 record it in model_schedule. We also compute the maximum pressure,
1772 MP, that was seen during this schedule.
1773
1774 - Instructions are added to the ready queue even if they require
1775 a stall. The length of the stall is instead computed as:
1776
1777 MAX (INSN_TICK (INSN) - clock_var, 0)
1778
1779 (= insn_delay). This allows rank_for_schedule to choose between
1780 introducing a deliberate stall or increasing pressure.
1781
1782 - Before sorting the ready queue, model_set_excess_costs assigns
1783 a pressure-based cost to each ready instruction in the queue.
1784 This is the instruction's INSN_REG_PRESSURE_EXCESS_COST_CHANGE
1785 (ECC for short) and is effectively measured in cycles.
1786
1787 - rank_for_schedule ranks instructions based on:
1788
1789 ECC (insn) + insn_delay (insn)
1790
1791 then as:
1792
1793 insn_delay (insn)
1794
1795 So, for example, an instruction X1 with an ECC of 1 that can issue
1796 now will win over an instruction X0 with an ECC of zero that would
1797 introduce a stall of one cycle. However, an instruction X2 with an
1798 ECC of 2 that can issue now will lose to both X0 and X1.
1799
1800 - When an instruction is scheduled, model_recompute updates the model
1801 schedule with the new pressures (some of which might now exceed the
1802 original maximum pressure MP). model_update_limit_points then searches
1803 for the new point of maximum pressure, if not already known. */
1804
1805 /* Used to separate high-verbosity debug information for SCHED_PRESSURE_MODEL
1806 from surrounding debug information. */
1807 #define MODEL_BAR \
1808 ";;\t\t+------------------------------------------------------\n"
1809
1810 /* Information about the pressure on a particular register class at a
1811 particular point of the model schedule. */
1812 struct model_pressure_data {
1813 /* The pressure at this point of the model schedule, or -1 if the
1814 point is associated with an instruction that has already been
1815 scheduled. */
1816 int ref_pressure;
1817
1818 /* The maximum pressure during or after this point of the model schedule. */
1819 int max_pressure;
1820 };
1821
1822 /* Per-instruction information that is used while building the model
1823 schedule. Here, "schedule" refers to the model schedule rather
1824 than the main schedule. */
1825 struct model_insn_info {
1826 /* The instruction itself. */
1827 rtx insn;
1828
1829 /* If this instruction is in model_worklist, these fields link to the
1830 previous (higher-priority) and next (lower-priority) instructions
1831 in the list. */
1832 struct model_insn_info *prev;
1833 struct model_insn_info *next;
1834
1835 /* While constructing the schedule, QUEUE_INDEX describes whether an
1836 instruction has already been added to the schedule (QUEUE_SCHEDULED),
1837 is in model_worklist (QUEUE_READY), or neither (QUEUE_NOWHERE).
1838 old_queue records the value that QUEUE_INDEX had before scheduling
1839 started, so that we can restore it once the schedule is complete. */
1840 int old_queue;
1841
1842 /* The relative importance of an unscheduled instruction. Higher
1843 values indicate greater importance. */
1844 unsigned int model_priority;
1845
1846 /* The length of the longest path of satisfied true dependencies
1847 that leads to this instruction. */
1848 unsigned int depth;
1849
1850 /* The length of the longest path of dependencies of any kind
1851 that leads from this instruction. */
1852 unsigned int alap;
1853
1854 /* The number of predecessor nodes that must still be scheduled. */
1855 int unscheduled_preds;
1856 };
1857
1858 /* Information about the pressure limit for a particular register class.
1859 This structure is used when applying a model schedule to the main
1860 schedule. */
1861 struct model_pressure_limit {
1862 /* The maximum register pressure seen in the original model schedule. */
1863 int orig_pressure;
1864
1865 /* The maximum register pressure seen in the current model schedule
1866 (which excludes instructions that have already been scheduled). */
1867 int pressure;
1868
1869 /* The point of the current model schedule at which PRESSURE is first
1870 reached. It is set to -1 if the value needs to be recomputed. */
1871 int point;
1872 };
1873
1874 /* Describes a particular way of measuring register pressure. */
1875 struct model_pressure_group {
1876 /* Index PCI describes the maximum pressure on ira_pressure_classes[PCI]. */
1877 struct model_pressure_limit limits[N_REG_CLASSES];
1878
1879 /* Index (POINT * ira_num_pressure_classes + PCI) describes the pressure
1880 on register class ira_pressure_classes[PCI] at point POINT of the
1881 current model schedule. A POINT of model_num_insns describes the
1882 pressure at the end of the schedule. */
1883 struct model_pressure_data *model;
1884 };
1885
1886 /* Index POINT gives the instruction at point POINT of the model schedule.
1887 This array doesn't change during main scheduling. */
1888 static vec<rtx> model_schedule;
1889
1890 /* The list of instructions in the model worklist, sorted in order of
1891 decreasing priority. */
1892 static struct model_insn_info *model_worklist;
1893
1894 /* Index I describes the instruction with INSN_LUID I. */
1895 static struct model_insn_info *model_insns;
1896
1897 /* The number of instructions in the model schedule. */
1898 static int model_num_insns;
1899
1900 /* The index of the first instruction in model_schedule that hasn't yet been
1901 added to the main schedule, or model_num_insns if all of them have. */
1902 static int model_curr_point;
1903
1904 /* Describes the pressure before each instruction in the model schedule. */
1905 static struct model_pressure_group model_before_pressure;
1906
1907 /* The first unused model_priority value (as used in model_insn_info). */
1908 static unsigned int model_next_priority;
1909
1910
1911 /* The model_pressure_data for ira_pressure_classes[PCI] in GROUP
1912 at point POINT of the model schedule. */
1913 #define MODEL_PRESSURE_DATA(GROUP, POINT, PCI) \
1914 (&(GROUP)->model[(POINT) * ira_pressure_classes_num + (PCI)])
1915
1916 /* The maximum pressure on ira_pressure_classes[PCI] in GROUP at or
1917 after point POINT of the model schedule. */
1918 #define MODEL_MAX_PRESSURE(GROUP, POINT, PCI) \
1919 (MODEL_PRESSURE_DATA (GROUP, POINT, PCI)->max_pressure)
1920
1921 /* The pressure on ira_pressure_classes[PCI] in GROUP at point POINT
1922 of the model schedule. */
1923 #define MODEL_REF_PRESSURE(GROUP, POINT, PCI) \
1924 (MODEL_PRESSURE_DATA (GROUP, POINT, PCI)->ref_pressure)
1925
1926 /* Information about INSN that is used when creating the model schedule. */
1927 #define MODEL_INSN_INFO(INSN) \
1928 (&model_insns[INSN_LUID (INSN)])
1929
1930 /* The instruction at point POINT of the model schedule. */
1931 #define MODEL_INSN(POINT) \
1932 (model_schedule[POINT])
1933
1934
1935 /* Return INSN's index in the model schedule, or model_num_insns if it
1936 doesn't belong to that schedule. */
1937
1938 static int
1939 model_index (rtx insn)
1940 {
1941 if (INSN_MODEL_INDEX (insn) == 0)
1942 return model_num_insns;
1943 return INSN_MODEL_INDEX (insn) - 1;
1944 }
1945
1946 /* Make sure that GROUP->limits is up-to-date for the current point
1947 of the model schedule. */
1948
1949 static void
1950 model_update_limit_points_in_group (struct model_pressure_group *group)
1951 {
1952 int pci, max_pressure, point;
1953
1954 for (pci = 0; pci < ira_pressure_classes_num; pci++)
1955 {
1956 /* We may have passed the final point at which the pressure in
1957 group->limits[pci].pressure was reached. Update the limit if so. */
1958 max_pressure = MODEL_MAX_PRESSURE (group, model_curr_point, pci);
1959 group->limits[pci].pressure = max_pressure;
1960
1961 /* Find the point at which MAX_PRESSURE is first reached. We need
1962 to search in three cases:
1963
1964 - We've already moved past the previous pressure point.
1965 In this case we search forward from model_curr_point.
1966
1967 - We scheduled the previous point of maximum pressure ahead of
1968 its position in the model schedule, but doing so didn't bring
1969 the pressure point earlier. In this case we search forward
1970 from that previous pressure point.
1971
1972 - Scheduling an instruction early caused the maximum pressure
1973 to decrease. In this case we will have set the pressure
1974 point to -1, and we search forward from model_curr_point. */
1975 point = MAX (group->limits[pci].point, model_curr_point);
1976 while (point < model_num_insns
1977 && MODEL_REF_PRESSURE (group, point, pci) < max_pressure)
1978 point++;
1979 group->limits[pci].point = point;
1980
1981 gcc_assert (MODEL_REF_PRESSURE (group, point, pci) == max_pressure);
1982 gcc_assert (MODEL_MAX_PRESSURE (group, point, pci) == max_pressure);
1983 }
1984 }
1985
1986 /* Make sure that all register-pressure limits are up-to-date for the
1987 current position in the model schedule. */
1988
1989 static void
1990 model_update_limit_points (void)
1991 {
1992 model_update_limit_points_in_group (&model_before_pressure);
1993 }
1994
1995 /* Return the model_index of the last unscheduled use in chain USE
1996 outside of USE's instruction. Return -1 if there are no other uses,
1997 or model_num_insns if the register is live at the end of the block. */
1998
1999 static int
2000 model_last_use_except (struct reg_use_data *use)
2001 {
2002 struct reg_use_data *next;
2003 int last, index;
2004
2005 last = -1;
2006 for (next = use->next_regno_use; next != use; next = next->next_regno_use)
2007 if (NONDEBUG_INSN_P (next->insn)
2008 && QUEUE_INDEX (next->insn) != QUEUE_SCHEDULED)
2009 {
2010 index = model_index (next->insn);
2011 if (index == model_num_insns)
2012 return model_num_insns;
2013 if (last < index)
2014 last = index;
2015 }
2016 return last;
2017 }
2018
2019 /* An instruction with model_index POINT has just been scheduled, and it
2020 adds DELTA to the pressure on ira_pressure_classes[PCI] after POINT - 1.
2021 Update MODEL_REF_PRESSURE (GROUP, POINT, PCI) and
2022 MODEL_MAX_PRESSURE (GROUP, POINT, PCI) accordingly. */
2023
2024 static void
2025 model_start_update_pressure (struct model_pressure_group *group,
2026 int point, int pci, int delta)
2027 {
2028 int next_max_pressure;
2029
2030 if (point == model_num_insns)
2031 {
2032 /* The instruction wasn't part of the model schedule; it was moved
2033 from a different block. Update the pressure for the end of
2034 the model schedule. */
2035 MODEL_REF_PRESSURE (group, point, pci) += delta;
2036 MODEL_MAX_PRESSURE (group, point, pci) += delta;
2037 }
2038 else
2039 {
2040 /* Record that this instruction has been scheduled. Nothing now
2041 changes between POINT and POINT + 1, so get the maximum pressure
2042 from the latter. If the maximum pressure decreases, the new
2043 pressure point may be before POINT. */
2044 MODEL_REF_PRESSURE (group, point, pci) = -1;
2045 next_max_pressure = MODEL_MAX_PRESSURE (group, point + 1, pci);
2046 if (MODEL_MAX_PRESSURE (group, point, pci) > next_max_pressure)
2047 {
2048 MODEL_MAX_PRESSURE (group, point, pci) = next_max_pressure;
2049 if (group->limits[pci].point == point)
2050 group->limits[pci].point = -1;
2051 }
2052 }
2053 }
2054
2055 /* Record that scheduling a later instruction has changed the pressure
2056 at point POINT of the model schedule by DELTA (which might be 0).
2057 Update GROUP accordingly. Return nonzero if these changes might
2058 trigger changes to previous points as well. */
2059
2060 static int
2061 model_update_pressure (struct model_pressure_group *group,
2062 int point, int pci, int delta)
2063 {
2064 int ref_pressure, max_pressure, next_max_pressure;
2065
2066 /* If POINT hasn't yet been scheduled, update its pressure. */
2067 ref_pressure = MODEL_REF_PRESSURE (group, point, pci);
2068 if (ref_pressure >= 0 && delta != 0)
2069 {
2070 ref_pressure += delta;
2071 MODEL_REF_PRESSURE (group, point, pci) = ref_pressure;
2072
2073 /* Check whether the maximum pressure in the overall schedule
2074 has increased. (This means that the MODEL_MAX_PRESSURE of
2075 every point <= POINT will need to increae too; see below.) */
2076 if (group->limits[pci].pressure < ref_pressure)
2077 group->limits[pci].pressure = ref_pressure;
2078
2079 /* If we are at maximum pressure, and the maximum pressure
2080 point was previously unknown or later than POINT,
2081 bring it forward. */
2082 if (group->limits[pci].pressure == ref_pressure
2083 && !IN_RANGE (group->limits[pci].point, 0, point))
2084 group->limits[pci].point = point;
2085
2086 /* If POINT used to be the point of maximum pressure, but isn't
2087 any longer, we need to recalculate it using a forward walk. */
2088 if (group->limits[pci].pressure > ref_pressure
2089 && group->limits[pci].point == point)
2090 group->limits[pci].point = -1;
2091 }
2092
2093 /* Update the maximum pressure at POINT. Changes here might also
2094 affect the maximum pressure at POINT - 1. */
2095 next_max_pressure = MODEL_MAX_PRESSURE (group, point + 1, pci);
2096 max_pressure = MAX (ref_pressure, next_max_pressure);
2097 if (MODEL_MAX_PRESSURE (group, point, pci) != max_pressure)
2098 {
2099 MODEL_MAX_PRESSURE (group, point, pci) = max_pressure;
2100 return 1;
2101 }
2102 return 0;
2103 }
2104
2105 /* INSN has just been scheduled. Update the model schedule accordingly. */
2106
2107 static void
2108 model_recompute (rtx insn)
2109 {
2110 struct {
2111 int last_use;
2112 int regno;
2113 } uses[FIRST_PSEUDO_REGISTER + MAX_RECOG_OPERANDS];
2114 struct reg_use_data *use;
2115 struct reg_pressure_data *reg_pressure;
2116 int delta[N_REG_CLASSES];
2117 int pci, point, mix, new_last, cl, ref_pressure, queue;
2118 unsigned int i, num_uses, num_pending_births;
2119 bool print_p;
2120
2121 /* The destinations of INSN were previously live from POINT onwards, but are
2122 now live from model_curr_point onwards. Set up DELTA accordingly. */
2123 point = model_index (insn);
2124 reg_pressure = INSN_REG_PRESSURE (insn);
2125 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2126 {
2127 cl = ira_pressure_classes[pci];
2128 delta[cl] = reg_pressure[pci].set_increase;
2129 }
2130
2131 /* Record which registers previously died at POINT, but which now die
2132 before POINT. Adjust DELTA so that it represents the effect of
2133 this change after POINT - 1. Set NUM_PENDING_BIRTHS to the number of
2134 registers that will be born in the range [model_curr_point, POINT). */
2135 num_uses = 0;
2136 num_pending_births = 0;
2137 for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
2138 {
2139 new_last = model_last_use_except (use);
2140 if (new_last < point)
2141 {
2142 gcc_assert (num_uses < ARRAY_SIZE (uses));
2143 uses[num_uses].last_use = new_last;
2144 uses[num_uses].regno = use->regno;
2145 /* This register is no longer live after POINT - 1. */
2146 mark_regno_birth_or_death (NULL, delta, use->regno, false);
2147 num_uses++;
2148 if (new_last >= 0)
2149 num_pending_births++;
2150 }
2151 }
2152
2153 /* Update the MODEL_REF_PRESSURE and MODEL_MAX_PRESSURE for POINT.
2154 Also set each group pressure limit for POINT. */
2155 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2156 {
2157 cl = ira_pressure_classes[pci];
2158 model_start_update_pressure (&model_before_pressure,
2159 point, pci, delta[cl]);
2160 }
2161
2162 /* Walk the model schedule backwards, starting immediately before POINT. */
2163 print_p = false;
2164 if (point != model_curr_point)
2165 do
2166 {
2167 point--;
2168 insn = MODEL_INSN (point);
2169 queue = QUEUE_INDEX (insn);
2170
2171 if (queue != QUEUE_SCHEDULED)
2172 {
2173 /* DELTA describes the effect of the move on the register pressure
2174 after POINT. Make it describe the effect on the pressure
2175 before POINT. */
2176 i = 0;
2177 while (i < num_uses)
2178 {
2179 if (uses[i].last_use == point)
2180 {
2181 /* This register is now live again. */
2182 mark_regno_birth_or_death (NULL, delta,
2183 uses[i].regno, true);
2184
2185 /* Remove this use from the array. */
2186 uses[i] = uses[num_uses - 1];
2187 num_uses--;
2188 num_pending_births--;
2189 }
2190 else
2191 i++;
2192 }
2193
2194 if (sched_verbose >= 5)
2195 {
2196 if (!print_p)
2197 {
2198 fprintf (sched_dump, MODEL_BAR);
2199 fprintf (sched_dump, ";;\t\t| New pressure for model"
2200 " schedule\n");
2201 fprintf (sched_dump, MODEL_BAR);
2202 print_p = true;
2203 }
2204
2205 fprintf (sched_dump, ";;\t\t| %3d %4d %-30s ",
2206 point, INSN_UID (insn),
2207 str_pattern_slim (PATTERN (insn)));
2208 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2209 {
2210 cl = ira_pressure_classes[pci];
2211 ref_pressure = MODEL_REF_PRESSURE (&model_before_pressure,
2212 point, pci);
2213 fprintf (sched_dump, " %s:[%d->%d]",
2214 reg_class_names[ira_pressure_classes[pci]],
2215 ref_pressure, ref_pressure + delta[cl]);
2216 }
2217 fprintf (sched_dump, "\n");
2218 }
2219 }
2220
2221 /* Adjust the pressure at POINT. Set MIX to nonzero if POINT - 1
2222 might have changed as well. */
2223 mix = num_pending_births;
2224 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2225 {
2226 cl = ira_pressure_classes[pci];
2227 mix |= delta[cl];
2228 mix |= model_update_pressure (&model_before_pressure,
2229 point, pci, delta[cl]);
2230 }
2231 }
2232 while (mix && point > model_curr_point);
2233
2234 if (print_p)
2235 fprintf (sched_dump, MODEL_BAR);
2236 }
2237
2238 /* After DEP, which was cancelled, has been resolved for insn NEXT,
2239 check whether the insn's pattern needs restoring. */
2240 static bool
2241 must_restore_pattern_p (rtx next, dep_t dep)
2242 {
2243 if (QUEUE_INDEX (next) == QUEUE_SCHEDULED)
2244 return false;
2245
2246 if (DEP_TYPE (dep) == REG_DEP_CONTROL)
2247 {
2248 gcc_assert (ORIG_PAT (next) != NULL_RTX);
2249 gcc_assert (next == DEP_CON (dep));
2250 }
2251 else
2252 {
2253 struct dep_replacement *desc = DEP_REPLACE (dep);
2254 if (desc->insn != next)
2255 {
2256 gcc_assert (*desc->loc == desc->orig);
2257 return false;
2258 }
2259 }
2260 return true;
2261 }
2262 \f
2263 /* model_spill_cost (CL, P, P') returns the cost of increasing the
2264 pressure on CL from P to P'. We use this to calculate a "base ECC",
2265 baseECC (CL, X), for each pressure class CL and each instruction X.
2266 Supposing X changes the pressure on CL from P to P', and that the
2267 maximum pressure on CL in the current model schedule is MP', then:
2268
2269 * if X occurs before or at the next point of maximum pressure in
2270 the model schedule and P' > MP', then:
2271
2272 baseECC (CL, X) = model_spill_cost (CL, MP, P')
2273
2274 The idea is that the pressure after scheduling a fixed set of
2275 instructions -- in this case, the set up to and including the
2276 next maximum pressure point -- is going to be the same regardless
2277 of the order; we simply want to keep the intermediate pressure
2278 under control. Thus X has a cost of zero unless scheduling it
2279 now would exceed MP'.
2280
2281 If all increases in the set are by the same amount, no zero-cost
2282 instruction will ever cause the pressure to exceed MP'. However,
2283 if X is instead moved past an instruction X' with pressure in the
2284 range (MP' - (P' - P), MP'), the pressure at X' will increase
2285 beyond MP'. Since baseECC is very much a heuristic anyway,
2286 it doesn't seem worth the overhead of tracking cases like these.
2287
2288 The cost of exceeding MP' is always based on the original maximum
2289 pressure MP. This is so that going 2 registers over the original
2290 limit has the same cost regardless of whether it comes from two
2291 separate +1 deltas or from a single +2 delta.
2292
2293 * if X occurs after the next point of maximum pressure in the model
2294 schedule and P' > P, then:
2295
2296 baseECC (CL, X) = model_spill_cost (CL, MP, MP' + (P' - P))
2297
2298 That is, if we move X forward across a point of maximum pressure,
2299 and if X increases the pressure by P' - P, then we conservatively
2300 assume that scheduling X next would increase the maximum pressure
2301 by P' - P. Again, the cost of doing this is based on the original
2302 maximum pressure MP, for the same reason as above.
2303
2304 * if P' < P, P > MP, and X occurs at or after the next point of
2305 maximum pressure, then:
2306
2307 baseECC (CL, X) = -model_spill_cost (CL, MAX (MP, P'), P)
2308
2309 That is, if we have already exceeded the original maximum pressure MP,
2310 and if X might reduce the maximum pressure again -- or at least push
2311 it further back, and thus allow more scheduling freedom -- it is given
2312 a negative cost to reflect the improvement.
2313
2314 * otherwise,
2315
2316 baseECC (CL, X) = 0
2317
2318 In this case, X is not expected to affect the maximum pressure MP',
2319 so it has zero cost.
2320
2321 We then create a combined value baseECC (X) that is the sum of
2322 baseECC (CL, X) for each pressure class CL.
2323
2324 baseECC (X) could itself be used as the ECC value described above.
2325 However, this is often too conservative, in the sense that it
2326 tends to make high-priority instructions that increase pressure
2327 wait too long in cases where introducing a spill would be better.
2328 For this reason the final ECC is a priority-adjusted form of
2329 baseECC (X). Specifically, we calculate:
2330
2331 P (X) = INSN_PRIORITY (X) - insn_delay (X) - baseECC (X)
2332 baseP = MAX { P (X) | baseECC (X) <= 0 }
2333
2334 Then:
2335
2336 ECC (X) = MAX (MIN (baseP - P (X), baseECC (X)), 0)
2337
2338 Thus an instruction's effect on pressure is ignored if it has a high
2339 enough priority relative to the ones that don't increase pressure.
2340 Negative values of baseECC (X) do not increase the priority of X
2341 itself, but they do make it harder for other instructions to
2342 increase the pressure further.
2343
2344 This pressure cost is deliberately timid. The intention has been
2345 to choose a heuristic that rarely interferes with the normal list
2346 scheduler in cases where that scheduler would produce good code.
2347 We simply want to curb some of its worst excesses. */
2348
2349 /* Return the cost of increasing the pressure in class CL from FROM to TO.
2350
2351 Here we use the very simplistic cost model that every register above
2352 ira_class_hard_regs_num[CL] has a spill cost of 1. We could use other
2353 measures instead, such as one based on MEMORY_MOVE_COST. However:
2354
2355 (1) In order for an instruction to be scheduled, the higher cost
2356 would need to be justified in a single saving of that many stalls.
2357 This is overly pessimistic, because the benefit of spilling is
2358 often to avoid a sequence of several short stalls rather than
2359 a single long one.
2360
2361 (2) The cost is still arbitrary. Because we are not allocating
2362 registers during scheduling, we have no way of knowing for
2363 sure how many memory accesses will be required by each spill,
2364 where the spills will be placed within the block, or even
2365 which block(s) will contain the spills.
2366
2367 So a higher cost than 1 is often too conservative in practice,
2368 forcing blocks to contain unnecessary stalls instead of spill code.
2369 The simple cost below seems to be the best compromise. It reduces
2370 the interference with the normal list scheduler, which helps make
2371 it more suitable for a default-on option. */
2372
2373 static int
2374 model_spill_cost (int cl, int from, int to)
2375 {
2376 from = MAX (from, ira_class_hard_regs_num[cl]);
2377 return MAX (to, from) - from;
2378 }
2379
2380 /* Return baseECC (ira_pressure_classes[PCI], POINT), given that
2381 P = curr_reg_pressure[ira_pressure_classes[PCI]] and that
2382 P' = P + DELTA. */
2383
2384 static int
2385 model_excess_group_cost (struct model_pressure_group *group,
2386 int point, int pci, int delta)
2387 {
2388 int pressure, cl;
2389
2390 cl = ira_pressure_classes[pci];
2391 if (delta < 0 && point >= group->limits[pci].point)
2392 {
2393 pressure = MAX (group->limits[pci].orig_pressure,
2394 curr_reg_pressure[cl] + delta);
2395 return -model_spill_cost (cl, pressure, curr_reg_pressure[cl]);
2396 }
2397
2398 if (delta > 0)
2399 {
2400 if (point > group->limits[pci].point)
2401 pressure = group->limits[pci].pressure + delta;
2402 else
2403 pressure = curr_reg_pressure[cl] + delta;
2404
2405 if (pressure > group->limits[pci].pressure)
2406 return model_spill_cost (cl, group->limits[pci].orig_pressure,
2407 pressure);
2408 }
2409
2410 return 0;
2411 }
2412
2413 /* Return baseECC (MODEL_INSN (INSN)). Dump the costs to sched_dump
2414 if PRINT_P. */
2415
2416 static int
2417 model_excess_cost (rtx insn, bool print_p)
2418 {
2419 int point, pci, cl, cost, this_cost, delta;
2420 struct reg_pressure_data *insn_reg_pressure;
2421 int insn_death[N_REG_CLASSES];
2422
2423 calculate_reg_deaths (insn, insn_death);
2424 point = model_index (insn);
2425 insn_reg_pressure = INSN_REG_PRESSURE (insn);
2426 cost = 0;
2427
2428 if (print_p)
2429 fprintf (sched_dump, ";;\t\t| %3d %4d | %4d %+3d |", point,
2430 INSN_UID (insn), INSN_PRIORITY (insn), insn_delay (insn));
2431
2432 /* Sum up the individual costs for each register class. */
2433 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2434 {
2435 cl = ira_pressure_classes[pci];
2436 delta = insn_reg_pressure[pci].set_increase - insn_death[cl];
2437 this_cost = model_excess_group_cost (&model_before_pressure,
2438 point, pci, delta);
2439 cost += this_cost;
2440 if (print_p)
2441 fprintf (sched_dump, " %s:[%d base cost %d]",
2442 reg_class_names[cl], delta, this_cost);
2443 }
2444
2445 if (print_p)
2446 fprintf (sched_dump, "\n");
2447
2448 return cost;
2449 }
2450
2451 /* Dump the next points of maximum pressure for GROUP. */
2452
2453 static void
2454 model_dump_pressure_points (struct model_pressure_group *group)
2455 {
2456 int pci, cl;
2457
2458 fprintf (sched_dump, ";;\t\t| pressure points");
2459 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2460 {
2461 cl = ira_pressure_classes[pci];
2462 fprintf (sched_dump, " %s:[%d->%d at ", reg_class_names[cl],
2463 curr_reg_pressure[cl], group->limits[pci].pressure);
2464 if (group->limits[pci].point < model_num_insns)
2465 fprintf (sched_dump, "%d:%d]", group->limits[pci].point,
2466 INSN_UID (MODEL_INSN (group->limits[pci].point)));
2467 else
2468 fprintf (sched_dump, "end]");
2469 }
2470 fprintf (sched_dump, "\n");
2471 }
2472
2473 /* Set INSN_REG_PRESSURE_EXCESS_COST_CHANGE for INSNS[0...COUNT-1]. */
2474
2475 static void
2476 model_set_excess_costs (rtx *insns, int count)
2477 {
2478 int i, cost, priority_base, priority;
2479 bool print_p;
2480
2481 /* Record the baseECC value for each instruction in the model schedule,
2482 except that negative costs are converted to zero ones now rather thatn
2483 later. Do not assign a cost to debug instructions, since they must
2484 not change code-generation decisions. Experiments suggest we also
2485 get better results by not assigning a cost to instructions from
2486 a different block.
2487
2488 Set PRIORITY_BASE to baseP in the block comment above. This is the
2489 maximum priority of the "cheap" instructions, which should always
2490 include the next model instruction. */
2491 priority_base = 0;
2492 print_p = false;
2493 for (i = 0; i < count; i++)
2494 if (INSN_MODEL_INDEX (insns[i]))
2495 {
2496 if (sched_verbose >= 6 && !print_p)
2497 {
2498 fprintf (sched_dump, MODEL_BAR);
2499 fprintf (sched_dump, ";;\t\t| Pressure costs for ready queue\n");
2500 model_dump_pressure_points (&model_before_pressure);
2501 fprintf (sched_dump, MODEL_BAR);
2502 print_p = true;
2503 }
2504 cost = model_excess_cost (insns[i], print_p);
2505 if (cost <= 0)
2506 {
2507 priority = INSN_PRIORITY (insns[i]) - insn_delay (insns[i]) - cost;
2508 priority_base = MAX (priority_base, priority);
2509 cost = 0;
2510 }
2511 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i]) = cost;
2512 }
2513 if (print_p)
2514 fprintf (sched_dump, MODEL_BAR);
2515
2516 /* Use MAX (baseECC, 0) and baseP to calculcate ECC for each
2517 instruction. */
2518 for (i = 0; i < count; i++)
2519 {
2520 cost = INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i]);
2521 priority = INSN_PRIORITY (insns[i]) - insn_delay (insns[i]);
2522 if (cost > 0 && priority > priority_base)
2523 {
2524 cost += priority_base - priority;
2525 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i]) = MAX (cost, 0);
2526 }
2527 }
2528 }
2529 \f
2530 /* Returns a positive value if x is preferred; returns a negative value if
2531 y is preferred. Should never return 0, since that will make the sort
2532 unstable. */
2533
2534 static int
2535 rank_for_schedule (const void *x, const void *y)
2536 {
2537 rtx tmp = *(const rtx *) y;
2538 rtx tmp2 = *(const rtx *) x;
2539 int tmp_class, tmp2_class;
2540 int val, priority_val, info_val, diff;
2541
2542 if (MAY_HAVE_DEBUG_INSNS)
2543 {
2544 /* Schedule debug insns as early as possible. */
2545 if (DEBUG_INSN_P (tmp) && !DEBUG_INSN_P (tmp2))
2546 return -1;
2547 else if (!DEBUG_INSN_P (tmp) && DEBUG_INSN_P (tmp2))
2548 return 1;
2549 else if (DEBUG_INSN_P (tmp) && DEBUG_INSN_P (tmp2))
2550 return INSN_LUID (tmp) - INSN_LUID (tmp2);
2551 }
2552
2553 if (live_range_shrinkage_p)
2554 {
2555 /* Don't use SCHED_PRESSURE_MODEL -- it results in much worse
2556 code. */
2557 gcc_assert (sched_pressure == SCHED_PRESSURE_WEIGHTED);
2558 if ((INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp) < 0
2559 || INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2) < 0)
2560 && (diff = (INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp)
2561 - INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2))) != 0)
2562 return diff;
2563 /* Sort by INSN_LUID (original insn order), so that we make the
2564 sort stable. This minimizes instruction movement, thus
2565 minimizing sched's effect on debugging and cross-jumping. */
2566 return INSN_LUID (tmp) - INSN_LUID (tmp2);
2567 }
2568
2569 /* The insn in a schedule group should be issued the first. */
2570 if (flag_sched_group_heuristic &&
2571 SCHED_GROUP_P (tmp) != SCHED_GROUP_P (tmp2))
2572 return SCHED_GROUP_P (tmp2) ? 1 : -1;
2573
2574 /* Make sure that priority of TMP and TMP2 are initialized. */
2575 gcc_assert (INSN_PRIORITY_KNOWN (tmp) && INSN_PRIORITY_KNOWN (tmp2));
2576
2577 if (sched_pressure != SCHED_PRESSURE_NONE)
2578 {
2579 /* Prefer insn whose scheduling results in the smallest register
2580 pressure excess. */
2581 if ((diff = (INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp)
2582 + insn_delay (tmp)
2583 - INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2)
2584 - insn_delay (tmp2))))
2585 return diff;
2586 }
2587
2588 if (sched_pressure != SCHED_PRESSURE_NONE
2589 && (INSN_TICK (tmp2) > clock_var || INSN_TICK (tmp) > clock_var))
2590 {
2591 if (INSN_TICK (tmp) <= clock_var)
2592 return -1;
2593 else if (INSN_TICK (tmp2) <= clock_var)
2594 return 1;
2595 else
2596 return INSN_TICK (tmp) - INSN_TICK (tmp2);
2597 }
2598
2599 /* If we are doing backtracking in this schedule, prefer insns that
2600 have forward dependencies with negative cost against an insn that
2601 was already scheduled. */
2602 if (current_sched_info->flags & DO_BACKTRACKING)
2603 {
2604 priority_val = FEEDS_BACKTRACK_INSN (tmp2) - FEEDS_BACKTRACK_INSN (tmp);
2605 if (priority_val)
2606 return priority_val;
2607 }
2608
2609 /* Prefer insn with higher priority. */
2610 priority_val = INSN_PRIORITY (tmp2) - INSN_PRIORITY (tmp);
2611
2612 if (flag_sched_critical_path_heuristic && priority_val)
2613 return priority_val;
2614
2615 /* Prefer speculative insn with greater dependencies weakness. */
2616 if (flag_sched_spec_insn_heuristic && spec_info)
2617 {
2618 ds_t ds1, ds2;
2619 dw_t dw1, dw2;
2620 int dw;
2621
2622 ds1 = TODO_SPEC (tmp) & SPECULATIVE;
2623 if (ds1)
2624 dw1 = ds_weak (ds1);
2625 else
2626 dw1 = NO_DEP_WEAK;
2627
2628 ds2 = TODO_SPEC (tmp2) & SPECULATIVE;
2629 if (ds2)
2630 dw2 = ds_weak (ds2);
2631 else
2632 dw2 = NO_DEP_WEAK;
2633
2634 dw = dw2 - dw1;
2635 if (dw > (NO_DEP_WEAK / 8) || dw < -(NO_DEP_WEAK / 8))
2636 return dw;
2637 }
2638
2639 info_val = (*current_sched_info->rank) (tmp, tmp2);
2640 if (flag_sched_rank_heuristic && info_val)
2641 return info_val;
2642
2643 /* Compare insns based on their relation to the last scheduled
2644 non-debug insn. */
2645 if (flag_sched_last_insn_heuristic && last_nondebug_scheduled_insn)
2646 {
2647 dep_t dep1;
2648 dep_t dep2;
2649 rtx last = last_nondebug_scheduled_insn;
2650
2651 /* Classify the instructions into three classes:
2652 1) Data dependent on last schedule insn.
2653 2) Anti/Output dependent on last scheduled insn.
2654 3) Independent of last scheduled insn, or has latency of one.
2655 Choose the insn from the highest numbered class if different. */
2656 dep1 = sd_find_dep_between (last, tmp, true);
2657
2658 if (dep1 == NULL || dep_cost (dep1) == 1)
2659 tmp_class = 3;
2660 else if (/* Data dependence. */
2661 DEP_TYPE (dep1) == REG_DEP_TRUE)
2662 tmp_class = 1;
2663 else
2664 tmp_class = 2;
2665
2666 dep2 = sd_find_dep_between (last, tmp2, true);
2667
2668 if (dep2 == NULL || dep_cost (dep2) == 1)
2669 tmp2_class = 3;
2670 else if (/* Data dependence. */
2671 DEP_TYPE (dep2) == REG_DEP_TRUE)
2672 tmp2_class = 1;
2673 else
2674 tmp2_class = 2;
2675
2676 if ((val = tmp2_class - tmp_class))
2677 return val;
2678 }
2679
2680 /* Prefer instructions that occur earlier in the model schedule. */
2681 if (sched_pressure == SCHED_PRESSURE_MODEL)
2682 {
2683 int diff;
2684
2685 diff = model_index (tmp) - model_index (tmp2);
2686 if (diff != 0)
2687 return diff;
2688 }
2689
2690 /* Prefer the insn which has more later insns that depend on it.
2691 This gives the scheduler more freedom when scheduling later
2692 instructions at the expense of added register pressure. */
2693
2694 val = (dep_list_size (tmp2, SD_LIST_FORW)
2695 - dep_list_size (tmp, SD_LIST_FORW));
2696
2697 if (flag_sched_dep_count_heuristic && val != 0)
2698 return val;
2699
2700 /* If insns are equally good, sort by INSN_LUID (original insn order),
2701 so that we make the sort stable. This minimizes instruction movement,
2702 thus minimizing sched's effect on debugging and cross-jumping. */
2703 return INSN_LUID (tmp) - INSN_LUID (tmp2);
2704 }
2705
2706 /* Resort the array A in which only element at index N may be out of order. */
2707
2708 HAIFA_INLINE static void
2709 swap_sort (rtx *a, int n)
2710 {
2711 rtx insn = a[n - 1];
2712 int i = n - 2;
2713
2714 while (i >= 0 && rank_for_schedule (a + i, &insn) >= 0)
2715 {
2716 a[i + 1] = a[i];
2717 i -= 1;
2718 }
2719 a[i + 1] = insn;
2720 }
2721
2722 /* Add INSN to the insn queue so that it can be executed at least
2723 N_CYCLES after the currently executing insn. Preserve insns
2724 chain for debugging purposes. REASON will be printed in debugging
2725 output. */
2726
2727 HAIFA_INLINE static void
2728 queue_insn (rtx insn, int n_cycles, const char *reason)
2729 {
2730 int next_q = NEXT_Q_AFTER (q_ptr, n_cycles);
2731 rtx link = alloc_INSN_LIST (insn, insn_queue[next_q]);
2732 int new_tick;
2733
2734 gcc_assert (n_cycles <= max_insn_queue_index);
2735 gcc_assert (!DEBUG_INSN_P (insn));
2736
2737 insn_queue[next_q] = link;
2738 q_size += 1;
2739
2740 if (sched_verbose >= 2)
2741 {
2742 fprintf (sched_dump, ";;\t\tReady-->Q: insn %s: ",
2743 (*current_sched_info->print_insn) (insn, 0));
2744
2745 fprintf (sched_dump, "queued for %d cycles (%s).\n", n_cycles, reason);
2746 }
2747
2748 QUEUE_INDEX (insn) = next_q;
2749
2750 if (current_sched_info->flags & DO_BACKTRACKING)
2751 {
2752 new_tick = clock_var + n_cycles;
2753 if (INSN_TICK (insn) == INVALID_TICK || INSN_TICK (insn) < new_tick)
2754 INSN_TICK (insn) = new_tick;
2755
2756 if (INSN_EXACT_TICK (insn) != INVALID_TICK
2757 && INSN_EXACT_TICK (insn) < clock_var + n_cycles)
2758 {
2759 must_backtrack = true;
2760 if (sched_verbose >= 2)
2761 fprintf (sched_dump, ";;\t\tcausing a backtrack.\n");
2762 }
2763 }
2764 }
2765
2766 /* Remove INSN from queue. */
2767 static void
2768 queue_remove (rtx insn)
2769 {
2770 gcc_assert (QUEUE_INDEX (insn) >= 0);
2771 remove_free_INSN_LIST_elem (insn, &insn_queue[QUEUE_INDEX (insn)]);
2772 q_size--;
2773 QUEUE_INDEX (insn) = QUEUE_NOWHERE;
2774 }
2775
2776 /* Return a pointer to the bottom of the ready list, i.e. the insn
2777 with the lowest priority. */
2778
2779 rtx *
2780 ready_lastpos (struct ready_list *ready)
2781 {
2782 gcc_assert (ready->n_ready >= 1);
2783 return ready->vec + ready->first - ready->n_ready + 1;
2784 }
2785
2786 /* Add an element INSN to the ready list so that it ends up with the
2787 lowest/highest priority depending on FIRST_P. */
2788
2789 HAIFA_INLINE static void
2790 ready_add (struct ready_list *ready, rtx insn, bool first_p)
2791 {
2792 if (!first_p)
2793 {
2794 if (ready->first == ready->n_ready)
2795 {
2796 memmove (ready->vec + ready->veclen - ready->n_ready,
2797 ready_lastpos (ready),
2798 ready->n_ready * sizeof (rtx));
2799 ready->first = ready->veclen - 1;
2800 }
2801 ready->vec[ready->first - ready->n_ready] = insn;
2802 }
2803 else
2804 {
2805 if (ready->first == ready->veclen - 1)
2806 {
2807 if (ready->n_ready)
2808 /* ready_lastpos() fails when called with (ready->n_ready == 0). */
2809 memmove (ready->vec + ready->veclen - ready->n_ready - 1,
2810 ready_lastpos (ready),
2811 ready->n_ready * sizeof (rtx));
2812 ready->first = ready->veclen - 2;
2813 }
2814 ready->vec[++(ready->first)] = insn;
2815 }
2816
2817 ready->n_ready++;
2818 if (DEBUG_INSN_P (insn))
2819 ready->n_debug++;
2820
2821 gcc_assert (QUEUE_INDEX (insn) != QUEUE_READY);
2822 QUEUE_INDEX (insn) = QUEUE_READY;
2823
2824 if (INSN_EXACT_TICK (insn) != INVALID_TICK
2825 && INSN_EXACT_TICK (insn) < clock_var)
2826 {
2827 must_backtrack = true;
2828 }
2829 }
2830
2831 /* Remove the element with the highest priority from the ready list and
2832 return it. */
2833
2834 HAIFA_INLINE static rtx
2835 ready_remove_first (struct ready_list *ready)
2836 {
2837 rtx t;
2838
2839 gcc_assert (ready->n_ready);
2840 t = ready->vec[ready->first--];
2841 ready->n_ready--;
2842 if (DEBUG_INSN_P (t))
2843 ready->n_debug--;
2844 /* If the queue becomes empty, reset it. */
2845 if (ready->n_ready == 0)
2846 ready->first = ready->veclen - 1;
2847
2848 gcc_assert (QUEUE_INDEX (t) == QUEUE_READY);
2849 QUEUE_INDEX (t) = QUEUE_NOWHERE;
2850
2851 return t;
2852 }
2853
2854 /* The following code implements multi-pass scheduling for the first
2855 cycle. In other words, we will try to choose ready insn which
2856 permits to start maximum number of insns on the same cycle. */
2857
2858 /* Return a pointer to the element INDEX from the ready. INDEX for
2859 insn with the highest priority is 0, and the lowest priority has
2860 N_READY - 1. */
2861
2862 rtx
2863 ready_element (struct ready_list *ready, int index)
2864 {
2865 gcc_assert (ready->n_ready && index < ready->n_ready);
2866
2867 return ready->vec[ready->first - index];
2868 }
2869
2870 /* Remove the element INDEX from the ready list and return it. INDEX
2871 for insn with the highest priority is 0, and the lowest priority
2872 has N_READY - 1. */
2873
2874 HAIFA_INLINE static rtx
2875 ready_remove (struct ready_list *ready, int index)
2876 {
2877 rtx t;
2878 int i;
2879
2880 if (index == 0)
2881 return ready_remove_first (ready);
2882 gcc_assert (ready->n_ready && index < ready->n_ready);
2883 t = ready->vec[ready->first - index];
2884 ready->n_ready--;
2885 if (DEBUG_INSN_P (t))
2886 ready->n_debug--;
2887 for (i = index; i < ready->n_ready; i++)
2888 ready->vec[ready->first - i] = ready->vec[ready->first - i - 1];
2889 QUEUE_INDEX (t) = QUEUE_NOWHERE;
2890 return t;
2891 }
2892
2893 /* Remove INSN from the ready list. */
2894 static void
2895 ready_remove_insn (rtx insn)
2896 {
2897 int i;
2898
2899 for (i = 0; i < readyp->n_ready; i++)
2900 if (ready_element (readyp, i) == insn)
2901 {
2902 ready_remove (readyp, i);
2903 return;
2904 }
2905 gcc_unreachable ();
2906 }
2907
2908 /* Sort the ready list READY by ascending priority, using the SCHED_SORT
2909 macro. */
2910
2911 void
2912 ready_sort (struct ready_list *ready)
2913 {
2914 int i;
2915 rtx *first = ready_lastpos (ready);
2916
2917 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
2918 {
2919 for (i = 0; i < ready->n_ready; i++)
2920 if (!DEBUG_INSN_P (first[i]))
2921 setup_insn_reg_pressure_info (first[i]);
2922 }
2923 if (sched_pressure == SCHED_PRESSURE_MODEL
2924 && model_curr_point < model_num_insns)
2925 model_set_excess_costs (first, ready->n_ready);
2926 SCHED_SORT (first, ready->n_ready);
2927 }
2928
2929 /* PREV is an insn that is ready to execute. Adjust its priority if that
2930 will help shorten or lengthen register lifetimes as appropriate. Also
2931 provide a hook for the target to tweak itself. */
2932
2933 HAIFA_INLINE static void
2934 adjust_priority (rtx prev)
2935 {
2936 /* ??? There used to be code here to try and estimate how an insn
2937 affected register lifetimes, but it did it by looking at REG_DEAD
2938 notes, which we removed in schedule_region. Nor did it try to
2939 take into account register pressure or anything useful like that.
2940
2941 Revisit when we have a machine model to work with and not before. */
2942
2943 if (targetm.sched.adjust_priority)
2944 INSN_PRIORITY (prev) =
2945 targetm.sched.adjust_priority (prev, INSN_PRIORITY (prev));
2946 }
2947
2948 /* Advance DFA state STATE on one cycle. */
2949 void
2950 advance_state (state_t state)
2951 {
2952 if (targetm.sched.dfa_pre_advance_cycle)
2953 targetm.sched.dfa_pre_advance_cycle ();
2954
2955 if (targetm.sched.dfa_pre_cycle_insn)
2956 state_transition (state,
2957 targetm.sched.dfa_pre_cycle_insn ());
2958
2959 state_transition (state, NULL);
2960
2961 if (targetm.sched.dfa_post_cycle_insn)
2962 state_transition (state,
2963 targetm.sched.dfa_post_cycle_insn ());
2964
2965 if (targetm.sched.dfa_post_advance_cycle)
2966 targetm.sched.dfa_post_advance_cycle ();
2967 }
2968
2969 /* Advance time on one cycle. */
2970 HAIFA_INLINE static void
2971 advance_one_cycle (void)
2972 {
2973 advance_state (curr_state);
2974 if (sched_verbose >= 6)
2975 fprintf (sched_dump, ";;\tAdvanced a state.\n");
2976 }
2977
2978 /* Update register pressure after scheduling INSN. */
2979 static void
2980 update_register_pressure (rtx insn)
2981 {
2982 struct reg_use_data *use;
2983 struct reg_set_data *set;
2984
2985 gcc_checking_assert (!DEBUG_INSN_P (insn));
2986
2987 for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
2988 if (dying_use_p (use))
2989 mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure,
2990 use->regno, false);
2991 for (set = INSN_REG_SET_LIST (insn); set != NULL; set = set->next_insn_set)
2992 mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure,
2993 set->regno, true);
2994 }
2995
2996 /* Set up or update (if UPDATE_P) max register pressure (see its
2997 meaning in sched-int.h::_haifa_insn_data) for all current BB insns
2998 after insn AFTER. */
2999 static void
3000 setup_insn_max_reg_pressure (rtx after, bool update_p)
3001 {
3002 int i, p;
3003 bool eq_p;
3004 rtx insn;
3005 static int max_reg_pressure[N_REG_CLASSES];
3006
3007 save_reg_pressure ();
3008 for (i = 0; i < ira_pressure_classes_num; i++)
3009 max_reg_pressure[ira_pressure_classes[i]]
3010 = curr_reg_pressure[ira_pressure_classes[i]];
3011 for (insn = NEXT_INSN (after);
3012 insn != NULL_RTX && ! BARRIER_P (insn)
3013 && BLOCK_FOR_INSN (insn) == BLOCK_FOR_INSN (after);
3014 insn = NEXT_INSN (insn))
3015 if (NONDEBUG_INSN_P (insn))
3016 {
3017 eq_p = true;
3018 for (i = 0; i < ira_pressure_classes_num; i++)
3019 {
3020 p = max_reg_pressure[ira_pressure_classes[i]];
3021 if (INSN_MAX_REG_PRESSURE (insn)[i] != p)
3022 {
3023 eq_p = false;
3024 INSN_MAX_REG_PRESSURE (insn)[i]
3025 = max_reg_pressure[ira_pressure_classes[i]];
3026 }
3027 }
3028 if (update_p && eq_p)
3029 break;
3030 update_register_pressure (insn);
3031 for (i = 0; i < ira_pressure_classes_num; i++)
3032 if (max_reg_pressure[ira_pressure_classes[i]]
3033 < curr_reg_pressure[ira_pressure_classes[i]])
3034 max_reg_pressure[ira_pressure_classes[i]]
3035 = curr_reg_pressure[ira_pressure_classes[i]];
3036 }
3037 restore_reg_pressure ();
3038 }
3039
3040 /* Update the current register pressure after scheduling INSN. Update
3041 also max register pressure for unscheduled insns of the current
3042 BB. */
3043 static void
3044 update_reg_and_insn_max_reg_pressure (rtx insn)
3045 {
3046 int i;
3047 int before[N_REG_CLASSES];
3048
3049 for (i = 0; i < ira_pressure_classes_num; i++)
3050 before[i] = curr_reg_pressure[ira_pressure_classes[i]];
3051 update_register_pressure (insn);
3052 for (i = 0; i < ira_pressure_classes_num; i++)
3053 if (curr_reg_pressure[ira_pressure_classes[i]] != before[i])
3054 break;
3055 if (i < ira_pressure_classes_num)
3056 setup_insn_max_reg_pressure (insn, true);
3057 }
3058
3059 /* Set up register pressure at the beginning of basic block BB whose
3060 insns starting after insn AFTER. Set up also max register pressure
3061 for all insns of the basic block. */
3062 void
3063 sched_setup_bb_reg_pressure_info (basic_block bb, rtx after)
3064 {
3065 gcc_assert (sched_pressure == SCHED_PRESSURE_WEIGHTED);
3066 initiate_bb_reg_pressure_info (bb);
3067 setup_insn_max_reg_pressure (after, false);
3068 }
3069 \f
3070 /* If doing predication while scheduling, verify whether INSN, which
3071 has just been scheduled, clobbers the conditions of any
3072 instructions that must be predicated in order to break their
3073 dependencies. If so, remove them from the queues so that they will
3074 only be scheduled once their control dependency is resolved. */
3075
3076 static void
3077 check_clobbered_conditions (rtx insn)
3078 {
3079 HARD_REG_SET t;
3080 int i;
3081
3082 if ((current_sched_info->flags & DO_PREDICATION) == 0)
3083 return;
3084
3085 find_all_hard_reg_sets (insn, &t);
3086
3087 restart:
3088 for (i = 0; i < ready.n_ready; i++)
3089 {
3090 rtx x = ready_element (&ready, i);
3091 if (TODO_SPEC (x) == DEP_CONTROL && cond_clobbered_p (x, t))
3092 {
3093 ready_remove_insn (x);
3094 goto restart;
3095 }
3096 }
3097 for (i = 0; i <= max_insn_queue_index; i++)
3098 {
3099 rtx link;
3100 int q = NEXT_Q_AFTER (q_ptr, i);
3101
3102 restart_queue:
3103 for (link = insn_queue[q]; link; link = XEXP (link, 1))
3104 {
3105 rtx x = XEXP (link, 0);
3106 if (TODO_SPEC (x) == DEP_CONTROL && cond_clobbered_p (x, t))
3107 {
3108 queue_remove (x);
3109 goto restart_queue;
3110 }
3111 }
3112 }
3113 }
3114 \f
3115 /* Return (in order):
3116
3117 - positive if INSN adversely affects the pressure on one
3118 register class
3119
3120 - negative if INSN reduces the pressure on one register class
3121
3122 - 0 if INSN doesn't affect the pressure on any register class. */
3123
3124 static int
3125 model_classify_pressure (struct model_insn_info *insn)
3126 {
3127 struct reg_pressure_data *reg_pressure;
3128 int death[N_REG_CLASSES];
3129 int pci, cl, sum;
3130
3131 calculate_reg_deaths (insn->insn, death);
3132 reg_pressure = INSN_REG_PRESSURE (insn->insn);
3133 sum = 0;
3134 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3135 {
3136 cl = ira_pressure_classes[pci];
3137 if (death[cl] < reg_pressure[pci].set_increase)
3138 return 1;
3139 sum += reg_pressure[pci].set_increase - death[cl];
3140 }
3141 return sum;
3142 }
3143
3144 /* Return true if INSN1 should come before INSN2 in the model schedule. */
3145
3146 static int
3147 model_order_p (struct model_insn_info *insn1, struct model_insn_info *insn2)
3148 {
3149 unsigned int height1, height2;
3150 unsigned int priority1, priority2;
3151
3152 /* Prefer instructions with a higher model priority. */
3153 if (insn1->model_priority != insn2->model_priority)
3154 return insn1->model_priority > insn2->model_priority;
3155
3156 /* Combine the length of the longest path of satisfied true dependencies
3157 that leads to each instruction (depth) with the length of the longest
3158 path of any dependencies that leads from the instruction (alap).
3159 Prefer instructions with the greatest combined length. If the combined
3160 lengths are equal, prefer instructions with the greatest depth.
3161
3162 The idea is that, if we have a set S of "equal" instructions that each
3163 have ALAP value X, and we pick one such instruction I, any true-dependent
3164 successors of I that have ALAP value X - 1 should be preferred over S.
3165 This encourages the schedule to be "narrow" rather than "wide".
3166 However, if I is a low-priority instruction that we decided to
3167 schedule because of its model_classify_pressure, and if there
3168 is a set of higher-priority instructions T, the aforementioned
3169 successors of I should not have the edge over T. */
3170 height1 = insn1->depth + insn1->alap;
3171 height2 = insn2->depth + insn2->alap;
3172 if (height1 != height2)
3173 return height1 > height2;
3174 if (insn1->depth != insn2->depth)
3175 return insn1->depth > insn2->depth;
3176
3177 /* We have no real preference between INSN1 an INSN2 as far as attempts
3178 to reduce pressure go. Prefer instructions with higher priorities. */
3179 priority1 = INSN_PRIORITY (insn1->insn);
3180 priority2 = INSN_PRIORITY (insn2->insn);
3181 if (priority1 != priority2)
3182 return priority1 > priority2;
3183
3184 /* Use the original rtl sequence as a tie-breaker. */
3185 return insn1 < insn2;
3186 }
3187
3188 /* Add INSN to the model worklist immediately after PREV. Add it to the
3189 beginning of the list if PREV is null. */
3190
3191 static void
3192 model_add_to_worklist_at (struct model_insn_info *insn,
3193 struct model_insn_info *prev)
3194 {
3195 gcc_assert (QUEUE_INDEX (insn->insn) == QUEUE_NOWHERE);
3196 QUEUE_INDEX (insn->insn) = QUEUE_READY;
3197
3198 insn->prev = prev;
3199 if (prev)
3200 {
3201 insn->next = prev->next;
3202 prev->next = insn;
3203 }
3204 else
3205 {
3206 insn->next = model_worklist;
3207 model_worklist = insn;
3208 }
3209 if (insn->next)
3210 insn->next->prev = insn;
3211 }
3212
3213 /* Remove INSN from the model worklist. */
3214
3215 static void
3216 model_remove_from_worklist (struct model_insn_info *insn)
3217 {
3218 gcc_assert (QUEUE_INDEX (insn->insn) == QUEUE_READY);
3219 QUEUE_INDEX (insn->insn) = QUEUE_NOWHERE;
3220
3221 if (insn->prev)
3222 insn->prev->next = insn->next;
3223 else
3224 model_worklist = insn->next;
3225 if (insn->next)
3226 insn->next->prev = insn->prev;
3227 }
3228
3229 /* Add INSN to the model worklist. Start looking for a suitable position
3230 between neighbors PREV and NEXT, testing at most MAX_SCHED_READY_INSNS
3231 insns either side. A null PREV indicates the beginning of the list and
3232 a null NEXT indicates the end. */
3233
3234 static void
3235 model_add_to_worklist (struct model_insn_info *insn,
3236 struct model_insn_info *prev,
3237 struct model_insn_info *next)
3238 {
3239 int count;
3240
3241 count = MAX_SCHED_READY_INSNS;
3242 if (count > 0 && prev && model_order_p (insn, prev))
3243 do
3244 {
3245 count--;
3246 prev = prev->prev;
3247 }
3248 while (count > 0 && prev && model_order_p (insn, prev));
3249 else
3250 while (count > 0 && next && model_order_p (next, insn))
3251 {
3252 count--;
3253 prev = next;
3254 next = next->next;
3255 }
3256 model_add_to_worklist_at (insn, prev);
3257 }
3258
3259 /* INSN may now have a higher priority (in the model_order_p sense)
3260 than before. Move it up the worklist if necessary. */
3261
3262 static void
3263 model_promote_insn (struct model_insn_info *insn)
3264 {
3265 struct model_insn_info *prev;
3266 int count;
3267
3268 prev = insn->prev;
3269 count = MAX_SCHED_READY_INSNS;
3270 while (count > 0 && prev && model_order_p (insn, prev))
3271 {
3272 count--;
3273 prev = prev->prev;
3274 }
3275 if (prev != insn->prev)
3276 {
3277 model_remove_from_worklist (insn);
3278 model_add_to_worklist_at (insn, prev);
3279 }
3280 }
3281
3282 /* Add INSN to the end of the model schedule. */
3283
3284 static void
3285 model_add_to_schedule (rtx insn)
3286 {
3287 unsigned int point;
3288
3289 gcc_assert (QUEUE_INDEX (insn) == QUEUE_NOWHERE);
3290 QUEUE_INDEX (insn) = QUEUE_SCHEDULED;
3291
3292 point = model_schedule.length ();
3293 model_schedule.quick_push (insn);
3294 INSN_MODEL_INDEX (insn) = point + 1;
3295 }
3296
3297 /* Analyze the instructions that are to be scheduled, setting up
3298 MODEL_INSN_INFO (...) and model_num_insns accordingly. Add ready
3299 instructions to model_worklist. */
3300
3301 static void
3302 model_analyze_insns (void)
3303 {
3304 rtx start, end, iter;
3305 sd_iterator_def sd_it;
3306 dep_t dep;
3307 struct model_insn_info *insn, *con;
3308
3309 model_num_insns = 0;
3310 start = PREV_INSN (current_sched_info->next_tail);
3311 end = current_sched_info->prev_head;
3312 for (iter = start; iter != end; iter = PREV_INSN (iter))
3313 if (NONDEBUG_INSN_P (iter))
3314 {
3315 insn = MODEL_INSN_INFO (iter);
3316 insn->insn = iter;
3317 FOR_EACH_DEP (iter, SD_LIST_FORW, sd_it, dep)
3318 {
3319 con = MODEL_INSN_INFO (DEP_CON (dep));
3320 if (con->insn && insn->alap < con->alap + 1)
3321 insn->alap = con->alap + 1;
3322 }
3323
3324 insn->old_queue = QUEUE_INDEX (iter);
3325 QUEUE_INDEX (iter) = QUEUE_NOWHERE;
3326
3327 insn->unscheduled_preds = dep_list_size (iter, SD_LIST_HARD_BACK);
3328 if (insn->unscheduled_preds == 0)
3329 model_add_to_worklist (insn, NULL, model_worklist);
3330
3331 model_num_insns++;
3332 }
3333 }
3334
3335 /* The global state describes the register pressure at the start of the
3336 model schedule. Initialize GROUP accordingly. */
3337
3338 static void
3339 model_init_pressure_group (struct model_pressure_group *group)
3340 {
3341 int pci, cl;
3342
3343 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3344 {
3345 cl = ira_pressure_classes[pci];
3346 group->limits[pci].pressure = curr_reg_pressure[cl];
3347 group->limits[pci].point = 0;
3348 }
3349 /* Use index model_num_insns to record the state after the last
3350 instruction in the model schedule. */
3351 group->model = XNEWVEC (struct model_pressure_data,
3352 (model_num_insns + 1) * ira_pressure_classes_num);
3353 }
3354
3355 /* Record that MODEL_REF_PRESSURE (GROUP, POINT, PCI) is PRESSURE.
3356 Update the maximum pressure for the whole schedule. */
3357
3358 static void
3359 model_record_pressure (struct model_pressure_group *group,
3360 int point, int pci, int pressure)
3361 {
3362 MODEL_REF_PRESSURE (group, point, pci) = pressure;
3363 if (group->limits[pci].pressure < pressure)
3364 {
3365 group->limits[pci].pressure = pressure;
3366 group->limits[pci].point = point;
3367 }
3368 }
3369
3370 /* INSN has just been added to the end of the model schedule. Record its
3371 register-pressure information. */
3372
3373 static void
3374 model_record_pressures (struct model_insn_info *insn)
3375 {
3376 struct reg_pressure_data *reg_pressure;
3377 int point, pci, cl, delta;
3378 int death[N_REG_CLASSES];
3379
3380 point = model_index (insn->insn);
3381 if (sched_verbose >= 2)
3382 {
3383 if (point == 0)
3384 {
3385 fprintf (sched_dump, "\n;;\tModel schedule:\n;;\n");
3386 fprintf (sched_dump, ";;\t| idx insn | mpri hght dpth prio |\n");
3387 }
3388 fprintf (sched_dump, ";;\t| %3d %4d | %4d %4d %4d %4d | %-30s ",
3389 point, INSN_UID (insn->insn), insn->model_priority,
3390 insn->depth + insn->alap, insn->depth,
3391 INSN_PRIORITY (insn->insn),
3392 str_pattern_slim (PATTERN (insn->insn)));
3393 }
3394 calculate_reg_deaths (insn->insn, death);
3395 reg_pressure = INSN_REG_PRESSURE (insn->insn);
3396 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3397 {
3398 cl = ira_pressure_classes[pci];
3399 delta = reg_pressure[pci].set_increase - death[cl];
3400 if (sched_verbose >= 2)
3401 fprintf (sched_dump, " %s:[%d,%+d]", reg_class_names[cl],
3402 curr_reg_pressure[cl], delta);
3403 model_record_pressure (&model_before_pressure, point, pci,
3404 curr_reg_pressure[cl]);
3405 }
3406 if (sched_verbose >= 2)
3407 fprintf (sched_dump, "\n");
3408 }
3409
3410 /* All instructions have been added to the model schedule. Record the
3411 final register pressure in GROUP and set up all MODEL_MAX_PRESSUREs. */
3412
3413 static void
3414 model_record_final_pressures (struct model_pressure_group *group)
3415 {
3416 int point, pci, max_pressure, ref_pressure, cl;
3417
3418 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3419 {
3420 /* Record the final pressure for this class. */
3421 cl = ira_pressure_classes[pci];
3422 point = model_num_insns;
3423 ref_pressure = curr_reg_pressure[cl];
3424 model_record_pressure (group, point, pci, ref_pressure);
3425
3426 /* Record the original maximum pressure. */
3427 group->limits[pci].orig_pressure = group->limits[pci].pressure;
3428
3429 /* Update the MODEL_MAX_PRESSURE for every point of the schedule. */
3430 max_pressure = ref_pressure;
3431 MODEL_MAX_PRESSURE (group, point, pci) = max_pressure;
3432 while (point > 0)
3433 {
3434 point--;
3435 ref_pressure = MODEL_REF_PRESSURE (group, point, pci);
3436 max_pressure = MAX (max_pressure, ref_pressure);
3437 MODEL_MAX_PRESSURE (group, point, pci) = max_pressure;
3438 }
3439 }
3440 }
3441
3442 /* Update all successors of INSN, given that INSN has just been scheduled. */
3443
3444 static void
3445 model_add_successors_to_worklist (struct model_insn_info *insn)
3446 {
3447 sd_iterator_def sd_it;
3448 struct model_insn_info *con;
3449 dep_t dep;
3450
3451 FOR_EACH_DEP (insn->insn, SD_LIST_FORW, sd_it, dep)
3452 {
3453 con = MODEL_INSN_INFO (DEP_CON (dep));
3454 /* Ignore debug instructions, and instructions from other blocks. */
3455 if (con->insn)
3456 {
3457 con->unscheduled_preds--;
3458
3459 /* Update the depth field of each true-dependent successor.
3460 Increasing the depth gives them a higher priority than
3461 before. */
3462 if (DEP_TYPE (dep) == REG_DEP_TRUE && con->depth < insn->depth + 1)
3463 {
3464 con->depth = insn->depth + 1;
3465 if (QUEUE_INDEX (con->insn) == QUEUE_READY)
3466 model_promote_insn (con);
3467 }
3468
3469 /* If this is a true dependency, or if there are no remaining
3470 dependencies for CON (meaning that CON only had non-true
3471 dependencies), make sure that CON is on the worklist.
3472 We don't bother otherwise because it would tend to fill the
3473 worklist with a lot of low-priority instructions that are not
3474 yet ready to issue. */
3475 if ((con->depth > 0 || con->unscheduled_preds == 0)
3476 && QUEUE_INDEX (con->insn) == QUEUE_NOWHERE)
3477 model_add_to_worklist (con, insn, insn->next);
3478 }
3479 }
3480 }
3481
3482 /* Give INSN a higher priority than any current instruction, then give
3483 unscheduled predecessors of INSN a higher priority still. If any of
3484 those predecessors are not on the model worklist, do the same for its
3485 predecessors, and so on. */
3486
3487 static void
3488 model_promote_predecessors (struct model_insn_info *insn)
3489 {
3490 struct model_insn_info *pro, *first;
3491 sd_iterator_def sd_it;
3492 dep_t dep;
3493
3494 if (sched_verbose >= 7)
3495 fprintf (sched_dump, ";;\t+--- priority of %d = %d, priority of",
3496 INSN_UID (insn->insn), model_next_priority);
3497 insn->model_priority = model_next_priority++;
3498 model_remove_from_worklist (insn);
3499 model_add_to_worklist_at (insn, NULL);
3500
3501 first = NULL;
3502 for (;;)
3503 {
3504 FOR_EACH_DEP (insn->insn, SD_LIST_HARD_BACK, sd_it, dep)
3505 {
3506 pro = MODEL_INSN_INFO (DEP_PRO (dep));
3507 /* The first test is to ignore debug instructions, and instructions
3508 from other blocks. */
3509 if (pro->insn
3510 && pro->model_priority != model_next_priority
3511 && QUEUE_INDEX (pro->insn) != QUEUE_SCHEDULED)
3512 {
3513 pro->model_priority = model_next_priority;
3514 if (sched_verbose >= 7)
3515 fprintf (sched_dump, " %d", INSN_UID (pro->insn));
3516 if (QUEUE_INDEX (pro->insn) == QUEUE_READY)
3517 {
3518 /* PRO is already in the worklist, but it now has
3519 a higher priority than before. Move it at the
3520 appropriate place. */
3521 model_remove_from_worklist (pro);
3522 model_add_to_worklist (pro, NULL, model_worklist);
3523 }
3524 else
3525 {
3526 /* PRO isn't in the worklist. Recursively process
3527 its predecessors until we find one that is. */
3528 pro->next = first;
3529 first = pro;
3530 }
3531 }
3532 }
3533 if (!first)
3534 break;
3535 insn = first;
3536 first = insn->next;
3537 }
3538 if (sched_verbose >= 7)
3539 fprintf (sched_dump, " = %d\n", model_next_priority);
3540 model_next_priority++;
3541 }
3542
3543 /* Pick one instruction from model_worklist and process it. */
3544
3545 static void
3546 model_choose_insn (void)
3547 {
3548 struct model_insn_info *insn, *fallback;
3549 int count;
3550
3551 if (sched_verbose >= 7)
3552 {
3553 fprintf (sched_dump, ";;\t+--- worklist:\n");
3554 insn = model_worklist;
3555 count = MAX_SCHED_READY_INSNS;
3556 while (count > 0 && insn)
3557 {
3558 fprintf (sched_dump, ";;\t+--- %d [%d, %d, %d, %d]\n",
3559 INSN_UID (insn->insn), insn->model_priority,
3560 insn->depth + insn->alap, insn->depth,
3561 INSN_PRIORITY (insn->insn));
3562 count--;
3563 insn = insn->next;
3564 }
3565 }
3566
3567 /* Look for a ready instruction whose model_classify_priority is zero
3568 or negative, picking the highest-priority one. Adding such an
3569 instruction to the schedule now should do no harm, and may actually
3570 do some good.
3571
3572 Failing that, see whether there is an instruction with the highest
3573 extant model_priority that is not yet ready, but which would reduce
3574 pressure if it became ready. This is designed to catch cases like:
3575
3576 (set (mem (reg R1)) (reg R2))
3577
3578 where the instruction is the last remaining use of R1 and where the
3579 value of R2 is not yet available (or vice versa). The death of R1
3580 means that this instruction already reduces pressure. It is of
3581 course possible that the computation of R2 involves other registers
3582 that are hard to kill, but such cases are rare enough for this
3583 heuristic to be a win in general.
3584
3585 Failing that, just pick the highest-priority instruction in the
3586 worklist. */
3587 count = MAX_SCHED_READY_INSNS;
3588 insn = model_worklist;
3589 fallback = 0;
3590 for (;;)
3591 {
3592 if (count == 0 || !insn)
3593 {
3594 insn = fallback ? fallback : model_worklist;
3595 break;
3596 }
3597 if (insn->unscheduled_preds)
3598 {
3599 if (model_worklist->model_priority == insn->model_priority
3600 && !fallback
3601 && model_classify_pressure (insn) < 0)
3602 fallback = insn;
3603 }
3604 else
3605 {
3606 if (model_classify_pressure (insn) <= 0)
3607 break;
3608 }
3609 count--;
3610 insn = insn->next;
3611 }
3612
3613 if (sched_verbose >= 7 && insn != model_worklist)
3614 {
3615 if (insn->unscheduled_preds)
3616 fprintf (sched_dump, ";;\t+--- promoting insn %d, with dependencies\n",
3617 INSN_UID (insn->insn));
3618 else
3619 fprintf (sched_dump, ";;\t+--- promoting insn %d, which is ready\n",
3620 INSN_UID (insn->insn));
3621 }
3622 if (insn->unscheduled_preds)
3623 /* INSN isn't yet ready to issue. Give all its predecessors the
3624 highest priority. */
3625 model_promote_predecessors (insn);
3626 else
3627 {
3628 /* INSN is ready. Add it to the end of model_schedule and
3629 process its successors. */
3630 model_add_successors_to_worklist (insn);
3631 model_remove_from_worklist (insn);
3632 model_add_to_schedule (insn->insn);
3633 model_record_pressures (insn);
3634 update_register_pressure (insn->insn);
3635 }
3636 }
3637
3638 /* Restore all QUEUE_INDEXs to the values that they had before
3639 model_start_schedule was called. */
3640
3641 static void
3642 model_reset_queue_indices (void)
3643 {
3644 unsigned int i;
3645 rtx insn;
3646
3647 FOR_EACH_VEC_ELT (model_schedule, i, insn)
3648 QUEUE_INDEX (insn) = MODEL_INSN_INFO (insn)->old_queue;
3649 }
3650
3651 /* We have calculated the model schedule and spill costs. Print a summary
3652 to sched_dump. */
3653
3654 static void
3655 model_dump_pressure_summary (void)
3656 {
3657 int pci, cl;
3658
3659 fprintf (sched_dump, ";; Pressure summary:");
3660 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3661 {
3662 cl = ira_pressure_classes[pci];
3663 fprintf (sched_dump, " %s:%d", reg_class_names[cl],
3664 model_before_pressure.limits[pci].pressure);
3665 }
3666 fprintf (sched_dump, "\n\n");
3667 }
3668
3669 /* Initialize the SCHED_PRESSURE_MODEL information for the current
3670 scheduling region. */
3671
3672 static void
3673 model_start_schedule (void)
3674 {
3675 basic_block bb;
3676
3677 model_next_priority = 1;
3678 model_schedule.create (sched_max_luid);
3679 model_insns = XCNEWVEC (struct model_insn_info, sched_max_luid);
3680
3681 bb = BLOCK_FOR_INSN (NEXT_INSN (current_sched_info->prev_head));
3682 initiate_reg_pressure_info (df_get_live_in (bb));
3683
3684 model_analyze_insns ();
3685 model_init_pressure_group (&model_before_pressure);
3686 while (model_worklist)
3687 model_choose_insn ();
3688 gcc_assert (model_num_insns == (int) model_schedule.length ());
3689 if (sched_verbose >= 2)
3690 fprintf (sched_dump, "\n");
3691
3692 model_record_final_pressures (&model_before_pressure);
3693 model_reset_queue_indices ();
3694
3695 XDELETEVEC (model_insns);
3696
3697 model_curr_point = 0;
3698 initiate_reg_pressure_info (df_get_live_in (bb));
3699 if (sched_verbose >= 1)
3700 model_dump_pressure_summary ();
3701 }
3702
3703 /* Free the information associated with GROUP. */
3704
3705 static void
3706 model_finalize_pressure_group (struct model_pressure_group *group)
3707 {
3708 XDELETEVEC (group->model);
3709 }
3710
3711 /* Free the information created by model_start_schedule. */
3712
3713 static void
3714 model_end_schedule (void)
3715 {
3716 model_finalize_pressure_group (&model_before_pressure);
3717 model_schedule.release ();
3718 }
3719 \f
3720 /* A structure that holds local state for the loop in schedule_block. */
3721 struct sched_block_state
3722 {
3723 /* True if no real insns have been scheduled in the current cycle. */
3724 bool first_cycle_insn_p;
3725 /* True if a shadow insn has been scheduled in the current cycle, which
3726 means that no more normal insns can be issued. */
3727 bool shadows_only_p;
3728 /* True if we're winding down a modulo schedule, which means that we only
3729 issue insns with INSN_EXACT_TICK set. */
3730 bool modulo_epilogue;
3731 /* Initialized with the machine's issue rate every cycle, and updated
3732 by calls to the variable_issue hook. */
3733 int can_issue_more;
3734 };
3735
3736 /* INSN is the "currently executing insn". Launch each insn which was
3737 waiting on INSN. READY is the ready list which contains the insns
3738 that are ready to fire. CLOCK is the current cycle. The function
3739 returns necessary cycle advance after issuing the insn (it is not
3740 zero for insns in a schedule group). */
3741
3742 static int
3743 schedule_insn (rtx insn)
3744 {
3745 sd_iterator_def sd_it;
3746 dep_t dep;
3747 int i;
3748 int advance = 0;
3749
3750 if (sched_verbose >= 1)
3751 {
3752 struct reg_pressure_data *pressure_info;
3753 fprintf (sched_dump, ";;\t%3i--> %s%-40s:",
3754 clock_var, (*current_sched_info->print_insn) (insn, 1),
3755 str_pattern_slim (PATTERN (insn)));
3756
3757 if (recog_memoized (insn) < 0)
3758 fprintf (sched_dump, "nothing");
3759 else
3760 print_reservation (sched_dump, insn);
3761 pressure_info = INSN_REG_PRESSURE (insn);
3762 if (pressure_info != NULL)
3763 {
3764 fputc (':', sched_dump);
3765 for (i = 0; i < ira_pressure_classes_num; i++)
3766 fprintf (sched_dump, "%s%s%+d(%d)",
3767 scheduled_insns.length () > 1
3768 && INSN_LUID (insn)
3769 < INSN_LUID (scheduled_insns[scheduled_insns.length () - 2]) ? "@" : "",
3770 reg_class_names[ira_pressure_classes[i]],
3771 pressure_info[i].set_increase, pressure_info[i].change);
3772 }
3773 if (sched_pressure == SCHED_PRESSURE_MODEL
3774 && model_curr_point < model_num_insns
3775 && model_index (insn) == model_curr_point)
3776 fprintf (sched_dump, ":model %d", model_curr_point);
3777 fputc ('\n', sched_dump);
3778 }
3779
3780 if (sched_pressure == SCHED_PRESSURE_WEIGHTED && !DEBUG_INSN_P (insn))
3781 update_reg_and_insn_max_reg_pressure (insn);
3782
3783 /* Scheduling instruction should have all its dependencies resolved and
3784 should have been removed from the ready list. */
3785 gcc_assert (sd_lists_empty_p (insn, SD_LIST_HARD_BACK));
3786
3787 /* Reset debug insns invalidated by moving this insn. */
3788 if (MAY_HAVE_DEBUG_INSNS && !DEBUG_INSN_P (insn))
3789 for (sd_it = sd_iterator_start (insn, SD_LIST_BACK);
3790 sd_iterator_cond (&sd_it, &dep);)
3791 {
3792 rtx dbg = DEP_PRO (dep);
3793 struct reg_use_data *use, *next;
3794
3795 if (DEP_STATUS (dep) & DEP_CANCELLED)
3796 {
3797 sd_iterator_next (&sd_it);
3798 continue;
3799 }
3800
3801 gcc_assert (DEBUG_INSN_P (dbg));
3802
3803 if (sched_verbose >= 6)
3804 fprintf (sched_dump, ";;\t\tresetting: debug insn %d\n",
3805 INSN_UID (dbg));
3806
3807 /* ??? Rather than resetting the debug insn, we might be able
3808 to emit a debug temp before the just-scheduled insn, but
3809 this would involve checking that the expression at the
3810 point of the debug insn is equivalent to the expression
3811 before the just-scheduled insn. They might not be: the
3812 expression in the debug insn may depend on other insns not
3813 yet scheduled that set MEMs, REGs or even other debug
3814 insns. It's not clear that attempting to preserve debug
3815 information in these cases is worth the effort, given how
3816 uncommon these resets are and the likelihood that the debug
3817 temps introduced won't survive the schedule change. */
3818 INSN_VAR_LOCATION_LOC (dbg) = gen_rtx_UNKNOWN_VAR_LOC ();
3819 df_insn_rescan (dbg);
3820
3821 /* Unknown location doesn't use any registers. */
3822 for (use = INSN_REG_USE_LIST (dbg); use != NULL; use = next)
3823 {
3824 struct reg_use_data *prev = use;
3825
3826 /* Remove use from the cyclic next_regno_use chain first. */
3827 while (prev->next_regno_use != use)
3828 prev = prev->next_regno_use;
3829 prev->next_regno_use = use->next_regno_use;
3830 next = use->next_insn_use;
3831 free (use);
3832 }
3833 INSN_REG_USE_LIST (dbg) = NULL;
3834
3835 /* We delete rather than resolve these deps, otherwise we
3836 crash in sched_free_deps(), because forward deps are
3837 expected to be released before backward deps. */
3838 sd_delete_dep (sd_it);
3839 }
3840
3841 gcc_assert (QUEUE_INDEX (insn) == QUEUE_NOWHERE);
3842 QUEUE_INDEX (insn) = QUEUE_SCHEDULED;
3843
3844 if (sched_pressure == SCHED_PRESSURE_MODEL
3845 && model_curr_point < model_num_insns
3846 && NONDEBUG_INSN_P (insn))
3847 {
3848 if (model_index (insn) == model_curr_point)
3849 do
3850 model_curr_point++;
3851 while (model_curr_point < model_num_insns
3852 && (QUEUE_INDEX (MODEL_INSN (model_curr_point))
3853 == QUEUE_SCHEDULED));
3854 else
3855 model_recompute (insn);
3856 model_update_limit_points ();
3857 update_register_pressure (insn);
3858 if (sched_verbose >= 2)
3859 print_curr_reg_pressure ();
3860 }
3861
3862 gcc_assert (INSN_TICK (insn) >= MIN_TICK);
3863 if (INSN_TICK (insn) > clock_var)
3864 /* INSN has been prematurely moved from the queue to the ready list.
3865 This is possible only if following flag is set. */
3866 gcc_assert (flag_sched_stalled_insns);
3867
3868 /* ??? Probably, if INSN is scheduled prematurely, we should leave
3869 INSN_TICK untouched. This is a machine-dependent issue, actually. */
3870 INSN_TICK (insn) = clock_var;
3871
3872 check_clobbered_conditions (insn);
3873
3874 /* Update dependent instructions. First, see if by scheduling this insn
3875 now we broke a dependence in a way that requires us to change another
3876 insn. */
3877 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
3878 sd_iterator_cond (&sd_it, &dep); sd_iterator_next (&sd_it))
3879 {
3880 struct dep_replacement *desc = DEP_REPLACE (dep);
3881 rtx pro = DEP_PRO (dep);
3882 if (QUEUE_INDEX (pro) != QUEUE_SCHEDULED
3883 && desc != NULL && desc->insn == pro)
3884 apply_replacement (dep, false);
3885 }
3886
3887 /* Go through and resolve forward dependencies. */
3888 for (sd_it = sd_iterator_start (insn, SD_LIST_FORW);
3889 sd_iterator_cond (&sd_it, &dep);)
3890 {
3891 rtx next = DEP_CON (dep);
3892 bool cancelled = (DEP_STATUS (dep) & DEP_CANCELLED) != 0;
3893
3894 /* Resolve the dependence between INSN and NEXT.
3895 sd_resolve_dep () moves current dep to another list thus
3896 advancing the iterator. */
3897 sd_resolve_dep (sd_it);
3898
3899 if (cancelled)
3900 {
3901 if (must_restore_pattern_p (next, dep))
3902 restore_pattern (dep, false);
3903 continue;
3904 }
3905
3906 /* Don't bother trying to mark next as ready if insn is a debug
3907 insn. If insn is the last hard dependency, it will have
3908 already been discounted. */
3909 if (DEBUG_INSN_P (insn) && !DEBUG_INSN_P (next))
3910 continue;
3911
3912 if (!IS_SPECULATION_BRANCHY_CHECK_P (insn))
3913 {
3914 int effective_cost;
3915
3916 effective_cost = try_ready (next);
3917
3918 if (effective_cost >= 0
3919 && SCHED_GROUP_P (next)
3920 && advance < effective_cost)
3921 advance = effective_cost;
3922 }
3923 else
3924 /* Check always has only one forward dependence (to the first insn in
3925 the recovery block), therefore, this will be executed only once. */
3926 {
3927 gcc_assert (sd_lists_empty_p (insn, SD_LIST_FORW));
3928 fix_recovery_deps (RECOVERY_BLOCK (insn));
3929 }
3930 }
3931
3932 /* Annotate the instruction with issue information -- TImode
3933 indicates that the instruction is expected not to be able
3934 to issue on the same cycle as the previous insn. A machine
3935 may use this information to decide how the instruction should
3936 be aligned. */
3937 if (issue_rate > 1
3938 && GET_CODE (PATTERN (insn)) != USE
3939 && GET_CODE (PATTERN (insn)) != CLOBBER
3940 && !DEBUG_INSN_P (insn))
3941 {
3942 if (reload_completed)
3943 PUT_MODE (insn, clock_var > last_clock_var ? TImode : VOIDmode);
3944 last_clock_var = clock_var;
3945 }
3946
3947 return advance;
3948 }
3949
3950 /* Functions for handling of notes. */
3951
3952 /* Add note list that ends on FROM_END to the end of TO_ENDP. */
3953 void
3954 concat_note_lists (rtx from_end, rtx *to_endp)
3955 {
3956 rtx from_start;
3957
3958 /* It's easy when have nothing to concat. */
3959 if (from_end == NULL)
3960 return;
3961
3962 /* It's also easy when destination is empty. */
3963 if (*to_endp == NULL)
3964 {
3965 *to_endp = from_end;
3966 return;
3967 }
3968
3969 from_start = from_end;
3970 while (PREV_INSN (from_start) != NULL)
3971 from_start = PREV_INSN (from_start);
3972
3973 PREV_INSN (from_start) = *to_endp;
3974 NEXT_INSN (*to_endp) = from_start;
3975 *to_endp = from_end;
3976 }
3977
3978 /* Delete notes between HEAD and TAIL and put them in the chain
3979 of notes ended by NOTE_LIST. */
3980 void
3981 remove_notes (rtx head, rtx tail)
3982 {
3983 rtx next_tail, insn, next;
3984
3985 note_list = 0;
3986 if (head == tail && !INSN_P (head))
3987 return;
3988
3989 next_tail = NEXT_INSN (tail);
3990 for (insn = head; insn != next_tail; insn = next)
3991 {
3992 next = NEXT_INSN (insn);
3993 if (!NOTE_P (insn))
3994 continue;
3995
3996 switch (NOTE_KIND (insn))
3997 {
3998 case NOTE_INSN_BASIC_BLOCK:
3999 continue;
4000
4001 case NOTE_INSN_EPILOGUE_BEG:
4002 if (insn != tail)
4003 {
4004 remove_insn (insn);
4005 add_reg_note (next, REG_SAVE_NOTE,
4006 GEN_INT (NOTE_INSN_EPILOGUE_BEG));
4007 break;
4008 }
4009 /* FALLTHRU */
4010
4011 default:
4012 remove_insn (insn);
4013
4014 /* Add the note to list that ends at NOTE_LIST. */
4015 PREV_INSN (insn) = note_list;
4016 NEXT_INSN (insn) = NULL_RTX;
4017 if (note_list)
4018 NEXT_INSN (note_list) = insn;
4019 note_list = insn;
4020 break;
4021 }
4022
4023 gcc_assert ((sel_sched_p () || insn != tail) && insn != head);
4024 }
4025 }
4026
4027 /* A structure to record enough data to allow us to backtrack the scheduler to
4028 a previous state. */
4029 struct haifa_saved_data
4030 {
4031 /* Next entry on the list. */
4032 struct haifa_saved_data *next;
4033
4034 /* Backtracking is associated with scheduling insns that have delay slots.
4035 DELAY_PAIR points to the structure that contains the insns involved, and
4036 the number of cycles between them. */
4037 struct delay_pair *delay_pair;
4038
4039 /* Data used by the frontend (e.g. sched-ebb or sched-rgn). */
4040 void *fe_saved_data;
4041 /* Data used by the backend. */
4042 void *be_saved_data;
4043
4044 /* Copies of global state. */
4045 int clock_var, last_clock_var;
4046 struct ready_list ready;
4047 state_t curr_state;
4048
4049 rtx last_scheduled_insn;
4050 rtx last_nondebug_scheduled_insn;
4051 int cycle_issued_insns;
4052
4053 /* Copies of state used in the inner loop of schedule_block. */
4054 struct sched_block_state sched_block;
4055
4056 /* We don't need to save q_ptr, as its value is arbitrary and we can set it
4057 to 0 when restoring. */
4058 int q_size;
4059 rtx *insn_queue;
4060
4061 /* Describe pattern replacements that occurred since this backtrack point
4062 was queued. */
4063 vec<dep_t> replacement_deps;
4064 vec<int> replace_apply;
4065
4066 /* A copy of the next-cycle replacement vectors at the time of the backtrack
4067 point. */
4068 vec<dep_t> next_cycle_deps;
4069 vec<int> next_cycle_apply;
4070 };
4071
4072 /* A record, in reverse order, of all scheduled insns which have delay slots
4073 and may require backtracking. */
4074 static struct haifa_saved_data *backtrack_queue;
4075
4076 /* For every dependency of INSN, set the FEEDS_BACKTRACK_INSN bit according
4077 to SET_P. */
4078 static void
4079 mark_backtrack_feeds (rtx insn, int set_p)
4080 {
4081 sd_iterator_def sd_it;
4082 dep_t dep;
4083 FOR_EACH_DEP (insn, SD_LIST_HARD_BACK, sd_it, dep)
4084 {
4085 FEEDS_BACKTRACK_INSN (DEP_PRO (dep)) = set_p;
4086 }
4087 }
4088
4089 /* Save the current scheduler state so that we can backtrack to it
4090 later if necessary. PAIR gives the insns that make it necessary to
4091 save this point. SCHED_BLOCK is the local state of schedule_block
4092 that need to be saved. */
4093 static void
4094 save_backtrack_point (struct delay_pair *pair,
4095 struct sched_block_state sched_block)
4096 {
4097 int i;
4098 struct haifa_saved_data *save = XNEW (struct haifa_saved_data);
4099
4100 save->curr_state = xmalloc (dfa_state_size);
4101 memcpy (save->curr_state, curr_state, dfa_state_size);
4102
4103 save->ready.first = ready.first;
4104 save->ready.n_ready = ready.n_ready;
4105 save->ready.n_debug = ready.n_debug;
4106 save->ready.veclen = ready.veclen;
4107 save->ready.vec = XNEWVEC (rtx, ready.veclen);
4108 memcpy (save->ready.vec, ready.vec, ready.veclen * sizeof (rtx));
4109
4110 save->insn_queue = XNEWVEC (rtx, max_insn_queue_index + 1);
4111 save->q_size = q_size;
4112 for (i = 0; i <= max_insn_queue_index; i++)
4113 {
4114 int q = NEXT_Q_AFTER (q_ptr, i);
4115 save->insn_queue[i] = copy_INSN_LIST (insn_queue[q]);
4116 }
4117
4118 save->clock_var = clock_var;
4119 save->last_clock_var = last_clock_var;
4120 save->cycle_issued_insns = cycle_issued_insns;
4121 save->last_scheduled_insn = last_scheduled_insn;
4122 save->last_nondebug_scheduled_insn = last_nondebug_scheduled_insn;
4123
4124 save->sched_block = sched_block;
4125
4126 save->replacement_deps.create (0);
4127 save->replace_apply.create (0);
4128 save->next_cycle_deps = next_cycle_replace_deps.copy ();
4129 save->next_cycle_apply = next_cycle_apply.copy ();
4130
4131 if (current_sched_info->save_state)
4132 save->fe_saved_data = (*current_sched_info->save_state) ();
4133
4134 if (targetm.sched.alloc_sched_context)
4135 {
4136 save->be_saved_data = targetm.sched.alloc_sched_context ();
4137 targetm.sched.init_sched_context (save->be_saved_data, false);
4138 }
4139 else
4140 save->be_saved_data = NULL;
4141
4142 save->delay_pair = pair;
4143
4144 save->next = backtrack_queue;
4145 backtrack_queue = save;
4146
4147 while (pair)
4148 {
4149 mark_backtrack_feeds (pair->i2, 1);
4150 INSN_TICK (pair->i2) = INVALID_TICK;
4151 INSN_EXACT_TICK (pair->i2) = clock_var + pair_delay (pair);
4152 SHADOW_P (pair->i2) = pair->stages == 0;
4153 pair = pair->next_same_i1;
4154 }
4155 }
4156
4157 /* Walk the ready list and all queues. If any insns have unresolved backwards
4158 dependencies, these must be cancelled deps, broken by predication. Set or
4159 clear (depending on SET) the DEP_CANCELLED bit in DEP_STATUS. */
4160
4161 static void
4162 toggle_cancelled_flags (bool set)
4163 {
4164 int i;
4165 sd_iterator_def sd_it;
4166 dep_t dep;
4167
4168 if (ready.n_ready > 0)
4169 {
4170 rtx *first = ready_lastpos (&ready);
4171 for (i = 0; i < ready.n_ready; i++)
4172 FOR_EACH_DEP (first[i], SD_LIST_BACK, sd_it, dep)
4173 if (!DEBUG_INSN_P (DEP_PRO (dep)))
4174 {
4175 if (set)
4176 DEP_STATUS (dep) |= DEP_CANCELLED;
4177 else
4178 DEP_STATUS (dep) &= ~DEP_CANCELLED;
4179 }
4180 }
4181 for (i = 0; i <= max_insn_queue_index; i++)
4182 {
4183 int q = NEXT_Q_AFTER (q_ptr, i);
4184 rtx link;
4185 for (link = insn_queue[q]; link; link = XEXP (link, 1))
4186 {
4187 rtx insn = XEXP (link, 0);
4188 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
4189 if (!DEBUG_INSN_P (DEP_PRO (dep)))
4190 {
4191 if (set)
4192 DEP_STATUS (dep) |= DEP_CANCELLED;
4193 else
4194 DEP_STATUS (dep) &= ~DEP_CANCELLED;
4195 }
4196 }
4197 }
4198 }
4199
4200 /* Undo the replacements that have occurred after backtrack point SAVE
4201 was placed. */
4202 static void
4203 undo_replacements_for_backtrack (struct haifa_saved_data *save)
4204 {
4205 while (!save->replacement_deps.is_empty ())
4206 {
4207 dep_t dep = save->replacement_deps.pop ();
4208 int apply_p = save->replace_apply.pop ();
4209
4210 if (apply_p)
4211 restore_pattern (dep, true);
4212 else
4213 apply_replacement (dep, true);
4214 }
4215 save->replacement_deps.release ();
4216 save->replace_apply.release ();
4217 }
4218
4219 /* Pop entries from the SCHEDULED_INSNS vector up to and including INSN.
4220 Restore their dependencies to an unresolved state, and mark them as
4221 queued nowhere. */
4222
4223 static void
4224 unschedule_insns_until (rtx insn)
4225 {
4226 auto_vec<rtx> recompute_vec;
4227
4228 /* Make two passes over the insns to be unscheduled. First, we clear out
4229 dependencies and other trivial bookkeeping. */
4230 for (;;)
4231 {
4232 rtx last;
4233 sd_iterator_def sd_it;
4234 dep_t dep;
4235
4236 last = scheduled_insns.pop ();
4237
4238 /* This will be changed by restore_backtrack_point if the insn is in
4239 any queue. */
4240 QUEUE_INDEX (last) = QUEUE_NOWHERE;
4241 if (last != insn)
4242 INSN_TICK (last) = INVALID_TICK;
4243
4244 if (modulo_ii > 0 && INSN_UID (last) < modulo_iter0_max_uid)
4245 modulo_insns_scheduled--;
4246
4247 for (sd_it = sd_iterator_start (last, SD_LIST_RES_FORW);
4248 sd_iterator_cond (&sd_it, &dep);)
4249 {
4250 rtx con = DEP_CON (dep);
4251 sd_unresolve_dep (sd_it);
4252 if (!MUST_RECOMPUTE_SPEC_P (con))
4253 {
4254 MUST_RECOMPUTE_SPEC_P (con) = 1;
4255 recompute_vec.safe_push (con);
4256 }
4257 }
4258
4259 if (last == insn)
4260 break;
4261 }
4262
4263 /* A second pass, to update ready and speculation status for insns
4264 depending on the unscheduled ones. The first pass must have
4265 popped the scheduled_insns vector up to the point where we
4266 restart scheduling, as recompute_todo_spec requires it to be
4267 up-to-date. */
4268 while (!recompute_vec.is_empty ())
4269 {
4270 rtx con;
4271
4272 con = recompute_vec.pop ();
4273 MUST_RECOMPUTE_SPEC_P (con) = 0;
4274 if (!sd_lists_empty_p (con, SD_LIST_HARD_BACK))
4275 {
4276 TODO_SPEC (con) = HARD_DEP;
4277 INSN_TICK (con) = INVALID_TICK;
4278 if (PREDICATED_PAT (con) != NULL_RTX)
4279 haifa_change_pattern (con, ORIG_PAT (con));
4280 }
4281 else if (QUEUE_INDEX (con) != QUEUE_SCHEDULED)
4282 TODO_SPEC (con) = recompute_todo_spec (con, true);
4283 }
4284 }
4285
4286 /* Restore scheduler state from the topmost entry on the backtracking queue.
4287 PSCHED_BLOCK_P points to the local data of schedule_block that we must
4288 overwrite with the saved data.
4289 The caller must already have called unschedule_insns_until. */
4290
4291 static void
4292 restore_last_backtrack_point (struct sched_block_state *psched_block)
4293 {
4294 rtx link;
4295 int i;
4296 struct haifa_saved_data *save = backtrack_queue;
4297
4298 backtrack_queue = save->next;
4299
4300 if (current_sched_info->restore_state)
4301 (*current_sched_info->restore_state) (save->fe_saved_data);
4302
4303 if (targetm.sched.alloc_sched_context)
4304 {
4305 targetm.sched.set_sched_context (save->be_saved_data);
4306 targetm.sched.free_sched_context (save->be_saved_data);
4307 }
4308
4309 /* Do this first since it clobbers INSN_TICK of the involved
4310 instructions. */
4311 undo_replacements_for_backtrack (save);
4312
4313 /* Clear the QUEUE_INDEX of everything in the ready list or one
4314 of the queues. */
4315 if (ready.n_ready > 0)
4316 {
4317 rtx *first = ready_lastpos (&ready);
4318 for (i = 0; i < ready.n_ready; i++)
4319 {
4320 rtx insn = first[i];
4321 QUEUE_INDEX (insn) = QUEUE_NOWHERE;
4322 INSN_TICK (insn) = INVALID_TICK;
4323 }
4324 }
4325 for (i = 0; i <= max_insn_queue_index; i++)
4326 {
4327 int q = NEXT_Q_AFTER (q_ptr, i);
4328
4329 for (link = insn_queue[q]; link; link = XEXP (link, 1))
4330 {
4331 rtx x = XEXP (link, 0);
4332 QUEUE_INDEX (x) = QUEUE_NOWHERE;
4333 INSN_TICK (x) = INVALID_TICK;
4334 }
4335 free_INSN_LIST_list (&insn_queue[q]);
4336 }
4337
4338 free (ready.vec);
4339 ready = save->ready;
4340
4341 if (ready.n_ready > 0)
4342 {
4343 rtx *first = ready_lastpos (&ready);
4344 for (i = 0; i < ready.n_ready; i++)
4345 {
4346 rtx insn = first[i];
4347 QUEUE_INDEX (insn) = QUEUE_READY;
4348 TODO_SPEC (insn) = recompute_todo_spec (insn, true);
4349 INSN_TICK (insn) = save->clock_var;
4350 }
4351 }
4352
4353 q_ptr = 0;
4354 q_size = save->q_size;
4355 for (i = 0; i <= max_insn_queue_index; i++)
4356 {
4357 int q = NEXT_Q_AFTER (q_ptr, i);
4358
4359 insn_queue[q] = save->insn_queue[q];
4360
4361 for (link = insn_queue[q]; link; link = XEXP (link, 1))
4362 {
4363 rtx x = XEXP (link, 0);
4364 QUEUE_INDEX (x) = i;
4365 TODO_SPEC (x) = recompute_todo_spec (x, true);
4366 INSN_TICK (x) = save->clock_var + i;
4367 }
4368 }
4369 free (save->insn_queue);
4370
4371 toggle_cancelled_flags (true);
4372
4373 clock_var = save->clock_var;
4374 last_clock_var = save->last_clock_var;
4375 cycle_issued_insns = save->cycle_issued_insns;
4376 last_scheduled_insn = save->last_scheduled_insn;
4377 last_nondebug_scheduled_insn = save->last_nondebug_scheduled_insn;
4378
4379 *psched_block = save->sched_block;
4380
4381 memcpy (curr_state, save->curr_state, dfa_state_size);
4382 free (save->curr_state);
4383
4384 mark_backtrack_feeds (save->delay_pair->i2, 0);
4385
4386 gcc_assert (next_cycle_replace_deps.is_empty ());
4387 next_cycle_replace_deps = save->next_cycle_deps.copy ();
4388 next_cycle_apply = save->next_cycle_apply.copy ();
4389
4390 free (save);
4391
4392 for (save = backtrack_queue; save; save = save->next)
4393 {
4394 mark_backtrack_feeds (save->delay_pair->i2, 1);
4395 }
4396 }
4397
4398 /* Discard all data associated with the topmost entry in the backtrack
4399 queue. If RESET_TICK is false, we just want to free the data. If true,
4400 we are doing this because we discovered a reason to backtrack. In the
4401 latter case, also reset the INSN_TICK for the shadow insn. */
4402 static void
4403 free_topmost_backtrack_point (bool reset_tick)
4404 {
4405 struct haifa_saved_data *save = backtrack_queue;
4406 int i;
4407
4408 backtrack_queue = save->next;
4409
4410 if (reset_tick)
4411 {
4412 struct delay_pair *pair = save->delay_pair;
4413 while (pair)
4414 {
4415 INSN_TICK (pair->i2) = INVALID_TICK;
4416 INSN_EXACT_TICK (pair->i2) = INVALID_TICK;
4417 pair = pair->next_same_i1;
4418 }
4419 undo_replacements_for_backtrack (save);
4420 }
4421 else
4422 {
4423 save->replacement_deps.release ();
4424 save->replace_apply.release ();
4425 }
4426
4427 if (targetm.sched.free_sched_context)
4428 targetm.sched.free_sched_context (save->be_saved_data);
4429 if (current_sched_info->restore_state)
4430 free (save->fe_saved_data);
4431 for (i = 0; i <= max_insn_queue_index; i++)
4432 free_INSN_LIST_list (&save->insn_queue[i]);
4433 free (save->insn_queue);
4434 free (save->curr_state);
4435 free (save->ready.vec);
4436 free (save);
4437 }
4438
4439 /* Free the entire backtrack queue. */
4440 static void
4441 free_backtrack_queue (void)
4442 {
4443 while (backtrack_queue)
4444 free_topmost_backtrack_point (false);
4445 }
4446
4447 /* Apply a replacement described by DESC. If IMMEDIATELY is false, we
4448 may have to postpone the replacement until the start of the next cycle,
4449 at which point we will be called again with IMMEDIATELY true. This is
4450 only done for machines which have instruction packets with explicit
4451 parallelism however. */
4452 static void
4453 apply_replacement (dep_t dep, bool immediately)
4454 {
4455 struct dep_replacement *desc = DEP_REPLACE (dep);
4456 if (!immediately && targetm.sched.exposed_pipeline && reload_completed)
4457 {
4458 next_cycle_replace_deps.safe_push (dep);
4459 next_cycle_apply.safe_push (1);
4460 }
4461 else
4462 {
4463 bool success;
4464
4465 if (QUEUE_INDEX (desc->insn) == QUEUE_SCHEDULED)
4466 return;
4467
4468 if (sched_verbose >= 5)
4469 fprintf (sched_dump, "applying replacement for insn %d\n",
4470 INSN_UID (desc->insn));
4471
4472 success = validate_change (desc->insn, desc->loc, desc->newval, 0);
4473 gcc_assert (success);
4474
4475 update_insn_after_change (desc->insn);
4476 if ((TODO_SPEC (desc->insn) & (HARD_DEP | DEP_POSTPONED)) == 0)
4477 fix_tick_ready (desc->insn);
4478
4479 if (backtrack_queue != NULL)
4480 {
4481 backtrack_queue->replacement_deps.safe_push (dep);
4482 backtrack_queue->replace_apply.safe_push (1);
4483 }
4484 }
4485 }
4486
4487 /* We have determined that a pattern involved in DEP must be restored.
4488 If IMMEDIATELY is false, we may have to postpone the replacement
4489 until the start of the next cycle, at which point we will be called
4490 again with IMMEDIATELY true. */
4491 static void
4492 restore_pattern (dep_t dep, bool immediately)
4493 {
4494 rtx next = DEP_CON (dep);
4495 int tick = INSN_TICK (next);
4496
4497 /* If we already scheduled the insn, the modified version is
4498 correct. */
4499 if (QUEUE_INDEX (next) == QUEUE_SCHEDULED)
4500 return;
4501
4502 if (!immediately && targetm.sched.exposed_pipeline && reload_completed)
4503 {
4504 next_cycle_replace_deps.safe_push (dep);
4505 next_cycle_apply.safe_push (0);
4506 return;
4507 }
4508
4509
4510 if (DEP_TYPE (dep) == REG_DEP_CONTROL)
4511 {
4512 if (sched_verbose >= 5)
4513 fprintf (sched_dump, "restoring pattern for insn %d\n",
4514 INSN_UID (next));
4515 haifa_change_pattern (next, ORIG_PAT (next));
4516 }
4517 else
4518 {
4519 struct dep_replacement *desc = DEP_REPLACE (dep);
4520 bool success;
4521
4522 if (sched_verbose >= 5)
4523 fprintf (sched_dump, "restoring pattern for insn %d\n",
4524 INSN_UID (desc->insn));
4525 tick = INSN_TICK (desc->insn);
4526
4527 success = validate_change (desc->insn, desc->loc, desc->orig, 0);
4528 gcc_assert (success);
4529 update_insn_after_change (desc->insn);
4530 if (backtrack_queue != NULL)
4531 {
4532 backtrack_queue->replacement_deps.safe_push (dep);
4533 backtrack_queue->replace_apply.safe_push (0);
4534 }
4535 }
4536 INSN_TICK (next) = tick;
4537 if (TODO_SPEC (next) == DEP_POSTPONED)
4538 return;
4539
4540 if (sd_lists_empty_p (next, SD_LIST_BACK))
4541 TODO_SPEC (next) = 0;
4542 else if (!sd_lists_empty_p (next, SD_LIST_HARD_BACK))
4543 TODO_SPEC (next) = HARD_DEP;
4544 }
4545
4546 /* Perform pattern replacements that were queued up until the next
4547 cycle. */
4548 static void
4549 perform_replacements_new_cycle (void)
4550 {
4551 int i;
4552 dep_t dep;
4553 FOR_EACH_VEC_ELT (next_cycle_replace_deps, i, dep)
4554 {
4555 int apply_p = next_cycle_apply[i];
4556 if (apply_p)
4557 apply_replacement (dep, true);
4558 else
4559 restore_pattern (dep, true);
4560 }
4561 next_cycle_replace_deps.truncate (0);
4562 next_cycle_apply.truncate (0);
4563 }
4564
4565 /* Compute INSN_TICK_ESTIMATE for INSN. PROCESSED is a bitmap of
4566 instructions we've previously encountered, a set bit prevents
4567 recursion. BUDGET is a limit on how far ahead we look, it is
4568 reduced on recursive calls. Return true if we produced a good
4569 estimate, or false if we exceeded the budget. */
4570 static bool
4571 estimate_insn_tick (bitmap processed, rtx insn, int budget)
4572 {
4573 sd_iterator_def sd_it;
4574 dep_t dep;
4575 int earliest = INSN_TICK (insn);
4576
4577 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
4578 {
4579 rtx pro = DEP_PRO (dep);
4580 int t;
4581
4582 if (DEP_STATUS (dep) & DEP_CANCELLED)
4583 continue;
4584
4585 if (QUEUE_INDEX (pro) == QUEUE_SCHEDULED)
4586 gcc_assert (INSN_TICK (pro) + dep_cost (dep) <= INSN_TICK (insn));
4587 else
4588 {
4589 int cost = dep_cost (dep);
4590 if (cost >= budget)
4591 return false;
4592 if (!bitmap_bit_p (processed, INSN_LUID (pro)))
4593 {
4594 if (!estimate_insn_tick (processed, pro, budget - cost))
4595 return false;
4596 }
4597 gcc_assert (INSN_TICK_ESTIMATE (pro) != INVALID_TICK);
4598 t = INSN_TICK_ESTIMATE (pro) + cost;
4599 if (earliest == INVALID_TICK || t > earliest)
4600 earliest = t;
4601 }
4602 }
4603 bitmap_set_bit (processed, INSN_LUID (insn));
4604 INSN_TICK_ESTIMATE (insn) = earliest;
4605 return true;
4606 }
4607
4608 /* Examine the pair of insns in P, and estimate (optimistically, assuming
4609 infinite resources) the cycle in which the delayed shadow can be issued.
4610 Return the number of cycles that must pass before the real insn can be
4611 issued in order to meet this constraint. */
4612 static int
4613 estimate_shadow_tick (struct delay_pair *p)
4614 {
4615 bitmap_head processed;
4616 int t;
4617 bool cutoff;
4618 bitmap_initialize (&processed, 0);
4619
4620 cutoff = !estimate_insn_tick (&processed, p->i2,
4621 max_insn_queue_index + pair_delay (p));
4622 bitmap_clear (&processed);
4623 if (cutoff)
4624 return max_insn_queue_index;
4625 t = INSN_TICK_ESTIMATE (p->i2) - (clock_var + pair_delay (p) + 1);
4626 if (t > 0)
4627 return t;
4628 return 0;
4629 }
4630
4631 /* If INSN has no unresolved backwards dependencies, add it to the schedule and
4632 recursively resolve all its forward dependencies. */
4633 static void
4634 resolve_dependencies (rtx insn)
4635 {
4636 sd_iterator_def sd_it;
4637 dep_t dep;
4638
4639 /* Don't use sd_lists_empty_p; it ignores debug insns. */
4640 if (DEPS_LIST_FIRST (INSN_HARD_BACK_DEPS (insn)) != NULL
4641 || DEPS_LIST_FIRST (INSN_SPEC_BACK_DEPS (insn)) != NULL)
4642 return;
4643
4644 if (sched_verbose >= 4)
4645 fprintf (sched_dump, ";;\tquickly resolving %d\n", INSN_UID (insn));
4646
4647 if (QUEUE_INDEX (insn) >= 0)
4648 queue_remove (insn);
4649
4650 scheduled_insns.safe_push (insn);
4651
4652 /* Update dependent instructions. */
4653 for (sd_it = sd_iterator_start (insn, SD_LIST_FORW);
4654 sd_iterator_cond (&sd_it, &dep);)
4655 {
4656 rtx next = DEP_CON (dep);
4657
4658 if (sched_verbose >= 4)
4659 fprintf (sched_dump, ";;\t\tdep %d against %d\n", INSN_UID (insn),
4660 INSN_UID (next));
4661
4662 /* Resolve the dependence between INSN and NEXT.
4663 sd_resolve_dep () moves current dep to another list thus
4664 advancing the iterator. */
4665 sd_resolve_dep (sd_it);
4666
4667 if (!IS_SPECULATION_BRANCHY_CHECK_P (insn))
4668 {
4669 resolve_dependencies (next);
4670 }
4671 else
4672 /* Check always has only one forward dependence (to the first insn in
4673 the recovery block), therefore, this will be executed only once. */
4674 {
4675 gcc_assert (sd_lists_empty_p (insn, SD_LIST_FORW));
4676 }
4677 }
4678 }
4679
4680
4681 /* Return the head and tail pointers of ebb starting at BEG and ending
4682 at END. */
4683 void
4684 get_ebb_head_tail (basic_block beg, basic_block end, rtx *headp, rtx *tailp)
4685 {
4686 rtx beg_head = BB_HEAD (beg);
4687 rtx beg_tail = BB_END (beg);
4688 rtx end_head = BB_HEAD (end);
4689 rtx end_tail = BB_END (end);
4690
4691 /* Don't include any notes or labels at the beginning of the BEG
4692 basic block, or notes at the end of the END basic blocks. */
4693
4694 if (LABEL_P (beg_head))
4695 beg_head = NEXT_INSN (beg_head);
4696
4697 while (beg_head != beg_tail)
4698 if (NOTE_P (beg_head))
4699 beg_head = NEXT_INSN (beg_head);
4700 else if (DEBUG_INSN_P (beg_head))
4701 {
4702 rtx note, next;
4703
4704 for (note = NEXT_INSN (beg_head);
4705 note != beg_tail;
4706 note = next)
4707 {
4708 next = NEXT_INSN (note);
4709 if (NOTE_P (note))
4710 {
4711 if (sched_verbose >= 9)
4712 fprintf (sched_dump, "reorder %i\n", INSN_UID (note));
4713
4714 reorder_insns_nobb (note, note, PREV_INSN (beg_head));
4715
4716 if (BLOCK_FOR_INSN (note) != beg)
4717 df_insn_change_bb (note, beg);
4718 }
4719 else if (!DEBUG_INSN_P (note))
4720 break;
4721 }
4722
4723 break;
4724 }
4725 else
4726 break;
4727
4728 *headp = beg_head;
4729
4730 if (beg == end)
4731 end_head = beg_head;
4732 else if (LABEL_P (end_head))
4733 end_head = NEXT_INSN (end_head);
4734
4735 while (end_head != end_tail)
4736 if (NOTE_P (end_tail))
4737 end_tail = PREV_INSN (end_tail);
4738 else if (DEBUG_INSN_P (end_tail))
4739 {
4740 rtx note, prev;
4741
4742 for (note = PREV_INSN (end_tail);
4743 note != end_head;
4744 note = prev)
4745 {
4746 prev = PREV_INSN (note);
4747 if (NOTE_P (note))
4748 {
4749 if (sched_verbose >= 9)
4750 fprintf (sched_dump, "reorder %i\n", INSN_UID (note));
4751
4752 reorder_insns_nobb (note, note, end_tail);
4753
4754 if (end_tail == BB_END (end))
4755 BB_END (end) = note;
4756
4757 if (BLOCK_FOR_INSN (note) != end)
4758 df_insn_change_bb (note, end);
4759 }
4760 else if (!DEBUG_INSN_P (note))
4761 break;
4762 }
4763
4764 break;
4765 }
4766 else
4767 break;
4768
4769 *tailp = end_tail;
4770 }
4771
4772 /* Return nonzero if there are no real insns in the range [ HEAD, TAIL ]. */
4773
4774 int
4775 no_real_insns_p (const_rtx head, const_rtx tail)
4776 {
4777 while (head != NEXT_INSN (tail))
4778 {
4779 if (!NOTE_P (head) && !LABEL_P (head))
4780 return 0;
4781 head = NEXT_INSN (head);
4782 }
4783 return 1;
4784 }
4785
4786 /* Restore-other-notes: NOTE_LIST is the end of a chain of notes
4787 previously found among the insns. Insert them just before HEAD. */
4788 rtx
4789 restore_other_notes (rtx head, basic_block head_bb)
4790 {
4791 if (note_list != 0)
4792 {
4793 rtx note_head = note_list;
4794
4795 if (head)
4796 head_bb = BLOCK_FOR_INSN (head);
4797 else
4798 head = NEXT_INSN (bb_note (head_bb));
4799
4800 while (PREV_INSN (note_head))
4801 {
4802 set_block_for_insn (note_head, head_bb);
4803 note_head = PREV_INSN (note_head);
4804 }
4805 /* In the above cycle we've missed this note. */
4806 set_block_for_insn (note_head, head_bb);
4807
4808 PREV_INSN (note_head) = PREV_INSN (head);
4809 NEXT_INSN (PREV_INSN (head)) = note_head;
4810 PREV_INSN (head) = note_list;
4811 NEXT_INSN (note_list) = head;
4812
4813 if (BLOCK_FOR_INSN (head) != head_bb)
4814 BB_END (head_bb) = note_list;
4815
4816 head = note_head;
4817 }
4818
4819 return head;
4820 }
4821
4822 /* When we know we are going to discard the schedule due to a failed attempt
4823 at modulo scheduling, undo all replacements. */
4824 static void
4825 undo_all_replacements (void)
4826 {
4827 rtx insn;
4828 int i;
4829
4830 FOR_EACH_VEC_ELT (scheduled_insns, i, insn)
4831 {
4832 sd_iterator_def sd_it;
4833 dep_t dep;
4834
4835 /* See if we must undo a replacement. */
4836 for (sd_it = sd_iterator_start (insn, SD_LIST_RES_FORW);
4837 sd_iterator_cond (&sd_it, &dep); sd_iterator_next (&sd_it))
4838 {
4839 struct dep_replacement *desc = DEP_REPLACE (dep);
4840 if (desc != NULL)
4841 validate_change (desc->insn, desc->loc, desc->orig, 0);
4842 }
4843 }
4844 }
4845
4846 /* Move insns that became ready to fire from queue to ready list. */
4847
4848 static void
4849 queue_to_ready (struct ready_list *ready)
4850 {
4851 rtx insn;
4852 rtx link;
4853 rtx skip_insn;
4854
4855 q_ptr = NEXT_Q (q_ptr);
4856
4857 if (dbg_cnt (sched_insn) == false)
4858 {
4859 /* If debug counter is activated do not requeue the first
4860 nonscheduled insn. */
4861 skip_insn = nonscheduled_insns_begin;
4862 do
4863 {
4864 skip_insn = next_nonnote_nondebug_insn (skip_insn);
4865 }
4866 while (QUEUE_INDEX (skip_insn) == QUEUE_SCHEDULED);
4867 }
4868 else
4869 skip_insn = NULL_RTX;
4870
4871 /* Add all pending insns that can be scheduled without stalls to the
4872 ready list. */
4873 for (link = insn_queue[q_ptr]; link; link = XEXP (link, 1))
4874 {
4875 insn = XEXP (link, 0);
4876 q_size -= 1;
4877
4878 if (sched_verbose >= 2)
4879 fprintf (sched_dump, ";;\t\tQ-->Ready: insn %s: ",
4880 (*current_sched_info->print_insn) (insn, 0));
4881
4882 /* If the ready list is full, delay the insn for 1 cycle.
4883 See the comment in schedule_block for the rationale. */
4884 if (!reload_completed
4885 && (ready->n_ready - ready->n_debug > MAX_SCHED_READY_INSNS
4886 || (sched_pressure == SCHED_PRESSURE_MODEL
4887 /* Limit pressure recalculations to MAX_SCHED_READY_INSNS
4888 instructions too. */
4889 && model_index (insn) > (model_curr_point
4890 + MAX_SCHED_READY_INSNS)))
4891 && !(sched_pressure == SCHED_PRESSURE_MODEL
4892 && model_curr_point < model_num_insns
4893 /* Always allow the next model instruction to issue. */
4894 && model_index (insn) == model_curr_point)
4895 && !SCHED_GROUP_P (insn)
4896 && insn != skip_insn)
4897 queue_insn (insn, 1, "ready full");
4898 else
4899 {
4900 ready_add (ready, insn, false);
4901 if (sched_verbose >= 2)
4902 fprintf (sched_dump, "moving to ready without stalls\n");
4903 }
4904 }
4905 free_INSN_LIST_list (&insn_queue[q_ptr]);
4906
4907 /* If there are no ready insns, stall until one is ready and add all
4908 of the pending insns at that point to the ready list. */
4909 if (ready->n_ready == 0)
4910 {
4911 int stalls;
4912
4913 for (stalls = 1; stalls <= max_insn_queue_index; stalls++)
4914 {
4915 if ((link = insn_queue[NEXT_Q_AFTER (q_ptr, stalls)]))
4916 {
4917 for (; link; link = XEXP (link, 1))
4918 {
4919 insn = XEXP (link, 0);
4920 q_size -= 1;
4921
4922 if (sched_verbose >= 2)
4923 fprintf (sched_dump, ";;\t\tQ-->Ready: insn %s: ",
4924 (*current_sched_info->print_insn) (insn, 0));
4925
4926 ready_add (ready, insn, false);
4927 if (sched_verbose >= 2)
4928 fprintf (sched_dump, "moving to ready with %d stalls\n", stalls);
4929 }
4930 free_INSN_LIST_list (&insn_queue[NEXT_Q_AFTER (q_ptr, stalls)]);
4931
4932 advance_one_cycle ();
4933
4934 break;
4935 }
4936
4937 advance_one_cycle ();
4938 }
4939
4940 q_ptr = NEXT_Q_AFTER (q_ptr, stalls);
4941 clock_var += stalls;
4942 }
4943 }
4944
4945 /* Used by early_queue_to_ready. Determines whether it is "ok" to
4946 prematurely move INSN from the queue to the ready list. Currently,
4947 if a target defines the hook 'is_costly_dependence', this function
4948 uses the hook to check whether there exist any dependences which are
4949 considered costly by the target, between INSN and other insns that
4950 have already been scheduled. Dependences are checked up to Y cycles
4951 back, with default Y=1; The flag -fsched-stalled-insns-dep=Y allows
4952 controlling this value.
4953 (Other considerations could be taken into account instead (or in
4954 addition) depending on user flags and target hooks. */
4955
4956 static bool
4957 ok_for_early_queue_removal (rtx insn)
4958 {
4959 if (targetm.sched.is_costly_dependence)
4960 {
4961 rtx prev_insn;
4962 int n_cycles;
4963 int i = scheduled_insns.length ();
4964 for (n_cycles = flag_sched_stalled_insns_dep; n_cycles; n_cycles--)
4965 {
4966 while (i-- > 0)
4967 {
4968 int cost;
4969
4970 prev_insn = scheduled_insns[i];
4971
4972 if (!NOTE_P (prev_insn))
4973 {
4974 dep_t dep;
4975
4976 dep = sd_find_dep_between (prev_insn, insn, true);
4977
4978 if (dep != NULL)
4979 {
4980 cost = dep_cost (dep);
4981
4982 if (targetm.sched.is_costly_dependence (dep, cost,
4983 flag_sched_stalled_insns_dep - n_cycles))
4984 return false;
4985 }
4986 }
4987
4988 if (GET_MODE (prev_insn) == TImode) /* end of dispatch group */
4989 break;
4990 }
4991
4992 if (i == 0)
4993 break;
4994 }
4995 }
4996
4997 return true;
4998 }
4999
5000
5001 /* Remove insns from the queue, before they become "ready" with respect
5002 to FU latency considerations. */
5003
5004 static int
5005 early_queue_to_ready (state_t state, struct ready_list *ready)
5006 {
5007 rtx insn;
5008 rtx link;
5009 rtx next_link;
5010 rtx prev_link;
5011 bool move_to_ready;
5012 int cost;
5013 state_t temp_state = alloca (dfa_state_size);
5014 int stalls;
5015 int insns_removed = 0;
5016
5017 /*
5018 Flag '-fsched-stalled-insns=X' determines the aggressiveness of this
5019 function:
5020
5021 X == 0: There is no limit on how many queued insns can be removed
5022 prematurely. (flag_sched_stalled_insns = -1).
5023
5024 X >= 1: Only X queued insns can be removed prematurely in each
5025 invocation. (flag_sched_stalled_insns = X).
5026
5027 Otherwise: Early queue removal is disabled.
5028 (flag_sched_stalled_insns = 0)
5029 */
5030
5031 if (! flag_sched_stalled_insns)
5032 return 0;
5033
5034 for (stalls = 0; stalls <= max_insn_queue_index; stalls++)
5035 {
5036 if ((link = insn_queue[NEXT_Q_AFTER (q_ptr, stalls)]))
5037 {
5038 if (sched_verbose > 6)
5039 fprintf (sched_dump, ";; look at index %d + %d\n", q_ptr, stalls);
5040
5041 prev_link = 0;
5042 while (link)
5043 {
5044 next_link = XEXP (link, 1);
5045 insn = XEXP (link, 0);
5046 if (insn && sched_verbose > 6)
5047 print_rtl_single (sched_dump, insn);
5048
5049 memcpy (temp_state, state, dfa_state_size);
5050 if (recog_memoized (insn) < 0)
5051 /* non-negative to indicate that it's not ready
5052 to avoid infinite Q->R->Q->R... */
5053 cost = 0;
5054 else
5055 cost = state_transition (temp_state, insn);
5056
5057 if (sched_verbose >= 6)
5058 fprintf (sched_dump, "transition cost = %d\n", cost);
5059
5060 move_to_ready = false;
5061 if (cost < 0)
5062 {
5063 move_to_ready = ok_for_early_queue_removal (insn);
5064 if (move_to_ready == true)
5065 {
5066 /* move from Q to R */
5067 q_size -= 1;
5068 ready_add (ready, insn, false);
5069
5070 if (prev_link)
5071 XEXP (prev_link, 1) = next_link;
5072 else
5073 insn_queue[NEXT_Q_AFTER (q_ptr, stalls)] = next_link;
5074
5075 free_INSN_LIST_node (link);
5076
5077 if (sched_verbose >= 2)
5078 fprintf (sched_dump, ";;\t\tEarly Q-->Ready: insn %s\n",
5079 (*current_sched_info->print_insn) (insn, 0));
5080
5081 insns_removed++;
5082 if (insns_removed == flag_sched_stalled_insns)
5083 /* Remove no more than flag_sched_stalled_insns insns
5084 from Q at a time. */
5085 return insns_removed;
5086 }
5087 }
5088
5089 if (move_to_ready == false)
5090 prev_link = link;
5091
5092 link = next_link;
5093 } /* while link */
5094 } /* if link */
5095
5096 } /* for stalls.. */
5097
5098 return insns_removed;
5099 }
5100
5101
5102 /* Print the ready list for debugging purposes. Callable from debugger. */
5103
5104 static void
5105 debug_ready_list (struct ready_list *ready)
5106 {
5107 rtx *p;
5108 int i;
5109
5110 if (ready->n_ready == 0)
5111 {
5112 fprintf (sched_dump, "\n");
5113 return;
5114 }
5115
5116 p = ready_lastpos (ready);
5117 for (i = 0; i < ready->n_ready; i++)
5118 {
5119 fprintf (sched_dump, " %s:%d",
5120 (*current_sched_info->print_insn) (p[i], 0),
5121 INSN_LUID (p[i]));
5122 if (sched_pressure != SCHED_PRESSURE_NONE)
5123 fprintf (sched_dump, "(cost=%d",
5124 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (p[i]));
5125 if (INSN_TICK (p[i]) > clock_var)
5126 fprintf (sched_dump, ":delay=%d", INSN_TICK (p[i]) - clock_var);
5127 if (sched_pressure != SCHED_PRESSURE_NONE)
5128 fprintf (sched_dump, ")");
5129 }
5130 fprintf (sched_dump, "\n");
5131 }
5132
5133 /* Search INSN for REG_SAVE_NOTE notes and convert them back into insn
5134 NOTEs. This is used for NOTE_INSN_EPILOGUE_BEG, so that sched-ebb
5135 replaces the epilogue note in the correct basic block. */
5136 void
5137 reemit_notes (rtx insn)
5138 {
5139 rtx note, last = insn;
5140
5141 for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
5142 {
5143 if (REG_NOTE_KIND (note) == REG_SAVE_NOTE)
5144 {
5145 enum insn_note note_type = (enum insn_note) INTVAL (XEXP (note, 0));
5146
5147 last = emit_note_before (note_type, last);
5148 remove_note (insn, note);
5149 }
5150 }
5151 }
5152
5153 /* Move INSN. Reemit notes if needed. Update CFG, if needed. */
5154 static void
5155 move_insn (rtx insn, rtx last, rtx nt)
5156 {
5157 if (PREV_INSN (insn) != last)
5158 {
5159 basic_block bb;
5160 rtx note;
5161 int jump_p = 0;
5162
5163 bb = BLOCK_FOR_INSN (insn);
5164
5165 /* BB_HEAD is either LABEL or NOTE. */
5166 gcc_assert (BB_HEAD (bb) != insn);
5167
5168 if (BB_END (bb) == insn)
5169 /* If this is last instruction in BB, move end marker one
5170 instruction up. */
5171 {
5172 /* Jumps are always placed at the end of basic block. */
5173 jump_p = control_flow_insn_p (insn);
5174
5175 gcc_assert (!jump_p
5176 || ((common_sched_info->sched_pass_id == SCHED_RGN_PASS)
5177 && IS_SPECULATION_BRANCHY_CHECK_P (insn))
5178 || (common_sched_info->sched_pass_id
5179 == SCHED_EBB_PASS));
5180
5181 gcc_assert (BLOCK_FOR_INSN (PREV_INSN (insn)) == bb);
5182
5183 BB_END (bb) = PREV_INSN (insn);
5184 }
5185
5186 gcc_assert (BB_END (bb) != last);
5187
5188 if (jump_p)
5189 /* We move the block note along with jump. */
5190 {
5191 gcc_assert (nt);
5192
5193 note = NEXT_INSN (insn);
5194 while (NOTE_NOT_BB_P (note) && note != nt)
5195 note = NEXT_INSN (note);
5196
5197 if (note != nt
5198 && (LABEL_P (note)
5199 || BARRIER_P (note)))
5200 note = NEXT_INSN (note);
5201
5202 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
5203 }
5204 else
5205 note = insn;
5206
5207 NEXT_INSN (PREV_INSN (insn)) = NEXT_INSN (note);
5208 PREV_INSN (NEXT_INSN (note)) = PREV_INSN (insn);
5209
5210 NEXT_INSN (note) = NEXT_INSN (last);
5211 PREV_INSN (NEXT_INSN (last)) = note;
5212
5213 NEXT_INSN (last) = insn;
5214 PREV_INSN (insn) = last;
5215
5216 bb = BLOCK_FOR_INSN (last);
5217
5218 if (jump_p)
5219 {
5220 fix_jump_move (insn);
5221
5222 if (BLOCK_FOR_INSN (insn) != bb)
5223 move_block_after_check (insn);
5224
5225 gcc_assert (BB_END (bb) == last);
5226 }
5227
5228 df_insn_change_bb (insn, bb);
5229
5230 /* Update BB_END, if needed. */
5231 if (BB_END (bb) == last)
5232 BB_END (bb) = insn;
5233 }
5234
5235 SCHED_GROUP_P (insn) = 0;
5236 }
5237
5238 /* Return true if scheduling INSN will finish current clock cycle. */
5239 static bool
5240 insn_finishes_cycle_p (rtx insn)
5241 {
5242 if (SCHED_GROUP_P (insn))
5243 /* After issuing INSN, rest of the sched_group will be forced to issue
5244 in order. Don't make any plans for the rest of cycle. */
5245 return true;
5246
5247 /* Finishing the block will, apparently, finish the cycle. */
5248 if (current_sched_info->insn_finishes_block_p
5249 && current_sched_info->insn_finishes_block_p (insn))
5250 return true;
5251
5252 return false;
5253 }
5254
5255 /* Define type for target data used in multipass scheduling. */
5256 #ifndef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DATA_T
5257 # define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DATA_T int
5258 #endif
5259 typedef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DATA_T first_cycle_multipass_data_t;
5260
5261 /* The following structure describe an entry of the stack of choices. */
5262 struct choice_entry
5263 {
5264 /* Ordinal number of the issued insn in the ready queue. */
5265 int index;
5266 /* The number of the rest insns whose issues we should try. */
5267 int rest;
5268 /* The number of issued essential insns. */
5269 int n;
5270 /* State after issuing the insn. */
5271 state_t state;
5272 /* Target-specific data. */
5273 first_cycle_multipass_data_t target_data;
5274 };
5275
5276 /* The following array is used to implement a stack of choices used in
5277 function max_issue. */
5278 static struct choice_entry *choice_stack;
5279
5280 /* This holds the value of the target dfa_lookahead hook. */
5281 int dfa_lookahead;
5282
5283 /* The following variable value is maximal number of tries of issuing
5284 insns for the first cycle multipass insn scheduling. We define
5285 this value as constant*(DFA_LOOKAHEAD**ISSUE_RATE). We would not
5286 need this constraint if all real insns (with non-negative codes)
5287 had reservations because in this case the algorithm complexity is
5288 O(DFA_LOOKAHEAD**ISSUE_RATE). Unfortunately, the dfa descriptions
5289 might be incomplete and such insn might occur. For such
5290 descriptions, the complexity of algorithm (without the constraint)
5291 could achieve DFA_LOOKAHEAD ** N , where N is the queue length. */
5292 static int max_lookahead_tries;
5293
5294 /* The following value is value of hook
5295 `first_cycle_multipass_dfa_lookahead' at the last call of
5296 `max_issue'. */
5297 static int cached_first_cycle_multipass_dfa_lookahead = 0;
5298
5299 /* The following value is value of `issue_rate' at the last call of
5300 `sched_init'. */
5301 static int cached_issue_rate = 0;
5302
5303 /* The following function returns maximal (or close to maximal) number
5304 of insns which can be issued on the same cycle and one of which
5305 insns is insns with the best rank (the first insn in READY). To
5306 make this function tries different samples of ready insns. READY
5307 is current queue `ready'. Global array READY_TRY reflects what
5308 insns are already issued in this try. The function stops immediately,
5309 if it reached the such a solution, that all instruction can be issued.
5310 INDEX will contain index of the best insn in READY. The following
5311 function is used only for first cycle multipass scheduling.
5312
5313 PRIVILEGED_N >= 0
5314
5315 This function expects recognized insns only. All USEs,
5316 CLOBBERs, etc must be filtered elsewhere. */
5317 int
5318 max_issue (struct ready_list *ready, int privileged_n, state_t state,
5319 bool first_cycle_insn_p, int *index)
5320 {
5321 int n, i, all, n_ready, best, delay, tries_num;
5322 int more_issue;
5323 struct choice_entry *top;
5324 rtx insn;
5325
5326 n_ready = ready->n_ready;
5327 gcc_assert (dfa_lookahead >= 1 && privileged_n >= 0
5328 && privileged_n <= n_ready);
5329
5330 /* Init MAX_LOOKAHEAD_TRIES. */
5331 if (cached_first_cycle_multipass_dfa_lookahead != dfa_lookahead)
5332 {
5333 cached_first_cycle_multipass_dfa_lookahead = dfa_lookahead;
5334 max_lookahead_tries = 100;
5335 for (i = 0; i < issue_rate; i++)
5336 max_lookahead_tries *= dfa_lookahead;
5337 }
5338
5339 /* Init max_points. */
5340 more_issue = issue_rate - cycle_issued_insns;
5341 gcc_assert (more_issue >= 0);
5342
5343 /* The number of the issued insns in the best solution. */
5344 best = 0;
5345
5346 top = choice_stack;
5347
5348 /* Set initial state of the search. */
5349 memcpy (top->state, state, dfa_state_size);
5350 top->rest = dfa_lookahead;
5351 top->n = 0;
5352 if (targetm.sched.first_cycle_multipass_begin)
5353 targetm.sched.first_cycle_multipass_begin (&top->target_data,
5354 ready_try, n_ready,
5355 first_cycle_insn_p);
5356
5357 /* Count the number of the insns to search among. */
5358 for (all = i = 0; i < n_ready; i++)
5359 if (!ready_try [i])
5360 all++;
5361
5362 /* I is the index of the insn to try next. */
5363 i = 0;
5364 tries_num = 0;
5365 for (;;)
5366 {
5367 if (/* If we've reached a dead end or searched enough of what we have
5368 been asked... */
5369 top->rest == 0
5370 /* or have nothing else to try... */
5371 || i >= n_ready
5372 /* or should not issue more. */
5373 || top->n >= more_issue)
5374 {
5375 /* ??? (... || i == n_ready). */
5376 gcc_assert (i <= n_ready);
5377
5378 /* We should not issue more than issue_rate instructions. */
5379 gcc_assert (top->n <= more_issue);
5380
5381 if (top == choice_stack)
5382 break;
5383
5384 if (best < top - choice_stack)
5385 {
5386 if (privileged_n)
5387 {
5388 n = privileged_n;
5389 /* Try to find issued privileged insn. */
5390 while (n && !ready_try[--n])
5391 ;
5392 }
5393
5394 if (/* If all insns are equally good... */
5395 privileged_n == 0
5396 /* Or a privileged insn will be issued. */
5397 || ready_try[n])
5398 /* Then we have a solution. */
5399 {
5400 best = top - choice_stack;
5401 /* This is the index of the insn issued first in this
5402 solution. */
5403 *index = choice_stack [1].index;
5404 if (top->n == more_issue || best == all)
5405 break;
5406 }
5407 }
5408
5409 /* Set ready-list index to point to the last insn
5410 ('i++' below will advance it to the next insn). */
5411 i = top->index;
5412
5413 /* Backtrack. */
5414 ready_try [i] = 0;
5415
5416 if (targetm.sched.first_cycle_multipass_backtrack)
5417 targetm.sched.first_cycle_multipass_backtrack (&top->target_data,
5418 ready_try, n_ready);
5419
5420 top--;
5421 memcpy (state, top->state, dfa_state_size);
5422 }
5423 else if (!ready_try [i])
5424 {
5425 tries_num++;
5426 if (tries_num > max_lookahead_tries)
5427 break;
5428 insn = ready_element (ready, i);
5429 delay = state_transition (state, insn);
5430 if (delay < 0)
5431 {
5432 if (state_dead_lock_p (state)
5433 || insn_finishes_cycle_p (insn))
5434 /* We won't issue any more instructions in the next
5435 choice_state. */
5436 top->rest = 0;
5437 else
5438 top->rest--;
5439
5440 n = top->n;
5441 if (memcmp (top->state, state, dfa_state_size) != 0)
5442 n++;
5443
5444 /* Advance to the next choice_entry. */
5445 top++;
5446 /* Initialize it. */
5447 top->rest = dfa_lookahead;
5448 top->index = i;
5449 top->n = n;
5450 memcpy (top->state, state, dfa_state_size);
5451 ready_try [i] = 1;
5452
5453 if (targetm.sched.first_cycle_multipass_issue)
5454 targetm.sched.first_cycle_multipass_issue (&top->target_data,
5455 ready_try, n_ready,
5456 insn,
5457 &((top - 1)
5458 ->target_data));
5459
5460 i = -1;
5461 }
5462 }
5463
5464 /* Increase ready-list index. */
5465 i++;
5466 }
5467
5468 if (targetm.sched.first_cycle_multipass_end)
5469 targetm.sched.first_cycle_multipass_end (best != 0
5470 ? &choice_stack[1].target_data
5471 : NULL);
5472
5473 /* Restore the original state of the DFA. */
5474 memcpy (state, choice_stack->state, dfa_state_size);
5475
5476 return best;
5477 }
5478
5479 /* The following function chooses insn from READY and modifies
5480 READY. The following function is used only for first
5481 cycle multipass scheduling.
5482 Return:
5483 -1 if cycle should be advanced,
5484 0 if INSN_PTR is set to point to the desirable insn,
5485 1 if choose_ready () should be restarted without advancing the cycle. */
5486 static int
5487 choose_ready (struct ready_list *ready, bool first_cycle_insn_p,
5488 rtx *insn_ptr)
5489 {
5490 int lookahead;
5491
5492 if (dbg_cnt (sched_insn) == false)
5493 {
5494 rtx insn = nonscheduled_insns_begin;
5495 do
5496 {
5497 insn = next_nonnote_insn (insn);
5498 }
5499 while (QUEUE_INDEX (insn) == QUEUE_SCHEDULED);
5500
5501 if (QUEUE_INDEX (insn) == QUEUE_READY)
5502 /* INSN is in the ready_list. */
5503 {
5504 nonscheduled_insns_begin = insn;
5505 ready_remove_insn (insn);
5506 *insn_ptr = insn;
5507 return 0;
5508 }
5509
5510 /* INSN is in the queue. Advance cycle to move it to the ready list. */
5511 return -1;
5512 }
5513
5514 lookahead = 0;
5515
5516 if (targetm.sched.first_cycle_multipass_dfa_lookahead)
5517 lookahead = targetm.sched.first_cycle_multipass_dfa_lookahead ();
5518 if (lookahead <= 0 || SCHED_GROUP_P (ready_element (ready, 0))
5519 || DEBUG_INSN_P (ready_element (ready, 0)))
5520 {
5521 if (targetm.sched.dispatch (NULL_RTX, IS_DISPATCH_ON))
5522 *insn_ptr = ready_remove_first_dispatch (ready);
5523 else
5524 *insn_ptr = ready_remove_first (ready);
5525
5526 return 0;
5527 }
5528 else
5529 {
5530 /* Try to choose the better insn. */
5531 int index = 0, i, n;
5532 rtx insn;
5533 int try_data = 1, try_control = 1;
5534 ds_t ts;
5535
5536 insn = ready_element (ready, 0);
5537 if (INSN_CODE (insn) < 0)
5538 {
5539 *insn_ptr = ready_remove_first (ready);
5540 return 0;
5541 }
5542
5543 if (spec_info
5544 && spec_info->flags & (PREFER_NON_DATA_SPEC
5545 | PREFER_NON_CONTROL_SPEC))
5546 {
5547 for (i = 0, n = ready->n_ready; i < n; i++)
5548 {
5549 rtx x;
5550 ds_t s;
5551
5552 x = ready_element (ready, i);
5553 s = TODO_SPEC (x);
5554
5555 if (spec_info->flags & PREFER_NON_DATA_SPEC
5556 && !(s & DATA_SPEC))
5557 {
5558 try_data = 0;
5559 if (!(spec_info->flags & PREFER_NON_CONTROL_SPEC)
5560 || !try_control)
5561 break;
5562 }
5563
5564 if (spec_info->flags & PREFER_NON_CONTROL_SPEC
5565 && !(s & CONTROL_SPEC))
5566 {
5567 try_control = 0;
5568 if (!(spec_info->flags & PREFER_NON_DATA_SPEC) || !try_data)
5569 break;
5570 }
5571 }
5572 }
5573
5574 ts = TODO_SPEC (insn);
5575 if ((ts & SPECULATIVE)
5576 && (((!try_data && (ts & DATA_SPEC))
5577 || (!try_control && (ts & CONTROL_SPEC)))
5578 || (targetm.sched.first_cycle_multipass_dfa_lookahead_guard_spec
5579 && !targetm.sched
5580 .first_cycle_multipass_dfa_lookahead_guard_spec (insn))))
5581 /* Discard speculative instruction that stands first in the ready
5582 list. */
5583 {
5584 change_queue_index (insn, 1);
5585 return 1;
5586 }
5587
5588 ready_try[0] = 0;
5589
5590 for (i = 1; i < ready->n_ready; i++)
5591 {
5592 insn = ready_element (ready, i);
5593
5594 ready_try [i]
5595 = ((!try_data && (TODO_SPEC (insn) & DATA_SPEC))
5596 || (!try_control && (TODO_SPEC (insn) & CONTROL_SPEC)));
5597 }
5598
5599 /* Let the target filter the search space. */
5600 for (i = 1; i < ready->n_ready; i++)
5601 if (!ready_try[i])
5602 {
5603 insn = ready_element (ready, i);
5604
5605 /* If this insn is recognizable we should have already
5606 recognized it earlier.
5607 ??? Not very clear where this is supposed to be done.
5608 See dep_cost_1. */
5609 gcc_checking_assert (INSN_CODE (insn) >= 0
5610 || recog_memoized (insn) < 0);
5611
5612 ready_try [i]
5613 = (/* INSN_CODE check can be omitted here as it is also done later
5614 in max_issue (). */
5615 INSN_CODE (insn) < 0
5616 || (targetm.sched.first_cycle_multipass_dfa_lookahead_guard
5617 && !targetm.sched.first_cycle_multipass_dfa_lookahead_guard
5618 (insn)));
5619 }
5620
5621 if (max_issue (ready, 1, curr_state, first_cycle_insn_p, &index) == 0)
5622 {
5623 *insn_ptr = ready_remove_first (ready);
5624 if (sched_verbose >= 4)
5625 fprintf (sched_dump, ";;\t\tChosen insn (but can't issue) : %s \n",
5626 (*current_sched_info->print_insn) (*insn_ptr, 0));
5627 return 0;
5628 }
5629 else
5630 {
5631 if (sched_verbose >= 4)
5632 fprintf (sched_dump, ";;\t\tChosen insn : %s\n",
5633 (*current_sched_info->print_insn)
5634 (ready_element (ready, index), 0));
5635
5636 *insn_ptr = ready_remove (ready, index);
5637 return 0;
5638 }
5639 }
5640 }
5641
5642 /* This function is called when we have successfully scheduled a
5643 block. It uses the schedule stored in the scheduled_insns vector
5644 to rearrange the RTL. PREV_HEAD is used as the anchor to which we
5645 append the scheduled insns; TAIL is the insn after the scheduled
5646 block. TARGET_BB is the argument passed to schedule_block. */
5647
5648 static void
5649 commit_schedule (rtx prev_head, rtx tail, basic_block *target_bb)
5650 {
5651 unsigned int i;
5652 rtx insn;
5653
5654 last_scheduled_insn = prev_head;
5655 for (i = 0;
5656 scheduled_insns.iterate (i, &insn);
5657 i++)
5658 {
5659 if (control_flow_insn_p (last_scheduled_insn)
5660 || current_sched_info->advance_target_bb (*target_bb, insn))
5661 {
5662 *target_bb = current_sched_info->advance_target_bb (*target_bb, 0);
5663
5664 if (sched_verbose)
5665 {
5666 rtx x;
5667
5668 x = next_real_insn (last_scheduled_insn);
5669 gcc_assert (x);
5670 dump_new_block_header (1, *target_bb, x, tail);
5671 }
5672
5673 last_scheduled_insn = bb_note (*target_bb);
5674 }
5675
5676 if (current_sched_info->begin_move_insn)
5677 (*current_sched_info->begin_move_insn) (insn, last_scheduled_insn);
5678 move_insn (insn, last_scheduled_insn,
5679 current_sched_info->next_tail);
5680 if (!DEBUG_INSN_P (insn))
5681 reemit_notes (insn);
5682 last_scheduled_insn = insn;
5683 }
5684
5685 scheduled_insns.truncate (0);
5686 }
5687
5688 /* Examine all insns on the ready list and queue those which can't be
5689 issued in this cycle. TEMP_STATE is temporary scheduler state we
5690 can use as scratch space. If FIRST_CYCLE_INSN_P is true, no insns
5691 have been issued for the current cycle, which means it is valid to
5692 issue an asm statement.
5693
5694 If SHADOWS_ONLY_P is true, we eliminate all real insns and only
5695 leave those for which SHADOW_P is true. If MODULO_EPILOGUE is true,
5696 we only leave insns which have an INSN_EXACT_TICK. */
5697
5698 static void
5699 prune_ready_list (state_t temp_state, bool first_cycle_insn_p,
5700 bool shadows_only_p, bool modulo_epilogue_p)
5701 {
5702 int i, pass;
5703 bool sched_group_found = false;
5704 int min_cost_group = 1;
5705
5706 for (i = 0; i < ready.n_ready; i++)
5707 {
5708 rtx insn = ready_element (&ready, i);
5709 if (SCHED_GROUP_P (insn))
5710 {
5711 sched_group_found = true;
5712 break;
5713 }
5714 }
5715
5716 /* Make two passes if there's a SCHED_GROUP_P insn; make sure to handle
5717 such an insn first and note its cost, then schedule all other insns
5718 for one cycle later. */
5719 for (pass = sched_group_found ? 0 : 1; pass < 2; )
5720 {
5721 int n = ready.n_ready;
5722 for (i = 0; i < n; i++)
5723 {
5724 rtx insn = ready_element (&ready, i);
5725 int cost = 0;
5726 const char *reason = "resource conflict";
5727
5728 if (DEBUG_INSN_P (insn))
5729 continue;
5730
5731 if (sched_group_found && !SCHED_GROUP_P (insn))
5732 {
5733 if (pass == 0)
5734 continue;
5735 cost = min_cost_group;
5736 reason = "not in sched group";
5737 }
5738 else if (modulo_epilogue_p
5739 && INSN_EXACT_TICK (insn) == INVALID_TICK)
5740 {
5741 cost = max_insn_queue_index;
5742 reason = "not an epilogue insn";
5743 }
5744 else if (shadows_only_p && !SHADOW_P (insn))
5745 {
5746 cost = 1;
5747 reason = "not a shadow";
5748 }
5749 else if (recog_memoized (insn) < 0)
5750 {
5751 if (!first_cycle_insn_p
5752 && (GET_CODE (PATTERN (insn)) == ASM_INPUT
5753 || asm_noperands (PATTERN (insn)) >= 0))
5754 cost = 1;
5755 reason = "asm";
5756 }
5757 else if (sched_pressure != SCHED_PRESSURE_NONE)
5758 {
5759 if (sched_pressure == SCHED_PRESSURE_MODEL
5760 && INSN_TICK (insn) <= clock_var)
5761 {
5762 memcpy (temp_state, curr_state, dfa_state_size);
5763 if (state_transition (temp_state, insn) >= 0)
5764 INSN_TICK (insn) = clock_var + 1;
5765 }
5766 cost = 0;
5767 }
5768 else
5769 {
5770 int delay_cost = 0;
5771
5772 if (delay_htab.is_created ())
5773 {
5774 struct delay_pair *delay_entry;
5775 delay_entry
5776 = delay_htab.find_with_hash (insn,
5777 htab_hash_pointer (insn));
5778 while (delay_entry && delay_cost == 0)
5779 {
5780 delay_cost = estimate_shadow_tick (delay_entry);
5781 if (delay_cost > max_insn_queue_index)
5782 delay_cost = max_insn_queue_index;
5783 delay_entry = delay_entry->next_same_i1;
5784 }
5785 }
5786
5787 memcpy (temp_state, curr_state, dfa_state_size);
5788 cost = state_transition (temp_state, insn);
5789 if (cost < 0)
5790 cost = 0;
5791 else if (cost == 0)
5792 cost = 1;
5793 if (cost < delay_cost)
5794 {
5795 cost = delay_cost;
5796 reason = "shadow tick";
5797 }
5798 }
5799 if (cost >= 1)
5800 {
5801 if (SCHED_GROUP_P (insn) && cost > min_cost_group)
5802 min_cost_group = cost;
5803 ready_remove (&ready, i);
5804 queue_insn (insn, cost, reason);
5805 if (i + 1 < n)
5806 break;
5807 }
5808 }
5809 if (i == n)
5810 pass++;
5811 }
5812 }
5813
5814 /* Called when we detect that the schedule is impossible. We examine the
5815 backtrack queue to find the earliest insn that caused this condition. */
5816
5817 static struct haifa_saved_data *
5818 verify_shadows (void)
5819 {
5820 struct haifa_saved_data *save, *earliest_fail = NULL;
5821 for (save = backtrack_queue; save; save = save->next)
5822 {
5823 int t;
5824 struct delay_pair *pair = save->delay_pair;
5825 rtx i1 = pair->i1;
5826
5827 for (; pair; pair = pair->next_same_i1)
5828 {
5829 rtx i2 = pair->i2;
5830
5831 if (QUEUE_INDEX (i2) == QUEUE_SCHEDULED)
5832 continue;
5833
5834 t = INSN_TICK (i1) + pair_delay (pair);
5835 if (t < clock_var)
5836 {
5837 if (sched_verbose >= 2)
5838 fprintf (sched_dump,
5839 ";;\t\tfailed delay requirements for %d/%d (%d->%d)"
5840 ", not ready\n",
5841 INSN_UID (pair->i1), INSN_UID (pair->i2),
5842 INSN_TICK (pair->i1), INSN_EXACT_TICK (pair->i2));
5843 earliest_fail = save;
5844 break;
5845 }
5846 if (QUEUE_INDEX (i2) >= 0)
5847 {
5848 int queued_for = INSN_TICK (i2);
5849
5850 if (t < queued_for)
5851 {
5852 if (sched_verbose >= 2)
5853 fprintf (sched_dump,
5854 ";;\t\tfailed delay requirements for %d/%d"
5855 " (%d->%d), queued too late\n",
5856 INSN_UID (pair->i1), INSN_UID (pair->i2),
5857 INSN_TICK (pair->i1), INSN_EXACT_TICK (pair->i2));
5858 earliest_fail = save;
5859 break;
5860 }
5861 }
5862 }
5863 }
5864
5865 return earliest_fail;
5866 }
5867
5868 /* Use forward list scheduling to rearrange insns of block pointed to by
5869 TARGET_BB, possibly bringing insns from subsequent blocks in the same
5870 region. */
5871
5872 bool
5873 schedule_block (basic_block *target_bb, state_t init_state)
5874 {
5875 int i;
5876 bool success = modulo_ii == 0;
5877 struct sched_block_state ls;
5878 state_t temp_state = NULL; /* It is used for multipass scheduling. */
5879 int sort_p, advance, start_clock_var;
5880
5881 /* Head/tail info for this block. */
5882 rtx prev_head = current_sched_info->prev_head;
5883 rtx next_tail = current_sched_info->next_tail;
5884 rtx head = NEXT_INSN (prev_head);
5885 rtx tail = PREV_INSN (next_tail);
5886
5887 if ((current_sched_info->flags & DONT_BREAK_DEPENDENCIES) == 0
5888 && sched_pressure != SCHED_PRESSURE_MODEL)
5889 find_modifiable_mems (head, tail);
5890
5891 /* We used to have code to avoid getting parameters moved from hard
5892 argument registers into pseudos.
5893
5894 However, it was removed when it proved to be of marginal benefit
5895 and caused problems because schedule_block and compute_forward_dependences
5896 had different notions of what the "head" insn was. */
5897
5898 gcc_assert (head != tail || INSN_P (head));
5899
5900 haifa_recovery_bb_recently_added_p = false;
5901
5902 backtrack_queue = NULL;
5903
5904 /* Debug info. */
5905 if (sched_verbose)
5906 dump_new_block_header (0, *target_bb, head, tail);
5907
5908 if (init_state == NULL)
5909 state_reset (curr_state);
5910 else
5911 memcpy (curr_state, init_state, dfa_state_size);
5912
5913 /* Clear the ready list. */
5914 ready.first = ready.veclen - 1;
5915 ready.n_ready = 0;
5916 ready.n_debug = 0;
5917
5918 /* It is used for first cycle multipass scheduling. */
5919 temp_state = alloca (dfa_state_size);
5920
5921 if (targetm.sched.init)
5922 targetm.sched.init (sched_dump, sched_verbose, ready.veclen);
5923
5924 /* We start inserting insns after PREV_HEAD. */
5925 last_scheduled_insn = nonscheduled_insns_begin = prev_head;
5926 last_nondebug_scheduled_insn = NULL_RTX;
5927
5928 gcc_assert ((NOTE_P (last_scheduled_insn)
5929 || DEBUG_INSN_P (last_scheduled_insn))
5930 && BLOCK_FOR_INSN (last_scheduled_insn) == *target_bb);
5931
5932 /* Initialize INSN_QUEUE. Q_SIZE is the total number of insns in the
5933 queue. */
5934 q_ptr = 0;
5935 q_size = 0;
5936
5937 insn_queue = XALLOCAVEC (rtx, max_insn_queue_index + 1);
5938 memset (insn_queue, 0, (max_insn_queue_index + 1) * sizeof (rtx));
5939
5940 /* Start just before the beginning of time. */
5941 clock_var = -1;
5942
5943 /* We need queue and ready lists and clock_var be initialized
5944 in try_ready () (which is called through init_ready_list ()). */
5945 (*current_sched_info->init_ready_list) ();
5946
5947 if (sched_pressure == SCHED_PRESSURE_MODEL)
5948 model_start_schedule ();
5949
5950 /* The algorithm is O(n^2) in the number of ready insns at any given
5951 time in the worst case. Before reload we are more likely to have
5952 big lists so truncate them to a reasonable size. */
5953 if (!reload_completed
5954 && ready.n_ready - ready.n_debug > MAX_SCHED_READY_INSNS)
5955 {
5956 ready_sort (&ready);
5957
5958 /* Find first free-standing insn past MAX_SCHED_READY_INSNS.
5959 If there are debug insns, we know they're first. */
5960 for (i = MAX_SCHED_READY_INSNS + ready.n_debug; i < ready.n_ready; i++)
5961 if (!SCHED_GROUP_P (ready_element (&ready, i)))
5962 break;
5963
5964 if (sched_verbose >= 2)
5965 {
5966 fprintf (sched_dump,
5967 ";;\t\tReady list on entry: %d insns\n", ready.n_ready);
5968 fprintf (sched_dump,
5969 ";;\t\t before reload => truncated to %d insns\n", i);
5970 }
5971
5972 /* Delay all insns past it for 1 cycle. If debug counter is
5973 activated make an exception for the insn right after
5974 nonscheduled_insns_begin. */
5975 {
5976 rtx skip_insn;
5977
5978 if (dbg_cnt (sched_insn) == false)
5979 skip_insn = next_nonnote_insn (nonscheduled_insns_begin);
5980 else
5981 skip_insn = NULL_RTX;
5982
5983 while (i < ready.n_ready)
5984 {
5985 rtx insn;
5986
5987 insn = ready_remove (&ready, i);
5988
5989 if (insn != skip_insn)
5990 queue_insn (insn, 1, "list truncated");
5991 }
5992 if (skip_insn)
5993 ready_add (&ready, skip_insn, true);
5994 }
5995 }
5996
5997 /* Now we can restore basic block notes and maintain precise cfg. */
5998 restore_bb_notes (*target_bb);
5999
6000 last_clock_var = -1;
6001
6002 advance = 0;
6003
6004 gcc_assert (scheduled_insns.length () == 0);
6005 sort_p = TRUE;
6006 must_backtrack = false;
6007 modulo_insns_scheduled = 0;
6008
6009 ls.modulo_epilogue = false;
6010
6011 /* Loop until all the insns in BB are scheduled. */
6012 while ((*current_sched_info->schedule_more_p) ())
6013 {
6014 perform_replacements_new_cycle ();
6015 do
6016 {
6017 start_clock_var = clock_var;
6018
6019 clock_var++;
6020
6021 advance_one_cycle ();
6022
6023 /* Add to the ready list all pending insns that can be issued now.
6024 If there are no ready insns, increment clock until one
6025 is ready and add all pending insns at that point to the ready
6026 list. */
6027 queue_to_ready (&ready);
6028
6029 gcc_assert (ready.n_ready);
6030
6031 if (sched_verbose >= 2)
6032 {
6033 fprintf (sched_dump, ";;\t\tReady list after queue_to_ready: ");
6034 debug_ready_list (&ready);
6035 }
6036 advance -= clock_var - start_clock_var;
6037 }
6038 while (advance > 0);
6039
6040 if (ls.modulo_epilogue)
6041 {
6042 int stage = clock_var / modulo_ii;
6043 if (stage > modulo_last_stage * 2 + 2)
6044 {
6045 if (sched_verbose >= 2)
6046 fprintf (sched_dump,
6047 ";;\t\tmodulo scheduled succeeded at II %d\n",
6048 modulo_ii);
6049 success = true;
6050 goto end_schedule;
6051 }
6052 }
6053 else if (modulo_ii > 0)
6054 {
6055 int stage = clock_var / modulo_ii;
6056 if (stage > modulo_max_stages)
6057 {
6058 if (sched_verbose >= 2)
6059 fprintf (sched_dump,
6060 ";;\t\tfailing schedule due to excessive stages\n");
6061 goto end_schedule;
6062 }
6063 if (modulo_n_insns == modulo_insns_scheduled
6064 && stage > modulo_last_stage)
6065 {
6066 if (sched_verbose >= 2)
6067 fprintf (sched_dump,
6068 ";;\t\tfound kernel after %d stages, II %d\n",
6069 stage, modulo_ii);
6070 ls.modulo_epilogue = true;
6071 }
6072 }
6073
6074 prune_ready_list (temp_state, true, false, ls.modulo_epilogue);
6075 if (ready.n_ready == 0)
6076 continue;
6077 if (must_backtrack)
6078 goto do_backtrack;
6079
6080 ls.first_cycle_insn_p = true;
6081 ls.shadows_only_p = false;
6082 cycle_issued_insns = 0;
6083 ls.can_issue_more = issue_rate;
6084 for (;;)
6085 {
6086 rtx insn;
6087 int cost;
6088 bool asm_p;
6089
6090 if (sort_p && ready.n_ready > 0)
6091 {
6092 /* Sort the ready list based on priority. This must be
6093 done every iteration through the loop, as schedule_insn
6094 may have readied additional insns that will not be
6095 sorted correctly. */
6096 ready_sort (&ready);
6097
6098 if (sched_verbose >= 2)
6099 {
6100 fprintf (sched_dump, ";;\t\tReady list after ready_sort: ");
6101 debug_ready_list (&ready);
6102 }
6103 }
6104
6105 /* We don't want md sched reorder to even see debug isns, so put
6106 them out right away. */
6107 if (ready.n_ready && DEBUG_INSN_P (ready_element (&ready, 0))
6108 && (*current_sched_info->schedule_more_p) ())
6109 {
6110 while (ready.n_ready && DEBUG_INSN_P (ready_element (&ready, 0)))
6111 {
6112 rtx insn = ready_remove_first (&ready);
6113 gcc_assert (DEBUG_INSN_P (insn));
6114 (*current_sched_info->begin_schedule_ready) (insn);
6115 scheduled_insns.safe_push (insn);
6116 last_scheduled_insn = insn;
6117 advance = schedule_insn (insn);
6118 gcc_assert (advance == 0);
6119 if (ready.n_ready > 0)
6120 ready_sort (&ready);
6121 }
6122 }
6123
6124 if (ls.first_cycle_insn_p && !ready.n_ready)
6125 break;
6126
6127 resume_after_backtrack:
6128 /* Allow the target to reorder the list, typically for
6129 better instruction bundling. */
6130 if (sort_p
6131 && (ready.n_ready == 0
6132 || !SCHED_GROUP_P (ready_element (&ready, 0))))
6133 {
6134 if (ls.first_cycle_insn_p && targetm.sched.reorder)
6135 ls.can_issue_more
6136 = targetm.sched.reorder (sched_dump, sched_verbose,
6137 ready_lastpos (&ready),
6138 &ready.n_ready, clock_var);
6139 else if (!ls.first_cycle_insn_p && targetm.sched.reorder2)
6140 ls.can_issue_more
6141 = targetm.sched.reorder2 (sched_dump, sched_verbose,
6142 ready.n_ready
6143 ? ready_lastpos (&ready) : NULL,
6144 &ready.n_ready, clock_var);
6145 }
6146
6147 restart_choose_ready:
6148 if (sched_verbose >= 2)
6149 {
6150 fprintf (sched_dump, ";;\tReady list (t = %3d): ",
6151 clock_var);
6152 debug_ready_list (&ready);
6153 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
6154 print_curr_reg_pressure ();
6155 }
6156
6157 if (ready.n_ready == 0
6158 && ls.can_issue_more
6159 && reload_completed)
6160 {
6161 /* Allow scheduling insns directly from the queue in case
6162 there's nothing better to do (ready list is empty) but
6163 there are still vacant dispatch slots in the current cycle. */
6164 if (sched_verbose >= 6)
6165 fprintf (sched_dump,";;\t\tSecond chance\n");
6166 memcpy (temp_state, curr_state, dfa_state_size);
6167 if (early_queue_to_ready (temp_state, &ready))
6168 ready_sort (&ready);
6169 }
6170
6171 if (ready.n_ready == 0
6172 || !ls.can_issue_more
6173 || state_dead_lock_p (curr_state)
6174 || !(*current_sched_info->schedule_more_p) ())
6175 break;
6176
6177 /* Select and remove the insn from the ready list. */
6178 if (sort_p)
6179 {
6180 int res;
6181
6182 insn = NULL_RTX;
6183 res = choose_ready (&ready, ls.first_cycle_insn_p, &insn);
6184
6185 if (res < 0)
6186 /* Finish cycle. */
6187 break;
6188 if (res > 0)
6189 goto restart_choose_ready;
6190
6191 gcc_assert (insn != NULL_RTX);
6192 }
6193 else
6194 insn = ready_remove_first (&ready);
6195
6196 if (sched_pressure != SCHED_PRESSURE_NONE
6197 && INSN_TICK (insn) > clock_var)
6198 {
6199 ready_add (&ready, insn, true);
6200 advance = 1;
6201 break;
6202 }
6203
6204 if (targetm.sched.dfa_new_cycle
6205 && targetm.sched.dfa_new_cycle (sched_dump, sched_verbose,
6206 insn, last_clock_var,
6207 clock_var, &sort_p))
6208 /* SORT_P is used by the target to override sorting
6209 of the ready list. This is needed when the target
6210 has modified its internal structures expecting that
6211 the insn will be issued next. As we need the insn
6212 to have the highest priority (so it will be returned by
6213 the ready_remove_first call above), we invoke
6214 ready_add (&ready, insn, true).
6215 But, still, there is one issue: INSN can be later
6216 discarded by scheduler's front end through
6217 current_sched_info->can_schedule_ready_p, hence, won't
6218 be issued next. */
6219 {
6220 ready_add (&ready, insn, true);
6221 break;
6222 }
6223
6224 sort_p = TRUE;
6225
6226 if (current_sched_info->can_schedule_ready_p
6227 && ! (*current_sched_info->can_schedule_ready_p) (insn))
6228 /* We normally get here only if we don't want to move
6229 insn from the split block. */
6230 {
6231 TODO_SPEC (insn) = DEP_POSTPONED;
6232 goto restart_choose_ready;
6233 }
6234
6235 if (delay_htab.is_created ())
6236 {
6237 /* If this insn is the first part of a delay-slot pair, record a
6238 backtrack point. */
6239 struct delay_pair *delay_entry;
6240 delay_entry
6241 = delay_htab.find_with_hash (insn, htab_hash_pointer (insn));
6242 if (delay_entry)
6243 {
6244 save_backtrack_point (delay_entry, ls);
6245 if (sched_verbose >= 2)
6246 fprintf (sched_dump, ";;\t\tsaving backtrack point\n");
6247 }
6248 }
6249
6250 /* DECISION is made. */
6251
6252 if (modulo_ii > 0 && INSN_UID (insn) < modulo_iter0_max_uid)
6253 {
6254 modulo_insns_scheduled++;
6255 modulo_last_stage = clock_var / modulo_ii;
6256 }
6257 if (TODO_SPEC (insn) & SPECULATIVE)
6258 generate_recovery_code (insn);
6259
6260 if (targetm.sched.dispatch (NULL_RTX, IS_DISPATCH_ON))
6261 targetm.sched.dispatch_do (insn, ADD_TO_DISPATCH_WINDOW);
6262
6263 /* Update counters, etc in the scheduler's front end. */
6264 (*current_sched_info->begin_schedule_ready) (insn);
6265 scheduled_insns.safe_push (insn);
6266 gcc_assert (NONDEBUG_INSN_P (insn));
6267 last_nondebug_scheduled_insn = last_scheduled_insn = insn;
6268
6269 if (recog_memoized (insn) >= 0)
6270 {
6271 memcpy (temp_state, curr_state, dfa_state_size);
6272 cost = state_transition (curr_state, insn);
6273 if (sched_pressure != SCHED_PRESSURE_WEIGHTED)
6274 gcc_assert (cost < 0);
6275 if (memcmp (temp_state, curr_state, dfa_state_size) != 0)
6276 cycle_issued_insns++;
6277 asm_p = false;
6278 }
6279 else
6280 asm_p = (GET_CODE (PATTERN (insn)) == ASM_INPUT
6281 || asm_noperands (PATTERN (insn)) >= 0);
6282
6283 if (targetm.sched.variable_issue)
6284 ls.can_issue_more =
6285 targetm.sched.variable_issue (sched_dump, sched_verbose,
6286 insn, ls.can_issue_more);
6287 /* A naked CLOBBER or USE generates no instruction, so do
6288 not count them against the issue rate. */
6289 else if (GET_CODE (PATTERN (insn)) != USE
6290 && GET_CODE (PATTERN (insn)) != CLOBBER)
6291 ls.can_issue_more--;
6292 advance = schedule_insn (insn);
6293
6294 if (SHADOW_P (insn))
6295 ls.shadows_only_p = true;
6296
6297 /* After issuing an asm insn we should start a new cycle. */
6298 if (advance == 0 && asm_p)
6299 advance = 1;
6300
6301 if (must_backtrack)
6302 break;
6303
6304 if (advance != 0)
6305 break;
6306
6307 ls.first_cycle_insn_p = false;
6308 if (ready.n_ready > 0)
6309 prune_ready_list (temp_state, false, ls.shadows_only_p,
6310 ls.modulo_epilogue);
6311 }
6312
6313 do_backtrack:
6314 if (!must_backtrack)
6315 for (i = 0; i < ready.n_ready; i++)
6316 {
6317 rtx insn = ready_element (&ready, i);
6318 if (INSN_EXACT_TICK (insn) == clock_var)
6319 {
6320 must_backtrack = true;
6321 clock_var++;
6322 break;
6323 }
6324 }
6325 if (must_backtrack && modulo_ii > 0)
6326 {
6327 if (modulo_backtracks_left == 0)
6328 goto end_schedule;
6329 modulo_backtracks_left--;
6330 }
6331 while (must_backtrack)
6332 {
6333 struct haifa_saved_data *failed;
6334 rtx failed_insn;
6335
6336 must_backtrack = false;
6337 failed = verify_shadows ();
6338 gcc_assert (failed);
6339
6340 failed_insn = failed->delay_pair->i1;
6341 /* Clear these queues. */
6342 perform_replacements_new_cycle ();
6343 toggle_cancelled_flags (false);
6344 unschedule_insns_until (failed_insn);
6345 while (failed != backtrack_queue)
6346 free_topmost_backtrack_point (true);
6347 restore_last_backtrack_point (&ls);
6348 if (sched_verbose >= 2)
6349 fprintf (sched_dump, ";;\t\trewind to cycle %d\n", clock_var);
6350 /* Delay by at least a cycle. This could cause additional
6351 backtracking. */
6352 queue_insn (failed_insn, 1, "backtracked");
6353 advance = 0;
6354 if (must_backtrack)
6355 continue;
6356 if (ready.n_ready > 0)
6357 goto resume_after_backtrack;
6358 else
6359 {
6360 if (clock_var == 0 && ls.first_cycle_insn_p)
6361 goto end_schedule;
6362 advance = 1;
6363 break;
6364 }
6365 }
6366 }
6367 if (ls.modulo_epilogue)
6368 success = true;
6369 end_schedule:
6370 advance_one_cycle ();
6371 perform_replacements_new_cycle ();
6372 if (modulo_ii > 0)
6373 {
6374 /* Once again, debug insn suckiness: they can be on the ready list
6375 even if they have unresolved dependencies. To make our view
6376 of the world consistent, remove such "ready" insns. */
6377 restart_debug_insn_loop:
6378 for (i = ready.n_ready - 1; i >= 0; i--)
6379 {
6380 rtx x;
6381
6382 x = ready_element (&ready, i);
6383 if (DEPS_LIST_FIRST (INSN_HARD_BACK_DEPS (x)) != NULL
6384 || DEPS_LIST_FIRST (INSN_SPEC_BACK_DEPS (x)) != NULL)
6385 {
6386 ready_remove (&ready, i);
6387 goto restart_debug_insn_loop;
6388 }
6389 }
6390 for (i = ready.n_ready - 1; i >= 0; i--)
6391 {
6392 rtx x;
6393
6394 x = ready_element (&ready, i);
6395 resolve_dependencies (x);
6396 }
6397 for (i = 0; i <= max_insn_queue_index; i++)
6398 {
6399 rtx link;
6400 while ((link = insn_queue[i]) != NULL)
6401 {
6402 rtx x = XEXP (link, 0);
6403 insn_queue[i] = XEXP (link, 1);
6404 QUEUE_INDEX (x) = QUEUE_NOWHERE;
6405 free_INSN_LIST_node (link);
6406 resolve_dependencies (x);
6407 }
6408 }
6409 }
6410
6411 if (!success)
6412 undo_all_replacements ();
6413
6414 /* Debug info. */
6415 if (sched_verbose)
6416 {
6417 fprintf (sched_dump, ";;\tReady list (final): ");
6418 debug_ready_list (&ready);
6419 }
6420
6421 if (modulo_ii == 0 && current_sched_info->queue_must_finish_empty)
6422 /* Sanity check -- queue must be empty now. Meaningless if region has
6423 multiple bbs. */
6424 gcc_assert (!q_size && !ready.n_ready && !ready.n_debug);
6425 else if (modulo_ii == 0)
6426 {
6427 /* We must maintain QUEUE_INDEX between blocks in region. */
6428 for (i = ready.n_ready - 1; i >= 0; i--)
6429 {
6430 rtx x;
6431
6432 x = ready_element (&ready, i);
6433 QUEUE_INDEX (x) = QUEUE_NOWHERE;
6434 TODO_SPEC (x) = HARD_DEP;
6435 }
6436
6437 if (q_size)
6438 for (i = 0; i <= max_insn_queue_index; i++)
6439 {
6440 rtx link;
6441 for (link = insn_queue[i]; link; link = XEXP (link, 1))
6442 {
6443 rtx x;
6444
6445 x = XEXP (link, 0);
6446 QUEUE_INDEX (x) = QUEUE_NOWHERE;
6447 TODO_SPEC (x) = HARD_DEP;
6448 }
6449 free_INSN_LIST_list (&insn_queue[i]);
6450 }
6451 }
6452
6453 if (sched_pressure == SCHED_PRESSURE_MODEL)
6454 model_end_schedule ();
6455
6456 if (success)
6457 {
6458 commit_schedule (prev_head, tail, target_bb);
6459 if (sched_verbose)
6460 fprintf (sched_dump, ";; total time = %d\n", clock_var);
6461 }
6462 else
6463 last_scheduled_insn = tail;
6464
6465 scheduled_insns.truncate (0);
6466
6467 if (!current_sched_info->queue_must_finish_empty
6468 || haifa_recovery_bb_recently_added_p)
6469 {
6470 /* INSN_TICK (minimum clock tick at which the insn becomes
6471 ready) may be not correct for the insn in the subsequent
6472 blocks of the region. We should use a correct value of
6473 `clock_var' or modify INSN_TICK. It is better to keep
6474 clock_var value equal to 0 at the start of a basic block.
6475 Therefore we modify INSN_TICK here. */
6476 fix_inter_tick (NEXT_INSN (prev_head), last_scheduled_insn);
6477 }
6478
6479 if (targetm.sched.finish)
6480 {
6481 targetm.sched.finish (sched_dump, sched_verbose);
6482 /* Target might have added some instructions to the scheduled block
6483 in its md_finish () hook. These new insns don't have any data
6484 initialized and to identify them we extend h_i_d so that they'll
6485 get zero luids. */
6486 sched_extend_luids ();
6487 }
6488
6489 if (sched_verbose)
6490 fprintf (sched_dump, ";; new head = %d\n;; new tail = %d\n\n",
6491 INSN_UID (head), INSN_UID (tail));
6492
6493 /* Update head/tail boundaries. */
6494 head = NEXT_INSN (prev_head);
6495 tail = last_scheduled_insn;
6496
6497 head = restore_other_notes (head, NULL);
6498
6499 current_sched_info->head = head;
6500 current_sched_info->tail = tail;
6501
6502 free_backtrack_queue ();
6503
6504 return success;
6505 }
6506 \f
6507 /* Set_priorities: compute priority of each insn in the block. */
6508
6509 int
6510 set_priorities (rtx head, rtx tail)
6511 {
6512 rtx insn;
6513 int n_insn;
6514 int sched_max_insns_priority =
6515 current_sched_info->sched_max_insns_priority;
6516 rtx prev_head;
6517
6518 if (head == tail && ! INSN_P (head))
6519 gcc_unreachable ();
6520
6521 n_insn = 0;
6522
6523 prev_head = PREV_INSN (head);
6524 for (insn = tail; insn != prev_head; insn = PREV_INSN (insn))
6525 {
6526 if (!INSN_P (insn))
6527 continue;
6528
6529 n_insn++;
6530 (void) priority (insn);
6531
6532 gcc_assert (INSN_PRIORITY_KNOWN (insn));
6533
6534 sched_max_insns_priority = MAX (sched_max_insns_priority,
6535 INSN_PRIORITY (insn));
6536 }
6537
6538 current_sched_info->sched_max_insns_priority = sched_max_insns_priority;
6539
6540 return n_insn;
6541 }
6542
6543 /* Set dump and sched_verbose for the desired debugging output. If no
6544 dump-file was specified, but -fsched-verbose=N (any N), print to stderr.
6545 For -fsched-verbose=N, N>=10, print everything to stderr. */
6546 void
6547 setup_sched_dump (void)
6548 {
6549 sched_verbose = sched_verbose_param;
6550 if (sched_verbose_param == 0 && dump_file)
6551 sched_verbose = 1;
6552 sched_dump = ((sched_verbose_param >= 10 || !dump_file)
6553 ? stderr : dump_file);
6554 }
6555
6556 /* Try to group comparison and the following conditional jump INSN if
6557 they're already adjacent. This is to prevent scheduler from scheduling
6558 them apart. */
6559
6560 static void
6561 try_group_insn (rtx insn)
6562 {
6563 unsigned int condreg1, condreg2;
6564 rtx cc_reg_1;
6565 rtx prev;
6566
6567 if (!any_condjump_p (insn))
6568 return;
6569
6570 targetm.fixed_condition_code_regs (&condreg1, &condreg2);
6571 cc_reg_1 = gen_rtx_REG (CCmode, condreg1);
6572 prev = prev_nonnote_nondebug_insn (insn);
6573 if (!reg_referenced_p (cc_reg_1, PATTERN (insn))
6574 || !prev
6575 || !modified_in_p (cc_reg_1, prev))
6576 return;
6577
6578 /* Different microarchitectures support macro fusions for different
6579 combinations of insn pairs. */
6580 if (!targetm.sched.macro_fusion_pair_p
6581 || !targetm.sched.macro_fusion_pair_p (prev, insn))
6582 return;
6583
6584 SCHED_GROUP_P (insn) = 1;
6585 }
6586
6587 /* If the last cond jump and the cond register defining insn are consecutive
6588 before scheduling, we want them to be in a schedule group. This is good
6589 for performance on microarchitectures supporting macro-fusion. */
6590
6591 static void
6592 group_insns_for_macro_fusion ()
6593 {
6594 basic_block bb;
6595
6596 FOR_EACH_BB (bb)
6597 try_group_insn (BB_END (bb));
6598 }
6599
6600 /* Initialize some global state for the scheduler. This function works
6601 with the common data shared between all the schedulers. It is called
6602 from the scheduler specific initialization routine. */
6603
6604 void
6605 sched_init (void)
6606 {
6607 /* Disable speculative loads in their presence if cc0 defined. */
6608 #ifdef HAVE_cc0
6609 flag_schedule_speculative_load = 0;
6610 #endif
6611
6612 if (targetm.sched.dispatch (NULL_RTX, IS_DISPATCH_ON))
6613 targetm.sched.dispatch_do (NULL_RTX, DISPATCH_INIT);
6614
6615 if (live_range_shrinkage_p)
6616 sched_pressure = SCHED_PRESSURE_WEIGHTED;
6617 else if (flag_sched_pressure
6618 && !reload_completed
6619 && common_sched_info->sched_pass_id == SCHED_RGN_PASS)
6620 sched_pressure = ((enum sched_pressure_algorithm)
6621 PARAM_VALUE (PARAM_SCHED_PRESSURE_ALGORITHM));
6622 else
6623 sched_pressure = SCHED_PRESSURE_NONE;
6624
6625 if (sched_pressure != SCHED_PRESSURE_NONE)
6626 ira_setup_eliminable_regset ();
6627
6628 /* Initialize SPEC_INFO. */
6629 if (targetm.sched.set_sched_flags)
6630 {
6631 spec_info = &spec_info_var;
6632 targetm.sched.set_sched_flags (spec_info);
6633
6634 if (spec_info->mask != 0)
6635 {
6636 spec_info->data_weakness_cutoff =
6637 (PARAM_VALUE (PARAM_SCHED_SPEC_PROB_CUTOFF) * MAX_DEP_WEAK) / 100;
6638 spec_info->control_weakness_cutoff =
6639 (PARAM_VALUE (PARAM_SCHED_SPEC_PROB_CUTOFF)
6640 * REG_BR_PROB_BASE) / 100;
6641 }
6642 else
6643 /* So we won't read anything accidentally. */
6644 spec_info = NULL;
6645
6646 }
6647 else
6648 /* So we won't read anything accidentally. */
6649 spec_info = 0;
6650
6651 /* Initialize issue_rate. */
6652 if (targetm.sched.issue_rate)
6653 issue_rate = targetm.sched.issue_rate ();
6654 else
6655 issue_rate = 1;
6656
6657 if (cached_issue_rate != issue_rate)
6658 {
6659 cached_issue_rate = issue_rate;
6660 /* To invalidate max_lookahead_tries: */
6661 cached_first_cycle_multipass_dfa_lookahead = 0;
6662 }
6663
6664 if (targetm.sched.first_cycle_multipass_dfa_lookahead)
6665 dfa_lookahead = targetm.sched.first_cycle_multipass_dfa_lookahead ();
6666 else
6667 dfa_lookahead = 0;
6668
6669 if (targetm.sched.init_dfa_pre_cycle_insn)
6670 targetm.sched.init_dfa_pre_cycle_insn ();
6671
6672 if (targetm.sched.init_dfa_post_cycle_insn)
6673 targetm.sched.init_dfa_post_cycle_insn ();
6674
6675 dfa_start ();
6676 dfa_state_size = state_size ();
6677
6678 init_alias_analysis ();
6679
6680 if (!sched_no_dce)
6681 df_set_flags (DF_LR_RUN_DCE);
6682 df_note_add_problem ();
6683
6684 /* More problems needed for interloop dep calculation in SMS. */
6685 if (common_sched_info->sched_pass_id == SCHED_SMS_PASS)
6686 {
6687 df_rd_add_problem ();
6688 df_chain_add_problem (DF_DU_CHAIN + DF_UD_CHAIN);
6689 }
6690
6691 df_analyze ();
6692
6693 /* Do not run DCE after reload, as this can kill nops inserted
6694 by bundling. */
6695 if (reload_completed)
6696 df_clear_flags (DF_LR_RUN_DCE);
6697
6698 regstat_compute_calls_crossed ();
6699
6700 if (targetm.sched.init_global)
6701 targetm.sched.init_global (sched_dump, sched_verbose, get_max_uid () + 1);
6702
6703 if (sched_pressure != SCHED_PRESSURE_NONE)
6704 {
6705 int i, max_regno = max_reg_num ();
6706
6707 if (sched_dump != NULL)
6708 /* We need info about pseudos for rtl dumps about pseudo
6709 classes and costs. */
6710 regstat_init_n_sets_and_refs ();
6711 ira_set_pseudo_classes (true, sched_verbose ? sched_dump : NULL);
6712 sched_regno_pressure_class
6713 = (enum reg_class *) xmalloc (max_regno * sizeof (enum reg_class));
6714 for (i = 0; i < max_regno; i++)
6715 sched_regno_pressure_class[i]
6716 = (i < FIRST_PSEUDO_REGISTER
6717 ? ira_pressure_class_translate[REGNO_REG_CLASS (i)]
6718 : ira_pressure_class_translate[reg_allocno_class (i)]);
6719 curr_reg_live = BITMAP_ALLOC (NULL);
6720 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
6721 {
6722 saved_reg_live = BITMAP_ALLOC (NULL);
6723 region_ref_regs = BITMAP_ALLOC (NULL);
6724 }
6725 }
6726
6727 curr_state = xmalloc (dfa_state_size);
6728
6729 /* Group compare and branch insns for macro-fusion. */
6730 if (targetm.sched.macro_fusion_p
6731 && targetm.sched.macro_fusion_p ())
6732 group_insns_for_macro_fusion ();
6733 }
6734
6735 static void haifa_init_only_bb (basic_block, basic_block);
6736
6737 /* Initialize data structures specific to the Haifa scheduler. */
6738 void
6739 haifa_sched_init (void)
6740 {
6741 setup_sched_dump ();
6742 sched_init ();
6743
6744 scheduled_insns.create (0);
6745
6746 if (spec_info != NULL)
6747 {
6748 sched_deps_info->use_deps_list = 1;
6749 sched_deps_info->generate_spec_deps = 1;
6750 }
6751
6752 /* Initialize luids, dependency caches, target and h_i_d for the
6753 whole function. */
6754 {
6755 bb_vec_t bbs;
6756 bbs.create (n_basic_blocks_for_fn (cfun));
6757 basic_block bb;
6758
6759 sched_init_bbs ();
6760
6761 FOR_EACH_BB (bb)
6762 bbs.quick_push (bb);
6763 sched_init_luids (bbs);
6764 sched_deps_init (true);
6765 sched_extend_target ();
6766 haifa_init_h_i_d (bbs);
6767
6768 bbs.release ();
6769 }
6770
6771 sched_init_only_bb = haifa_init_only_bb;
6772 sched_split_block = sched_split_block_1;
6773 sched_create_empty_bb = sched_create_empty_bb_1;
6774 haifa_recovery_bb_ever_added_p = false;
6775
6776 nr_begin_data = nr_begin_control = nr_be_in_data = nr_be_in_control = 0;
6777 before_recovery = 0;
6778 after_recovery = 0;
6779
6780 modulo_ii = 0;
6781 }
6782
6783 /* Finish work with the data specific to the Haifa scheduler. */
6784 void
6785 haifa_sched_finish (void)
6786 {
6787 sched_create_empty_bb = NULL;
6788 sched_split_block = NULL;
6789 sched_init_only_bb = NULL;
6790
6791 if (spec_info && spec_info->dump)
6792 {
6793 char c = reload_completed ? 'a' : 'b';
6794
6795 fprintf (spec_info->dump,
6796 ";; %s:\n", current_function_name ());
6797
6798 fprintf (spec_info->dump,
6799 ";; Procedure %cr-begin-data-spec motions == %d\n",
6800 c, nr_begin_data);
6801 fprintf (spec_info->dump,
6802 ";; Procedure %cr-be-in-data-spec motions == %d\n",
6803 c, nr_be_in_data);
6804 fprintf (spec_info->dump,
6805 ";; Procedure %cr-begin-control-spec motions == %d\n",
6806 c, nr_begin_control);
6807 fprintf (spec_info->dump,
6808 ";; Procedure %cr-be-in-control-spec motions == %d\n",
6809 c, nr_be_in_control);
6810 }
6811
6812 scheduled_insns.release ();
6813
6814 /* Finalize h_i_d, dependency caches, and luids for the whole
6815 function. Target will be finalized in md_global_finish (). */
6816 sched_deps_finish ();
6817 sched_finish_luids ();
6818 current_sched_info = NULL;
6819 sched_finish ();
6820 }
6821
6822 /* Free global data used during insn scheduling. This function works with
6823 the common data shared between the schedulers. */
6824
6825 void
6826 sched_finish (void)
6827 {
6828 haifa_finish_h_i_d ();
6829 if (sched_pressure != SCHED_PRESSURE_NONE)
6830 {
6831 if (regstat_n_sets_and_refs != NULL)
6832 regstat_free_n_sets_and_refs ();
6833 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
6834 {
6835 BITMAP_FREE (region_ref_regs);
6836 BITMAP_FREE (saved_reg_live);
6837 }
6838 BITMAP_FREE (curr_reg_live);
6839 free (sched_regno_pressure_class);
6840 }
6841 free (curr_state);
6842
6843 if (targetm.sched.finish_global)
6844 targetm.sched.finish_global (sched_dump, sched_verbose);
6845
6846 end_alias_analysis ();
6847
6848 regstat_free_calls_crossed ();
6849
6850 dfa_finish ();
6851 }
6852
6853 /* Free all delay_pair structures that were recorded. */
6854 void
6855 free_delay_pairs (void)
6856 {
6857 if (delay_htab.is_created ())
6858 {
6859 delay_htab.empty ();
6860 delay_htab_i2.empty ();
6861 }
6862 }
6863
6864 /* Fix INSN_TICKs of the instructions in the current block as well as
6865 INSN_TICKs of their dependents.
6866 HEAD and TAIL are the begin and the end of the current scheduled block. */
6867 static void
6868 fix_inter_tick (rtx head, rtx tail)
6869 {
6870 /* Set of instructions with corrected INSN_TICK. */
6871 bitmap_head processed;
6872 /* ??? It is doubtful if we should assume that cycle advance happens on
6873 basic block boundaries. Basically insns that are unconditionally ready
6874 on the start of the block are more preferable then those which have
6875 a one cycle dependency over insn from the previous block. */
6876 int next_clock = clock_var + 1;
6877
6878 bitmap_initialize (&processed, 0);
6879
6880 /* Iterates over scheduled instructions and fix their INSN_TICKs and
6881 INSN_TICKs of dependent instructions, so that INSN_TICKs are consistent
6882 across different blocks. */
6883 for (tail = NEXT_INSN (tail); head != tail; head = NEXT_INSN (head))
6884 {
6885 if (INSN_P (head))
6886 {
6887 int tick;
6888 sd_iterator_def sd_it;
6889 dep_t dep;
6890
6891 tick = INSN_TICK (head);
6892 gcc_assert (tick >= MIN_TICK);
6893
6894 /* Fix INSN_TICK of instruction from just scheduled block. */
6895 if (bitmap_set_bit (&processed, INSN_LUID (head)))
6896 {
6897 tick -= next_clock;
6898
6899 if (tick < MIN_TICK)
6900 tick = MIN_TICK;
6901
6902 INSN_TICK (head) = tick;
6903 }
6904
6905 if (DEBUG_INSN_P (head))
6906 continue;
6907
6908 FOR_EACH_DEP (head, SD_LIST_RES_FORW, sd_it, dep)
6909 {
6910 rtx next;
6911
6912 next = DEP_CON (dep);
6913 tick = INSN_TICK (next);
6914
6915 if (tick != INVALID_TICK
6916 /* If NEXT has its INSN_TICK calculated, fix it.
6917 If not - it will be properly calculated from
6918 scratch later in fix_tick_ready. */
6919 && bitmap_set_bit (&processed, INSN_LUID (next)))
6920 {
6921 tick -= next_clock;
6922
6923 if (tick < MIN_TICK)
6924 tick = MIN_TICK;
6925
6926 if (tick > INTER_TICK (next))
6927 INTER_TICK (next) = tick;
6928 else
6929 tick = INTER_TICK (next);
6930
6931 INSN_TICK (next) = tick;
6932 }
6933 }
6934 }
6935 }
6936 bitmap_clear (&processed);
6937 }
6938
6939 /* Check if NEXT is ready to be added to the ready or queue list.
6940 If "yes", add it to the proper list.
6941 Returns:
6942 -1 - is not ready yet,
6943 0 - added to the ready list,
6944 0 < N - queued for N cycles. */
6945 int
6946 try_ready (rtx next)
6947 {
6948 ds_t old_ts, new_ts;
6949
6950 old_ts = TODO_SPEC (next);
6951
6952 gcc_assert (!(old_ts & ~(SPECULATIVE | HARD_DEP | DEP_CONTROL | DEP_POSTPONED))
6953 && (old_ts == HARD_DEP
6954 || old_ts == DEP_POSTPONED
6955 || (old_ts & SPECULATIVE)
6956 || old_ts == DEP_CONTROL));
6957
6958 new_ts = recompute_todo_spec (next, false);
6959
6960 if (new_ts & (HARD_DEP | DEP_POSTPONED))
6961 gcc_assert (new_ts == old_ts
6962 && QUEUE_INDEX (next) == QUEUE_NOWHERE);
6963 else if (current_sched_info->new_ready)
6964 new_ts = current_sched_info->new_ready (next, new_ts);
6965
6966 /* * if !(old_ts & SPECULATIVE) (e.g. HARD_DEP or 0), then insn might
6967 have its original pattern or changed (speculative) one. This is due
6968 to changing ebb in region scheduling.
6969 * But if (old_ts & SPECULATIVE), then we are pretty sure that insn
6970 has speculative pattern.
6971
6972 We can't assert (!(new_ts & HARD_DEP) || new_ts == old_ts) here because
6973 control-speculative NEXT could have been discarded by sched-rgn.c
6974 (the same case as when discarded by can_schedule_ready_p ()). */
6975
6976 if ((new_ts & SPECULATIVE)
6977 /* If (old_ts == new_ts), then (old_ts & SPECULATIVE) and we don't
6978 need to change anything. */
6979 && new_ts != old_ts)
6980 {
6981 int res;
6982 rtx new_pat;
6983
6984 gcc_assert ((new_ts & SPECULATIVE) && !(new_ts & ~SPECULATIVE));
6985
6986 res = haifa_speculate_insn (next, new_ts, &new_pat);
6987
6988 switch (res)
6989 {
6990 case -1:
6991 /* It would be nice to change DEP_STATUS of all dependences,
6992 which have ((DEP_STATUS & SPECULATIVE) == new_ts) to HARD_DEP,
6993 so we won't reanalyze anything. */
6994 new_ts = HARD_DEP;
6995 break;
6996
6997 case 0:
6998 /* We follow the rule, that every speculative insn
6999 has non-null ORIG_PAT. */
7000 if (!ORIG_PAT (next))
7001 ORIG_PAT (next) = PATTERN (next);
7002 break;
7003
7004 case 1:
7005 if (!ORIG_PAT (next))
7006 /* If we gonna to overwrite the original pattern of insn,
7007 save it. */
7008 ORIG_PAT (next) = PATTERN (next);
7009
7010 res = haifa_change_pattern (next, new_pat);
7011 gcc_assert (res);
7012 break;
7013
7014 default:
7015 gcc_unreachable ();
7016 }
7017 }
7018
7019 /* We need to restore pattern only if (new_ts == 0), because otherwise it is
7020 either correct (new_ts & SPECULATIVE),
7021 or we simply don't care (new_ts & HARD_DEP). */
7022
7023 gcc_assert (!ORIG_PAT (next)
7024 || !IS_SPECULATION_BRANCHY_CHECK_P (next));
7025
7026 TODO_SPEC (next) = new_ts;
7027
7028 if (new_ts & (HARD_DEP | DEP_POSTPONED))
7029 {
7030 /* We can't assert (QUEUE_INDEX (next) == QUEUE_NOWHERE) here because
7031 control-speculative NEXT could have been discarded by sched-rgn.c
7032 (the same case as when discarded by can_schedule_ready_p ()). */
7033 /*gcc_assert (QUEUE_INDEX (next) == QUEUE_NOWHERE);*/
7034
7035 change_queue_index (next, QUEUE_NOWHERE);
7036
7037 return -1;
7038 }
7039 else if (!(new_ts & BEGIN_SPEC)
7040 && ORIG_PAT (next) && PREDICATED_PAT (next) == NULL_RTX
7041 && !IS_SPECULATION_CHECK_P (next))
7042 /* We should change pattern of every previously speculative
7043 instruction - and we determine if NEXT was speculative by using
7044 ORIG_PAT field. Except one case - speculation checks have ORIG_PAT
7045 pat too, so skip them. */
7046 {
7047 bool success = haifa_change_pattern (next, ORIG_PAT (next));
7048 gcc_assert (success);
7049 ORIG_PAT (next) = 0;
7050 }
7051
7052 if (sched_verbose >= 2)
7053 {
7054 fprintf (sched_dump, ";;\t\tdependencies resolved: insn %s",
7055 (*current_sched_info->print_insn) (next, 0));
7056
7057 if (spec_info && spec_info->dump)
7058 {
7059 if (new_ts & BEGIN_DATA)
7060 fprintf (spec_info->dump, "; data-spec;");
7061 if (new_ts & BEGIN_CONTROL)
7062 fprintf (spec_info->dump, "; control-spec;");
7063 if (new_ts & BE_IN_CONTROL)
7064 fprintf (spec_info->dump, "; in-control-spec;");
7065 }
7066 if (TODO_SPEC (next) & DEP_CONTROL)
7067 fprintf (sched_dump, " predicated");
7068 fprintf (sched_dump, "\n");
7069 }
7070
7071 adjust_priority (next);
7072
7073 return fix_tick_ready (next);
7074 }
7075
7076 /* Calculate INSN_TICK of NEXT and add it to either ready or queue list. */
7077 static int
7078 fix_tick_ready (rtx next)
7079 {
7080 int tick, delay;
7081
7082 if (!DEBUG_INSN_P (next) && !sd_lists_empty_p (next, SD_LIST_RES_BACK))
7083 {
7084 int full_p;
7085 sd_iterator_def sd_it;
7086 dep_t dep;
7087
7088 tick = INSN_TICK (next);
7089 /* if tick is not equal to INVALID_TICK, then update
7090 INSN_TICK of NEXT with the most recent resolved dependence
7091 cost. Otherwise, recalculate from scratch. */
7092 full_p = (tick == INVALID_TICK);
7093
7094 FOR_EACH_DEP (next, SD_LIST_RES_BACK, sd_it, dep)
7095 {
7096 rtx pro = DEP_PRO (dep);
7097 int tick1;
7098
7099 gcc_assert (INSN_TICK (pro) >= MIN_TICK);
7100
7101 tick1 = INSN_TICK (pro) + dep_cost (dep);
7102 if (tick1 > tick)
7103 tick = tick1;
7104
7105 if (!full_p)
7106 break;
7107 }
7108 }
7109 else
7110 tick = -1;
7111
7112 INSN_TICK (next) = tick;
7113
7114 delay = tick - clock_var;
7115 if (delay <= 0 || sched_pressure != SCHED_PRESSURE_NONE)
7116 delay = QUEUE_READY;
7117
7118 change_queue_index (next, delay);
7119
7120 return delay;
7121 }
7122
7123 /* Move NEXT to the proper queue list with (DELAY >= 1),
7124 or add it to the ready list (DELAY == QUEUE_READY),
7125 or remove it from ready and queue lists at all (DELAY == QUEUE_NOWHERE). */
7126 static void
7127 change_queue_index (rtx next, int delay)
7128 {
7129 int i = QUEUE_INDEX (next);
7130
7131 gcc_assert (QUEUE_NOWHERE <= delay && delay <= max_insn_queue_index
7132 && delay != 0);
7133 gcc_assert (i != QUEUE_SCHEDULED);
7134
7135 if ((delay > 0 && NEXT_Q_AFTER (q_ptr, delay) == i)
7136 || (delay < 0 && delay == i))
7137 /* We have nothing to do. */
7138 return;
7139
7140 /* Remove NEXT from wherever it is now. */
7141 if (i == QUEUE_READY)
7142 ready_remove_insn (next);
7143 else if (i >= 0)
7144 queue_remove (next);
7145
7146 /* Add it to the proper place. */
7147 if (delay == QUEUE_READY)
7148 ready_add (readyp, next, false);
7149 else if (delay >= 1)
7150 queue_insn (next, delay, "change queue index");
7151
7152 if (sched_verbose >= 2)
7153 {
7154 fprintf (sched_dump, ";;\t\ttick updated: insn %s",
7155 (*current_sched_info->print_insn) (next, 0));
7156
7157 if (delay == QUEUE_READY)
7158 fprintf (sched_dump, " into ready\n");
7159 else if (delay >= 1)
7160 fprintf (sched_dump, " into queue with cost=%d\n", delay);
7161 else
7162 fprintf (sched_dump, " removed from ready or queue lists\n");
7163 }
7164 }
7165
7166 static int sched_ready_n_insns = -1;
7167
7168 /* Initialize per region data structures. */
7169 void
7170 sched_extend_ready_list (int new_sched_ready_n_insns)
7171 {
7172 int i;
7173
7174 if (sched_ready_n_insns == -1)
7175 /* At the first call we need to initialize one more choice_stack
7176 entry. */
7177 {
7178 i = 0;
7179 sched_ready_n_insns = 0;
7180 scheduled_insns.reserve (new_sched_ready_n_insns);
7181 }
7182 else
7183 i = sched_ready_n_insns + 1;
7184
7185 ready.veclen = new_sched_ready_n_insns + issue_rate;
7186 ready.vec = XRESIZEVEC (rtx, ready.vec, ready.veclen);
7187
7188 gcc_assert (new_sched_ready_n_insns >= sched_ready_n_insns);
7189
7190 ready_try = (char *) xrecalloc (ready_try, new_sched_ready_n_insns,
7191 sched_ready_n_insns, sizeof (*ready_try));
7192
7193 /* We allocate +1 element to save initial state in the choice_stack[0]
7194 entry. */
7195 choice_stack = XRESIZEVEC (struct choice_entry, choice_stack,
7196 new_sched_ready_n_insns + 1);
7197
7198 for (; i <= new_sched_ready_n_insns; i++)
7199 {
7200 choice_stack[i].state = xmalloc (dfa_state_size);
7201
7202 if (targetm.sched.first_cycle_multipass_init)
7203 targetm.sched.first_cycle_multipass_init (&(choice_stack[i]
7204 .target_data));
7205 }
7206
7207 sched_ready_n_insns = new_sched_ready_n_insns;
7208 }
7209
7210 /* Free per region data structures. */
7211 void
7212 sched_finish_ready_list (void)
7213 {
7214 int i;
7215
7216 free (ready.vec);
7217 ready.vec = NULL;
7218 ready.veclen = 0;
7219
7220 free (ready_try);
7221 ready_try = NULL;
7222
7223 for (i = 0; i <= sched_ready_n_insns; i++)
7224 {
7225 if (targetm.sched.first_cycle_multipass_fini)
7226 targetm.sched.first_cycle_multipass_fini (&(choice_stack[i]
7227 .target_data));
7228
7229 free (choice_stack [i].state);
7230 }
7231 free (choice_stack);
7232 choice_stack = NULL;
7233
7234 sched_ready_n_insns = -1;
7235 }
7236
7237 static int
7238 haifa_luid_for_non_insn (rtx x)
7239 {
7240 gcc_assert (NOTE_P (x) || LABEL_P (x));
7241
7242 return 0;
7243 }
7244
7245 /* Generates recovery code for INSN. */
7246 static void
7247 generate_recovery_code (rtx insn)
7248 {
7249 if (TODO_SPEC (insn) & BEGIN_SPEC)
7250 begin_speculative_block (insn);
7251
7252 /* Here we have insn with no dependencies to
7253 instructions other then CHECK_SPEC ones. */
7254
7255 if (TODO_SPEC (insn) & BE_IN_SPEC)
7256 add_to_speculative_block (insn);
7257 }
7258
7259 /* Helper function.
7260 Tries to add speculative dependencies of type FS between instructions
7261 in deps_list L and TWIN. */
7262 static void
7263 process_insn_forw_deps_be_in_spec (rtx insn, rtx twin, ds_t fs)
7264 {
7265 sd_iterator_def sd_it;
7266 dep_t dep;
7267
7268 FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep)
7269 {
7270 ds_t ds;
7271 rtx consumer;
7272
7273 consumer = DEP_CON (dep);
7274
7275 ds = DEP_STATUS (dep);
7276
7277 if (/* If we want to create speculative dep. */
7278 fs
7279 /* And we can do that because this is a true dep. */
7280 && (ds & DEP_TYPES) == DEP_TRUE)
7281 {
7282 gcc_assert (!(ds & BE_IN_SPEC));
7283
7284 if (/* If this dep can be overcome with 'begin speculation'. */
7285 ds & BEGIN_SPEC)
7286 /* Then we have a choice: keep the dep 'begin speculative'
7287 or transform it into 'be in speculative'. */
7288 {
7289 if (/* In try_ready we assert that if insn once became ready
7290 it can be removed from the ready (or queue) list only
7291 due to backend decision. Hence we can't let the
7292 probability of the speculative dep to decrease. */
7293 ds_weak (ds) <= ds_weak (fs))
7294 {
7295 ds_t new_ds;
7296
7297 new_ds = (ds & ~BEGIN_SPEC) | fs;
7298
7299 if (/* consumer can 'be in speculative'. */
7300 sched_insn_is_legitimate_for_speculation_p (consumer,
7301 new_ds))
7302 /* Transform it to be in speculative. */
7303 ds = new_ds;
7304 }
7305 }
7306 else
7307 /* Mark the dep as 'be in speculative'. */
7308 ds |= fs;
7309 }
7310
7311 {
7312 dep_def _new_dep, *new_dep = &_new_dep;
7313
7314 init_dep_1 (new_dep, twin, consumer, DEP_TYPE (dep), ds);
7315 sd_add_dep (new_dep, false);
7316 }
7317 }
7318 }
7319
7320 /* Generates recovery code for BEGIN speculative INSN. */
7321 static void
7322 begin_speculative_block (rtx insn)
7323 {
7324 if (TODO_SPEC (insn) & BEGIN_DATA)
7325 nr_begin_data++;
7326 if (TODO_SPEC (insn) & BEGIN_CONTROL)
7327 nr_begin_control++;
7328
7329 create_check_block_twin (insn, false);
7330
7331 TODO_SPEC (insn) &= ~BEGIN_SPEC;
7332 }
7333
7334 static void haifa_init_insn (rtx);
7335
7336 /* Generates recovery code for BE_IN speculative INSN. */
7337 static void
7338 add_to_speculative_block (rtx insn)
7339 {
7340 ds_t ts;
7341 sd_iterator_def sd_it;
7342 dep_t dep;
7343 rtx twins = NULL;
7344 rtx_vec_t priorities_roots;
7345
7346 ts = TODO_SPEC (insn);
7347 gcc_assert (!(ts & ~BE_IN_SPEC));
7348
7349 if (ts & BE_IN_DATA)
7350 nr_be_in_data++;
7351 if (ts & BE_IN_CONTROL)
7352 nr_be_in_control++;
7353
7354 TODO_SPEC (insn) &= ~BE_IN_SPEC;
7355 gcc_assert (!TODO_SPEC (insn));
7356
7357 DONE_SPEC (insn) |= ts;
7358
7359 /* First we convert all simple checks to branchy. */
7360 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
7361 sd_iterator_cond (&sd_it, &dep);)
7362 {
7363 rtx check = DEP_PRO (dep);
7364
7365 if (IS_SPECULATION_SIMPLE_CHECK_P (check))
7366 {
7367 create_check_block_twin (check, true);
7368
7369 /* Restart search. */
7370 sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
7371 }
7372 else
7373 /* Continue search. */
7374 sd_iterator_next (&sd_it);
7375 }
7376
7377 priorities_roots.create (0);
7378 clear_priorities (insn, &priorities_roots);
7379
7380 while (1)
7381 {
7382 rtx check, twin;
7383 basic_block rec;
7384
7385 /* Get the first backward dependency of INSN. */
7386 sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
7387 if (!sd_iterator_cond (&sd_it, &dep))
7388 /* INSN has no backward dependencies left. */
7389 break;
7390
7391 gcc_assert ((DEP_STATUS (dep) & BEGIN_SPEC) == 0
7392 && (DEP_STATUS (dep) & BE_IN_SPEC) != 0
7393 && (DEP_STATUS (dep) & DEP_TYPES) == DEP_TRUE);
7394
7395 check = DEP_PRO (dep);
7396
7397 gcc_assert (!IS_SPECULATION_CHECK_P (check) && !ORIG_PAT (check)
7398 && QUEUE_INDEX (check) == QUEUE_NOWHERE);
7399
7400 rec = BLOCK_FOR_INSN (check);
7401
7402 twin = emit_insn_before (copy_insn (PATTERN (insn)), BB_END (rec));
7403 haifa_init_insn (twin);
7404
7405 sd_copy_back_deps (twin, insn, true);
7406
7407 if (sched_verbose && spec_info->dump)
7408 /* INSN_BB (insn) isn't determined for twin insns yet.
7409 So we can't use current_sched_info->print_insn. */
7410 fprintf (spec_info->dump, ";;\t\tGenerated twin insn : %d/rec%d\n",
7411 INSN_UID (twin), rec->index);
7412
7413 twins = alloc_INSN_LIST (twin, twins);
7414
7415 /* Add dependences between TWIN and all appropriate
7416 instructions from REC. */
7417 FOR_EACH_DEP (insn, SD_LIST_SPEC_BACK, sd_it, dep)
7418 {
7419 rtx pro = DEP_PRO (dep);
7420
7421 gcc_assert (DEP_TYPE (dep) == REG_DEP_TRUE);
7422
7423 /* INSN might have dependencies from the instructions from
7424 several recovery blocks. At this iteration we process those
7425 producers that reside in REC. */
7426 if (BLOCK_FOR_INSN (pro) == rec)
7427 {
7428 dep_def _new_dep, *new_dep = &_new_dep;
7429
7430 init_dep (new_dep, pro, twin, REG_DEP_TRUE);
7431 sd_add_dep (new_dep, false);
7432 }
7433 }
7434
7435 process_insn_forw_deps_be_in_spec (insn, twin, ts);
7436
7437 /* Remove all dependencies between INSN and insns in REC. */
7438 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
7439 sd_iterator_cond (&sd_it, &dep);)
7440 {
7441 rtx pro = DEP_PRO (dep);
7442
7443 if (BLOCK_FOR_INSN (pro) == rec)
7444 sd_delete_dep (sd_it);
7445 else
7446 sd_iterator_next (&sd_it);
7447 }
7448 }
7449
7450 /* We couldn't have added the dependencies between INSN and TWINS earlier
7451 because that would make TWINS appear in the INSN_BACK_DEPS (INSN). */
7452 while (twins)
7453 {
7454 rtx twin;
7455
7456 twin = XEXP (twins, 0);
7457
7458 {
7459 dep_def _new_dep, *new_dep = &_new_dep;
7460
7461 init_dep (new_dep, insn, twin, REG_DEP_OUTPUT);
7462 sd_add_dep (new_dep, false);
7463 }
7464
7465 twin = XEXP (twins, 1);
7466 free_INSN_LIST_node (twins);
7467 twins = twin;
7468 }
7469
7470 calc_priorities (priorities_roots);
7471 priorities_roots.release ();
7472 }
7473
7474 /* Extends and fills with zeros (only the new part) array pointed to by P. */
7475 void *
7476 xrecalloc (void *p, size_t new_nmemb, size_t old_nmemb, size_t size)
7477 {
7478 gcc_assert (new_nmemb >= old_nmemb);
7479 p = XRESIZEVAR (void, p, new_nmemb * size);
7480 memset (((char *) p) + old_nmemb * size, 0, (new_nmemb - old_nmemb) * size);
7481 return p;
7482 }
7483
7484 /* Helper function.
7485 Find fallthru edge from PRED. */
7486 edge
7487 find_fallthru_edge_from (basic_block pred)
7488 {
7489 edge e;
7490 basic_block succ;
7491
7492 succ = pred->next_bb;
7493 gcc_assert (succ->prev_bb == pred);
7494
7495 if (EDGE_COUNT (pred->succs) <= EDGE_COUNT (succ->preds))
7496 {
7497 e = find_fallthru_edge (pred->succs);
7498
7499 if (e)
7500 {
7501 gcc_assert (e->dest == succ);
7502 return e;
7503 }
7504 }
7505 else
7506 {
7507 e = find_fallthru_edge (succ->preds);
7508
7509 if (e)
7510 {
7511 gcc_assert (e->src == pred);
7512 return e;
7513 }
7514 }
7515
7516 return NULL;
7517 }
7518
7519 /* Extend per basic block data structures. */
7520 static void
7521 sched_extend_bb (void)
7522 {
7523 /* The following is done to keep current_sched_info->next_tail non null. */
7524 rtx end = BB_END (EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb);
7525 rtx insn = DEBUG_INSN_P (end) ? prev_nondebug_insn (end) : end;
7526 if (NEXT_INSN (end) == 0
7527 || (!NOTE_P (insn)
7528 && !LABEL_P (insn)
7529 /* Don't emit a NOTE if it would end up before a BARRIER. */
7530 && !BARRIER_P (NEXT_INSN (end))))
7531 {
7532 rtx note = emit_note_after (NOTE_INSN_DELETED, end);
7533 /* Make note appear outside BB. */
7534 set_block_for_insn (note, NULL);
7535 BB_END (EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb) = end;
7536 }
7537 }
7538
7539 /* Init per basic block data structures. */
7540 void
7541 sched_init_bbs (void)
7542 {
7543 sched_extend_bb ();
7544 }
7545
7546 /* Initialize BEFORE_RECOVERY variable. */
7547 static void
7548 init_before_recovery (basic_block *before_recovery_ptr)
7549 {
7550 basic_block last;
7551 edge e;
7552
7553 last = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
7554 e = find_fallthru_edge_from (last);
7555
7556 if (e)
7557 {
7558 /* We create two basic blocks:
7559 1. Single instruction block is inserted right after E->SRC
7560 and has jump to
7561 2. Empty block right before EXIT_BLOCK.
7562 Between these two blocks recovery blocks will be emitted. */
7563
7564 basic_block single, empty;
7565 rtx x, label;
7566
7567 /* If the fallthrough edge to exit we've found is from the block we've
7568 created before, don't do anything more. */
7569 if (last == after_recovery)
7570 return;
7571
7572 adding_bb_to_current_region_p = false;
7573
7574 single = sched_create_empty_bb (last);
7575 empty = sched_create_empty_bb (single);
7576
7577 /* Add new blocks to the root loop. */
7578 if (current_loops != NULL)
7579 {
7580 add_bb_to_loop (single, (*current_loops->larray)[0]);
7581 add_bb_to_loop (empty, (*current_loops->larray)[0]);
7582 }
7583
7584 single->count = last->count;
7585 empty->count = last->count;
7586 single->frequency = last->frequency;
7587 empty->frequency = last->frequency;
7588 BB_COPY_PARTITION (single, last);
7589 BB_COPY_PARTITION (empty, last);
7590
7591 redirect_edge_succ (e, single);
7592 make_single_succ_edge (single, empty, 0);
7593 make_single_succ_edge (empty, EXIT_BLOCK_PTR_FOR_FN (cfun),
7594 EDGE_FALLTHRU);
7595
7596 label = block_label (empty);
7597 x = emit_jump_insn_after (gen_jump (label), BB_END (single));
7598 JUMP_LABEL (x) = label;
7599 LABEL_NUSES (label)++;
7600 haifa_init_insn (x);
7601
7602 emit_barrier_after (x);
7603
7604 sched_init_only_bb (empty, NULL);
7605 sched_init_only_bb (single, NULL);
7606 sched_extend_bb ();
7607
7608 adding_bb_to_current_region_p = true;
7609 before_recovery = single;
7610 after_recovery = empty;
7611
7612 if (before_recovery_ptr)
7613 *before_recovery_ptr = before_recovery;
7614
7615 if (sched_verbose >= 2 && spec_info->dump)
7616 fprintf (spec_info->dump,
7617 ";;\t\tFixed fallthru to EXIT : %d->>%d->%d->>EXIT\n",
7618 last->index, single->index, empty->index);
7619 }
7620 else
7621 before_recovery = last;
7622 }
7623
7624 /* Returns new recovery block. */
7625 basic_block
7626 sched_create_recovery_block (basic_block *before_recovery_ptr)
7627 {
7628 rtx label;
7629 rtx barrier;
7630 basic_block rec;
7631
7632 haifa_recovery_bb_recently_added_p = true;
7633 haifa_recovery_bb_ever_added_p = true;
7634
7635 init_before_recovery (before_recovery_ptr);
7636
7637 barrier = get_last_bb_insn (before_recovery);
7638 gcc_assert (BARRIER_P (barrier));
7639
7640 label = emit_label_after (gen_label_rtx (), barrier);
7641
7642 rec = create_basic_block (label, label, before_recovery);
7643
7644 /* A recovery block always ends with an unconditional jump. */
7645 emit_barrier_after (BB_END (rec));
7646
7647 if (BB_PARTITION (before_recovery) != BB_UNPARTITIONED)
7648 BB_SET_PARTITION (rec, BB_COLD_PARTITION);
7649
7650 if (sched_verbose && spec_info->dump)
7651 fprintf (spec_info->dump, ";;\t\tGenerated recovery block rec%d\n",
7652 rec->index);
7653
7654 return rec;
7655 }
7656
7657 /* Create edges: FIRST_BB -> REC; FIRST_BB -> SECOND_BB; REC -> SECOND_BB
7658 and emit necessary jumps. */
7659 void
7660 sched_create_recovery_edges (basic_block first_bb, basic_block rec,
7661 basic_block second_bb)
7662 {
7663 rtx label;
7664 rtx jump;
7665 int edge_flags;
7666
7667 /* This is fixing of incoming edge. */
7668 /* ??? Which other flags should be specified? */
7669 if (BB_PARTITION (first_bb) != BB_PARTITION (rec))
7670 /* Partition type is the same, if it is "unpartitioned". */
7671 edge_flags = EDGE_CROSSING;
7672 else
7673 edge_flags = 0;
7674
7675 make_edge (first_bb, rec, edge_flags);
7676 label = block_label (second_bb);
7677 jump = emit_jump_insn_after (gen_jump (label), BB_END (rec));
7678 JUMP_LABEL (jump) = label;
7679 LABEL_NUSES (label)++;
7680
7681 if (BB_PARTITION (second_bb) != BB_PARTITION (rec))
7682 /* Partition type is the same, if it is "unpartitioned". */
7683 {
7684 /* Rewritten from cfgrtl.c. */
7685 if (flag_reorder_blocks_and_partition
7686 && targetm_common.have_named_sections)
7687 {
7688 /* We don't need the same note for the check because
7689 any_condjump_p (check) == true. */
7690 add_reg_note (jump, REG_CROSSING_JUMP, NULL_RTX);
7691 }
7692 edge_flags = EDGE_CROSSING;
7693 }
7694 else
7695 edge_flags = 0;
7696
7697 make_single_succ_edge (rec, second_bb, edge_flags);
7698 if (dom_info_available_p (CDI_DOMINATORS))
7699 set_immediate_dominator (CDI_DOMINATORS, rec, first_bb);
7700 }
7701
7702 /* This function creates recovery code for INSN. If MUTATE_P is nonzero,
7703 INSN is a simple check, that should be converted to branchy one. */
7704 static void
7705 create_check_block_twin (rtx insn, bool mutate_p)
7706 {
7707 basic_block rec;
7708 rtx label, check, twin;
7709 ds_t fs;
7710 sd_iterator_def sd_it;
7711 dep_t dep;
7712 dep_def _new_dep, *new_dep = &_new_dep;
7713 ds_t todo_spec;
7714
7715 gcc_assert (ORIG_PAT (insn) != NULL_RTX);
7716
7717 if (!mutate_p)
7718 todo_spec = TODO_SPEC (insn);
7719 else
7720 {
7721 gcc_assert (IS_SPECULATION_SIMPLE_CHECK_P (insn)
7722 && (TODO_SPEC (insn) & SPECULATIVE) == 0);
7723
7724 todo_spec = CHECK_SPEC (insn);
7725 }
7726
7727 todo_spec &= SPECULATIVE;
7728
7729 /* Create recovery block. */
7730 if (mutate_p || targetm.sched.needs_block_p (todo_spec))
7731 {
7732 rec = sched_create_recovery_block (NULL);
7733 label = BB_HEAD (rec);
7734 }
7735 else
7736 {
7737 rec = EXIT_BLOCK_PTR_FOR_FN (cfun);
7738 label = NULL_RTX;
7739 }
7740
7741 /* Emit CHECK. */
7742 check = targetm.sched.gen_spec_check (insn, label, todo_spec);
7743
7744 if (rec != EXIT_BLOCK_PTR_FOR_FN (cfun))
7745 {
7746 /* To have mem_reg alive at the beginning of second_bb,
7747 we emit check BEFORE insn, so insn after splitting
7748 insn will be at the beginning of second_bb, which will
7749 provide us with the correct life information. */
7750 check = emit_jump_insn_before (check, insn);
7751 JUMP_LABEL (check) = label;
7752 LABEL_NUSES (label)++;
7753 }
7754 else
7755 check = emit_insn_before (check, insn);
7756
7757 /* Extend data structures. */
7758 haifa_init_insn (check);
7759
7760 /* CHECK is being added to current region. Extend ready list. */
7761 gcc_assert (sched_ready_n_insns != -1);
7762 sched_extend_ready_list (sched_ready_n_insns + 1);
7763
7764 if (current_sched_info->add_remove_insn)
7765 current_sched_info->add_remove_insn (insn, 0);
7766
7767 RECOVERY_BLOCK (check) = rec;
7768
7769 if (sched_verbose && spec_info->dump)
7770 fprintf (spec_info->dump, ";;\t\tGenerated check insn : %s\n",
7771 (*current_sched_info->print_insn) (check, 0));
7772
7773 gcc_assert (ORIG_PAT (insn));
7774
7775 /* Initialize TWIN (twin is a duplicate of original instruction
7776 in the recovery block). */
7777 if (rec != EXIT_BLOCK_PTR_FOR_FN (cfun))
7778 {
7779 sd_iterator_def sd_it;
7780 dep_t dep;
7781
7782 FOR_EACH_DEP (insn, SD_LIST_RES_BACK, sd_it, dep)
7783 if ((DEP_STATUS (dep) & DEP_OUTPUT) != 0)
7784 {
7785 struct _dep _dep2, *dep2 = &_dep2;
7786
7787 init_dep (dep2, DEP_PRO (dep), check, REG_DEP_TRUE);
7788
7789 sd_add_dep (dep2, true);
7790 }
7791
7792 twin = emit_insn_after (ORIG_PAT (insn), BB_END (rec));
7793 haifa_init_insn (twin);
7794
7795 if (sched_verbose && spec_info->dump)
7796 /* INSN_BB (insn) isn't determined for twin insns yet.
7797 So we can't use current_sched_info->print_insn. */
7798 fprintf (spec_info->dump, ";;\t\tGenerated twin insn : %d/rec%d\n",
7799 INSN_UID (twin), rec->index);
7800 }
7801 else
7802 {
7803 ORIG_PAT (check) = ORIG_PAT (insn);
7804 HAS_INTERNAL_DEP (check) = 1;
7805 twin = check;
7806 /* ??? We probably should change all OUTPUT dependencies to
7807 (TRUE | OUTPUT). */
7808 }
7809
7810 /* Copy all resolved back dependencies of INSN to TWIN. This will
7811 provide correct value for INSN_TICK (TWIN). */
7812 sd_copy_back_deps (twin, insn, true);
7813
7814 if (rec != EXIT_BLOCK_PTR_FOR_FN (cfun))
7815 /* In case of branchy check, fix CFG. */
7816 {
7817 basic_block first_bb, second_bb;
7818 rtx jump;
7819
7820 first_bb = BLOCK_FOR_INSN (check);
7821 second_bb = sched_split_block (first_bb, check);
7822
7823 sched_create_recovery_edges (first_bb, rec, second_bb);
7824
7825 sched_init_only_bb (second_bb, first_bb);
7826 sched_init_only_bb (rec, EXIT_BLOCK_PTR_FOR_FN (cfun));
7827
7828 jump = BB_END (rec);
7829 haifa_init_insn (jump);
7830 }
7831
7832 /* Move backward dependences from INSN to CHECK and
7833 move forward dependences from INSN to TWIN. */
7834
7835 /* First, create dependencies between INSN's producers and CHECK & TWIN. */
7836 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
7837 {
7838 rtx pro = DEP_PRO (dep);
7839 ds_t ds;
7840
7841 /* If BEGIN_DATA: [insn ~~TRUE~~> producer]:
7842 check --TRUE--> producer ??? or ANTI ???
7843 twin --TRUE--> producer
7844 twin --ANTI--> check
7845
7846 If BEGIN_CONTROL: [insn ~~ANTI~~> producer]:
7847 check --ANTI--> producer
7848 twin --ANTI--> producer
7849 twin --ANTI--> check
7850
7851 If BE_IN_SPEC: [insn ~~TRUE~~> producer]:
7852 check ~~TRUE~~> producer
7853 twin ~~TRUE~~> producer
7854 twin --ANTI--> check */
7855
7856 ds = DEP_STATUS (dep);
7857
7858 if (ds & BEGIN_SPEC)
7859 {
7860 gcc_assert (!mutate_p);
7861 ds &= ~BEGIN_SPEC;
7862 }
7863
7864 init_dep_1 (new_dep, pro, check, DEP_TYPE (dep), ds);
7865 sd_add_dep (new_dep, false);
7866
7867 if (rec != EXIT_BLOCK_PTR_FOR_FN (cfun))
7868 {
7869 DEP_CON (new_dep) = twin;
7870 sd_add_dep (new_dep, false);
7871 }
7872 }
7873
7874 /* Second, remove backward dependencies of INSN. */
7875 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
7876 sd_iterator_cond (&sd_it, &dep);)
7877 {
7878 if ((DEP_STATUS (dep) & BEGIN_SPEC)
7879 || mutate_p)
7880 /* We can delete this dep because we overcome it with
7881 BEGIN_SPECULATION. */
7882 sd_delete_dep (sd_it);
7883 else
7884 sd_iterator_next (&sd_it);
7885 }
7886
7887 /* Future Speculations. Determine what BE_IN speculations will be like. */
7888 fs = 0;
7889
7890 /* Fields (DONE_SPEC (x) & BEGIN_SPEC) and CHECK_SPEC (x) are set only
7891 here. */
7892
7893 gcc_assert (!DONE_SPEC (insn));
7894
7895 if (!mutate_p)
7896 {
7897 ds_t ts = TODO_SPEC (insn);
7898
7899 DONE_SPEC (insn) = ts & BEGIN_SPEC;
7900 CHECK_SPEC (check) = ts & BEGIN_SPEC;
7901
7902 /* Luckiness of future speculations solely depends upon initial
7903 BEGIN speculation. */
7904 if (ts & BEGIN_DATA)
7905 fs = set_dep_weak (fs, BE_IN_DATA, get_dep_weak (ts, BEGIN_DATA));
7906 if (ts & BEGIN_CONTROL)
7907 fs = set_dep_weak (fs, BE_IN_CONTROL,
7908 get_dep_weak (ts, BEGIN_CONTROL));
7909 }
7910 else
7911 CHECK_SPEC (check) = CHECK_SPEC (insn);
7912
7913 /* Future speculations: call the helper. */
7914 process_insn_forw_deps_be_in_spec (insn, twin, fs);
7915
7916 if (rec != EXIT_BLOCK_PTR_FOR_FN (cfun))
7917 {
7918 /* Which types of dependencies should we use here is,
7919 generally, machine-dependent question... But, for now,
7920 it is not. */
7921
7922 if (!mutate_p)
7923 {
7924 init_dep (new_dep, insn, check, REG_DEP_TRUE);
7925 sd_add_dep (new_dep, false);
7926
7927 init_dep (new_dep, insn, twin, REG_DEP_OUTPUT);
7928 sd_add_dep (new_dep, false);
7929 }
7930 else
7931 {
7932 if (spec_info->dump)
7933 fprintf (spec_info->dump, ";;\t\tRemoved simple check : %s\n",
7934 (*current_sched_info->print_insn) (insn, 0));
7935
7936 /* Remove all dependencies of the INSN. */
7937 {
7938 sd_it = sd_iterator_start (insn, (SD_LIST_FORW
7939 | SD_LIST_BACK
7940 | SD_LIST_RES_BACK));
7941 while (sd_iterator_cond (&sd_it, &dep))
7942 sd_delete_dep (sd_it);
7943 }
7944
7945 /* If former check (INSN) already was moved to the ready (or queue)
7946 list, add new check (CHECK) there too. */
7947 if (QUEUE_INDEX (insn) != QUEUE_NOWHERE)
7948 try_ready (check);
7949
7950 /* Remove old check from instruction stream and free its
7951 data. */
7952 sched_remove_insn (insn);
7953 }
7954
7955 init_dep (new_dep, check, twin, REG_DEP_ANTI);
7956 sd_add_dep (new_dep, false);
7957 }
7958 else
7959 {
7960 init_dep_1 (new_dep, insn, check, REG_DEP_TRUE, DEP_TRUE | DEP_OUTPUT);
7961 sd_add_dep (new_dep, false);
7962 }
7963
7964 if (!mutate_p)
7965 /* Fix priorities. If MUTATE_P is nonzero, this is not necessary,
7966 because it'll be done later in add_to_speculative_block. */
7967 {
7968 rtx_vec_t priorities_roots = rtx_vec_t ();
7969
7970 clear_priorities (twin, &priorities_roots);
7971 calc_priorities (priorities_roots);
7972 priorities_roots.release ();
7973 }
7974 }
7975
7976 /* Removes dependency between instructions in the recovery block REC
7977 and usual region instructions. It keeps inner dependences so it
7978 won't be necessary to recompute them. */
7979 static void
7980 fix_recovery_deps (basic_block rec)
7981 {
7982 rtx note, insn, jump, ready_list = 0;
7983 bitmap_head in_ready;
7984 rtx link;
7985
7986 bitmap_initialize (&in_ready, 0);
7987
7988 /* NOTE - a basic block note. */
7989 note = NEXT_INSN (BB_HEAD (rec));
7990 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
7991 insn = BB_END (rec);
7992 gcc_assert (JUMP_P (insn));
7993 insn = PREV_INSN (insn);
7994
7995 do
7996 {
7997 sd_iterator_def sd_it;
7998 dep_t dep;
7999
8000 for (sd_it = sd_iterator_start (insn, SD_LIST_FORW);
8001 sd_iterator_cond (&sd_it, &dep);)
8002 {
8003 rtx consumer = DEP_CON (dep);
8004
8005 if (BLOCK_FOR_INSN (consumer) != rec)
8006 {
8007 sd_delete_dep (sd_it);
8008
8009 if (bitmap_set_bit (&in_ready, INSN_LUID (consumer)))
8010 ready_list = alloc_INSN_LIST (consumer, ready_list);
8011 }
8012 else
8013 {
8014 gcc_assert ((DEP_STATUS (dep) & DEP_TYPES) == DEP_TRUE);
8015
8016 sd_iterator_next (&sd_it);
8017 }
8018 }
8019
8020 insn = PREV_INSN (insn);
8021 }
8022 while (insn != note);
8023
8024 bitmap_clear (&in_ready);
8025
8026 /* Try to add instructions to the ready or queue list. */
8027 for (link = ready_list; link; link = XEXP (link, 1))
8028 try_ready (XEXP (link, 0));
8029 free_INSN_LIST_list (&ready_list);
8030
8031 /* Fixing jump's dependences. */
8032 insn = BB_HEAD (rec);
8033 jump = BB_END (rec);
8034
8035 gcc_assert (LABEL_P (insn));
8036 insn = NEXT_INSN (insn);
8037
8038 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (insn));
8039 add_jump_dependencies (insn, jump);
8040 }
8041
8042 /* Change pattern of INSN to NEW_PAT. Invalidate cached haifa
8043 instruction data. */
8044 static bool
8045 haifa_change_pattern (rtx insn, rtx new_pat)
8046 {
8047 int t;
8048
8049 t = validate_change (insn, &PATTERN (insn), new_pat, 0);
8050 if (!t)
8051 return false;
8052
8053 update_insn_after_change (insn);
8054 return true;
8055 }
8056
8057 /* -1 - can't speculate,
8058 0 - for speculation with REQUEST mode it is OK to use
8059 current instruction pattern,
8060 1 - need to change pattern for *NEW_PAT to be speculative. */
8061 int
8062 sched_speculate_insn (rtx insn, ds_t request, rtx *new_pat)
8063 {
8064 gcc_assert (current_sched_info->flags & DO_SPECULATION
8065 && (request & SPECULATIVE)
8066 && sched_insn_is_legitimate_for_speculation_p (insn, request));
8067
8068 if ((request & spec_info->mask) != request)
8069 return -1;
8070
8071 if (request & BE_IN_SPEC
8072 && !(request & BEGIN_SPEC))
8073 return 0;
8074
8075 return targetm.sched.speculate_insn (insn, request, new_pat);
8076 }
8077
8078 static int
8079 haifa_speculate_insn (rtx insn, ds_t request, rtx *new_pat)
8080 {
8081 gcc_assert (sched_deps_info->generate_spec_deps
8082 && !IS_SPECULATION_CHECK_P (insn));
8083
8084 if (HAS_INTERNAL_DEP (insn)
8085 || SCHED_GROUP_P (insn))
8086 return -1;
8087
8088 return sched_speculate_insn (insn, request, new_pat);
8089 }
8090
8091 /* Print some information about block BB, which starts with HEAD and
8092 ends with TAIL, before scheduling it.
8093 I is zero, if scheduler is about to start with the fresh ebb. */
8094 static void
8095 dump_new_block_header (int i, basic_block bb, rtx head, rtx tail)
8096 {
8097 if (!i)
8098 fprintf (sched_dump,
8099 ";; ======================================================\n");
8100 else
8101 fprintf (sched_dump,
8102 ";; =====================ADVANCING TO=====================\n");
8103 fprintf (sched_dump,
8104 ";; -- basic block %d from %d to %d -- %s reload\n",
8105 bb->index, INSN_UID (head), INSN_UID (tail),
8106 (reload_completed ? "after" : "before"));
8107 fprintf (sched_dump,
8108 ";; ======================================================\n");
8109 fprintf (sched_dump, "\n");
8110 }
8111
8112 /* Unlink basic block notes and labels and saves them, so they
8113 can be easily restored. We unlink basic block notes in EBB to
8114 provide back-compatibility with the previous code, as target backends
8115 assume, that there'll be only instructions between
8116 current_sched_info->{head and tail}. We restore these notes as soon
8117 as we can.
8118 FIRST (LAST) is the first (last) basic block in the ebb.
8119 NB: In usual case (FIRST == LAST) nothing is really done. */
8120 void
8121 unlink_bb_notes (basic_block first, basic_block last)
8122 {
8123 /* We DON'T unlink basic block notes of the first block in the ebb. */
8124 if (first == last)
8125 return;
8126
8127 bb_header = XNEWVEC (rtx, last_basic_block);
8128
8129 /* Make a sentinel. */
8130 if (last->next_bb != EXIT_BLOCK_PTR_FOR_FN (cfun))
8131 bb_header[last->next_bb->index] = 0;
8132
8133 first = first->next_bb;
8134 do
8135 {
8136 rtx prev, label, note, next;
8137
8138 label = BB_HEAD (last);
8139 if (LABEL_P (label))
8140 note = NEXT_INSN (label);
8141 else
8142 note = label;
8143 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
8144
8145 prev = PREV_INSN (label);
8146 next = NEXT_INSN (note);
8147 gcc_assert (prev && next);
8148
8149 NEXT_INSN (prev) = next;
8150 PREV_INSN (next) = prev;
8151
8152 bb_header[last->index] = label;
8153
8154 if (last == first)
8155 break;
8156
8157 last = last->prev_bb;
8158 }
8159 while (1);
8160 }
8161
8162 /* Restore basic block notes.
8163 FIRST is the first basic block in the ebb. */
8164 static void
8165 restore_bb_notes (basic_block first)
8166 {
8167 if (!bb_header)
8168 return;
8169
8170 /* We DON'T unlink basic block notes of the first block in the ebb. */
8171 first = first->next_bb;
8172 /* Remember: FIRST is actually a second basic block in the ebb. */
8173
8174 while (first != EXIT_BLOCK_PTR_FOR_FN (cfun)
8175 && bb_header[first->index])
8176 {
8177 rtx prev, label, note, next;
8178
8179 label = bb_header[first->index];
8180 prev = PREV_INSN (label);
8181 next = NEXT_INSN (prev);
8182
8183 if (LABEL_P (label))
8184 note = NEXT_INSN (label);
8185 else
8186 note = label;
8187 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
8188
8189 bb_header[first->index] = 0;
8190
8191 NEXT_INSN (prev) = label;
8192 NEXT_INSN (note) = next;
8193 PREV_INSN (next) = note;
8194
8195 first = first->next_bb;
8196 }
8197
8198 free (bb_header);
8199 bb_header = 0;
8200 }
8201
8202 /* Helper function.
8203 Fix CFG after both in- and inter-block movement of
8204 control_flow_insn_p JUMP. */
8205 static void
8206 fix_jump_move (rtx jump)
8207 {
8208 basic_block bb, jump_bb, jump_bb_next;
8209
8210 bb = BLOCK_FOR_INSN (PREV_INSN (jump));
8211 jump_bb = BLOCK_FOR_INSN (jump);
8212 jump_bb_next = jump_bb->next_bb;
8213
8214 gcc_assert (common_sched_info->sched_pass_id == SCHED_EBB_PASS
8215 || IS_SPECULATION_BRANCHY_CHECK_P (jump));
8216
8217 if (!NOTE_INSN_BASIC_BLOCK_P (BB_END (jump_bb_next)))
8218 /* if jump_bb_next is not empty. */
8219 BB_END (jump_bb) = BB_END (jump_bb_next);
8220
8221 if (BB_END (bb) != PREV_INSN (jump))
8222 /* Then there are instruction after jump that should be placed
8223 to jump_bb_next. */
8224 BB_END (jump_bb_next) = BB_END (bb);
8225 else
8226 /* Otherwise jump_bb_next is empty. */
8227 BB_END (jump_bb_next) = NEXT_INSN (BB_HEAD (jump_bb_next));
8228
8229 /* To make assertion in move_insn happy. */
8230 BB_END (bb) = PREV_INSN (jump);
8231
8232 update_bb_for_insn (jump_bb_next);
8233 }
8234
8235 /* Fix CFG after interblock movement of control_flow_insn_p JUMP. */
8236 static void
8237 move_block_after_check (rtx jump)
8238 {
8239 basic_block bb, jump_bb, jump_bb_next;
8240 vec<edge, va_gc> *t;
8241
8242 bb = BLOCK_FOR_INSN (PREV_INSN (jump));
8243 jump_bb = BLOCK_FOR_INSN (jump);
8244 jump_bb_next = jump_bb->next_bb;
8245
8246 update_bb_for_insn (jump_bb);
8247
8248 gcc_assert (IS_SPECULATION_CHECK_P (jump)
8249 || IS_SPECULATION_CHECK_P (BB_END (jump_bb_next)));
8250
8251 unlink_block (jump_bb_next);
8252 link_block (jump_bb_next, bb);
8253
8254 t = bb->succs;
8255 bb->succs = 0;
8256 move_succs (&(jump_bb->succs), bb);
8257 move_succs (&(jump_bb_next->succs), jump_bb);
8258 move_succs (&t, jump_bb_next);
8259
8260 df_mark_solutions_dirty ();
8261
8262 common_sched_info->fix_recovery_cfg
8263 (bb->index, jump_bb->index, jump_bb_next->index);
8264 }
8265
8266 /* Helper function for move_block_after_check.
8267 This functions attaches edge vector pointed to by SUCCSP to
8268 block TO. */
8269 static void
8270 move_succs (vec<edge, va_gc> **succsp, basic_block to)
8271 {
8272 edge e;
8273 edge_iterator ei;
8274
8275 gcc_assert (to->succs == 0);
8276
8277 to->succs = *succsp;
8278
8279 FOR_EACH_EDGE (e, ei, to->succs)
8280 e->src = to;
8281
8282 *succsp = 0;
8283 }
8284
8285 /* Remove INSN from the instruction stream.
8286 INSN should have any dependencies. */
8287 static void
8288 sched_remove_insn (rtx insn)
8289 {
8290 sd_finish_insn (insn);
8291
8292 change_queue_index (insn, QUEUE_NOWHERE);
8293 current_sched_info->add_remove_insn (insn, 1);
8294 delete_insn (insn);
8295 }
8296
8297 /* Clear priorities of all instructions, that are forward dependent on INSN.
8298 Store in vector pointed to by ROOTS_PTR insns on which priority () should
8299 be invoked to initialize all cleared priorities. */
8300 static void
8301 clear_priorities (rtx insn, rtx_vec_t *roots_ptr)
8302 {
8303 sd_iterator_def sd_it;
8304 dep_t dep;
8305 bool insn_is_root_p = true;
8306
8307 gcc_assert (QUEUE_INDEX (insn) != QUEUE_SCHEDULED);
8308
8309 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
8310 {
8311 rtx pro = DEP_PRO (dep);
8312
8313 if (INSN_PRIORITY_STATUS (pro) >= 0
8314 && QUEUE_INDEX (insn) != QUEUE_SCHEDULED)
8315 {
8316 /* If DEP doesn't contribute to priority then INSN itself should
8317 be added to priority roots. */
8318 if (contributes_to_priority_p (dep))
8319 insn_is_root_p = false;
8320
8321 INSN_PRIORITY_STATUS (pro) = -1;
8322 clear_priorities (pro, roots_ptr);
8323 }
8324 }
8325
8326 if (insn_is_root_p)
8327 roots_ptr->safe_push (insn);
8328 }
8329
8330 /* Recompute priorities of instructions, whose priorities might have been
8331 changed. ROOTS is a vector of instructions whose priority computation will
8332 trigger initialization of all cleared priorities. */
8333 static void
8334 calc_priorities (rtx_vec_t roots)
8335 {
8336 int i;
8337 rtx insn;
8338
8339 FOR_EACH_VEC_ELT (roots, i, insn)
8340 priority (insn);
8341 }
8342
8343
8344 /* Add dependences between JUMP and other instructions in the recovery
8345 block. INSN is the first insn the recovery block. */
8346 static void
8347 add_jump_dependencies (rtx insn, rtx jump)
8348 {
8349 do
8350 {
8351 insn = NEXT_INSN (insn);
8352 if (insn == jump)
8353 break;
8354
8355 if (dep_list_size (insn, SD_LIST_FORW) == 0)
8356 {
8357 dep_def _new_dep, *new_dep = &_new_dep;
8358
8359 init_dep (new_dep, insn, jump, REG_DEP_ANTI);
8360 sd_add_dep (new_dep, false);
8361 }
8362 }
8363 while (1);
8364
8365 gcc_assert (!sd_lists_empty_p (jump, SD_LIST_BACK));
8366 }
8367
8368 /* Extend data structures for logical insn UID. */
8369 void
8370 sched_extend_luids (void)
8371 {
8372 int new_luids_max_uid = get_max_uid () + 1;
8373
8374 sched_luids.safe_grow_cleared (new_luids_max_uid);
8375 }
8376
8377 /* Initialize LUID for INSN. */
8378 void
8379 sched_init_insn_luid (rtx insn)
8380 {
8381 int i = INSN_P (insn) ? 1 : common_sched_info->luid_for_non_insn (insn);
8382 int luid;
8383
8384 if (i >= 0)
8385 {
8386 luid = sched_max_luid;
8387 sched_max_luid += i;
8388 }
8389 else
8390 luid = -1;
8391
8392 SET_INSN_LUID (insn, luid);
8393 }
8394
8395 /* Initialize luids for BBS.
8396 The hook common_sched_info->luid_for_non_insn () is used to determine
8397 if notes, labels, etc. need luids. */
8398 void
8399 sched_init_luids (bb_vec_t bbs)
8400 {
8401 int i;
8402 basic_block bb;
8403
8404 sched_extend_luids ();
8405 FOR_EACH_VEC_ELT (bbs, i, bb)
8406 {
8407 rtx insn;
8408
8409 FOR_BB_INSNS (bb, insn)
8410 sched_init_insn_luid (insn);
8411 }
8412 }
8413
8414 /* Free LUIDs. */
8415 void
8416 sched_finish_luids (void)
8417 {
8418 sched_luids.release ();
8419 sched_max_luid = 1;
8420 }
8421
8422 /* Return logical uid of INSN. Helpful while debugging. */
8423 int
8424 insn_luid (rtx insn)
8425 {
8426 return INSN_LUID (insn);
8427 }
8428
8429 /* Extend per insn data in the target. */
8430 void
8431 sched_extend_target (void)
8432 {
8433 if (targetm.sched.h_i_d_extended)
8434 targetm.sched.h_i_d_extended ();
8435 }
8436
8437 /* Extend global scheduler structures (those, that live across calls to
8438 schedule_block) to include information about just emitted INSN. */
8439 static void
8440 extend_h_i_d (void)
8441 {
8442 int reserve = (get_max_uid () + 1 - h_i_d.length ());
8443 if (reserve > 0
8444 && ! h_i_d.space (reserve))
8445 {
8446 h_i_d.safe_grow_cleared (3 * get_max_uid () / 2);
8447 sched_extend_target ();
8448 }
8449 }
8450
8451 /* Initialize h_i_d entry of the INSN with default values.
8452 Values, that are not explicitly initialized here, hold zero. */
8453 static void
8454 init_h_i_d (rtx insn)
8455 {
8456 if (INSN_LUID (insn) > 0)
8457 {
8458 INSN_COST (insn) = -1;
8459 QUEUE_INDEX (insn) = QUEUE_NOWHERE;
8460 INSN_TICK (insn) = INVALID_TICK;
8461 INSN_EXACT_TICK (insn) = INVALID_TICK;
8462 INTER_TICK (insn) = INVALID_TICK;
8463 TODO_SPEC (insn) = HARD_DEP;
8464 }
8465 }
8466
8467 /* Initialize haifa_insn_data for BBS. */
8468 void
8469 haifa_init_h_i_d (bb_vec_t bbs)
8470 {
8471 int i;
8472 basic_block bb;
8473
8474 extend_h_i_d ();
8475 FOR_EACH_VEC_ELT (bbs, i, bb)
8476 {
8477 rtx insn;
8478
8479 FOR_BB_INSNS (bb, insn)
8480 init_h_i_d (insn);
8481 }
8482 }
8483
8484 /* Finalize haifa_insn_data. */
8485 void
8486 haifa_finish_h_i_d (void)
8487 {
8488 int i;
8489 haifa_insn_data_t data;
8490 struct reg_use_data *use, *next;
8491
8492 FOR_EACH_VEC_ELT (h_i_d, i, data)
8493 {
8494 free (data->max_reg_pressure);
8495 free (data->reg_pressure);
8496 for (use = data->reg_use_list; use != NULL; use = next)
8497 {
8498 next = use->next_insn_use;
8499 free (use);
8500 }
8501 }
8502 h_i_d.release ();
8503 }
8504
8505 /* Init data for the new insn INSN. */
8506 static void
8507 haifa_init_insn (rtx insn)
8508 {
8509 gcc_assert (insn != NULL);
8510
8511 sched_extend_luids ();
8512 sched_init_insn_luid (insn);
8513 sched_extend_target ();
8514 sched_deps_init (false);
8515 extend_h_i_d ();
8516 init_h_i_d (insn);
8517
8518 if (adding_bb_to_current_region_p)
8519 {
8520 sd_init_insn (insn);
8521
8522 /* Extend dependency caches by one element. */
8523 extend_dependency_caches (1, false);
8524 }
8525 if (sched_pressure != SCHED_PRESSURE_NONE)
8526 init_insn_reg_pressure_info (insn);
8527 }
8528
8529 /* Init data for the new basic block BB which comes after AFTER. */
8530 static void
8531 haifa_init_only_bb (basic_block bb, basic_block after)
8532 {
8533 gcc_assert (bb != NULL);
8534
8535 sched_init_bbs ();
8536
8537 if (common_sched_info->add_block)
8538 /* This changes only data structures of the front-end. */
8539 common_sched_info->add_block (bb, after);
8540 }
8541
8542 /* A generic version of sched_split_block (). */
8543 basic_block
8544 sched_split_block_1 (basic_block first_bb, rtx after)
8545 {
8546 edge e;
8547
8548 e = split_block (first_bb, after);
8549 gcc_assert (e->src == first_bb);
8550
8551 /* sched_split_block emits note if *check == BB_END. Probably it
8552 is better to rip that note off. */
8553
8554 return e->dest;
8555 }
8556
8557 /* A generic version of sched_create_empty_bb (). */
8558 basic_block
8559 sched_create_empty_bb_1 (basic_block after)
8560 {
8561 return create_empty_bb (after);
8562 }
8563
8564 /* Insert PAT as an INSN into the schedule and update the necessary data
8565 structures to account for it. */
8566 rtx
8567 sched_emit_insn (rtx pat)
8568 {
8569 rtx insn = emit_insn_before (pat, nonscheduled_insns_begin);
8570 haifa_init_insn (insn);
8571
8572 if (current_sched_info->add_remove_insn)
8573 current_sched_info->add_remove_insn (insn, 0);
8574
8575 (*current_sched_info->begin_schedule_ready) (insn);
8576 scheduled_insns.safe_push (insn);
8577
8578 last_scheduled_insn = insn;
8579 return insn;
8580 }
8581
8582 /* This function returns a candidate satisfying dispatch constraints from
8583 the ready list. */
8584
8585 static rtx
8586 ready_remove_first_dispatch (struct ready_list *ready)
8587 {
8588 int i;
8589 rtx insn = ready_element (ready, 0);
8590
8591 if (ready->n_ready == 1
8592 || !INSN_P (insn)
8593 || INSN_CODE (insn) < 0
8594 || !active_insn_p (insn)
8595 || targetm.sched.dispatch (insn, FITS_DISPATCH_WINDOW))
8596 return ready_remove_first (ready);
8597
8598 for (i = 1; i < ready->n_ready; i++)
8599 {
8600 insn = ready_element (ready, i);
8601
8602 if (!INSN_P (insn)
8603 || INSN_CODE (insn) < 0
8604 || !active_insn_p (insn))
8605 continue;
8606
8607 if (targetm.sched.dispatch (insn, FITS_DISPATCH_WINDOW))
8608 {
8609 /* Return ith element of ready. */
8610 insn = ready_remove (ready, i);
8611 return insn;
8612 }
8613 }
8614
8615 if (targetm.sched.dispatch (NULL_RTX, DISPATCH_VIOLATION))
8616 return ready_remove_first (ready);
8617
8618 for (i = 1; i < ready->n_ready; i++)
8619 {
8620 insn = ready_element (ready, i);
8621
8622 if (!INSN_P (insn)
8623 || INSN_CODE (insn) < 0
8624 || !active_insn_p (insn))
8625 continue;
8626
8627 /* Return i-th element of ready. */
8628 if (targetm.sched.dispatch (insn, IS_CMP))
8629 return ready_remove (ready, i);
8630 }
8631
8632 return ready_remove_first (ready);
8633 }
8634
8635 /* Get number of ready insn in the ready list. */
8636
8637 int
8638 number_in_ready (void)
8639 {
8640 return ready.n_ready;
8641 }
8642
8643 /* Get number of ready's in the ready list. */
8644
8645 rtx
8646 get_ready_element (int i)
8647 {
8648 return ready_element (&ready, i);
8649 }
8650
8651 #endif /* INSN_SCHEDULING */