Avoid is_constant calls in vectorizable_bswap
[official-gcc.git] / gcc / haifa-sched.c
blob4f0221f6f43714fdce9119b76ea2f9fb79b46873
1 /* Instruction scheduling pass.
2 Copyright (C) 1992-2018 Free Software Foundation, Inc.
3 Contributed by Michael Tiemann (tiemann@cygnus.com) Enhanced by,
4 and currently maintained by, Jim Wilson (wilson@cygnus.com)
6 This file is part of GCC.
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.
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.
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/>. */
22 /* Instruction scheduling pass. This file, along with sched-deps.c,
23 contains the generic parts. The actual entry point for
24 the normal instruction scheduling pass is found in sched-rgn.c.
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.
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.
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.
55 The following list shows the order in which we want to break ties
56 among insns in the ready list:
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.
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.
78 Before reload, an extended analysis of interblock data dependences
79 is required for interblock scheduling. This is performed in
80 compute_block_dependences ().
82 Dependencies set up by memory references are treated in exactly the
83 same way as other dependencies, by using insn backward dependences
84 INSN_BACK_DEPS. INSN_BACK_DEPS are translated into forward dependences
85 INSN_FORW_DEPS for the purpose of forward list scheduling.
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.
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.
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.
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.
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).
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.
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. */
125 #include "config.h"
126 #include "system.h"
127 #include "coretypes.h"
128 #include "backend.h"
129 #include "target.h"
130 #include "rtl.h"
131 #include "cfghooks.h"
132 #include "df.h"
133 #include "memmodel.h"
134 #include "tm_p.h"
135 #include "insn-config.h"
136 #include "regs.h"
137 #include "ira.h"
138 #include "recog.h"
139 #include "insn-attr.h"
140 #include "cfgrtl.h"
141 #include "cfgbuild.h"
142 #include "sched-int.h"
143 #include "common/common-target.h"
144 #include "params.h"
145 #include "dbgcnt.h"
146 #include "cfgloop.h"
147 #include "dumpfile.h"
148 #include "print-rtl.h"
150 #ifdef INSN_SCHEDULING
152 /* True if we do register pressure relief through live-range
153 shrinkage. */
154 static bool live_range_shrinkage_p;
156 /* Switch on live range shrinkage. */
157 void
158 initialize_live_range_shrinkage (void)
160 live_range_shrinkage_p = true;
163 /* Switch off live range shrinkage. */
164 void
165 finish_live_range_shrinkage (void)
167 live_range_shrinkage_p = false;
170 /* issue_rate is the number of insns that can be scheduled in the same
171 machine cycle. It can be defined in the config/mach/mach.h file,
172 otherwise we set it to 1. */
174 int issue_rate;
176 /* This can be set to true by a backend if the scheduler should not
177 enable a DCE pass. */
178 bool sched_no_dce;
180 /* The current initiation interval used when modulo scheduling. */
181 static int modulo_ii;
183 /* The maximum number of stages we are prepared to handle. */
184 static int modulo_max_stages;
186 /* The number of insns that exist in each iteration of the loop. We use this
187 to detect when we've scheduled all insns from the first iteration. */
188 static int modulo_n_insns;
190 /* The current count of insns in the first iteration of the loop that have
191 already been scheduled. */
192 static int modulo_insns_scheduled;
194 /* The maximum uid of insns from the first iteration of the loop. */
195 static int modulo_iter0_max_uid;
197 /* The number of times we should attempt to backtrack when modulo scheduling.
198 Decreased each time we have to backtrack. */
199 static int modulo_backtracks_left;
201 /* The stage in which the last insn from the original loop was
202 scheduled. */
203 static int modulo_last_stage;
205 /* sched-verbose controls the amount of debugging output the
206 scheduler prints. It is controlled by -fsched-verbose=N:
207 N=0: no debugging output.
208 N=1: default value.
209 N=2: bb's probabilities, detailed ready list info, unit/insn info.
210 N=3: rtl at abort point, control-flow, regions info.
211 N=5: dependences info. */
212 int sched_verbose = 0;
214 /* Debugging file. All printouts are sent to dump. */
215 FILE *sched_dump = 0;
217 /* This is a placeholder for the scheduler parameters common
218 to all schedulers. */
219 struct common_sched_info_def *common_sched_info;
221 #define INSN_TICK(INSN) (HID (INSN)->tick)
222 #define INSN_EXACT_TICK(INSN) (HID (INSN)->exact_tick)
223 #define INSN_TICK_ESTIMATE(INSN) (HID (INSN)->tick_estimate)
224 #define INTER_TICK(INSN) (HID (INSN)->inter_tick)
225 #define FEEDS_BACKTRACK_INSN(INSN) (HID (INSN)->feeds_backtrack_insn)
226 #define SHADOW_P(INSN) (HID (INSN)->shadow_p)
227 #define MUST_RECOMPUTE_SPEC_P(INSN) (HID (INSN)->must_recompute_spec)
228 /* Cached cost of the instruction. Use insn_sched_cost to get cost of the
229 insn. -1 here means that the field is not initialized. */
230 #define INSN_COST(INSN) (HID (INSN)->cost)
232 /* If INSN_TICK of an instruction is equal to INVALID_TICK,
233 then it should be recalculated from scratch. */
234 #define INVALID_TICK (-(max_insn_queue_index + 1))
235 /* The minimal value of the INSN_TICK of an instruction. */
236 #define MIN_TICK (-max_insn_queue_index)
238 /* Original order of insns in the ready list.
239 Used to keep order of normal insns while separating DEBUG_INSNs. */
240 #define INSN_RFS_DEBUG_ORIG_ORDER(INSN) (HID (INSN)->rfs_debug_orig_order)
242 /* The deciding reason for INSN's place in the ready list. */
243 #define INSN_LAST_RFS_WIN(INSN) (HID (INSN)->last_rfs_win)
245 /* List of important notes we must keep around. This is a pointer to the
246 last element in the list. */
247 rtx_insn *note_list;
249 static struct spec_info_def spec_info_var;
250 /* Description of the speculative part of the scheduling.
251 If NULL - no speculation. */
252 spec_info_t spec_info = NULL;
254 /* True, if recovery block was added during scheduling of current block.
255 Used to determine, if we need to fix INSN_TICKs. */
256 static bool haifa_recovery_bb_recently_added_p;
258 /* True, if recovery block was added during this scheduling pass.
259 Used to determine if we should have empty memory pools of dependencies
260 after finishing current region. */
261 bool haifa_recovery_bb_ever_added_p;
263 /* Counters of different types of speculative instructions. */
264 static int nr_begin_data, nr_be_in_data, nr_begin_control, nr_be_in_control;
266 /* Array used in {unlink, restore}_bb_notes. */
267 static rtx_insn **bb_header = 0;
269 /* Basic block after which recovery blocks will be created. */
270 static basic_block before_recovery;
272 /* Basic block just before the EXIT_BLOCK and after recovery, if we have
273 created it. */
274 basic_block after_recovery;
276 /* FALSE if we add bb to another region, so we don't need to initialize it. */
277 bool adding_bb_to_current_region_p = true;
279 /* Queues, etc. */
281 /* An instruction is ready to be scheduled when all insns preceding it
282 have already been scheduled. It is important to ensure that all
283 insns which use its result will not be executed until its result
284 has been computed. An insn is maintained in one of four structures:
286 (P) the "Pending" set of insns which cannot be scheduled until
287 their dependencies have been satisfied.
288 (Q) the "Queued" set of insns that can be scheduled when sufficient
289 time has passed.
290 (R) the "Ready" list of unscheduled, uncommitted insns.
291 (S) the "Scheduled" list of insns.
293 Initially, all insns are either "Pending" or "Ready" depending on
294 whether their dependencies are satisfied.
296 Insns move from the "Ready" list to the "Scheduled" list as they
297 are committed to the schedule. As this occurs, the insns in the
298 "Pending" list have their dependencies satisfied and move to either
299 the "Ready" list or the "Queued" set depending on whether
300 sufficient time has passed to make them ready. As time passes,
301 insns move from the "Queued" set to the "Ready" list.
303 The "Pending" list (P) are the insns in the INSN_FORW_DEPS of the
304 unscheduled insns, i.e., those that are ready, queued, and pending.
305 The "Queued" set (Q) is implemented by the variable `insn_queue'.
306 The "Ready" list (R) is implemented by the variables `ready' and
307 `n_ready'.
308 The "Scheduled" list (S) is the new insn chain built by this pass.
310 The transition (R->S) is implemented in the scheduling loop in
311 `schedule_block' when the best insn to schedule is chosen.
312 The transitions (P->R and P->Q) are implemented in `schedule_insn' as
313 insns move from the ready list to the scheduled list.
314 The transition (Q->R) is implemented in 'queue_to_insn' as time
315 passes or stalls are introduced. */
317 /* Implement a circular buffer to delay instructions until sufficient
318 time has passed. For the new pipeline description interface,
319 MAX_INSN_QUEUE_INDEX is a power of two minus one which is not less
320 than maximal time of instruction execution computed by genattr.c on
321 the base maximal time of functional unit reservations and getting a
322 result. This is the longest time an insn may be queued. */
324 static rtx_insn_list **insn_queue;
325 static int q_ptr = 0;
326 static int q_size = 0;
327 #define NEXT_Q(X) (((X)+1) & max_insn_queue_index)
328 #define NEXT_Q_AFTER(X, C) (((X)+C) & max_insn_queue_index)
330 #define QUEUE_SCHEDULED (-3)
331 #define QUEUE_NOWHERE (-2)
332 #define QUEUE_READY (-1)
333 /* QUEUE_SCHEDULED - INSN is scheduled.
334 QUEUE_NOWHERE - INSN isn't scheduled yet and is neither in
335 queue or ready list.
336 QUEUE_READY - INSN is in ready list.
337 N >= 0 - INSN queued for X [where NEXT_Q_AFTER (q_ptr, X) == N] cycles. */
339 #define QUEUE_INDEX(INSN) (HID (INSN)->queue_index)
341 /* The following variable value refers for all current and future
342 reservations of the processor units. */
343 state_t curr_state;
345 /* The following variable value is size of memory representing all
346 current and future reservations of the processor units. */
347 size_t dfa_state_size;
349 /* The following array is used to find the best insn from ready when
350 the automaton pipeline interface is used. */
351 signed char *ready_try = NULL;
353 /* The ready list. */
354 struct ready_list ready = {NULL, 0, 0, 0, 0};
356 /* The pointer to the ready list (to be removed). */
357 static struct ready_list *readyp = &ready;
359 /* Scheduling clock. */
360 static int clock_var;
362 /* Clock at which the previous instruction was issued. */
363 static int last_clock_var;
365 /* Set to true if, when queuing a shadow insn, we discover that it would be
366 scheduled too late. */
367 static bool must_backtrack;
369 /* The following variable value is number of essential insns issued on
370 the current cycle. An insn is essential one if it changes the
371 processors state. */
372 int cycle_issued_insns;
374 /* This records the actual schedule. It is built up during the main phase
375 of schedule_block, and afterwards used to reorder the insns in the RTL. */
376 static vec<rtx_insn *> scheduled_insns;
378 static int may_trap_exp (const_rtx, int);
380 /* Nonzero iff the address is comprised from at most 1 register. */
381 #define CONST_BASED_ADDRESS_P(x) \
382 (REG_P (x) \
383 || ((GET_CODE (x) == PLUS || GET_CODE (x) == MINUS \
384 || (GET_CODE (x) == LO_SUM)) \
385 && (CONSTANT_P (XEXP (x, 0)) \
386 || CONSTANT_P (XEXP (x, 1)))))
388 /* Returns a class that insn with GET_DEST(insn)=x may belong to,
389 as found by analyzing insn's expression. */
392 static int haifa_luid_for_non_insn (rtx x);
394 /* Haifa version of sched_info hooks common to all headers. */
395 const struct common_sched_info_def haifa_common_sched_info =
397 NULL, /* fix_recovery_cfg */
398 NULL, /* add_block */
399 NULL, /* estimate_number_of_insns */
400 haifa_luid_for_non_insn, /* luid_for_non_insn */
401 SCHED_PASS_UNKNOWN /* sched_pass_id */
404 /* Mapping from instruction UID to its Logical UID. */
405 vec<int> sched_luids;
407 /* Next LUID to assign to an instruction. */
408 int sched_max_luid = 1;
410 /* Haifa Instruction Data. */
411 vec<haifa_insn_data_def> h_i_d;
413 void (* sched_init_only_bb) (basic_block, basic_block);
415 /* Split block function. Different schedulers might use different functions
416 to handle their internal data consistent. */
417 basic_block (* sched_split_block) (basic_block, rtx);
419 /* Create empty basic block after the specified block. */
420 basic_block (* sched_create_empty_bb) (basic_block);
422 /* Return the number of cycles until INSN is expected to be ready.
423 Return zero if it already is. */
424 static int
425 insn_delay (rtx_insn *insn)
427 return MAX (INSN_TICK (insn) - clock_var, 0);
430 static int
431 may_trap_exp (const_rtx x, int is_store)
433 enum rtx_code code;
435 if (x == 0)
436 return TRAP_FREE;
437 code = GET_CODE (x);
438 if (is_store)
440 if (code == MEM && may_trap_p (x))
441 return TRAP_RISKY;
442 else
443 return TRAP_FREE;
445 if (code == MEM)
447 /* The insn uses memory: a volatile load. */
448 if (MEM_VOLATILE_P (x))
449 return IRISKY;
450 /* An exception-free load. */
451 if (!may_trap_p (x))
452 return IFREE;
453 /* A load with 1 base register, to be further checked. */
454 if (CONST_BASED_ADDRESS_P (XEXP (x, 0)))
455 return PFREE_CANDIDATE;
456 /* No info on the load, to be further checked. */
457 return PRISKY_CANDIDATE;
459 else
461 const char *fmt;
462 int i, insn_class = TRAP_FREE;
464 /* Neither store nor load, check if it may cause a trap. */
465 if (may_trap_p (x))
466 return TRAP_RISKY;
467 /* Recursive step: walk the insn... */
468 fmt = GET_RTX_FORMAT (code);
469 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
471 if (fmt[i] == 'e')
473 int tmp_class = may_trap_exp (XEXP (x, i), is_store);
474 insn_class = WORST_CLASS (insn_class, tmp_class);
476 else if (fmt[i] == 'E')
478 int j;
479 for (j = 0; j < XVECLEN (x, i); j++)
481 int tmp_class = may_trap_exp (XVECEXP (x, i, j), is_store);
482 insn_class = WORST_CLASS (insn_class, tmp_class);
483 if (insn_class == TRAP_RISKY || insn_class == IRISKY)
484 break;
487 if (insn_class == TRAP_RISKY || insn_class == IRISKY)
488 break;
490 return insn_class;
494 /* Classifies rtx X of an insn for the purpose of verifying that X can be
495 executed speculatively (and consequently the insn can be moved
496 speculatively), by examining X, returning:
497 TRAP_RISKY: store, or risky non-load insn (e.g. division by variable).
498 TRAP_FREE: non-load insn.
499 IFREE: load from a globally safe location.
500 IRISKY: volatile load.
501 PFREE_CANDIDATE, PRISKY_CANDIDATE: load that need to be checked for
502 being either PFREE or PRISKY. */
504 static int
505 haifa_classify_rtx (const_rtx x)
507 int tmp_class = TRAP_FREE;
508 int insn_class = TRAP_FREE;
509 enum rtx_code code;
511 if (GET_CODE (x) == PARALLEL)
513 int i, len = XVECLEN (x, 0);
515 for (i = len - 1; i >= 0; i--)
517 tmp_class = haifa_classify_rtx (XVECEXP (x, 0, i));
518 insn_class = WORST_CLASS (insn_class, tmp_class);
519 if (insn_class == TRAP_RISKY || insn_class == IRISKY)
520 break;
523 else
525 code = GET_CODE (x);
526 switch (code)
528 case CLOBBER:
529 /* Test if it is a 'store'. */
530 tmp_class = may_trap_exp (XEXP (x, 0), 1);
531 break;
532 case CLOBBER_HIGH:
533 gcc_assert (REG_P (XEXP (x, 0)));
534 break;
535 case SET:
536 /* Test if it is a store. */
537 tmp_class = may_trap_exp (SET_DEST (x), 1);
538 if (tmp_class == TRAP_RISKY)
539 break;
540 /* Test if it is a load. */
541 tmp_class =
542 WORST_CLASS (tmp_class,
543 may_trap_exp (SET_SRC (x), 0));
544 break;
545 case COND_EXEC:
546 tmp_class = haifa_classify_rtx (COND_EXEC_CODE (x));
547 if (tmp_class == TRAP_RISKY)
548 break;
549 tmp_class = WORST_CLASS (tmp_class,
550 may_trap_exp (COND_EXEC_TEST (x), 0));
551 break;
552 case TRAP_IF:
553 tmp_class = TRAP_RISKY;
554 break;
555 default:;
557 insn_class = tmp_class;
560 return insn_class;
564 haifa_classify_insn (const_rtx insn)
566 return haifa_classify_rtx (PATTERN (insn));
569 /* After the scheduler initialization function has been called, this function
570 can be called to enable modulo scheduling. II is the initiation interval
571 we should use, it affects the delays for delay_pairs that were recorded as
572 separated by a given number of stages.
574 MAX_STAGES provides us with a limit
575 after which we give up scheduling; the caller must have unrolled at least
576 as many copies of the loop body and recorded delay_pairs for them.
578 INSNS is the number of real (non-debug) insns in one iteration of
579 the loop. MAX_UID can be used to test whether an insn belongs to
580 the first iteration of the loop; all of them have a uid lower than
581 MAX_UID. */
582 void
583 set_modulo_params (int ii, int max_stages, int insns, int max_uid)
585 modulo_ii = ii;
586 modulo_max_stages = max_stages;
587 modulo_n_insns = insns;
588 modulo_iter0_max_uid = max_uid;
589 modulo_backtracks_left = PARAM_VALUE (PARAM_MAX_MODULO_BACKTRACK_ATTEMPTS);
592 /* A structure to record a pair of insns where the first one is a real
593 insn that has delay slots, and the second is its delayed shadow.
594 I1 is scheduled normally and will emit an assembly instruction,
595 while I2 describes the side effect that takes place at the
596 transition between cycles CYCLES and (CYCLES + 1) after I1. */
597 struct delay_pair
599 struct delay_pair *next_same_i1;
600 rtx_insn *i1, *i2;
601 int cycles;
602 /* When doing modulo scheduling, we a delay_pair can also be used to
603 show that I1 and I2 are the same insn in a different stage. If that
604 is the case, STAGES will be nonzero. */
605 int stages;
608 /* Helpers for delay hashing. */
610 struct delay_i1_hasher : nofree_ptr_hash <delay_pair>
612 typedef void *compare_type;
613 static inline hashval_t hash (const delay_pair *);
614 static inline bool equal (const delay_pair *, const void *);
617 /* Returns a hash value for X, based on hashing just I1. */
619 inline hashval_t
620 delay_i1_hasher::hash (const delay_pair *x)
622 return htab_hash_pointer (x->i1);
625 /* Return true if I1 of pair X is the same as that of pair Y. */
627 inline bool
628 delay_i1_hasher::equal (const delay_pair *x, const void *y)
630 return x->i1 == y;
633 struct delay_i2_hasher : free_ptr_hash <delay_pair>
635 typedef void *compare_type;
636 static inline hashval_t hash (const delay_pair *);
637 static inline bool equal (const delay_pair *, const void *);
640 /* Returns a hash value for X, based on hashing just I2. */
642 inline hashval_t
643 delay_i2_hasher::hash (const delay_pair *x)
645 return htab_hash_pointer (x->i2);
648 /* Return true if I2 of pair X is the same as that of pair Y. */
650 inline bool
651 delay_i2_hasher::equal (const delay_pair *x, const void *y)
653 return x->i2 == y;
656 /* Two hash tables to record delay_pairs, one indexed by I1 and the other
657 indexed by I2. */
658 static hash_table<delay_i1_hasher> *delay_htab;
659 static hash_table<delay_i2_hasher> *delay_htab_i2;
661 /* Called through htab_traverse. Walk the hashtable using I2 as
662 index, and delete all elements involving an UID higher than
663 that pointed to by *DATA. */
665 haifa_htab_i2_traverse (delay_pair **slot, int *data)
667 int maxuid = *data;
668 struct delay_pair *p = *slot;
669 if (INSN_UID (p->i2) >= maxuid || INSN_UID (p->i1) >= maxuid)
671 delay_htab_i2->clear_slot (slot);
673 return 1;
676 /* Called through htab_traverse. Walk the hashtable using I2 as
677 index, and delete all elements involving an UID higher than
678 that pointed to by *DATA. */
680 haifa_htab_i1_traverse (delay_pair **pslot, int *data)
682 int maxuid = *data;
683 struct delay_pair *p, *first, **pprev;
685 if (INSN_UID ((*pslot)->i1) >= maxuid)
687 delay_htab->clear_slot (pslot);
688 return 1;
690 pprev = &first;
691 for (p = *pslot; p; p = p->next_same_i1)
693 if (INSN_UID (p->i2) < maxuid)
695 *pprev = p;
696 pprev = &p->next_same_i1;
699 *pprev = NULL;
700 if (first == NULL)
701 delay_htab->clear_slot (pslot);
702 else
703 *pslot = first;
704 return 1;
707 /* Discard all delay pairs which involve an insn with an UID higher
708 than MAX_UID. */
709 void
710 discard_delay_pairs_above (int max_uid)
712 delay_htab->traverse <int *, haifa_htab_i1_traverse> (&max_uid);
713 delay_htab_i2->traverse <int *, haifa_htab_i2_traverse> (&max_uid);
716 /* This function can be called by a port just before it starts the final
717 scheduling pass. It records the fact that an instruction with delay
718 slots has been split into two insns, I1 and I2. The first one will be
719 scheduled normally and initiates the operation. The second one is a
720 shadow which must follow a specific number of cycles after I1; its only
721 purpose is to show the side effect that occurs at that cycle in the RTL.
722 If a JUMP_INSN or a CALL_INSN has been split, I1 should be a normal INSN,
723 while I2 retains the original insn type.
725 There are two ways in which the number of cycles can be specified,
726 involving the CYCLES and STAGES arguments to this function. If STAGES
727 is zero, we just use the value of CYCLES. Otherwise, STAGES is a factor
728 which is multiplied by MODULO_II to give the number of cycles. This is
729 only useful if the caller also calls set_modulo_params to enable modulo
730 scheduling. */
732 void
733 record_delay_slot_pair (rtx_insn *i1, rtx_insn *i2, int cycles, int stages)
735 struct delay_pair *p = XNEW (struct delay_pair);
736 struct delay_pair **slot;
738 p->i1 = i1;
739 p->i2 = i2;
740 p->cycles = cycles;
741 p->stages = stages;
743 if (!delay_htab)
745 delay_htab = new hash_table<delay_i1_hasher> (10);
746 delay_htab_i2 = new hash_table<delay_i2_hasher> (10);
748 slot = delay_htab->find_slot_with_hash (i1, htab_hash_pointer (i1), INSERT);
749 p->next_same_i1 = *slot;
750 *slot = p;
751 slot = delay_htab_i2->find_slot (p, INSERT);
752 *slot = p;
755 /* Examine the delay pair hashtable to see if INSN is a shadow for another,
756 and return the other insn if so. Return NULL otherwise. */
757 rtx_insn *
758 real_insn_for_shadow (rtx_insn *insn)
760 struct delay_pair *pair;
762 if (!delay_htab)
763 return NULL;
765 pair = delay_htab_i2->find_with_hash (insn, htab_hash_pointer (insn));
766 if (!pair || pair->stages > 0)
767 return NULL;
768 return pair->i1;
771 /* For a pair P of insns, return the fixed distance in cycles from the first
772 insn after which the second must be scheduled. */
773 static int
774 pair_delay (struct delay_pair *p)
776 if (p->stages == 0)
777 return p->cycles;
778 else
779 return p->stages * modulo_ii;
782 /* Given an insn INSN, add a dependence on its delayed shadow if it
783 has one. Also try to find situations where shadows depend on each other
784 and add dependencies to the real insns to limit the amount of backtracking
785 needed. */
786 void
787 add_delay_dependencies (rtx_insn *insn)
789 struct delay_pair *pair;
790 sd_iterator_def sd_it;
791 dep_t dep;
793 if (!delay_htab)
794 return;
796 pair = delay_htab_i2->find_with_hash (insn, htab_hash_pointer (insn));
797 if (!pair)
798 return;
799 add_dependence (insn, pair->i1, REG_DEP_ANTI);
800 if (pair->stages)
801 return;
803 FOR_EACH_DEP (pair->i2, SD_LIST_BACK, sd_it, dep)
805 rtx_insn *pro = DEP_PRO (dep);
806 struct delay_pair *other_pair
807 = delay_htab_i2->find_with_hash (pro, htab_hash_pointer (pro));
808 if (!other_pair || other_pair->stages)
809 continue;
810 if (pair_delay (other_pair) >= pair_delay (pair))
812 if (sched_verbose >= 4)
814 fprintf (sched_dump, ";;\tadding dependence %d <- %d\n",
815 INSN_UID (other_pair->i1),
816 INSN_UID (pair->i1));
817 fprintf (sched_dump, ";;\tpair1 %d <- %d, cost %d\n",
818 INSN_UID (pair->i1),
819 INSN_UID (pair->i2),
820 pair_delay (pair));
821 fprintf (sched_dump, ";;\tpair2 %d <- %d, cost %d\n",
822 INSN_UID (other_pair->i1),
823 INSN_UID (other_pair->i2),
824 pair_delay (other_pair));
826 add_dependence (pair->i1, other_pair->i1, REG_DEP_ANTI);
831 /* Forward declarations. */
833 static int priority (rtx_insn *);
834 static int autopref_rank_for_schedule (const rtx_insn *, const rtx_insn *);
835 static int rank_for_schedule (const void *, const void *);
836 static void swap_sort (rtx_insn **, int);
837 static void queue_insn (rtx_insn *, int, const char *);
838 static int schedule_insn (rtx_insn *);
839 static void adjust_priority (rtx_insn *);
840 static void advance_one_cycle (void);
841 static void extend_h_i_d (void);
844 /* Notes handling mechanism:
845 =========================
846 Generally, NOTES are saved before scheduling and restored after scheduling.
847 The scheduler distinguishes between two types of notes:
849 (1) LOOP_BEGIN, LOOP_END, SETJMP, EHREGION_BEG, EHREGION_END notes:
850 Before scheduling a region, a pointer to the note is added to the insn
851 that follows or precedes it. (This happens as part of the data dependence
852 computation). After scheduling an insn, the pointer contained in it is
853 used for regenerating the corresponding note (in reemit_notes).
855 (2) All other notes (e.g. INSN_DELETED): Before scheduling a block,
856 these notes are put in a list (in rm_other_notes() and
857 unlink_other_notes ()). After scheduling the block, these notes are
858 inserted at the beginning of the block (in schedule_block()). */
860 static void ready_add (struct ready_list *, rtx_insn *, bool);
861 static rtx_insn *ready_remove_first (struct ready_list *);
862 static rtx_insn *ready_remove_first_dispatch (struct ready_list *ready);
864 static void queue_to_ready (struct ready_list *);
865 static int early_queue_to_ready (state_t, struct ready_list *);
867 /* The following functions are used to implement multi-pass scheduling
868 on the first cycle. */
869 static rtx_insn *ready_remove (struct ready_list *, int);
870 static void ready_remove_insn (rtx_insn *);
872 static void fix_inter_tick (rtx_insn *, rtx_insn *);
873 static int fix_tick_ready (rtx_insn *);
874 static void change_queue_index (rtx_insn *, int);
876 /* The following functions are used to implement scheduling of data/control
877 speculative instructions. */
879 static void extend_h_i_d (void);
880 static void init_h_i_d (rtx_insn *);
881 static int haifa_speculate_insn (rtx_insn *, ds_t, rtx *);
882 static void generate_recovery_code (rtx_insn *);
883 static void process_insn_forw_deps_be_in_spec (rtx_insn *, rtx_insn *, ds_t);
884 static void begin_speculative_block (rtx_insn *);
885 static void add_to_speculative_block (rtx_insn *);
886 static void init_before_recovery (basic_block *);
887 static void create_check_block_twin (rtx_insn *, bool);
888 static void fix_recovery_deps (basic_block);
889 static bool haifa_change_pattern (rtx_insn *, rtx);
890 static void dump_new_block_header (int, basic_block, rtx_insn *, rtx_insn *);
891 static void restore_bb_notes (basic_block);
892 static void fix_jump_move (rtx_insn *);
893 static void move_block_after_check (rtx_insn *);
894 static void move_succs (vec<edge, va_gc> **, basic_block);
895 static void sched_remove_insn (rtx_insn *);
896 static void clear_priorities (rtx_insn *, rtx_vec_t *);
897 static void calc_priorities (rtx_vec_t);
898 static void add_jump_dependencies (rtx_insn *, rtx_insn *);
900 #endif /* INSN_SCHEDULING */
902 /* Point to state used for the current scheduling pass. */
903 struct haifa_sched_info *current_sched_info;
905 #ifndef INSN_SCHEDULING
906 void
907 schedule_insns (void)
910 #else
912 /* Do register pressure sensitive insn scheduling if the flag is set
913 up. */
914 enum sched_pressure_algorithm sched_pressure;
916 /* Map regno -> its pressure class. The map defined only when
917 SCHED_PRESSURE != SCHED_PRESSURE_NONE. */
918 enum reg_class *sched_regno_pressure_class;
920 /* The current register pressure. Only elements corresponding pressure
921 classes are defined. */
922 static int curr_reg_pressure[N_REG_CLASSES];
924 /* Saved value of the previous array. */
925 static int saved_reg_pressure[N_REG_CLASSES];
927 /* Register living at given scheduling point. */
928 static bitmap curr_reg_live;
930 /* Saved value of the previous array. */
931 static bitmap saved_reg_live;
933 /* Registers mentioned in the current region. */
934 static bitmap region_ref_regs;
936 /* Temporary bitmap used for SCHED_PRESSURE_MODEL. */
937 static bitmap tmp_bitmap;
939 /* Effective number of available registers of a given class (see comment
940 in sched_pressure_start_bb). */
941 static int sched_class_regs_num[N_REG_CLASSES];
942 /* Number of call_saved_regs and fixed_regs. Helpers for calculating of
943 sched_class_regs_num. */
944 static int call_saved_regs_num[N_REG_CLASSES];
945 static int fixed_regs_num[N_REG_CLASSES];
947 /* Initiate register pressure relative info for scheduling the current
948 region. Currently it is only clearing register mentioned in the
949 current region. */
950 void
951 sched_init_region_reg_pressure_info (void)
953 bitmap_clear (region_ref_regs);
956 /* PRESSURE[CL] describes the pressure on register class CL. Update it
957 for the birth (if BIRTH_P) or death (if !BIRTH_P) of register REGNO.
958 LIVE tracks the set of live registers; if it is null, assume that
959 every birth or death is genuine. */
960 static inline void
961 mark_regno_birth_or_death (bitmap live, int *pressure, int regno, bool birth_p)
963 enum reg_class pressure_class;
965 pressure_class = sched_regno_pressure_class[regno];
966 if (regno >= FIRST_PSEUDO_REGISTER)
968 if (pressure_class != NO_REGS)
970 if (birth_p)
972 if (!live || bitmap_set_bit (live, regno))
973 pressure[pressure_class]
974 += (ira_reg_class_max_nregs
975 [pressure_class][PSEUDO_REGNO_MODE (regno)]);
977 else
979 if (!live || bitmap_clear_bit (live, regno))
980 pressure[pressure_class]
981 -= (ira_reg_class_max_nregs
982 [pressure_class][PSEUDO_REGNO_MODE (regno)]);
986 else if (pressure_class != NO_REGS
987 && ! TEST_HARD_REG_BIT (ira_no_alloc_regs, regno))
989 if (birth_p)
991 if (!live || bitmap_set_bit (live, regno))
992 pressure[pressure_class]++;
994 else
996 if (!live || bitmap_clear_bit (live, regno))
997 pressure[pressure_class]--;
1002 /* Initiate current register pressure related info from living
1003 registers given by LIVE. */
1004 static void
1005 initiate_reg_pressure_info (bitmap live)
1007 int i;
1008 unsigned int j;
1009 bitmap_iterator bi;
1011 for (i = 0; i < ira_pressure_classes_num; i++)
1012 curr_reg_pressure[ira_pressure_classes[i]] = 0;
1013 bitmap_clear (curr_reg_live);
1014 EXECUTE_IF_SET_IN_BITMAP (live, 0, j, bi)
1015 if (sched_pressure == SCHED_PRESSURE_MODEL
1016 || current_nr_blocks == 1
1017 || bitmap_bit_p (region_ref_regs, j))
1018 mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure, j, true);
1021 /* Mark registers in X as mentioned in the current region. */
1022 static void
1023 setup_ref_regs (rtx x)
1025 int i, j;
1026 const RTX_CODE code = GET_CODE (x);
1027 const char *fmt;
1029 if (REG_P (x))
1031 bitmap_set_range (region_ref_regs, REGNO (x), REG_NREGS (x));
1032 return;
1034 fmt = GET_RTX_FORMAT (code);
1035 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1036 if (fmt[i] == 'e')
1037 setup_ref_regs (XEXP (x, i));
1038 else if (fmt[i] == 'E')
1040 for (j = 0; j < XVECLEN (x, i); j++)
1041 setup_ref_regs (XVECEXP (x, i, j));
1045 /* Initiate current register pressure related info at the start of
1046 basic block BB. */
1047 static void
1048 initiate_bb_reg_pressure_info (basic_block bb)
1050 unsigned int i ATTRIBUTE_UNUSED;
1051 rtx_insn *insn;
1053 if (current_nr_blocks > 1)
1054 FOR_BB_INSNS (bb, insn)
1055 if (NONDEBUG_INSN_P (insn))
1056 setup_ref_regs (PATTERN (insn));
1057 initiate_reg_pressure_info (df_get_live_in (bb));
1058 if (bb_has_eh_pred (bb))
1059 for (i = 0; ; ++i)
1061 unsigned int regno = EH_RETURN_DATA_REGNO (i);
1063 if (regno == INVALID_REGNUM)
1064 break;
1065 if (! bitmap_bit_p (df_get_live_in (bb), regno))
1066 mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure,
1067 regno, true);
1071 /* Save current register pressure related info. */
1072 static void
1073 save_reg_pressure (void)
1075 int i;
1077 for (i = 0; i < ira_pressure_classes_num; i++)
1078 saved_reg_pressure[ira_pressure_classes[i]]
1079 = curr_reg_pressure[ira_pressure_classes[i]];
1080 bitmap_copy (saved_reg_live, curr_reg_live);
1083 /* Restore saved register pressure related info. */
1084 static void
1085 restore_reg_pressure (void)
1087 int i;
1089 for (i = 0; i < ira_pressure_classes_num; i++)
1090 curr_reg_pressure[ira_pressure_classes[i]]
1091 = saved_reg_pressure[ira_pressure_classes[i]];
1092 bitmap_copy (curr_reg_live, saved_reg_live);
1095 /* Return TRUE if the register is dying after its USE. */
1096 static bool
1097 dying_use_p (struct reg_use_data *use)
1099 struct reg_use_data *next;
1101 for (next = use->next_regno_use; next != use; next = next->next_regno_use)
1102 if (NONDEBUG_INSN_P (next->insn)
1103 && QUEUE_INDEX (next->insn) != QUEUE_SCHEDULED)
1104 return false;
1105 return true;
1108 /* Print info about the current register pressure and its excess for
1109 each pressure class. */
1110 static void
1111 print_curr_reg_pressure (void)
1113 int i;
1114 enum reg_class cl;
1116 fprintf (sched_dump, ";;\t");
1117 for (i = 0; i < ira_pressure_classes_num; i++)
1119 cl = ira_pressure_classes[i];
1120 gcc_assert (curr_reg_pressure[cl] >= 0);
1121 fprintf (sched_dump, " %s:%d(%d)", reg_class_names[cl],
1122 curr_reg_pressure[cl],
1123 curr_reg_pressure[cl] - sched_class_regs_num[cl]);
1125 fprintf (sched_dump, "\n");
1128 /* Determine if INSN has a condition that is clobbered if a register
1129 in SET_REGS is modified. */
1130 static bool
1131 cond_clobbered_p (rtx_insn *insn, HARD_REG_SET set_regs)
1133 rtx pat = PATTERN (insn);
1134 gcc_assert (GET_CODE (pat) == COND_EXEC);
1135 if (TEST_HARD_REG_BIT (set_regs, REGNO (XEXP (COND_EXEC_TEST (pat), 0))))
1137 sd_iterator_def sd_it;
1138 dep_t dep;
1139 haifa_change_pattern (insn, ORIG_PAT (insn));
1140 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
1141 DEP_STATUS (dep) &= ~DEP_CANCELLED;
1142 TODO_SPEC (insn) = HARD_DEP;
1143 if (sched_verbose >= 2)
1144 fprintf (sched_dump,
1145 ";;\t\tdequeue insn %s because of clobbered condition\n",
1146 (*current_sched_info->print_insn) (insn, 0));
1147 return true;
1150 return false;
1153 /* This function should be called after modifying the pattern of INSN,
1154 to update scheduler data structures as needed. */
1155 static void
1156 update_insn_after_change (rtx_insn *insn)
1158 sd_iterator_def sd_it;
1159 dep_t dep;
1161 dfa_clear_single_insn_cache (insn);
1163 sd_it = sd_iterator_start (insn,
1164 SD_LIST_FORW | SD_LIST_BACK | SD_LIST_RES_BACK);
1165 while (sd_iterator_cond (&sd_it, &dep))
1167 DEP_COST (dep) = UNKNOWN_DEP_COST;
1168 sd_iterator_next (&sd_it);
1171 /* Invalidate INSN_COST, so it'll be recalculated. */
1172 INSN_COST (insn) = -1;
1173 /* Invalidate INSN_TICK, so it'll be recalculated. */
1174 INSN_TICK (insn) = INVALID_TICK;
1176 /* Invalidate autoprefetch data entry. */
1177 INSN_AUTOPREF_MULTIPASS_DATA (insn)[0].status
1178 = AUTOPREF_MULTIPASS_DATA_UNINITIALIZED;
1179 INSN_AUTOPREF_MULTIPASS_DATA (insn)[1].status
1180 = AUTOPREF_MULTIPASS_DATA_UNINITIALIZED;
1184 /* Two VECs, one to hold dependencies for which pattern replacements
1185 need to be applied or restored at the start of the next cycle, and
1186 another to hold an integer that is either one, to apply the
1187 corresponding replacement, or zero to restore it. */
1188 static vec<dep_t> next_cycle_replace_deps;
1189 static vec<int> next_cycle_apply;
1191 static void apply_replacement (dep_t, bool);
1192 static void restore_pattern (dep_t, bool);
1194 /* Look at the remaining dependencies for insn NEXT, and compute and return
1195 the TODO_SPEC value we should use for it. This is called after one of
1196 NEXT's dependencies has been resolved.
1197 We also perform pattern replacements for predication, and for broken
1198 replacement dependencies. The latter is only done if FOR_BACKTRACK is
1199 false. */
1201 static ds_t
1202 recompute_todo_spec (rtx_insn *next, bool for_backtrack)
1204 ds_t new_ds;
1205 sd_iterator_def sd_it;
1206 dep_t dep, modify_dep = NULL;
1207 int n_spec = 0;
1208 int n_control = 0;
1209 int n_replace = 0;
1210 bool first_p = true;
1212 if (sd_lists_empty_p (next, SD_LIST_BACK))
1213 /* NEXT has all its dependencies resolved. */
1214 return 0;
1216 if (!sd_lists_empty_p (next, SD_LIST_HARD_BACK))
1217 return HARD_DEP;
1219 /* If NEXT is intended to sit adjacent to this instruction, we don't
1220 want to try to break any dependencies. Treat it as a HARD_DEP. */
1221 if (SCHED_GROUP_P (next))
1222 return HARD_DEP;
1224 /* Now we've got NEXT with speculative deps only.
1225 1. Look at the deps to see what we have to do.
1226 2. Check if we can do 'todo'. */
1227 new_ds = 0;
1229 FOR_EACH_DEP (next, SD_LIST_BACK, sd_it, dep)
1231 rtx_insn *pro = DEP_PRO (dep);
1232 ds_t ds = DEP_STATUS (dep) & SPECULATIVE;
1234 if (DEBUG_INSN_P (pro) && !DEBUG_INSN_P (next))
1235 continue;
1237 if (ds)
1239 n_spec++;
1240 if (first_p)
1242 first_p = false;
1244 new_ds = ds;
1246 else
1247 new_ds = ds_merge (new_ds, ds);
1249 else if (DEP_TYPE (dep) == REG_DEP_CONTROL)
1251 if (QUEUE_INDEX (pro) != QUEUE_SCHEDULED)
1253 n_control++;
1254 modify_dep = dep;
1256 DEP_STATUS (dep) &= ~DEP_CANCELLED;
1258 else if (DEP_REPLACE (dep) != NULL)
1260 if (QUEUE_INDEX (pro) != QUEUE_SCHEDULED)
1262 n_replace++;
1263 modify_dep = dep;
1265 DEP_STATUS (dep) &= ~DEP_CANCELLED;
1269 if (n_replace > 0 && n_control == 0 && n_spec == 0)
1271 if (!dbg_cnt (sched_breakdep))
1272 return HARD_DEP;
1273 FOR_EACH_DEP (next, SD_LIST_BACK, sd_it, dep)
1275 struct dep_replacement *desc = DEP_REPLACE (dep);
1276 if (desc != NULL)
1278 if (desc->insn == next && !for_backtrack)
1280 gcc_assert (n_replace == 1);
1281 apply_replacement (dep, true);
1283 DEP_STATUS (dep) |= DEP_CANCELLED;
1286 return 0;
1289 else if (n_control == 1 && n_replace == 0 && n_spec == 0)
1291 rtx_insn *pro, *other;
1292 rtx new_pat;
1293 rtx cond = NULL_RTX;
1294 bool success;
1295 rtx_insn *prev = NULL;
1296 int i;
1297 unsigned regno;
1299 if ((current_sched_info->flags & DO_PREDICATION) == 0
1300 || (ORIG_PAT (next) != NULL_RTX
1301 && PREDICATED_PAT (next) == NULL_RTX))
1302 return HARD_DEP;
1304 pro = DEP_PRO (modify_dep);
1305 other = real_insn_for_shadow (pro);
1306 if (other != NULL_RTX)
1307 pro = other;
1309 cond = sched_get_reverse_condition_uncached (pro);
1310 regno = REGNO (XEXP (cond, 0));
1312 /* Find the last scheduled insn that modifies the condition register.
1313 We can stop looking once we find the insn we depend on through the
1314 REG_DEP_CONTROL; if the condition register isn't modified after it,
1315 we know that it still has the right value. */
1316 if (QUEUE_INDEX (pro) == QUEUE_SCHEDULED)
1317 FOR_EACH_VEC_ELT_REVERSE (scheduled_insns, i, prev)
1319 HARD_REG_SET t;
1321 find_all_hard_reg_sets (prev, &t, true);
1322 if (TEST_HARD_REG_BIT (t, regno))
1323 return HARD_DEP;
1324 if (prev == pro)
1325 break;
1327 if (ORIG_PAT (next) == NULL_RTX)
1329 ORIG_PAT (next) = PATTERN (next);
1331 new_pat = gen_rtx_COND_EXEC (VOIDmode, cond, PATTERN (next));
1332 success = haifa_change_pattern (next, new_pat);
1333 if (!success)
1334 return HARD_DEP;
1335 PREDICATED_PAT (next) = new_pat;
1337 else if (PATTERN (next) != PREDICATED_PAT (next))
1339 bool success = haifa_change_pattern (next,
1340 PREDICATED_PAT (next));
1341 gcc_assert (success);
1343 DEP_STATUS (modify_dep) |= DEP_CANCELLED;
1344 return DEP_CONTROL;
1347 if (PREDICATED_PAT (next) != NULL_RTX)
1349 int tick = INSN_TICK (next);
1350 bool success = haifa_change_pattern (next,
1351 ORIG_PAT (next));
1352 INSN_TICK (next) = tick;
1353 gcc_assert (success);
1356 /* We can't handle the case where there are both speculative and control
1357 dependencies, so we return HARD_DEP in such a case. Also fail if
1358 we have speculative dependencies with not enough points, or more than
1359 one control dependency. */
1360 if ((n_spec > 0 && (n_control > 0 || n_replace > 0))
1361 || (n_spec > 0
1362 /* Too few points? */
1363 && ds_weak (new_ds) < spec_info->data_weakness_cutoff)
1364 || n_control > 0
1365 || n_replace > 0)
1366 return HARD_DEP;
1368 return new_ds;
1371 /* Pointer to the last instruction scheduled. */
1372 static rtx_insn *last_scheduled_insn;
1374 /* Pointer to the last nondebug instruction scheduled within the
1375 block, or the prev_head of the scheduling block. Used by
1376 rank_for_schedule, so that insns independent of the last scheduled
1377 insn will be preferred over dependent instructions. */
1378 static rtx_insn *last_nondebug_scheduled_insn;
1380 /* Pointer that iterates through the list of unscheduled insns if we
1381 have a dbg_cnt enabled. It always points at an insn prior to the
1382 first unscheduled one. */
1383 static rtx_insn *nonscheduled_insns_begin;
1385 /* Compute cost of executing INSN.
1386 This is the number of cycles between instruction issue and
1387 instruction results. */
1389 insn_sched_cost (rtx_insn *insn)
1391 int cost;
1393 if (sched_fusion)
1394 return 0;
1396 if (sel_sched_p ())
1398 if (recog_memoized (insn) < 0)
1399 return 0;
1401 cost = insn_default_latency (insn);
1402 if (cost < 0)
1403 cost = 0;
1405 return cost;
1408 cost = INSN_COST (insn);
1410 if (cost < 0)
1412 /* A USE insn, or something else we don't need to
1413 understand. We can't pass these directly to
1414 result_ready_cost or insn_default_latency because it will
1415 trigger a fatal error for unrecognizable insns. */
1416 if (recog_memoized (insn) < 0)
1418 INSN_COST (insn) = 0;
1419 return 0;
1421 else
1423 cost = insn_default_latency (insn);
1424 if (cost < 0)
1425 cost = 0;
1427 INSN_COST (insn) = cost;
1431 return cost;
1434 /* Compute cost of dependence LINK.
1435 This is the number of cycles between instruction issue and
1436 instruction results.
1437 ??? We also use this function to call recog_memoized on all insns. */
1439 dep_cost_1 (dep_t link, dw_t dw)
1441 rtx_insn *insn = DEP_PRO (link);
1442 rtx_insn *used = DEP_CON (link);
1443 int cost;
1445 if (DEP_COST (link) != UNKNOWN_DEP_COST)
1446 return DEP_COST (link);
1448 if (delay_htab)
1450 struct delay_pair *delay_entry;
1451 delay_entry
1452 = delay_htab_i2->find_with_hash (used, htab_hash_pointer (used));
1453 if (delay_entry)
1455 if (delay_entry->i1 == insn)
1457 DEP_COST (link) = pair_delay (delay_entry);
1458 return DEP_COST (link);
1463 /* A USE insn should never require the value used to be computed.
1464 This allows the computation of a function's result and parameter
1465 values to overlap the return and call. We don't care about the
1466 dependence cost when only decreasing register pressure. */
1467 if (recog_memoized (used) < 0)
1469 cost = 0;
1470 recog_memoized (insn);
1472 else
1474 enum reg_note dep_type = DEP_TYPE (link);
1476 cost = insn_sched_cost (insn);
1478 if (INSN_CODE (insn) >= 0)
1480 if (dep_type == REG_DEP_ANTI)
1481 cost = 0;
1482 else if (dep_type == REG_DEP_OUTPUT)
1484 cost = (insn_default_latency (insn)
1485 - insn_default_latency (used));
1486 if (cost <= 0)
1487 cost = 1;
1489 else if (bypass_p (insn))
1490 cost = insn_latency (insn, used);
1494 if (targetm.sched.adjust_cost)
1495 cost = targetm.sched.adjust_cost (used, (int) dep_type, insn, cost,
1496 dw);
1498 if (cost < 0)
1499 cost = 0;
1502 DEP_COST (link) = cost;
1503 return cost;
1506 /* Compute cost of dependence LINK.
1507 This is the number of cycles between instruction issue and
1508 instruction results. */
1510 dep_cost (dep_t link)
1512 return dep_cost_1 (link, 0);
1515 /* Use this sel-sched.c friendly function in reorder2 instead of increasing
1516 INSN_PRIORITY explicitly. */
1517 void
1518 increase_insn_priority (rtx_insn *insn, int amount)
1520 if (!sel_sched_p ())
1522 /* We're dealing with haifa-sched.c INSN_PRIORITY. */
1523 if (INSN_PRIORITY_KNOWN (insn))
1524 INSN_PRIORITY (insn) += amount;
1526 else
1528 /* In sel-sched.c INSN_PRIORITY is not kept up to date.
1529 Use EXPR_PRIORITY instead. */
1530 sel_add_to_insn_priority (insn, amount);
1534 /* Return 'true' if DEP should be included in priority calculations. */
1535 static bool
1536 contributes_to_priority_p (dep_t dep)
1538 if (DEBUG_INSN_P (DEP_CON (dep))
1539 || DEBUG_INSN_P (DEP_PRO (dep)))
1540 return false;
1542 /* Critical path is meaningful in block boundaries only. */
1543 if (!current_sched_info->contributes_to_priority (DEP_CON (dep),
1544 DEP_PRO (dep)))
1545 return false;
1547 if (DEP_REPLACE (dep) != NULL)
1548 return false;
1550 /* If flag COUNT_SPEC_IN_CRITICAL_PATH is set,
1551 then speculative instructions will less likely be
1552 scheduled. That is because the priority of
1553 their producers will increase, and, thus, the
1554 producers will more likely be scheduled, thus,
1555 resolving the dependence. */
1556 if (sched_deps_info->generate_spec_deps
1557 && !(spec_info->flags & COUNT_SPEC_IN_CRITICAL_PATH)
1558 && (DEP_STATUS (dep) & SPECULATIVE))
1559 return false;
1561 return true;
1564 /* Compute the number of nondebug deps in list LIST for INSN. */
1566 static int
1567 dep_list_size (rtx_insn *insn, sd_list_types_def list)
1569 sd_iterator_def sd_it;
1570 dep_t dep;
1571 int dbgcount = 0, nodbgcount = 0;
1573 if (!MAY_HAVE_DEBUG_INSNS)
1574 return sd_lists_size (insn, list);
1576 FOR_EACH_DEP (insn, list, sd_it, dep)
1578 if (DEBUG_INSN_P (DEP_CON (dep)))
1579 dbgcount++;
1580 else if (!DEBUG_INSN_P (DEP_PRO (dep)))
1581 nodbgcount++;
1584 gcc_assert (dbgcount + nodbgcount == sd_lists_size (insn, list));
1586 return nodbgcount;
1589 bool sched_fusion;
1591 /* Compute the priority number for INSN. */
1592 static int
1593 priority (rtx_insn *insn)
1595 if (! INSN_P (insn))
1596 return 0;
1598 /* We should not be interested in priority of an already scheduled insn. */
1599 gcc_assert (QUEUE_INDEX (insn) != QUEUE_SCHEDULED);
1601 if (!INSN_PRIORITY_KNOWN (insn))
1603 int this_priority = -1;
1605 if (sched_fusion)
1607 int this_fusion_priority;
1609 targetm.sched.fusion_priority (insn, FUSION_MAX_PRIORITY,
1610 &this_fusion_priority, &this_priority);
1611 INSN_FUSION_PRIORITY (insn) = this_fusion_priority;
1613 else if (dep_list_size (insn, SD_LIST_FORW) == 0)
1614 /* ??? We should set INSN_PRIORITY to insn_sched_cost when and insn
1615 has some forward deps but all of them are ignored by
1616 contributes_to_priority hook. At the moment we set priority of
1617 such insn to 0. */
1618 this_priority = insn_sched_cost (insn);
1619 else
1621 rtx_insn *prev_first, *twin;
1622 basic_block rec;
1624 /* For recovery check instructions we calculate priority slightly
1625 different than that of normal instructions. Instead of walking
1626 through INSN_FORW_DEPS (check) list, we walk through
1627 INSN_FORW_DEPS list of each instruction in the corresponding
1628 recovery block. */
1630 /* Selective scheduling does not define RECOVERY_BLOCK macro. */
1631 rec = sel_sched_p () ? NULL : RECOVERY_BLOCK (insn);
1632 if (!rec || rec == EXIT_BLOCK_PTR_FOR_FN (cfun))
1634 prev_first = PREV_INSN (insn);
1635 twin = insn;
1637 else
1639 prev_first = NEXT_INSN (BB_HEAD (rec));
1640 twin = PREV_INSN (BB_END (rec));
1645 sd_iterator_def sd_it;
1646 dep_t dep;
1648 FOR_EACH_DEP (twin, SD_LIST_FORW, sd_it, dep)
1650 rtx_insn *next;
1651 int next_priority;
1653 next = DEP_CON (dep);
1655 if (BLOCK_FOR_INSN (next) != rec)
1657 int cost;
1659 if (!contributes_to_priority_p (dep))
1660 continue;
1662 if (twin == insn)
1663 cost = dep_cost (dep);
1664 else
1666 struct _dep _dep1, *dep1 = &_dep1;
1668 init_dep (dep1, insn, next, REG_DEP_ANTI);
1670 cost = dep_cost (dep1);
1673 next_priority = cost + priority (next);
1675 if (next_priority > this_priority)
1676 this_priority = next_priority;
1680 twin = PREV_INSN (twin);
1682 while (twin != prev_first);
1685 if (this_priority < 0)
1687 gcc_assert (this_priority == -1);
1689 this_priority = insn_sched_cost (insn);
1692 INSN_PRIORITY (insn) = this_priority;
1693 INSN_PRIORITY_STATUS (insn) = 1;
1696 return INSN_PRIORITY (insn);
1699 /* Macros and functions for keeping the priority queue sorted, and
1700 dealing with queuing and dequeuing of instructions. */
1702 /* For each pressure class CL, set DEATH[CL] to the number of registers
1703 in that class that die in INSN. */
1705 static void
1706 calculate_reg_deaths (rtx_insn *insn, int *death)
1708 int i;
1709 struct reg_use_data *use;
1711 for (i = 0; i < ira_pressure_classes_num; i++)
1712 death[ira_pressure_classes[i]] = 0;
1713 for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
1714 if (dying_use_p (use))
1715 mark_regno_birth_or_death (0, death, use->regno, true);
1718 /* Setup info about the current register pressure impact of scheduling
1719 INSN at the current scheduling point. */
1720 static void
1721 setup_insn_reg_pressure_info (rtx_insn *insn)
1723 int i, change, before, after, hard_regno;
1724 int excess_cost_change;
1725 machine_mode mode;
1726 enum reg_class cl;
1727 struct reg_pressure_data *pressure_info;
1728 int *max_reg_pressure;
1729 static int death[N_REG_CLASSES];
1731 gcc_checking_assert (!DEBUG_INSN_P (insn));
1733 excess_cost_change = 0;
1734 calculate_reg_deaths (insn, death);
1735 pressure_info = INSN_REG_PRESSURE (insn);
1736 max_reg_pressure = INSN_MAX_REG_PRESSURE (insn);
1737 gcc_assert (pressure_info != NULL && max_reg_pressure != NULL);
1738 for (i = 0; i < ira_pressure_classes_num; i++)
1740 cl = ira_pressure_classes[i];
1741 gcc_assert (curr_reg_pressure[cl] >= 0);
1742 change = (int) pressure_info[i].set_increase - death[cl];
1743 before = MAX (0, max_reg_pressure[i] - sched_class_regs_num[cl]);
1744 after = MAX (0, max_reg_pressure[i] + change
1745 - sched_class_regs_num[cl]);
1746 hard_regno = ira_class_hard_regs[cl][0];
1747 gcc_assert (hard_regno >= 0);
1748 mode = reg_raw_mode[hard_regno];
1749 excess_cost_change += ((after - before)
1750 * (ira_memory_move_cost[mode][cl][0]
1751 + ira_memory_move_cost[mode][cl][1]));
1753 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insn) = excess_cost_change;
1756 /* This is the first page of code related to SCHED_PRESSURE_MODEL.
1757 It tries to make the scheduler take register pressure into account
1758 without introducing too many unnecessary stalls. It hooks into the
1759 main scheduling algorithm at several points:
1761 - Before scheduling starts, model_start_schedule constructs a
1762 "model schedule" for the current block. This model schedule is
1763 chosen solely to keep register pressure down. It does not take the
1764 target's pipeline or the original instruction order into account,
1765 except as a tie-breaker. It also doesn't work to a particular
1766 pressure limit.
1768 This model schedule gives us an idea of what pressure can be
1769 achieved for the block and gives us an example of a schedule that
1770 keeps to that pressure. It also makes the final schedule less
1771 dependent on the original instruction order. This is important
1772 because the original order can either be "wide" (many values live
1773 at once, such as in user-scheduled code) or "narrow" (few values
1774 live at once, such as after loop unrolling, where several
1775 iterations are executed sequentially).
1777 We do not apply this model schedule to the rtx stream. We simply
1778 record it in model_schedule. We also compute the maximum pressure,
1779 MP, that was seen during this schedule.
1781 - Instructions are added to the ready queue even if they require
1782 a stall. The length of the stall is instead computed as:
1784 MAX (INSN_TICK (INSN) - clock_var, 0)
1786 (= insn_delay). This allows rank_for_schedule to choose between
1787 introducing a deliberate stall or increasing pressure.
1789 - Before sorting the ready queue, model_set_excess_costs assigns
1790 a pressure-based cost to each ready instruction in the queue.
1791 This is the instruction's INSN_REG_PRESSURE_EXCESS_COST_CHANGE
1792 (ECC for short) and is effectively measured in cycles.
1794 - rank_for_schedule ranks instructions based on:
1796 ECC (insn) + insn_delay (insn)
1798 then as:
1800 insn_delay (insn)
1802 So, for example, an instruction X1 with an ECC of 1 that can issue
1803 now will win over an instruction X0 with an ECC of zero that would
1804 introduce a stall of one cycle. However, an instruction X2 with an
1805 ECC of 2 that can issue now will lose to both X0 and X1.
1807 - When an instruction is scheduled, model_recompute updates the model
1808 schedule with the new pressures (some of which might now exceed the
1809 original maximum pressure MP). model_update_limit_points then searches
1810 for the new point of maximum pressure, if not already known. */
1812 /* Used to separate high-verbosity debug information for SCHED_PRESSURE_MODEL
1813 from surrounding debug information. */
1814 #define MODEL_BAR \
1815 ";;\t\t+------------------------------------------------------\n"
1817 /* Information about the pressure on a particular register class at a
1818 particular point of the model schedule. */
1819 struct model_pressure_data {
1820 /* The pressure at this point of the model schedule, or -1 if the
1821 point is associated with an instruction that has already been
1822 scheduled. */
1823 int ref_pressure;
1825 /* The maximum pressure during or after this point of the model schedule. */
1826 int max_pressure;
1829 /* Per-instruction information that is used while building the model
1830 schedule. Here, "schedule" refers to the model schedule rather
1831 than the main schedule. */
1832 struct model_insn_info {
1833 /* The instruction itself. */
1834 rtx_insn *insn;
1836 /* If this instruction is in model_worklist, these fields link to the
1837 previous (higher-priority) and next (lower-priority) instructions
1838 in the list. */
1839 struct model_insn_info *prev;
1840 struct model_insn_info *next;
1842 /* While constructing the schedule, QUEUE_INDEX describes whether an
1843 instruction has already been added to the schedule (QUEUE_SCHEDULED),
1844 is in model_worklist (QUEUE_READY), or neither (QUEUE_NOWHERE).
1845 old_queue records the value that QUEUE_INDEX had before scheduling
1846 started, so that we can restore it once the schedule is complete. */
1847 int old_queue;
1849 /* The relative importance of an unscheduled instruction. Higher
1850 values indicate greater importance. */
1851 unsigned int model_priority;
1853 /* The length of the longest path of satisfied true dependencies
1854 that leads to this instruction. */
1855 unsigned int depth;
1857 /* The length of the longest path of dependencies of any kind
1858 that leads from this instruction. */
1859 unsigned int alap;
1861 /* The number of predecessor nodes that must still be scheduled. */
1862 int unscheduled_preds;
1865 /* Information about the pressure limit for a particular register class.
1866 This structure is used when applying a model schedule to the main
1867 schedule. */
1868 struct model_pressure_limit {
1869 /* The maximum register pressure seen in the original model schedule. */
1870 int orig_pressure;
1872 /* The maximum register pressure seen in the current model schedule
1873 (which excludes instructions that have already been scheduled). */
1874 int pressure;
1876 /* The point of the current model schedule at which PRESSURE is first
1877 reached. It is set to -1 if the value needs to be recomputed. */
1878 int point;
1881 /* Describes a particular way of measuring register pressure. */
1882 struct model_pressure_group {
1883 /* Index PCI describes the maximum pressure on ira_pressure_classes[PCI]. */
1884 struct model_pressure_limit limits[N_REG_CLASSES];
1886 /* Index (POINT * ira_num_pressure_classes + PCI) describes the pressure
1887 on register class ira_pressure_classes[PCI] at point POINT of the
1888 current model schedule. A POINT of model_num_insns describes the
1889 pressure at the end of the schedule. */
1890 struct model_pressure_data *model;
1893 /* Index POINT gives the instruction at point POINT of the model schedule.
1894 This array doesn't change during main scheduling. */
1895 static vec<rtx_insn *> model_schedule;
1897 /* The list of instructions in the model worklist, sorted in order of
1898 decreasing priority. */
1899 static struct model_insn_info *model_worklist;
1901 /* Index I describes the instruction with INSN_LUID I. */
1902 static struct model_insn_info *model_insns;
1904 /* The number of instructions in the model schedule. */
1905 static int model_num_insns;
1907 /* The index of the first instruction in model_schedule that hasn't yet been
1908 added to the main schedule, or model_num_insns if all of them have. */
1909 static int model_curr_point;
1911 /* Describes the pressure before each instruction in the model schedule. */
1912 static struct model_pressure_group model_before_pressure;
1914 /* The first unused model_priority value (as used in model_insn_info). */
1915 static unsigned int model_next_priority;
1918 /* The model_pressure_data for ira_pressure_classes[PCI] in GROUP
1919 at point POINT of the model schedule. */
1920 #define MODEL_PRESSURE_DATA(GROUP, POINT, PCI) \
1921 (&(GROUP)->model[(POINT) * ira_pressure_classes_num + (PCI)])
1923 /* The maximum pressure on ira_pressure_classes[PCI] in GROUP at or
1924 after point POINT of the model schedule. */
1925 #define MODEL_MAX_PRESSURE(GROUP, POINT, PCI) \
1926 (MODEL_PRESSURE_DATA (GROUP, POINT, PCI)->max_pressure)
1928 /* The pressure on ira_pressure_classes[PCI] in GROUP at point POINT
1929 of the model schedule. */
1930 #define MODEL_REF_PRESSURE(GROUP, POINT, PCI) \
1931 (MODEL_PRESSURE_DATA (GROUP, POINT, PCI)->ref_pressure)
1933 /* Information about INSN that is used when creating the model schedule. */
1934 #define MODEL_INSN_INFO(INSN) \
1935 (&model_insns[INSN_LUID (INSN)])
1937 /* The instruction at point POINT of the model schedule. */
1938 #define MODEL_INSN(POINT) \
1939 (model_schedule[POINT])
1942 /* Return INSN's index in the model schedule, or model_num_insns if it
1943 doesn't belong to that schedule. */
1945 static int
1946 model_index (rtx_insn *insn)
1948 if (INSN_MODEL_INDEX (insn) == 0)
1949 return model_num_insns;
1950 return INSN_MODEL_INDEX (insn) - 1;
1953 /* Make sure that GROUP->limits is up-to-date for the current point
1954 of the model schedule. */
1956 static void
1957 model_update_limit_points_in_group (struct model_pressure_group *group)
1959 int pci, max_pressure, point;
1961 for (pci = 0; pci < ira_pressure_classes_num; pci++)
1963 /* We may have passed the final point at which the pressure in
1964 group->limits[pci].pressure was reached. Update the limit if so. */
1965 max_pressure = MODEL_MAX_PRESSURE (group, model_curr_point, pci);
1966 group->limits[pci].pressure = max_pressure;
1968 /* Find the point at which MAX_PRESSURE is first reached. We need
1969 to search in three cases:
1971 - We've already moved past the previous pressure point.
1972 In this case we search forward from model_curr_point.
1974 - We scheduled the previous point of maximum pressure ahead of
1975 its position in the model schedule, but doing so didn't bring
1976 the pressure point earlier. In this case we search forward
1977 from that previous pressure point.
1979 - Scheduling an instruction early caused the maximum pressure
1980 to decrease. In this case we will have set the pressure
1981 point to -1, and we search forward from model_curr_point. */
1982 point = MAX (group->limits[pci].point, model_curr_point);
1983 while (point < model_num_insns
1984 && MODEL_REF_PRESSURE (group, point, pci) < max_pressure)
1985 point++;
1986 group->limits[pci].point = point;
1988 gcc_assert (MODEL_REF_PRESSURE (group, point, pci) == max_pressure);
1989 gcc_assert (MODEL_MAX_PRESSURE (group, point, pci) == max_pressure);
1993 /* Make sure that all register-pressure limits are up-to-date for the
1994 current position in the model schedule. */
1996 static void
1997 model_update_limit_points (void)
1999 model_update_limit_points_in_group (&model_before_pressure);
2002 /* Return the model_index of the last unscheduled use in chain USE
2003 outside of USE's instruction. Return -1 if there are no other uses,
2004 or model_num_insns if the register is live at the end of the block. */
2006 static int
2007 model_last_use_except (struct reg_use_data *use)
2009 struct reg_use_data *next;
2010 int last, index;
2012 last = -1;
2013 for (next = use->next_regno_use; next != use; next = next->next_regno_use)
2014 if (NONDEBUG_INSN_P (next->insn)
2015 && QUEUE_INDEX (next->insn) != QUEUE_SCHEDULED)
2017 index = model_index (next->insn);
2018 if (index == model_num_insns)
2019 return model_num_insns;
2020 if (last < index)
2021 last = index;
2023 return last;
2026 /* An instruction with model_index POINT has just been scheduled, and it
2027 adds DELTA to the pressure on ira_pressure_classes[PCI] after POINT - 1.
2028 Update MODEL_REF_PRESSURE (GROUP, POINT, PCI) and
2029 MODEL_MAX_PRESSURE (GROUP, POINT, PCI) accordingly. */
2031 static void
2032 model_start_update_pressure (struct model_pressure_group *group,
2033 int point, int pci, int delta)
2035 int next_max_pressure;
2037 if (point == model_num_insns)
2039 /* The instruction wasn't part of the model schedule; it was moved
2040 from a different block. Update the pressure for the end of
2041 the model schedule. */
2042 MODEL_REF_PRESSURE (group, point, pci) += delta;
2043 MODEL_MAX_PRESSURE (group, point, pci) += delta;
2045 else
2047 /* Record that this instruction has been scheduled. Nothing now
2048 changes between POINT and POINT + 1, so get the maximum pressure
2049 from the latter. If the maximum pressure decreases, the new
2050 pressure point may be before POINT. */
2051 MODEL_REF_PRESSURE (group, point, pci) = -1;
2052 next_max_pressure = MODEL_MAX_PRESSURE (group, point + 1, pci);
2053 if (MODEL_MAX_PRESSURE (group, point, pci) > next_max_pressure)
2055 MODEL_MAX_PRESSURE (group, point, pci) = next_max_pressure;
2056 if (group->limits[pci].point == point)
2057 group->limits[pci].point = -1;
2062 /* Record that scheduling a later instruction has changed the pressure
2063 at point POINT of the model schedule by DELTA (which might be 0).
2064 Update GROUP accordingly. Return nonzero if these changes might
2065 trigger changes to previous points as well. */
2067 static int
2068 model_update_pressure (struct model_pressure_group *group,
2069 int point, int pci, int delta)
2071 int ref_pressure, max_pressure, next_max_pressure;
2073 /* If POINT hasn't yet been scheduled, update its pressure. */
2074 ref_pressure = MODEL_REF_PRESSURE (group, point, pci);
2075 if (ref_pressure >= 0 && delta != 0)
2077 ref_pressure += delta;
2078 MODEL_REF_PRESSURE (group, point, pci) = ref_pressure;
2080 /* Check whether the maximum pressure in the overall schedule
2081 has increased. (This means that the MODEL_MAX_PRESSURE of
2082 every point <= POINT will need to increase too; see below.) */
2083 if (group->limits[pci].pressure < ref_pressure)
2084 group->limits[pci].pressure = ref_pressure;
2086 /* If we are at maximum pressure, and the maximum pressure
2087 point was previously unknown or later than POINT,
2088 bring it forward. */
2089 if (group->limits[pci].pressure == ref_pressure
2090 && !IN_RANGE (group->limits[pci].point, 0, point))
2091 group->limits[pci].point = point;
2093 /* If POINT used to be the point of maximum pressure, but isn't
2094 any longer, we need to recalculate it using a forward walk. */
2095 if (group->limits[pci].pressure > ref_pressure
2096 && group->limits[pci].point == point)
2097 group->limits[pci].point = -1;
2100 /* Update the maximum pressure at POINT. Changes here might also
2101 affect the maximum pressure at POINT - 1. */
2102 next_max_pressure = MODEL_MAX_PRESSURE (group, point + 1, pci);
2103 max_pressure = MAX (ref_pressure, next_max_pressure);
2104 if (MODEL_MAX_PRESSURE (group, point, pci) != max_pressure)
2106 MODEL_MAX_PRESSURE (group, point, pci) = max_pressure;
2107 return 1;
2109 return 0;
2112 /* INSN has just been scheduled. Update the model schedule accordingly. */
2114 static void
2115 model_recompute (rtx_insn *insn)
2117 struct {
2118 int last_use;
2119 int regno;
2120 } uses[FIRST_PSEUDO_REGISTER + MAX_RECOG_OPERANDS];
2121 struct reg_use_data *use;
2122 struct reg_pressure_data *reg_pressure;
2123 int delta[N_REG_CLASSES];
2124 int pci, point, mix, new_last, cl, ref_pressure, queue;
2125 unsigned int i, num_uses, num_pending_births;
2126 bool print_p;
2128 /* The destinations of INSN were previously live from POINT onwards, but are
2129 now live from model_curr_point onwards. Set up DELTA accordingly. */
2130 point = model_index (insn);
2131 reg_pressure = INSN_REG_PRESSURE (insn);
2132 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2134 cl = ira_pressure_classes[pci];
2135 delta[cl] = reg_pressure[pci].set_increase;
2138 /* Record which registers previously died at POINT, but which now die
2139 before POINT. Adjust DELTA so that it represents the effect of
2140 this change after POINT - 1. Set NUM_PENDING_BIRTHS to the number of
2141 registers that will be born in the range [model_curr_point, POINT). */
2142 num_uses = 0;
2143 num_pending_births = 0;
2144 bitmap_clear (tmp_bitmap);
2145 for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
2147 new_last = model_last_use_except (use);
2148 if (new_last < point && bitmap_set_bit (tmp_bitmap, use->regno))
2150 gcc_assert (num_uses < ARRAY_SIZE (uses));
2151 uses[num_uses].last_use = new_last;
2152 uses[num_uses].regno = use->regno;
2153 /* This register is no longer live after POINT - 1. */
2154 mark_regno_birth_or_death (NULL, delta, use->regno, false);
2155 num_uses++;
2156 if (new_last >= 0)
2157 num_pending_births++;
2161 /* Update the MODEL_REF_PRESSURE and MODEL_MAX_PRESSURE for POINT.
2162 Also set each group pressure limit for POINT. */
2163 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2165 cl = ira_pressure_classes[pci];
2166 model_start_update_pressure (&model_before_pressure,
2167 point, pci, delta[cl]);
2170 /* Walk the model schedule backwards, starting immediately before POINT. */
2171 print_p = false;
2172 if (point != model_curr_point)
2175 point--;
2176 insn = MODEL_INSN (point);
2177 queue = QUEUE_INDEX (insn);
2179 if (queue != QUEUE_SCHEDULED)
2181 /* DELTA describes the effect of the move on the register pressure
2182 after POINT. Make it describe the effect on the pressure
2183 before POINT. */
2184 i = 0;
2185 while (i < num_uses)
2187 if (uses[i].last_use == point)
2189 /* This register is now live again. */
2190 mark_regno_birth_or_death (NULL, delta,
2191 uses[i].regno, true);
2193 /* Remove this use from the array. */
2194 uses[i] = uses[num_uses - 1];
2195 num_uses--;
2196 num_pending_births--;
2198 else
2199 i++;
2202 if (sched_verbose >= 5)
2204 if (!print_p)
2206 fprintf (sched_dump, MODEL_BAR);
2207 fprintf (sched_dump, ";;\t\t| New pressure for model"
2208 " schedule\n");
2209 fprintf (sched_dump, MODEL_BAR);
2210 print_p = true;
2213 fprintf (sched_dump, ";;\t\t| %3d %4d %-30s ",
2214 point, INSN_UID (insn),
2215 str_pattern_slim (PATTERN (insn)));
2216 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2218 cl = ira_pressure_classes[pci];
2219 ref_pressure = MODEL_REF_PRESSURE (&model_before_pressure,
2220 point, pci);
2221 fprintf (sched_dump, " %s:[%d->%d]",
2222 reg_class_names[ira_pressure_classes[pci]],
2223 ref_pressure, ref_pressure + delta[cl]);
2225 fprintf (sched_dump, "\n");
2229 /* Adjust the pressure at POINT. Set MIX to nonzero if POINT - 1
2230 might have changed as well. */
2231 mix = num_pending_births;
2232 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2234 cl = ira_pressure_classes[pci];
2235 mix |= delta[cl];
2236 mix |= model_update_pressure (&model_before_pressure,
2237 point, pci, delta[cl]);
2240 while (mix && point > model_curr_point);
2242 if (print_p)
2243 fprintf (sched_dump, MODEL_BAR);
2246 /* After DEP, which was cancelled, has been resolved for insn NEXT,
2247 check whether the insn's pattern needs restoring. */
2248 static bool
2249 must_restore_pattern_p (rtx_insn *next, dep_t dep)
2251 if (QUEUE_INDEX (next) == QUEUE_SCHEDULED)
2252 return false;
2254 if (DEP_TYPE (dep) == REG_DEP_CONTROL)
2256 gcc_assert (ORIG_PAT (next) != NULL_RTX);
2257 gcc_assert (next == DEP_CON (dep));
2259 else
2261 struct dep_replacement *desc = DEP_REPLACE (dep);
2262 if (desc->insn != next)
2264 gcc_assert (*desc->loc == desc->orig);
2265 return false;
2268 return true;
2271 /* model_spill_cost (CL, P, P') returns the cost of increasing the
2272 pressure on CL from P to P'. We use this to calculate a "base ECC",
2273 baseECC (CL, X), for each pressure class CL and each instruction X.
2274 Supposing X changes the pressure on CL from P to P', and that the
2275 maximum pressure on CL in the current model schedule is MP', then:
2277 * if X occurs before or at the next point of maximum pressure in
2278 the model schedule and P' > MP', then:
2280 baseECC (CL, X) = model_spill_cost (CL, MP, P')
2282 The idea is that the pressure after scheduling a fixed set of
2283 instructions -- in this case, the set up to and including the
2284 next maximum pressure point -- is going to be the same regardless
2285 of the order; we simply want to keep the intermediate pressure
2286 under control. Thus X has a cost of zero unless scheduling it
2287 now would exceed MP'.
2289 If all increases in the set are by the same amount, no zero-cost
2290 instruction will ever cause the pressure to exceed MP'. However,
2291 if X is instead moved past an instruction X' with pressure in the
2292 range (MP' - (P' - P), MP'), the pressure at X' will increase
2293 beyond MP'. Since baseECC is very much a heuristic anyway,
2294 it doesn't seem worth the overhead of tracking cases like these.
2296 The cost of exceeding MP' is always based on the original maximum
2297 pressure MP. This is so that going 2 registers over the original
2298 limit has the same cost regardless of whether it comes from two
2299 separate +1 deltas or from a single +2 delta.
2301 * if X occurs after the next point of maximum pressure in the model
2302 schedule and P' > P, then:
2304 baseECC (CL, X) = model_spill_cost (CL, MP, MP' + (P' - P))
2306 That is, if we move X forward across a point of maximum pressure,
2307 and if X increases the pressure by P' - P, then we conservatively
2308 assume that scheduling X next would increase the maximum pressure
2309 by P' - P. Again, the cost of doing this is based on the original
2310 maximum pressure MP, for the same reason as above.
2312 * if P' < P, P > MP, and X occurs at or after the next point of
2313 maximum pressure, then:
2315 baseECC (CL, X) = -model_spill_cost (CL, MAX (MP, P'), P)
2317 That is, if we have already exceeded the original maximum pressure MP,
2318 and if X might reduce the maximum pressure again -- or at least push
2319 it further back, and thus allow more scheduling freedom -- it is given
2320 a negative cost to reflect the improvement.
2322 * otherwise,
2324 baseECC (CL, X) = 0
2326 In this case, X is not expected to affect the maximum pressure MP',
2327 so it has zero cost.
2329 We then create a combined value baseECC (X) that is the sum of
2330 baseECC (CL, X) for each pressure class CL.
2332 baseECC (X) could itself be used as the ECC value described above.
2333 However, this is often too conservative, in the sense that it
2334 tends to make high-priority instructions that increase pressure
2335 wait too long in cases where introducing a spill would be better.
2336 For this reason the final ECC is a priority-adjusted form of
2337 baseECC (X). Specifically, we calculate:
2339 P (X) = INSN_PRIORITY (X) - insn_delay (X) - baseECC (X)
2340 baseP = MAX { P (X) | baseECC (X) <= 0 }
2342 Then:
2344 ECC (X) = MAX (MIN (baseP - P (X), baseECC (X)), 0)
2346 Thus an instruction's effect on pressure is ignored if it has a high
2347 enough priority relative to the ones that don't increase pressure.
2348 Negative values of baseECC (X) do not increase the priority of X
2349 itself, but they do make it harder for other instructions to
2350 increase the pressure further.
2352 This pressure cost is deliberately timid. The intention has been
2353 to choose a heuristic that rarely interferes with the normal list
2354 scheduler in cases where that scheduler would produce good code.
2355 We simply want to curb some of its worst excesses. */
2357 /* Return the cost of increasing the pressure in class CL from FROM to TO.
2359 Here we use the very simplistic cost model that every register above
2360 sched_class_regs_num[CL] has a spill cost of 1. We could use other
2361 measures instead, such as one based on MEMORY_MOVE_COST. However:
2363 (1) In order for an instruction to be scheduled, the higher cost
2364 would need to be justified in a single saving of that many stalls.
2365 This is overly pessimistic, because the benefit of spilling is
2366 often to avoid a sequence of several short stalls rather than
2367 a single long one.
2369 (2) The cost is still arbitrary. Because we are not allocating
2370 registers during scheduling, we have no way of knowing for
2371 sure how many memory accesses will be required by each spill,
2372 where the spills will be placed within the block, or even
2373 which block(s) will contain the spills.
2375 So a higher cost than 1 is often too conservative in practice,
2376 forcing blocks to contain unnecessary stalls instead of spill code.
2377 The simple cost below seems to be the best compromise. It reduces
2378 the interference with the normal list scheduler, which helps make
2379 it more suitable for a default-on option. */
2381 static int
2382 model_spill_cost (int cl, int from, int to)
2384 from = MAX (from, sched_class_regs_num[cl]);
2385 return MAX (to, from) - from;
2388 /* Return baseECC (ira_pressure_classes[PCI], POINT), given that
2389 P = curr_reg_pressure[ira_pressure_classes[PCI]] and that
2390 P' = P + DELTA. */
2392 static int
2393 model_excess_group_cost (struct model_pressure_group *group,
2394 int point, int pci, int delta)
2396 int pressure, cl;
2398 cl = ira_pressure_classes[pci];
2399 if (delta < 0 && point >= group->limits[pci].point)
2401 pressure = MAX (group->limits[pci].orig_pressure,
2402 curr_reg_pressure[cl] + delta);
2403 return -model_spill_cost (cl, pressure, curr_reg_pressure[cl]);
2406 if (delta > 0)
2408 if (point > group->limits[pci].point)
2409 pressure = group->limits[pci].pressure + delta;
2410 else
2411 pressure = curr_reg_pressure[cl] + delta;
2413 if (pressure > group->limits[pci].pressure)
2414 return model_spill_cost (cl, group->limits[pci].orig_pressure,
2415 pressure);
2418 return 0;
2421 /* Return baseECC (MODEL_INSN (INSN)). Dump the costs to sched_dump
2422 if PRINT_P. */
2424 static int
2425 model_excess_cost (rtx_insn *insn, bool print_p)
2427 int point, pci, cl, cost, this_cost, delta;
2428 struct reg_pressure_data *insn_reg_pressure;
2429 int insn_death[N_REG_CLASSES];
2431 calculate_reg_deaths (insn, insn_death);
2432 point = model_index (insn);
2433 insn_reg_pressure = INSN_REG_PRESSURE (insn);
2434 cost = 0;
2436 if (print_p)
2437 fprintf (sched_dump, ";;\t\t| %3d %4d | %4d %+3d |", point,
2438 INSN_UID (insn), INSN_PRIORITY (insn), insn_delay (insn));
2440 /* Sum up the individual costs for each register class. */
2441 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2443 cl = ira_pressure_classes[pci];
2444 delta = insn_reg_pressure[pci].set_increase - insn_death[cl];
2445 this_cost = model_excess_group_cost (&model_before_pressure,
2446 point, pci, delta);
2447 cost += this_cost;
2448 if (print_p)
2449 fprintf (sched_dump, " %s:[%d base cost %d]",
2450 reg_class_names[cl], delta, this_cost);
2453 if (print_p)
2454 fprintf (sched_dump, "\n");
2456 return cost;
2459 /* Dump the next points of maximum pressure for GROUP. */
2461 static void
2462 model_dump_pressure_points (struct model_pressure_group *group)
2464 int pci, cl;
2466 fprintf (sched_dump, ";;\t\t| pressure points");
2467 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2469 cl = ira_pressure_classes[pci];
2470 fprintf (sched_dump, " %s:[%d->%d at ", reg_class_names[cl],
2471 curr_reg_pressure[cl], group->limits[pci].pressure);
2472 if (group->limits[pci].point < model_num_insns)
2473 fprintf (sched_dump, "%d:%d]", group->limits[pci].point,
2474 INSN_UID (MODEL_INSN (group->limits[pci].point)));
2475 else
2476 fprintf (sched_dump, "end]");
2478 fprintf (sched_dump, "\n");
2481 /* Set INSN_REG_PRESSURE_EXCESS_COST_CHANGE for INSNS[0...COUNT-1]. */
2483 static void
2484 model_set_excess_costs (rtx_insn **insns, int count)
2486 int i, cost, priority_base, priority;
2487 bool print_p;
2489 /* Record the baseECC value for each instruction in the model schedule,
2490 except that negative costs are converted to zero ones now rather than
2491 later. Do not assign a cost to debug instructions, since they must
2492 not change code-generation decisions. Experiments suggest we also
2493 get better results by not assigning a cost to instructions from
2494 a different block.
2496 Set PRIORITY_BASE to baseP in the block comment above. This is the
2497 maximum priority of the "cheap" instructions, which should always
2498 include the next model instruction. */
2499 priority_base = 0;
2500 print_p = false;
2501 for (i = 0; i < count; i++)
2502 if (INSN_MODEL_INDEX (insns[i]))
2504 if (sched_verbose >= 6 && !print_p)
2506 fprintf (sched_dump, MODEL_BAR);
2507 fprintf (sched_dump, ";;\t\t| Pressure costs for ready queue\n");
2508 model_dump_pressure_points (&model_before_pressure);
2509 fprintf (sched_dump, MODEL_BAR);
2510 print_p = true;
2512 cost = model_excess_cost (insns[i], print_p);
2513 if (cost <= 0)
2515 priority = INSN_PRIORITY (insns[i]) - insn_delay (insns[i]) - cost;
2516 priority_base = MAX (priority_base, priority);
2517 cost = 0;
2519 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i]) = cost;
2521 if (print_p)
2522 fprintf (sched_dump, MODEL_BAR);
2524 /* Use MAX (baseECC, 0) and baseP to calculcate ECC for each
2525 instruction. */
2526 for (i = 0; i < count; i++)
2528 cost = INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i]);
2529 priority = INSN_PRIORITY (insns[i]) - insn_delay (insns[i]);
2530 if (cost > 0 && priority > priority_base)
2532 cost += priority_base - priority;
2533 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i]) = MAX (cost, 0);
2539 /* Enum of rank_for_schedule heuristic decisions. */
2540 enum rfs_decision {
2541 RFS_LIVE_RANGE_SHRINK1, RFS_LIVE_RANGE_SHRINK2,
2542 RFS_SCHED_GROUP, RFS_PRESSURE_DELAY, RFS_PRESSURE_TICK,
2543 RFS_FEEDS_BACKTRACK_INSN, RFS_PRIORITY, RFS_SPECULATION,
2544 RFS_SCHED_RANK, RFS_LAST_INSN, RFS_PRESSURE_INDEX,
2545 RFS_DEP_COUNT, RFS_TIE, RFS_FUSION, RFS_N };
2547 /* Corresponding strings for print outs. */
2548 static const char *rfs_str[RFS_N] = {
2549 "RFS_LIVE_RANGE_SHRINK1", "RFS_LIVE_RANGE_SHRINK2",
2550 "RFS_SCHED_GROUP", "RFS_PRESSURE_DELAY", "RFS_PRESSURE_TICK",
2551 "RFS_FEEDS_BACKTRACK_INSN", "RFS_PRIORITY", "RFS_SPECULATION",
2552 "RFS_SCHED_RANK", "RFS_LAST_INSN", "RFS_PRESSURE_INDEX",
2553 "RFS_DEP_COUNT", "RFS_TIE", "RFS_FUSION" };
2555 /* Statistical breakdown of rank_for_schedule decisions. */
2556 struct rank_for_schedule_stats_t { unsigned stats[RFS_N]; };
2557 static rank_for_schedule_stats_t rank_for_schedule_stats;
2559 /* Return the result of comparing insns TMP and TMP2 and update
2560 Rank_For_Schedule statistics. */
2561 static int
2562 rfs_result (enum rfs_decision decision, int result, rtx tmp, rtx tmp2)
2564 ++rank_for_schedule_stats.stats[decision];
2565 if (result < 0)
2566 INSN_LAST_RFS_WIN (tmp) = decision;
2567 else if (result > 0)
2568 INSN_LAST_RFS_WIN (tmp2) = decision;
2569 else
2570 gcc_unreachable ();
2571 return result;
2574 /* Sorting predicate to move DEBUG_INSNs to the top of ready list, while
2575 keeping normal insns in original order. */
2577 static int
2578 rank_for_schedule_debug (const void *x, const void *y)
2580 rtx_insn *tmp = *(rtx_insn * const *) y;
2581 rtx_insn *tmp2 = *(rtx_insn * const *) x;
2583 /* Schedule debug insns as early as possible. */
2584 if (DEBUG_INSN_P (tmp) && !DEBUG_INSN_P (tmp2))
2585 return -1;
2586 else if (!DEBUG_INSN_P (tmp) && DEBUG_INSN_P (tmp2))
2587 return 1;
2588 else if (DEBUG_INSN_P (tmp) && DEBUG_INSN_P (tmp2))
2589 return INSN_LUID (tmp) - INSN_LUID (tmp2);
2590 else
2591 return INSN_RFS_DEBUG_ORIG_ORDER (tmp2) - INSN_RFS_DEBUG_ORIG_ORDER (tmp);
2594 /* Returns a positive value if x is preferred; returns a negative value if
2595 y is preferred. Should never return 0, since that will make the sort
2596 unstable. */
2598 static int
2599 rank_for_schedule (const void *x, const void *y)
2601 rtx_insn *tmp = *(rtx_insn * const *) y;
2602 rtx_insn *tmp2 = *(rtx_insn * const *) x;
2603 int tmp_class, tmp2_class;
2604 int val, priority_val, info_val, diff;
2606 if (live_range_shrinkage_p)
2608 /* Don't use SCHED_PRESSURE_MODEL -- it results in much worse
2609 code. */
2610 gcc_assert (sched_pressure == SCHED_PRESSURE_WEIGHTED);
2611 if ((INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp) < 0
2612 || INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2) < 0)
2613 && (diff = (INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp)
2614 - INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2))) != 0)
2615 return rfs_result (RFS_LIVE_RANGE_SHRINK1, diff, tmp, tmp2);
2616 /* Sort by INSN_LUID (original insn order), so that we make the
2617 sort stable. This minimizes instruction movement, thus
2618 minimizing sched's effect on debugging and cross-jumping. */
2619 return rfs_result (RFS_LIVE_RANGE_SHRINK2,
2620 INSN_LUID (tmp) - INSN_LUID (tmp2), tmp, tmp2);
2623 /* The insn in a schedule group should be issued the first. */
2624 if (flag_sched_group_heuristic &&
2625 SCHED_GROUP_P (tmp) != SCHED_GROUP_P (tmp2))
2626 return rfs_result (RFS_SCHED_GROUP, SCHED_GROUP_P (tmp2) ? 1 : -1,
2627 tmp, tmp2);
2629 /* Make sure that priority of TMP and TMP2 are initialized. */
2630 gcc_assert (INSN_PRIORITY_KNOWN (tmp) && INSN_PRIORITY_KNOWN (tmp2));
2632 if (sched_fusion)
2634 /* The instruction that has the same fusion priority as the last
2635 instruction is the instruction we picked next. If that is not
2636 the case, we sort ready list firstly by fusion priority, then
2637 by priority, and at last by INSN_LUID. */
2638 int a = INSN_FUSION_PRIORITY (tmp);
2639 int b = INSN_FUSION_PRIORITY (tmp2);
2640 int last = -1;
2642 if (last_nondebug_scheduled_insn
2643 && !NOTE_P (last_nondebug_scheduled_insn)
2644 && BLOCK_FOR_INSN (tmp)
2645 == BLOCK_FOR_INSN (last_nondebug_scheduled_insn))
2646 last = INSN_FUSION_PRIORITY (last_nondebug_scheduled_insn);
2648 if (a != last && b != last)
2650 if (a == b)
2652 a = INSN_PRIORITY (tmp);
2653 b = INSN_PRIORITY (tmp2);
2655 if (a != b)
2656 return rfs_result (RFS_FUSION, b - a, tmp, tmp2);
2657 else
2658 return rfs_result (RFS_FUSION,
2659 INSN_LUID (tmp) - INSN_LUID (tmp2), tmp, tmp2);
2661 else if (a == b)
2663 gcc_assert (last_nondebug_scheduled_insn
2664 && !NOTE_P (last_nondebug_scheduled_insn));
2665 last = INSN_PRIORITY (last_nondebug_scheduled_insn);
2667 a = abs (INSN_PRIORITY (tmp) - last);
2668 b = abs (INSN_PRIORITY (tmp2) - last);
2669 if (a != b)
2670 return rfs_result (RFS_FUSION, a - b, tmp, tmp2);
2671 else
2672 return rfs_result (RFS_FUSION,
2673 INSN_LUID (tmp) - INSN_LUID (tmp2), tmp, tmp2);
2675 else if (a == last)
2676 return rfs_result (RFS_FUSION, -1, tmp, tmp2);
2677 else
2678 return rfs_result (RFS_FUSION, 1, tmp, tmp2);
2681 if (sched_pressure != SCHED_PRESSURE_NONE)
2683 /* Prefer insn whose scheduling results in the smallest register
2684 pressure excess. */
2685 if ((diff = (INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp)
2686 + insn_delay (tmp)
2687 - INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2)
2688 - insn_delay (tmp2))))
2689 return rfs_result (RFS_PRESSURE_DELAY, diff, tmp, tmp2);
2692 if (sched_pressure != SCHED_PRESSURE_NONE
2693 && (INSN_TICK (tmp2) > clock_var || INSN_TICK (tmp) > clock_var)
2694 && INSN_TICK (tmp2) != INSN_TICK (tmp))
2696 diff = INSN_TICK (tmp) - INSN_TICK (tmp2);
2697 return rfs_result (RFS_PRESSURE_TICK, diff, tmp, tmp2);
2700 /* If we are doing backtracking in this schedule, prefer insns that
2701 have forward dependencies with negative cost against an insn that
2702 was already scheduled. */
2703 if (current_sched_info->flags & DO_BACKTRACKING)
2705 priority_val = FEEDS_BACKTRACK_INSN (tmp2) - FEEDS_BACKTRACK_INSN (tmp);
2706 if (priority_val)
2707 return rfs_result (RFS_FEEDS_BACKTRACK_INSN, priority_val, tmp, tmp2);
2710 /* Prefer insn with higher priority. */
2711 priority_val = INSN_PRIORITY (tmp2) - INSN_PRIORITY (tmp);
2713 if (flag_sched_critical_path_heuristic && priority_val)
2714 return rfs_result (RFS_PRIORITY, priority_val, tmp, tmp2);
2716 if (PARAM_VALUE (PARAM_SCHED_AUTOPREF_QUEUE_DEPTH) >= 0)
2718 int autopref = autopref_rank_for_schedule (tmp, tmp2);
2719 if (autopref != 0)
2720 return autopref;
2723 /* Prefer speculative insn with greater dependencies weakness. */
2724 if (flag_sched_spec_insn_heuristic && spec_info)
2726 ds_t ds1, ds2;
2727 dw_t dw1, dw2;
2728 int dw;
2730 ds1 = TODO_SPEC (tmp) & SPECULATIVE;
2731 if (ds1)
2732 dw1 = ds_weak (ds1);
2733 else
2734 dw1 = NO_DEP_WEAK;
2736 ds2 = TODO_SPEC (tmp2) & SPECULATIVE;
2737 if (ds2)
2738 dw2 = ds_weak (ds2);
2739 else
2740 dw2 = NO_DEP_WEAK;
2742 dw = dw2 - dw1;
2743 if (dw > (NO_DEP_WEAK / 8) || dw < -(NO_DEP_WEAK / 8))
2744 return rfs_result (RFS_SPECULATION, dw, tmp, tmp2);
2747 info_val = (*current_sched_info->rank) (tmp, tmp2);
2748 if (flag_sched_rank_heuristic && info_val)
2749 return rfs_result (RFS_SCHED_RANK, info_val, tmp, tmp2);
2751 /* Compare insns based on their relation to the last scheduled
2752 non-debug insn. */
2753 if (flag_sched_last_insn_heuristic && last_nondebug_scheduled_insn)
2755 dep_t dep1;
2756 dep_t dep2;
2757 rtx_insn *last = last_nondebug_scheduled_insn;
2759 /* Classify the instructions into three classes:
2760 1) Data dependent on last schedule insn.
2761 2) Anti/Output dependent on last scheduled insn.
2762 3) Independent of last scheduled insn, or has latency of one.
2763 Choose the insn from the highest numbered class if different. */
2764 dep1 = sd_find_dep_between (last, tmp, true);
2766 if (dep1 == NULL || dep_cost (dep1) == 1)
2767 tmp_class = 3;
2768 else if (/* Data dependence. */
2769 DEP_TYPE (dep1) == REG_DEP_TRUE)
2770 tmp_class = 1;
2771 else
2772 tmp_class = 2;
2774 dep2 = sd_find_dep_between (last, tmp2, true);
2776 if (dep2 == NULL || dep_cost (dep2) == 1)
2777 tmp2_class = 3;
2778 else if (/* Data dependence. */
2779 DEP_TYPE (dep2) == REG_DEP_TRUE)
2780 tmp2_class = 1;
2781 else
2782 tmp2_class = 2;
2784 if ((val = tmp2_class - tmp_class))
2785 return rfs_result (RFS_LAST_INSN, val, tmp, tmp2);
2788 /* Prefer instructions that occur earlier in the model schedule. */
2789 if (sched_pressure == SCHED_PRESSURE_MODEL)
2791 diff = model_index (tmp) - model_index (tmp2);
2792 if (diff != 0)
2793 return rfs_result (RFS_PRESSURE_INDEX, diff, tmp, tmp2);
2796 /* Prefer the insn which has more later insns that depend on it.
2797 This gives the scheduler more freedom when scheduling later
2798 instructions at the expense of added register pressure. */
2800 val = (dep_list_size (tmp2, SD_LIST_FORW)
2801 - dep_list_size (tmp, SD_LIST_FORW));
2803 if (flag_sched_dep_count_heuristic && val != 0)
2804 return rfs_result (RFS_DEP_COUNT, val, tmp, tmp2);
2806 /* If insns are equally good, sort by INSN_LUID (original insn order),
2807 so that we make the sort stable. This minimizes instruction movement,
2808 thus minimizing sched's effect on debugging and cross-jumping. */
2809 return rfs_result (RFS_TIE, INSN_LUID (tmp) - INSN_LUID (tmp2), tmp, tmp2);
2812 /* Resort the array A in which only element at index N may be out of order. */
2814 HAIFA_INLINE static void
2815 swap_sort (rtx_insn **a, int n)
2817 rtx_insn *insn = a[n - 1];
2818 int i = n - 2;
2820 while (i >= 0 && rank_for_schedule (a + i, &insn) >= 0)
2822 a[i + 1] = a[i];
2823 i -= 1;
2825 a[i + 1] = insn;
2828 /* Add INSN to the insn queue so that it can be executed at least
2829 N_CYCLES after the currently executing insn. Preserve insns
2830 chain for debugging purposes. REASON will be printed in debugging
2831 output. */
2833 HAIFA_INLINE static void
2834 queue_insn (rtx_insn *insn, int n_cycles, const char *reason)
2836 int next_q = NEXT_Q_AFTER (q_ptr, n_cycles);
2837 rtx_insn_list *link = alloc_INSN_LIST (insn, insn_queue[next_q]);
2838 int new_tick;
2840 gcc_assert (n_cycles <= max_insn_queue_index);
2841 gcc_assert (!DEBUG_INSN_P (insn));
2843 insn_queue[next_q] = link;
2844 q_size += 1;
2846 if (sched_verbose >= 2)
2848 fprintf (sched_dump, ";;\t\tReady-->Q: insn %s: ",
2849 (*current_sched_info->print_insn) (insn, 0));
2851 fprintf (sched_dump, "queued for %d cycles (%s).\n", n_cycles, reason);
2854 QUEUE_INDEX (insn) = next_q;
2856 if (current_sched_info->flags & DO_BACKTRACKING)
2858 new_tick = clock_var + n_cycles;
2859 if (INSN_TICK (insn) == INVALID_TICK || INSN_TICK (insn) < new_tick)
2860 INSN_TICK (insn) = new_tick;
2862 if (INSN_EXACT_TICK (insn) != INVALID_TICK
2863 && INSN_EXACT_TICK (insn) < clock_var + n_cycles)
2865 must_backtrack = true;
2866 if (sched_verbose >= 2)
2867 fprintf (sched_dump, ";;\t\tcausing a backtrack.\n");
2872 /* Remove INSN from queue. */
2873 static void
2874 queue_remove (rtx_insn *insn)
2876 gcc_assert (QUEUE_INDEX (insn) >= 0);
2877 remove_free_INSN_LIST_elem (insn, &insn_queue[QUEUE_INDEX (insn)]);
2878 q_size--;
2879 QUEUE_INDEX (insn) = QUEUE_NOWHERE;
2882 /* Return a pointer to the bottom of the ready list, i.e. the insn
2883 with the lowest priority. */
2885 rtx_insn **
2886 ready_lastpos (struct ready_list *ready)
2888 gcc_assert (ready->n_ready >= 1);
2889 return ready->vec + ready->first - ready->n_ready + 1;
2892 /* Add an element INSN to the ready list so that it ends up with the
2893 lowest/highest priority depending on FIRST_P. */
2895 HAIFA_INLINE static void
2896 ready_add (struct ready_list *ready, rtx_insn *insn, bool first_p)
2898 if (!first_p)
2900 if (ready->first == ready->n_ready)
2902 memmove (ready->vec + ready->veclen - ready->n_ready,
2903 ready_lastpos (ready),
2904 ready->n_ready * sizeof (rtx));
2905 ready->first = ready->veclen - 1;
2907 ready->vec[ready->first - ready->n_ready] = insn;
2909 else
2911 if (ready->first == ready->veclen - 1)
2913 if (ready->n_ready)
2914 /* ready_lastpos() fails when called with (ready->n_ready == 0). */
2915 memmove (ready->vec + ready->veclen - ready->n_ready - 1,
2916 ready_lastpos (ready),
2917 ready->n_ready * sizeof (rtx));
2918 ready->first = ready->veclen - 2;
2920 ready->vec[++(ready->first)] = insn;
2923 ready->n_ready++;
2924 if (DEBUG_INSN_P (insn))
2925 ready->n_debug++;
2927 gcc_assert (QUEUE_INDEX (insn) != QUEUE_READY);
2928 QUEUE_INDEX (insn) = QUEUE_READY;
2930 if (INSN_EXACT_TICK (insn) != INVALID_TICK
2931 && INSN_EXACT_TICK (insn) < clock_var)
2933 must_backtrack = true;
2937 /* Remove the element with the highest priority from the ready list and
2938 return it. */
2940 HAIFA_INLINE static rtx_insn *
2941 ready_remove_first (struct ready_list *ready)
2943 rtx_insn *t;
2945 gcc_assert (ready->n_ready);
2946 t = ready->vec[ready->first--];
2947 ready->n_ready--;
2948 if (DEBUG_INSN_P (t))
2949 ready->n_debug--;
2950 /* If the queue becomes empty, reset it. */
2951 if (ready->n_ready == 0)
2952 ready->first = ready->veclen - 1;
2954 gcc_assert (QUEUE_INDEX (t) == QUEUE_READY);
2955 QUEUE_INDEX (t) = QUEUE_NOWHERE;
2957 return t;
2960 /* The following code implements multi-pass scheduling for the first
2961 cycle. In other words, we will try to choose ready insn which
2962 permits to start maximum number of insns on the same cycle. */
2964 /* Return a pointer to the element INDEX from the ready. INDEX for
2965 insn with the highest priority is 0, and the lowest priority has
2966 N_READY - 1. */
2968 rtx_insn *
2969 ready_element (struct ready_list *ready, int index)
2971 gcc_assert (ready->n_ready && index < ready->n_ready);
2973 return ready->vec[ready->first - index];
2976 /* Remove the element INDEX from the ready list and return it. INDEX
2977 for insn with the highest priority is 0, and the lowest priority
2978 has N_READY - 1. */
2980 HAIFA_INLINE static rtx_insn *
2981 ready_remove (struct ready_list *ready, int index)
2983 rtx_insn *t;
2984 int i;
2986 if (index == 0)
2987 return ready_remove_first (ready);
2988 gcc_assert (ready->n_ready && index < ready->n_ready);
2989 t = ready->vec[ready->first - index];
2990 ready->n_ready--;
2991 if (DEBUG_INSN_P (t))
2992 ready->n_debug--;
2993 for (i = index; i < ready->n_ready; i++)
2994 ready->vec[ready->first - i] = ready->vec[ready->first - i - 1];
2995 QUEUE_INDEX (t) = QUEUE_NOWHERE;
2996 return t;
2999 /* Remove INSN from the ready list. */
3000 static void
3001 ready_remove_insn (rtx_insn *insn)
3003 int i;
3005 for (i = 0; i < readyp->n_ready; i++)
3006 if (ready_element (readyp, i) == insn)
3008 ready_remove (readyp, i);
3009 return;
3011 gcc_unreachable ();
3014 /* Calculate difference of two statistics set WAS and NOW.
3015 Result returned in WAS. */
3016 static void
3017 rank_for_schedule_stats_diff (rank_for_schedule_stats_t *was,
3018 const rank_for_schedule_stats_t *now)
3020 for (int i = 0; i < RFS_N; ++i)
3021 was->stats[i] = now->stats[i] - was->stats[i];
3024 /* Print rank_for_schedule statistics. */
3025 static void
3026 print_rank_for_schedule_stats (const char *prefix,
3027 const rank_for_schedule_stats_t *stats,
3028 struct ready_list *ready)
3030 for (int i = 0; i < RFS_N; ++i)
3031 if (stats->stats[i])
3033 fprintf (sched_dump, "%s%20s: %u", prefix, rfs_str[i], stats->stats[i]);
3035 if (ready != NULL)
3036 /* Print out insns that won due to RFS_<I>. */
3038 rtx_insn **p = ready_lastpos (ready);
3040 fprintf (sched_dump, ":");
3041 /* Start with 1 since least-priority insn didn't have any wins. */
3042 for (int j = 1; j < ready->n_ready; ++j)
3043 if (INSN_LAST_RFS_WIN (p[j]) == i)
3044 fprintf (sched_dump, " %s",
3045 (*current_sched_info->print_insn) (p[j], 0));
3047 fprintf (sched_dump, "\n");
3051 /* Separate DEBUG_INSNS from normal insns. DEBUG_INSNs go to the end
3052 of array. */
3053 static void
3054 ready_sort_debug (struct ready_list *ready)
3056 int i;
3057 rtx_insn **first = ready_lastpos (ready);
3059 for (i = 0; i < ready->n_ready; ++i)
3060 if (!DEBUG_INSN_P (first[i]))
3061 INSN_RFS_DEBUG_ORIG_ORDER (first[i]) = i;
3063 qsort (first, ready->n_ready, sizeof (rtx), rank_for_schedule_debug);
3066 /* Sort non-debug insns in the ready list READY by ascending priority.
3067 Assumes that all debug insns are separated from the real insns. */
3068 static void
3069 ready_sort_real (struct ready_list *ready)
3071 int i;
3072 rtx_insn **first = ready_lastpos (ready);
3073 int n_ready_real = ready->n_ready - ready->n_debug;
3075 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
3076 for (i = 0; i < n_ready_real; ++i)
3077 setup_insn_reg_pressure_info (first[i]);
3078 else if (sched_pressure == SCHED_PRESSURE_MODEL
3079 && model_curr_point < model_num_insns)
3080 model_set_excess_costs (first, n_ready_real);
3082 rank_for_schedule_stats_t stats1;
3083 if (sched_verbose >= 4)
3084 stats1 = rank_for_schedule_stats;
3086 if (n_ready_real == 2)
3087 swap_sort (first, n_ready_real);
3088 else if (n_ready_real > 2)
3089 qsort (first, n_ready_real, sizeof (rtx), rank_for_schedule);
3091 if (sched_verbose >= 4)
3093 rank_for_schedule_stats_diff (&stats1, &rank_for_schedule_stats);
3094 print_rank_for_schedule_stats (";;\t\t", &stats1, ready);
3098 /* Sort the ready list READY by ascending priority. */
3099 static void
3100 ready_sort (struct ready_list *ready)
3102 if (ready->n_debug > 0)
3103 ready_sort_debug (ready);
3104 else
3105 ready_sort_real (ready);
3108 /* PREV is an insn that is ready to execute. Adjust its priority if that
3109 will help shorten or lengthen register lifetimes as appropriate. Also
3110 provide a hook for the target to tweak itself. */
3112 HAIFA_INLINE static void
3113 adjust_priority (rtx_insn *prev)
3115 /* ??? There used to be code here to try and estimate how an insn
3116 affected register lifetimes, but it did it by looking at REG_DEAD
3117 notes, which we removed in schedule_region. Nor did it try to
3118 take into account register pressure or anything useful like that.
3120 Revisit when we have a machine model to work with and not before. */
3122 if (targetm.sched.adjust_priority)
3123 INSN_PRIORITY (prev) =
3124 targetm.sched.adjust_priority (prev, INSN_PRIORITY (prev));
3127 /* Advance DFA state STATE on one cycle. */
3128 void
3129 advance_state (state_t state)
3131 if (targetm.sched.dfa_pre_advance_cycle)
3132 targetm.sched.dfa_pre_advance_cycle ();
3134 if (targetm.sched.dfa_pre_cycle_insn)
3135 state_transition (state,
3136 targetm.sched.dfa_pre_cycle_insn ());
3138 state_transition (state, NULL);
3140 if (targetm.sched.dfa_post_cycle_insn)
3141 state_transition (state,
3142 targetm.sched.dfa_post_cycle_insn ());
3144 if (targetm.sched.dfa_post_advance_cycle)
3145 targetm.sched.dfa_post_advance_cycle ();
3148 /* Advance time on one cycle. */
3149 HAIFA_INLINE static void
3150 advance_one_cycle (void)
3152 advance_state (curr_state);
3153 if (sched_verbose >= 4)
3154 fprintf (sched_dump, ";;\tAdvance the current state.\n");
3157 /* Update register pressure after scheduling INSN. */
3158 static void
3159 update_register_pressure (rtx_insn *insn)
3161 struct reg_use_data *use;
3162 struct reg_set_data *set;
3164 gcc_checking_assert (!DEBUG_INSN_P (insn));
3166 for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
3167 if (dying_use_p (use))
3168 mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure,
3169 use->regno, false);
3170 for (set = INSN_REG_SET_LIST (insn); set != NULL; set = set->next_insn_set)
3171 mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure,
3172 set->regno, true);
3175 /* Set up or update (if UPDATE_P) max register pressure (see its
3176 meaning in sched-int.h::_haifa_insn_data) for all current BB insns
3177 after insn AFTER. */
3178 static void
3179 setup_insn_max_reg_pressure (rtx_insn *after, bool update_p)
3181 int i, p;
3182 bool eq_p;
3183 rtx_insn *insn;
3184 static int max_reg_pressure[N_REG_CLASSES];
3186 save_reg_pressure ();
3187 for (i = 0; i < ira_pressure_classes_num; i++)
3188 max_reg_pressure[ira_pressure_classes[i]]
3189 = curr_reg_pressure[ira_pressure_classes[i]];
3190 for (insn = NEXT_INSN (after);
3191 insn != NULL_RTX && ! BARRIER_P (insn)
3192 && BLOCK_FOR_INSN (insn) == BLOCK_FOR_INSN (after);
3193 insn = NEXT_INSN (insn))
3194 if (NONDEBUG_INSN_P (insn))
3196 eq_p = true;
3197 for (i = 0; i < ira_pressure_classes_num; i++)
3199 p = max_reg_pressure[ira_pressure_classes[i]];
3200 if (INSN_MAX_REG_PRESSURE (insn)[i] != p)
3202 eq_p = false;
3203 INSN_MAX_REG_PRESSURE (insn)[i]
3204 = max_reg_pressure[ira_pressure_classes[i]];
3207 if (update_p && eq_p)
3208 break;
3209 update_register_pressure (insn);
3210 for (i = 0; i < ira_pressure_classes_num; i++)
3211 if (max_reg_pressure[ira_pressure_classes[i]]
3212 < curr_reg_pressure[ira_pressure_classes[i]])
3213 max_reg_pressure[ira_pressure_classes[i]]
3214 = curr_reg_pressure[ira_pressure_classes[i]];
3216 restore_reg_pressure ();
3219 /* Update the current register pressure after scheduling INSN. Update
3220 also max register pressure for unscheduled insns of the current
3221 BB. */
3222 static void
3223 update_reg_and_insn_max_reg_pressure (rtx_insn *insn)
3225 int i;
3226 int before[N_REG_CLASSES];
3228 for (i = 0; i < ira_pressure_classes_num; i++)
3229 before[i] = curr_reg_pressure[ira_pressure_classes[i]];
3230 update_register_pressure (insn);
3231 for (i = 0; i < ira_pressure_classes_num; i++)
3232 if (curr_reg_pressure[ira_pressure_classes[i]] != before[i])
3233 break;
3234 if (i < ira_pressure_classes_num)
3235 setup_insn_max_reg_pressure (insn, true);
3238 /* Set up register pressure at the beginning of basic block BB whose
3239 insns starting after insn AFTER. Set up also max register pressure
3240 for all insns of the basic block. */
3241 void
3242 sched_setup_bb_reg_pressure_info (basic_block bb, rtx_insn *after)
3244 gcc_assert (sched_pressure == SCHED_PRESSURE_WEIGHTED);
3245 initiate_bb_reg_pressure_info (bb);
3246 setup_insn_max_reg_pressure (after, false);
3249 /* If doing predication while scheduling, verify whether INSN, which
3250 has just been scheduled, clobbers the conditions of any
3251 instructions that must be predicated in order to break their
3252 dependencies. If so, remove them from the queues so that they will
3253 only be scheduled once their control dependency is resolved. */
3255 static void
3256 check_clobbered_conditions (rtx_insn *insn)
3258 HARD_REG_SET t;
3259 int i;
3261 if ((current_sched_info->flags & DO_PREDICATION) == 0)
3262 return;
3264 find_all_hard_reg_sets (insn, &t, true);
3266 restart:
3267 for (i = 0; i < ready.n_ready; i++)
3269 rtx_insn *x = ready_element (&ready, i);
3270 if (TODO_SPEC (x) == DEP_CONTROL && cond_clobbered_p (x, t))
3272 ready_remove_insn (x);
3273 goto restart;
3276 for (i = 0; i <= max_insn_queue_index; i++)
3278 rtx_insn_list *link;
3279 int q = NEXT_Q_AFTER (q_ptr, i);
3281 restart_queue:
3282 for (link = insn_queue[q]; link; link = link->next ())
3284 rtx_insn *x = link->insn ();
3285 if (TODO_SPEC (x) == DEP_CONTROL && cond_clobbered_p (x, t))
3287 queue_remove (x);
3288 goto restart_queue;
3294 /* Return (in order):
3296 - positive if INSN adversely affects the pressure on one
3297 register class
3299 - negative if INSN reduces the pressure on one register class
3301 - 0 if INSN doesn't affect the pressure on any register class. */
3303 static int
3304 model_classify_pressure (struct model_insn_info *insn)
3306 struct reg_pressure_data *reg_pressure;
3307 int death[N_REG_CLASSES];
3308 int pci, cl, sum;
3310 calculate_reg_deaths (insn->insn, death);
3311 reg_pressure = INSN_REG_PRESSURE (insn->insn);
3312 sum = 0;
3313 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3315 cl = ira_pressure_classes[pci];
3316 if (death[cl] < reg_pressure[pci].set_increase)
3317 return 1;
3318 sum += reg_pressure[pci].set_increase - death[cl];
3320 return sum;
3323 /* Return true if INSN1 should come before INSN2 in the model schedule. */
3325 static int
3326 model_order_p (struct model_insn_info *insn1, struct model_insn_info *insn2)
3328 unsigned int height1, height2;
3329 unsigned int priority1, priority2;
3331 /* Prefer instructions with a higher model priority. */
3332 if (insn1->model_priority != insn2->model_priority)
3333 return insn1->model_priority > insn2->model_priority;
3335 /* Combine the length of the longest path of satisfied true dependencies
3336 that leads to each instruction (depth) with the length of the longest
3337 path of any dependencies that leads from the instruction (alap).
3338 Prefer instructions with the greatest combined length. If the combined
3339 lengths are equal, prefer instructions with the greatest depth.
3341 The idea is that, if we have a set S of "equal" instructions that each
3342 have ALAP value X, and we pick one such instruction I, any true-dependent
3343 successors of I that have ALAP value X - 1 should be preferred over S.
3344 This encourages the schedule to be "narrow" rather than "wide".
3345 However, if I is a low-priority instruction that we decided to
3346 schedule because of its model_classify_pressure, and if there
3347 is a set of higher-priority instructions T, the aforementioned
3348 successors of I should not have the edge over T. */
3349 height1 = insn1->depth + insn1->alap;
3350 height2 = insn2->depth + insn2->alap;
3351 if (height1 != height2)
3352 return height1 > height2;
3353 if (insn1->depth != insn2->depth)
3354 return insn1->depth > insn2->depth;
3356 /* We have no real preference between INSN1 an INSN2 as far as attempts
3357 to reduce pressure go. Prefer instructions with higher priorities. */
3358 priority1 = INSN_PRIORITY (insn1->insn);
3359 priority2 = INSN_PRIORITY (insn2->insn);
3360 if (priority1 != priority2)
3361 return priority1 > priority2;
3363 /* Use the original rtl sequence as a tie-breaker. */
3364 return insn1 < insn2;
3367 /* Add INSN to the model worklist immediately after PREV. Add it to the
3368 beginning of the list if PREV is null. */
3370 static void
3371 model_add_to_worklist_at (struct model_insn_info *insn,
3372 struct model_insn_info *prev)
3374 gcc_assert (QUEUE_INDEX (insn->insn) == QUEUE_NOWHERE);
3375 QUEUE_INDEX (insn->insn) = QUEUE_READY;
3377 insn->prev = prev;
3378 if (prev)
3380 insn->next = prev->next;
3381 prev->next = insn;
3383 else
3385 insn->next = model_worklist;
3386 model_worklist = insn;
3388 if (insn->next)
3389 insn->next->prev = insn;
3392 /* Remove INSN from the model worklist. */
3394 static void
3395 model_remove_from_worklist (struct model_insn_info *insn)
3397 gcc_assert (QUEUE_INDEX (insn->insn) == QUEUE_READY);
3398 QUEUE_INDEX (insn->insn) = QUEUE_NOWHERE;
3400 if (insn->prev)
3401 insn->prev->next = insn->next;
3402 else
3403 model_worklist = insn->next;
3404 if (insn->next)
3405 insn->next->prev = insn->prev;
3408 /* Add INSN to the model worklist. Start looking for a suitable position
3409 between neighbors PREV and NEXT, testing at most MAX_SCHED_READY_INSNS
3410 insns either side. A null PREV indicates the beginning of the list and
3411 a null NEXT indicates the end. */
3413 static void
3414 model_add_to_worklist (struct model_insn_info *insn,
3415 struct model_insn_info *prev,
3416 struct model_insn_info *next)
3418 int count;
3420 count = MAX_SCHED_READY_INSNS;
3421 if (count > 0 && prev && model_order_p (insn, prev))
3424 count--;
3425 prev = prev->prev;
3427 while (count > 0 && prev && model_order_p (insn, prev));
3428 else
3429 while (count > 0 && next && model_order_p (next, insn))
3431 count--;
3432 prev = next;
3433 next = next->next;
3435 model_add_to_worklist_at (insn, prev);
3438 /* INSN may now have a higher priority (in the model_order_p sense)
3439 than before. Move it up the worklist if necessary. */
3441 static void
3442 model_promote_insn (struct model_insn_info *insn)
3444 struct model_insn_info *prev;
3445 int count;
3447 prev = insn->prev;
3448 count = MAX_SCHED_READY_INSNS;
3449 while (count > 0 && prev && model_order_p (insn, prev))
3451 count--;
3452 prev = prev->prev;
3454 if (prev != insn->prev)
3456 model_remove_from_worklist (insn);
3457 model_add_to_worklist_at (insn, prev);
3461 /* Add INSN to the end of the model schedule. */
3463 static void
3464 model_add_to_schedule (rtx_insn *insn)
3466 unsigned int point;
3468 gcc_assert (QUEUE_INDEX (insn) == QUEUE_NOWHERE);
3469 QUEUE_INDEX (insn) = QUEUE_SCHEDULED;
3471 point = model_schedule.length ();
3472 model_schedule.quick_push (insn);
3473 INSN_MODEL_INDEX (insn) = point + 1;
3476 /* Analyze the instructions that are to be scheduled, setting up
3477 MODEL_INSN_INFO (...) and model_num_insns accordingly. Add ready
3478 instructions to model_worklist. */
3480 static void
3481 model_analyze_insns (void)
3483 rtx_insn *start, *end, *iter;
3484 sd_iterator_def sd_it;
3485 dep_t dep;
3486 struct model_insn_info *insn, *con;
3488 model_num_insns = 0;
3489 start = PREV_INSN (current_sched_info->next_tail);
3490 end = current_sched_info->prev_head;
3491 for (iter = start; iter != end; iter = PREV_INSN (iter))
3492 if (NONDEBUG_INSN_P (iter))
3494 insn = MODEL_INSN_INFO (iter);
3495 insn->insn = iter;
3496 FOR_EACH_DEP (iter, SD_LIST_FORW, sd_it, dep)
3498 con = MODEL_INSN_INFO (DEP_CON (dep));
3499 if (con->insn && insn->alap < con->alap + 1)
3500 insn->alap = con->alap + 1;
3503 insn->old_queue = QUEUE_INDEX (iter);
3504 QUEUE_INDEX (iter) = QUEUE_NOWHERE;
3506 insn->unscheduled_preds = dep_list_size (iter, SD_LIST_HARD_BACK);
3507 if (insn->unscheduled_preds == 0)
3508 model_add_to_worklist (insn, NULL, model_worklist);
3510 model_num_insns++;
3514 /* The global state describes the register pressure at the start of the
3515 model schedule. Initialize GROUP accordingly. */
3517 static void
3518 model_init_pressure_group (struct model_pressure_group *group)
3520 int pci, cl;
3522 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3524 cl = ira_pressure_classes[pci];
3525 group->limits[pci].pressure = curr_reg_pressure[cl];
3526 group->limits[pci].point = 0;
3528 /* Use index model_num_insns to record the state after the last
3529 instruction in the model schedule. */
3530 group->model = XNEWVEC (struct model_pressure_data,
3531 (model_num_insns + 1) * ira_pressure_classes_num);
3534 /* Record that MODEL_REF_PRESSURE (GROUP, POINT, PCI) is PRESSURE.
3535 Update the maximum pressure for the whole schedule. */
3537 static void
3538 model_record_pressure (struct model_pressure_group *group,
3539 int point, int pci, int pressure)
3541 MODEL_REF_PRESSURE (group, point, pci) = pressure;
3542 if (group->limits[pci].pressure < pressure)
3544 group->limits[pci].pressure = pressure;
3545 group->limits[pci].point = point;
3549 /* INSN has just been added to the end of the model schedule. Record its
3550 register-pressure information. */
3552 static void
3553 model_record_pressures (struct model_insn_info *insn)
3555 struct reg_pressure_data *reg_pressure;
3556 int point, pci, cl, delta;
3557 int death[N_REG_CLASSES];
3559 point = model_index (insn->insn);
3560 if (sched_verbose >= 2)
3562 if (point == 0)
3564 fprintf (sched_dump, "\n;;\tModel schedule:\n;;\n");
3565 fprintf (sched_dump, ";;\t| idx insn | mpri hght dpth prio |\n");
3567 fprintf (sched_dump, ";;\t| %3d %4d | %4d %4d %4d %4d | %-30s ",
3568 point, INSN_UID (insn->insn), insn->model_priority,
3569 insn->depth + insn->alap, insn->depth,
3570 INSN_PRIORITY (insn->insn),
3571 str_pattern_slim (PATTERN (insn->insn)));
3573 calculate_reg_deaths (insn->insn, death);
3574 reg_pressure = INSN_REG_PRESSURE (insn->insn);
3575 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3577 cl = ira_pressure_classes[pci];
3578 delta = reg_pressure[pci].set_increase - death[cl];
3579 if (sched_verbose >= 2)
3580 fprintf (sched_dump, " %s:[%d,%+d]", reg_class_names[cl],
3581 curr_reg_pressure[cl], delta);
3582 model_record_pressure (&model_before_pressure, point, pci,
3583 curr_reg_pressure[cl]);
3585 if (sched_verbose >= 2)
3586 fprintf (sched_dump, "\n");
3589 /* All instructions have been added to the model schedule. Record the
3590 final register pressure in GROUP and set up all MODEL_MAX_PRESSUREs. */
3592 static void
3593 model_record_final_pressures (struct model_pressure_group *group)
3595 int point, pci, max_pressure, ref_pressure, cl;
3597 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3599 /* Record the final pressure for this class. */
3600 cl = ira_pressure_classes[pci];
3601 point = model_num_insns;
3602 ref_pressure = curr_reg_pressure[cl];
3603 model_record_pressure (group, point, pci, ref_pressure);
3605 /* Record the original maximum pressure. */
3606 group->limits[pci].orig_pressure = group->limits[pci].pressure;
3608 /* Update the MODEL_MAX_PRESSURE for every point of the schedule. */
3609 max_pressure = ref_pressure;
3610 MODEL_MAX_PRESSURE (group, point, pci) = max_pressure;
3611 while (point > 0)
3613 point--;
3614 ref_pressure = MODEL_REF_PRESSURE (group, point, pci);
3615 max_pressure = MAX (max_pressure, ref_pressure);
3616 MODEL_MAX_PRESSURE (group, point, pci) = max_pressure;
3621 /* Update all successors of INSN, given that INSN has just been scheduled. */
3623 static void
3624 model_add_successors_to_worklist (struct model_insn_info *insn)
3626 sd_iterator_def sd_it;
3627 struct model_insn_info *con;
3628 dep_t dep;
3630 FOR_EACH_DEP (insn->insn, SD_LIST_FORW, sd_it, dep)
3632 con = MODEL_INSN_INFO (DEP_CON (dep));
3633 /* Ignore debug instructions, and instructions from other blocks. */
3634 if (con->insn)
3636 con->unscheduled_preds--;
3638 /* Update the depth field of each true-dependent successor.
3639 Increasing the depth gives them a higher priority than
3640 before. */
3641 if (DEP_TYPE (dep) == REG_DEP_TRUE && con->depth < insn->depth + 1)
3643 con->depth = insn->depth + 1;
3644 if (QUEUE_INDEX (con->insn) == QUEUE_READY)
3645 model_promote_insn (con);
3648 /* If this is a true dependency, or if there are no remaining
3649 dependencies for CON (meaning that CON only had non-true
3650 dependencies), make sure that CON is on the worklist.
3651 We don't bother otherwise because it would tend to fill the
3652 worklist with a lot of low-priority instructions that are not
3653 yet ready to issue. */
3654 if ((con->depth > 0 || con->unscheduled_preds == 0)
3655 && QUEUE_INDEX (con->insn) == QUEUE_NOWHERE)
3656 model_add_to_worklist (con, insn, insn->next);
3661 /* Give INSN a higher priority than any current instruction, then give
3662 unscheduled predecessors of INSN a higher priority still. If any of
3663 those predecessors are not on the model worklist, do the same for its
3664 predecessors, and so on. */
3666 static void
3667 model_promote_predecessors (struct model_insn_info *insn)
3669 struct model_insn_info *pro, *first;
3670 sd_iterator_def sd_it;
3671 dep_t dep;
3673 if (sched_verbose >= 7)
3674 fprintf (sched_dump, ";;\t+--- priority of %d = %d, priority of",
3675 INSN_UID (insn->insn), model_next_priority);
3676 insn->model_priority = model_next_priority++;
3677 model_remove_from_worklist (insn);
3678 model_add_to_worklist_at (insn, NULL);
3680 first = NULL;
3681 for (;;)
3683 FOR_EACH_DEP (insn->insn, SD_LIST_HARD_BACK, sd_it, dep)
3685 pro = MODEL_INSN_INFO (DEP_PRO (dep));
3686 /* The first test is to ignore debug instructions, and instructions
3687 from other blocks. */
3688 if (pro->insn
3689 && pro->model_priority != model_next_priority
3690 && QUEUE_INDEX (pro->insn) != QUEUE_SCHEDULED)
3692 pro->model_priority = model_next_priority;
3693 if (sched_verbose >= 7)
3694 fprintf (sched_dump, " %d", INSN_UID (pro->insn));
3695 if (QUEUE_INDEX (pro->insn) == QUEUE_READY)
3697 /* PRO is already in the worklist, but it now has
3698 a higher priority than before. Move it at the
3699 appropriate place. */
3700 model_remove_from_worklist (pro);
3701 model_add_to_worklist (pro, NULL, model_worklist);
3703 else
3705 /* PRO isn't in the worklist. Recursively process
3706 its predecessors until we find one that is. */
3707 pro->next = first;
3708 first = pro;
3712 if (!first)
3713 break;
3714 insn = first;
3715 first = insn->next;
3717 if (sched_verbose >= 7)
3718 fprintf (sched_dump, " = %d\n", model_next_priority);
3719 model_next_priority++;
3722 /* Pick one instruction from model_worklist and process it. */
3724 static void
3725 model_choose_insn (void)
3727 struct model_insn_info *insn, *fallback;
3728 int count;
3730 if (sched_verbose >= 7)
3732 fprintf (sched_dump, ";;\t+--- worklist:\n");
3733 insn = model_worklist;
3734 count = MAX_SCHED_READY_INSNS;
3735 while (count > 0 && insn)
3737 fprintf (sched_dump, ";;\t+--- %d [%d, %d, %d, %d]\n",
3738 INSN_UID (insn->insn), insn->model_priority,
3739 insn->depth + insn->alap, insn->depth,
3740 INSN_PRIORITY (insn->insn));
3741 count--;
3742 insn = insn->next;
3746 /* Look for a ready instruction whose model_classify_priority is zero
3747 or negative, picking the highest-priority one. Adding such an
3748 instruction to the schedule now should do no harm, and may actually
3749 do some good.
3751 Failing that, see whether there is an instruction with the highest
3752 extant model_priority that is not yet ready, but which would reduce
3753 pressure if it became ready. This is designed to catch cases like:
3755 (set (mem (reg R1)) (reg R2))
3757 where the instruction is the last remaining use of R1 and where the
3758 value of R2 is not yet available (or vice versa). The death of R1
3759 means that this instruction already reduces pressure. It is of
3760 course possible that the computation of R2 involves other registers
3761 that are hard to kill, but such cases are rare enough for this
3762 heuristic to be a win in general.
3764 Failing that, just pick the highest-priority instruction in the
3765 worklist. */
3766 count = MAX_SCHED_READY_INSNS;
3767 insn = model_worklist;
3768 fallback = 0;
3769 for (;;)
3771 if (count == 0 || !insn)
3773 insn = fallback ? fallback : model_worklist;
3774 break;
3776 if (insn->unscheduled_preds)
3778 if (model_worklist->model_priority == insn->model_priority
3779 && !fallback
3780 && model_classify_pressure (insn) < 0)
3781 fallback = insn;
3783 else
3785 if (model_classify_pressure (insn) <= 0)
3786 break;
3788 count--;
3789 insn = insn->next;
3792 if (sched_verbose >= 7 && insn != model_worklist)
3794 if (insn->unscheduled_preds)
3795 fprintf (sched_dump, ";;\t+--- promoting insn %d, with dependencies\n",
3796 INSN_UID (insn->insn));
3797 else
3798 fprintf (sched_dump, ";;\t+--- promoting insn %d, which is ready\n",
3799 INSN_UID (insn->insn));
3801 if (insn->unscheduled_preds)
3802 /* INSN isn't yet ready to issue. Give all its predecessors the
3803 highest priority. */
3804 model_promote_predecessors (insn);
3805 else
3807 /* INSN is ready. Add it to the end of model_schedule and
3808 process its successors. */
3809 model_add_successors_to_worklist (insn);
3810 model_remove_from_worklist (insn);
3811 model_add_to_schedule (insn->insn);
3812 model_record_pressures (insn);
3813 update_register_pressure (insn->insn);
3817 /* Restore all QUEUE_INDEXs to the values that they had before
3818 model_start_schedule was called. */
3820 static void
3821 model_reset_queue_indices (void)
3823 unsigned int i;
3824 rtx_insn *insn;
3826 FOR_EACH_VEC_ELT (model_schedule, i, insn)
3827 QUEUE_INDEX (insn) = MODEL_INSN_INFO (insn)->old_queue;
3830 /* We have calculated the model schedule and spill costs. Print a summary
3831 to sched_dump. */
3833 static void
3834 model_dump_pressure_summary (void)
3836 int pci, cl;
3838 fprintf (sched_dump, ";; Pressure summary:");
3839 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3841 cl = ira_pressure_classes[pci];
3842 fprintf (sched_dump, " %s:%d", reg_class_names[cl],
3843 model_before_pressure.limits[pci].pressure);
3845 fprintf (sched_dump, "\n\n");
3848 /* Initialize the SCHED_PRESSURE_MODEL information for the current
3849 scheduling region. */
3851 static void
3852 model_start_schedule (basic_block bb)
3854 model_next_priority = 1;
3855 model_schedule.create (sched_max_luid);
3856 model_insns = XCNEWVEC (struct model_insn_info, sched_max_luid);
3858 gcc_assert (bb == BLOCK_FOR_INSN (NEXT_INSN (current_sched_info->prev_head)));
3859 initiate_reg_pressure_info (df_get_live_in (bb));
3861 model_analyze_insns ();
3862 model_init_pressure_group (&model_before_pressure);
3863 while (model_worklist)
3864 model_choose_insn ();
3865 gcc_assert (model_num_insns == (int) model_schedule.length ());
3866 if (sched_verbose >= 2)
3867 fprintf (sched_dump, "\n");
3869 model_record_final_pressures (&model_before_pressure);
3870 model_reset_queue_indices ();
3872 XDELETEVEC (model_insns);
3874 model_curr_point = 0;
3875 initiate_reg_pressure_info (df_get_live_in (bb));
3876 if (sched_verbose >= 1)
3877 model_dump_pressure_summary ();
3880 /* Free the information associated with GROUP. */
3882 static void
3883 model_finalize_pressure_group (struct model_pressure_group *group)
3885 XDELETEVEC (group->model);
3888 /* Free the information created by model_start_schedule. */
3890 static void
3891 model_end_schedule (void)
3893 model_finalize_pressure_group (&model_before_pressure);
3894 model_schedule.release ();
3897 /* Prepare reg pressure scheduling for basic block BB. */
3898 static void
3899 sched_pressure_start_bb (basic_block bb)
3901 /* Set the number of available registers for each class taking into account
3902 relative probability of current basic block versus function prologue and
3903 epilogue.
3904 * If the basic block executes much more often than the prologue/epilogue
3905 (e.g., inside a hot loop), then cost of spill in the prologue is close to
3906 nil, so the effective number of available registers is
3907 (ira_class_hard_regs_num[cl] - fixed_regs_num[cl] - 0).
3908 * If the basic block executes as often as the prologue/epilogue,
3909 then spill in the block is as costly as in the prologue, so the effective
3910 number of available registers is
3911 (ira_class_hard_regs_num[cl] - fixed_regs_num[cl]
3912 - call_saved_regs_num[cl]).
3913 Note that all-else-equal, we prefer to spill in the prologue, since that
3914 allows "extra" registers for other basic blocks of the function.
3915 * If the basic block is on the cold path of the function and executes
3916 rarely, then we should always prefer to spill in the block, rather than
3917 in the prologue/epilogue. The effective number of available register is
3918 (ira_class_hard_regs_num[cl] - fixed_regs_num[cl]
3919 - call_saved_regs_num[cl]). */
3921 int i;
3922 int entry_freq = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count.to_frequency (cfun);
3923 int bb_freq = bb->count.to_frequency (cfun);
3925 if (bb_freq == 0)
3927 if (entry_freq == 0)
3928 entry_freq = bb_freq = 1;
3930 if (bb_freq < entry_freq)
3931 bb_freq = entry_freq;
3933 for (i = 0; i < ira_pressure_classes_num; ++i)
3935 enum reg_class cl = ira_pressure_classes[i];
3936 sched_class_regs_num[cl] = ira_class_hard_regs_num[cl]
3937 - fixed_regs_num[cl];
3938 sched_class_regs_num[cl]
3939 -= (call_saved_regs_num[cl] * entry_freq) / bb_freq;
3943 if (sched_pressure == SCHED_PRESSURE_MODEL)
3944 model_start_schedule (bb);
3947 /* A structure that holds local state for the loop in schedule_block. */
3948 struct sched_block_state
3950 /* True if no real insns have been scheduled in the current cycle. */
3951 bool first_cycle_insn_p;
3952 /* True if a shadow insn has been scheduled in the current cycle, which
3953 means that no more normal insns can be issued. */
3954 bool shadows_only_p;
3955 /* True if we're winding down a modulo schedule, which means that we only
3956 issue insns with INSN_EXACT_TICK set. */
3957 bool modulo_epilogue;
3958 /* Initialized with the machine's issue rate every cycle, and updated
3959 by calls to the variable_issue hook. */
3960 int can_issue_more;
3963 /* INSN is the "currently executing insn". Launch each insn which was
3964 waiting on INSN. READY is the ready list which contains the insns
3965 that are ready to fire. CLOCK is the current cycle. The function
3966 returns necessary cycle advance after issuing the insn (it is not
3967 zero for insns in a schedule group). */
3969 static int
3970 schedule_insn (rtx_insn *insn)
3972 sd_iterator_def sd_it;
3973 dep_t dep;
3974 int i;
3975 int advance = 0;
3977 if (sched_verbose >= 1)
3979 struct reg_pressure_data *pressure_info;
3980 fprintf (sched_dump, ";;\t%3i--> %s %-40s:",
3981 clock_var, (*current_sched_info->print_insn) (insn, 1),
3982 str_pattern_slim (PATTERN (insn)));
3984 if (recog_memoized (insn) < 0)
3985 fprintf (sched_dump, "nothing");
3986 else
3987 print_reservation (sched_dump, insn);
3988 pressure_info = INSN_REG_PRESSURE (insn);
3989 if (pressure_info != NULL)
3991 fputc (':', sched_dump);
3992 for (i = 0; i < ira_pressure_classes_num; i++)
3993 fprintf (sched_dump, "%s%s%+d(%d)",
3994 scheduled_insns.length () > 1
3995 && INSN_LUID (insn)
3996 < INSN_LUID (scheduled_insns[scheduled_insns.length () - 2]) ? "@" : "",
3997 reg_class_names[ira_pressure_classes[i]],
3998 pressure_info[i].set_increase, pressure_info[i].change);
4000 if (sched_pressure == SCHED_PRESSURE_MODEL
4001 && model_curr_point < model_num_insns
4002 && model_index (insn) == model_curr_point)
4003 fprintf (sched_dump, ":model %d", model_curr_point);
4004 fputc ('\n', sched_dump);
4007 if (sched_pressure == SCHED_PRESSURE_WEIGHTED && !DEBUG_INSN_P (insn))
4008 update_reg_and_insn_max_reg_pressure (insn);
4010 /* Scheduling instruction should have all its dependencies resolved and
4011 should have been removed from the ready list. */
4012 gcc_assert (sd_lists_empty_p (insn, SD_LIST_HARD_BACK));
4014 /* Reset debug insns invalidated by moving this insn. */
4015 if (MAY_HAVE_DEBUG_BIND_INSNS && !DEBUG_INSN_P (insn))
4016 for (sd_it = sd_iterator_start (insn, SD_LIST_BACK);
4017 sd_iterator_cond (&sd_it, &dep);)
4019 rtx_insn *dbg = DEP_PRO (dep);
4020 struct reg_use_data *use, *next;
4022 if (DEP_STATUS (dep) & DEP_CANCELLED)
4024 sd_iterator_next (&sd_it);
4025 continue;
4028 gcc_assert (DEBUG_BIND_INSN_P (dbg));
4030 if (sched_verbose >= 6)
4031 fprintf (sched_dump, ";;\t\tresetting: debug insn %d\n",
4032 INSN_UID (dbg));
4034 /* ??? Rather than resetting the debug insn, we might be able
4035 to emit a debug temp before the just-scheduled insn, but
4036 this would involve checking that the expression at the
4037 point of the debug insn is equivalent to the expression
4038 before the just-scheduled insn. They might not be: the
4039 expression in the debug insn may depend on other insns not
4040 yet scheduled that set MEMs, REGs or even other debug
4041 insns. It's not clear that attempting to preserve debug
4042 information in these cases is worth the effort, given how
4043 uncommon these resets are and the likelihood that the debug
4044 temps introduced won't survive the schedule change. */
4045 INSN_VAR_LOCATION_LOC (dbg) = gen_rtx_UNKNOWN_VAR_LOC ();
4046 df_insn_rescan (dbg);
4048 /* Unknown location doesn't use any registers. */
4049 for (use = INSN_REG_USE_LIST (dbg); use != NULL; use = next)
4051 struct reg_use_data *prev = use;
4053 /* Remove use from the cyclic next_regno_use chain first. */
4054 while (prev->next_regno_use != use)
4055 prev = prev->next_regno_use;
4056 prev->next_regno_use = use->next_regno_use;
4057 next = use->next_insn_use;
4058 free (use);
4060 INSN_REG_USE_LIST (dbg) = NULL;
4062 /* We delete rather than resolve these deps, otherwise we
4063 crash in sched_free_deps(), because forward deps are
4064 expected to be released before backward deps. */
4065 sd_delete_dep (sd_it);
4068 gcc_assert (QUEUE_INDEX (insn) == QUEUE_NOWHERE);
4069 QUEUE_INDEX (insn) = QUEUE_SCHEDULED;
4071 if (sched_pressure == SCHED_PRESSURE_MODEL
4072 && model_curr_point < model_num_insns
4073 && NONDEBUG_INSN_P (insn))
4075 if (model_index (insn) == model_curr_point)
4077 model_curr_point++;
4078 while (model_curr_point < model_num_insns
4079 && (QUEUE_INDEX (MODEL_INSN (model_curr_point))
4080 == QUEUE_SCHEDULED));
4081 else
4082 model_recompute (insn);
4083 model_update_limit_points ();
4084 update_register_pressure (insn);
4085 if (sched_verbose >= 2)
4086 print_curr_reg_pressure ();
4089 gcc_assert (INSN_TICK (insn) >= MIN_TICK);
4090 if (INSN_TICK (insn) > clock_var)
4091 /* INSN has been prematurely moved from the queue to the ready list.
4092 This is possible only if following flags are set. */
4093 gcc_assert (flag_sched_stalled_insns || sched_fusion);
4095 /* ??? Probably, if INSN is scheduled prematurely, we should leave
4096 INSN_TICK untouched. This is a machine-dependent issue, actually. */
4097 INSN_TICK (insn) = clock_var;
4099 check_clobbered_conditions (insn);
4101 /* Update dependent instructions. First, see if by scheduling this insn
4102 now we broke a dependence in a way that requires us to change another
4103 insn. */
4104 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
4105 sd_iterator_cond (&sd_it, &dep); sd_iterator_next (&sd_it))
4107 struct dep_replacement *desc = DEP_REPLACE (dep);
4108 rtx_insn *pro = DEP_PRO (dep);
4109 if (QUEUE_INDEX (pro) != QUEUE_SCHEDULED
4110 && desc != NULL && desc->insn == pro)
4111 apply_replacement (dep, false);
4114 /* Go through and resolve forward dependencies. */
4115 for (sd_it = sd_iterator_start (insn, SD_LIST_FORW);
4116 sd_iterator_cond (&sd_it, &dep);)
4118 rtx_insn *next = DEP_CON (dep);
4119 bool cancelled = (DEP_STATUS (dep) & DEP_CANCELLED) != 0;
4121 /* Resolve the dependence between INSN and NEXT.
4122 sd_resolve_dep () moves current dep to another list thus
4123 advancing the iterator. */
4124 sd_resolve_dep (sd_it);
4126 if (cancelled)
4128 if (must_restore_pattern_p (next, dep))
4129 restore_pattern (dep, false);
4130 continue;
4133 /* Don't bother trying to mark next as ready if insn is a debug
4134 insn. If insn is the last hard dependency, it will have
4135 already been discounted. */
4136 if (DEBUG_INSN_P (insn) && !DEBUG_INSN_P (next))
4137 continue;
4139 if (!IS_SPECULATION_BRANCHY_CHECK_P (insn))
4141 int effective_cost;
4143 effective_cost = try_ready (next);
4145 if (effective_cost >= 0
4146 && SCHED_GROUP_P (next)
4147 && advance < effective_cost)
4148 advance = effective_cost;
4150 else
4151 /* Check always has only one forward dependence (to the first insn in
4152 the recovery block), therefore, this will be executed only once. */
4154 gcc_assert (sd_lists_empty_p (insn, SD_LIST_FORW));
4155 fix_recovery_deps (RECOVERY_BLOCK (insn));
4159 /* Annotate the instruction with issue information -- TImode
4160 indicates that the instruction is expected not to be able
4161 to issue on the same cycle as the previous insn. A machine
4162 may use this information to decide how the instruction should
4163 be aligned. */
4164 if (issue_rate > 1
4165 && GET_CODE (PATTERN (insn)) != USE
4166 && GET_CODE (PATTERN (insn)) != CLOBBER
4167 && !DEBUG_INSN_P (insn))
4169 if (reload_completed)
4170 PUT_MODE (insn, clock_var > last_clock_var ? TImode : VOIDmode);
4171 last_clock_var = clock_var;
4174 if (nonscheduled_insns_begin != NULL_RTX)
4175 /* Indicate to debug counters that INSN is scheduled. */
4176 nonscheduled_insns_begin = insn;
4178 return advance;
4181 /* Functions for handling of notes. */
4183 /* Add note list that ends on FROM_END to the end of TO_ENDP. */
4184 void
4185 concat_note_lists (rtx_insn *from_end, rtx_insn **to_endp)
4187 rtx_insn *from_start;
4189 /* It's easy when have nothing to concat. */
4190 if (from_end == NULL)
4191 return;
4193 /* It's also easy when destination is empty. */
4194 if (*to_endp == NULL)
4196 *to_endp = from_end;
4197 return;
4200 from_start = from_end;
4201 while (PREV_INSN (from_start) != NULL)
4202 from_start = PREV_INSN (from_start);
4204 SET_PREV_INSN (from_start) = *to_endp;
4205 SET_NEXT_INSN (*to_endp) = from_start;
4206 *to_endp = from_end;
4209 /* Delete notes between HEAD and TAIL and put them in the chain
4210 of notes ended by NOTE_LIST. */
4211 void
4212 remove_notes (rtx_insn *head, rtx_insn *tail)
4214 rtx_insn *next_tail, *insn, *next;
4216 note_list = 0;
4217 if (head == tail && !INSN_P (head))
4218 return;
4220 next_tail = NEXT_INSN (tail);
4221 for (insn = head; insn != next_tail; insn = next)
4223 next = NEXT_INSN (insn);
4224 if (!NOTE_P (insn))
4225 continue;
4227 switch (NOTE_KIND (insn))
4229 case NOTE_INSN_BASIC_BLOCK:
4230 continue;
4232 case NOTE_INSN_EPILOGUE_BEG:
4233 if (insn != tail)
4235 remove_insn (insn);
4236 add_reg_note (next, REG_SAVE_NOTE,
4237 GEN_INT (NOTE_INSN_EPILOGUE_BEG));
4238 break;
4240 /* FALLTHRU */
4242 default:
4243 remove_insn (insn);
4245 /* Add the note to list that ends at NOTE_LIST. */
4246 SET_PREV_INSN (insn) = note_list;
4247 SET_NEXT_INSN (insn) = NULL_RTX;
4248 if (note_list)
4249 SET_NEXT_INSN (note_list) = insn;
4250 note_list = insn;
4251 break;
4254 gcc_assert ((sel_sched_p () || insn != tail) && insn != head);
4258 /* A structure to record enough data to allow us to backtrack the scheduler to
4259 a previous state. */
4260 struct haifa_saved_data
4262 /* Next entry on the list. */
4263 struct haifa_saved_data *next;
4265 /* Backtracking is associated with scheduling insns that have delay slots.
4266 DELAY_PAIR points to the structure that contains the insns involved, and
4267 the number of cycles between them. */
4268 struct delay_pair *delay_pair;
4270 /* Data used by the frontend (e.g. sched-ebb or sched-rgn). */
4271 void *fe_saved_data;
4272 /* Data used by the backend. */
4273 void *be_saved_data;
4275 /* Copies of global state. */
4276 int clock_var, last_clock_var;
4277 struct ready_list ready;
4278 state_t curr_state;
4280 rtx_insn *last_scheduled_insn;
4281 rtx_insn *last_nondebug_scheduled_insn;
4282 rtx_insn *nonscheduled_insns_begin;
4283 int cycle_issued_insns;
4285 /* Copies of state used in the inner loop of schedule_block. */
4286 struct sched_block_state sched_block;
4288 /* We don't need to save q_ptr, as its value is arbitrary and we can set it
4289 to 0 when restoring. */
4290 int q_size;
4291 rtx_insn_list **insn_queue;
4293 /* Describe pattern replacements that occurred since this backtrack point
4294 was queued. */
4295 vec<dep_t> replacement_deps;
4296 vec<int> replace_apply;
4298 /* A copy of the next-cycle replacement vectors at the time of the backtrack
4299 point. */
4300 vec<dep_t> next_cycle_deps;
4301 vec<int> next_cycle_apply;
4304 /* A record, in reverse order, of all scheduled insns which have delay slots
4305 and may require backtracking. */
4306 static struct haifa_saved_data *backtrack_queue;
4308 /* For every dependency of INSN, set the FEEDS_BACKTRACK_INSN bit according
4309 to SET_P. */
4310 static void
4311 mark_backtrack_feeds (rtx_insn *insn, int set_p)
4313 sd_iterator_def sd_it;
4314 dep_t dep;
4315 FOR_EACH_DEP (insn, SD_LIST_HARD_BACK, sd_it, dep)
4317 FEEDS_BACKTRACK_INSN (DEP_PRO (dep)) = set_p;
4321 /* Save the current scheduler state so that we can backtrack to it
4322 later if necessary. PAIR gives the insns that make it necessary to
4323 save this point. SCHED_BLOCK is the local state of schedule_block
4324 that need to be saved. */
4325 static void
4326 save_backtrack_point (struct delay_pair *pair,
4327 struct sched_block_state sched_block)
4329 int i;
4330 struct haifa_saved_data *save = XNEW (struct haifa_saved_data);
4332 save->curr_state = xmalloc (dfa_state_size);
4333 memcpy (save->curr_state, curr_state, dfa_state_size);
4335 save->ready.first = ready.first;
4336 save->ready.n_ready = ready.n_ready;
4337 save->ready.n_debug = ready.n_debug;
4338 save->ready.veclen = ready.veclen;
4339 save->ready.vec = XNEWVEC (rtx_insn *, ready.veclen);
4340 memcpy (save->ready.vec, ready.vec, ready.veclen * sizeof (rtx));
4342 save->insn_queue = XNEWVEC (rtx_insn_list *, max_insn_queue_index + 1);
4343 save->q_size = q_size;
4344 for (i = 0; i <= max_insn_queue_index; i++)
4346 int q = NEXT_Q_AFTER (q_ptr, i);
4347 save->insn_queue[i] = copy_INSN_LIST (insn_queue[q]);
4350 save->clock_var = clock_var;
4351 save->last_clock_var = last_clock_var;
4352 save->cycle_issued_insns = cycle_issued_insns;
4353 save->last_scheduled_insn = last_scheduled_insn;
4354 save->last_nondebug_scheduled_insn = last_nondebug_scheduled_insn;
4355 save->nonscheduled_insns_begin = nonscheduled_insns_begin;
4357 save->sched_block = sched_block;
4359 save->replacement_deps.create (0);
4360 save->replace_apply.create (0);
4361 save->next_cycle_deps = next_cycle_replace_deps.copy ();
4362 save->next_cycle_apply = next_cycle_apply.copy ();
4364 if (current_sched_info->save_state)
4365 save->fe_saved_data = (*current_sched_info->save_state) ();
4367 if (targetm.sched.alloc_sched_context)
4369 save->be_saved_data = targetm.sched.alloc_sched_context ();
4370 targetm.sched.init_sched_context (save->be_saved_data, false);
4372 else
4373 save->be_saved_data = NULL;
4375 save->delay_pair = pair;
4377 save->next = backtrack_queue;
4378 backtrack_queue = save;
4380 while (pair)
4382 mark_backtrack_feeds (pair->i2, 1);
4383 INSN_TICK (pair->i2) = INVALID_TICK;
4384 INSN_EXACT_TICK (pair->i2) = clock_var + pair_delay (pair);
4385 SHADOW_P (pair->i2) = pair->stages == 0;
4386 pair = pair->next_same_i1;
4390 /* Walk the ready list and all queues. If any insns have unresolved backwards
4391 dependencies, these must be cancelled deps, broken by predication. Set or
4392 clear (depending on SET) the DEP_CANCELLED bit in DEP_STATUS. */
4394 static void
4395 toggle_cancelled_flags (bool set)
4397 int i;
4398 sd_iterator_def sd_it;
4399 dep_t dep;
4401 if (ready.n_ready > 0)
4403 rtx_insn **first = ready_lastpos (&ready);
4404 for (i = 0; i < ready.n_ready; i++)
4405 FOR_EACH_DEP (first[i], SD_LIST_BACK, sd_it, dep)
4406 if (!DEBUG_INSN_P (DEP_PRO (dep)))
4408 if (set)
4409 DEP_STATUS (dep) |= DEP_CANCELLED;
4410 else
4411 DEP_STATUS (dep) &= ~DEP_CANCELLED;
4414 for (i = 0; i <= max_insn_queue_index; i++)
4416 int q = NEXT_Q_AFTER (q_ptr, i);
4417 rtx_insn_list *link;
4418 for (link = insn_queue[q]; link; link = link->next ())
4420 rtx_insn *insn = link->insn ();
4421 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
4422 if (!DEBUG_INSN_P (DEP_PRO (dep)))
4424 if (set)
4425 DEP_STATUS (dep) |= DEP_CANCELLED;
4426 else
4427 DEP_STATUS (dep) &= ~DEP_CANCELLED;
4433 /* Undo the replacements that have occurred after backtrack point SAVE
4434 was placed. */
4435 static void
4436 undo_replacements_for_backtrack (struct haifa_saved_data *save)
4438 while (!save->replacement_deps.is_empty ())
4440 dep_t dep = save->replacement_deps.pop ();
4441 int apply_p = save->replace_apply.pop ();
4443 if (apply_p)
4444 restore_pattern (dep, true);
4445 else
4446 apply_replacement (dep, true);
4448 save->replacement_deps.release ();
4449 save->replace_apply.release ();
4452 /* Pop entries from the SCHEDULED_INSNS vector up to and including INSN.
4453 Restore their dependencies to an unresolved state, and mark them as
4454 queued nowhere. */
4456 static void
4457 unschedule_insns_until (rtx_insn *insn)
4459 auto_vec<rtx_insn *> recompute_vec;
4461 /* Make two passes over the insns to be unscheduled. First, we clear out
4462 dependencies and other trivial bookkeeping. */
4463 for (;;)
4465 rtx_insn *last;
4466 sd_iterator_def sd_it;
4467 dep_t dep;
4469 last = scheduled_insns.pop ();
4471 /* This will be changed by restore_backtrack_point if the insn is in
4472 any queue. */
4473 QUEUE_INDEX (last) = QUEUE_NOWHERE;
4474 if (last != insn)
4475 INSN_TICK (last) = INVALID_TICK;
4477 if (modulo_ii > 0 && INSN_UID (last) < modulo_iter0_max_uid)
4478 modulo_insns_scheduled--;
4480 for (sd_it = sd_iterator_start (last, SD_LIST_RES_FORW);
4481 sd_iterator_cond (&sd_it, &dep);)
4483 rtx_insn *con = DEP_CON (dep);
4484 sd_unresolve_dep (sd_it);
4485 if (!MUST_RECOMPUTE_SPEC_P (con))
4487 MUST_RECOMPUTE_SPEC_P (con) = 1;
4488 recompute_vec.safe_push (con);
4492 if (last == insn)
4493 break;
4496 /* A second pass, to update ready and speculation status for insns
4497 depending on the unscheduled ones. The first pass must have
4498 popped the scheduled_insns vector up to the point where we
4499 restart scheduling, as recompute_todo_spec requires it to be
4500 up-to-date. */
4501 while (!recompute_vec.is_empty ())
4503 rtx_insn *con;
4505 con = recompute_vec.pop ();
4506 MUST_RECOMPUTE_SPEC_P (con) = 0;
4507 if (!sd_lists_empty_p (con, SD_LIST_HARD_BACK))
4509 TODO_SPEC (con) = HARD_DEP;
4510 INSN_TICK (con) = INVALID_TICK;
4511 if (PREDICATED_PAT (con) != NULL_RTX)
4512 haifa_change_pattern (con, ORIG_PAT (con));
4514 else if (QUEUE_INDEX (con) != QUEUE_SCHEDULED)
4515 TODO_SPEC (con) = recompute_todo_spec (con, true);
4519 /* Restore scheduler state from the topmost entry on the backtracking queue.
4520 PSCHED_BLOCK_P points to the local data of schedule_block that we must
4521 overwrite with the saved data.
4522 The caller must already have called unschedule_insns_until. */
4524 static void
4525 restore_last_backtrack_point (struct sched_block_state *psched_block)
4527 int i;
4528 struct haifa_saved_data *save = backtrack_queue;
4530 backtrack_queue = save->next;
4532 if (current_sched_info->restore_state)
4533 (*current_sched_info->restore_state) (save->fe_saved_data);
4535 if (targetm.sched.alloc_sched_context)
4537 targetm.sched.set_sched_context (save->be_saved_data);
4538 targetm.sched.free_sched_context (save->be_saved_data);
4541 /* Do this first since it clobbers INSN_TICK of the involved
4542 instructions. */
4543 undo_replacements_for_backtrack (save);
4545 /* Clear the QUEUE_INDEX of everything in the ready list or one
4546 of the queues. */
4547 if (ready.n_ready > 0)
4549 rtx_insn **first = ready_lastpos (&ready);
4550 for (i = 0; i < ready.n_ready; i++)
4552 rtx_insn *insn = first[i];
4553 QUEUE_INDEX (insn) = QUEUE_NOWHERE;
4554 INSN_TICK (insn) = INVALID_TICK;
4557 for (i = 0; i <= max_insn_queue_index; i++)
4559 int q = NEXT_Q_AFTER (q_ptr, i);
4561 for (rtx_insn_list *link = insn_queue[q]; link; link = link->next ())
4563 rtx_insn *x = link->insn ();
4564 QUEUE_INDEX (x) = QUEUE_NOWHERE;
4565 INSN_TICK (x) = INVALID_TICK;
4567 free_INSN_LIST_list (&insn_queue[q]);
4570 free (ready.vec);
4571 ready = save->ready;
4573 if (ready.n_ready > 0)
4575 rtx_insn **first = ready_lastpos (&ready);
4576 for (i = 0; i < ready.n_ready; i++)
4578 rtx_insn *insn = first[i];
4579 QUEUE_INDEX (insn) = QUEUE_READY;
4580 TODO_SPEC (insn) = recompute_todo_spec (insn, true);
4581 INSN_TICK (insn) = save->clock_var;
4585 q_ptr = 0;
4586 q_size = save->q_size;
4587 for (i = 0; i <= max_insn_queue_index; i++)
4589 int q = NEXT_Q_AFTER (q_ptr, i);
4591 insn_queue[q] = save->insn_queue[q];
4593 for (rtx_insn_list *link = insn_queue[q]; link; link = link->next ())
4595 rtx_insn *x = link->insn ();
4596 QUEUE_INDEX (x) = i;
4597 TODO_SPEC (x) = recompute_todo_spec (x, true);
4598 INSN_TICK (x) = save->clock_var + i;
4601 free (save->insn_queue);
4603 toggle_cancelled_flags (true);
4605 clock_var = save->clock_var;
4606 last_clock_var = save->last_clock_var;
4607 cycle_issued_insns = save->cycle_issued_insns;
4608 last_scheduled_insn = save->last_scheduled_insn;
4609 last_nondebug_scheduled_insn = save->last_nondebug_scheduled_insn;
4610 nonscheduled_insns_begin = save->nonscheduled_insns_begin;
4612 *psched_block = save->sched_block;
4614 memcpy (curr_state, save->curr_state, dfa_state_size);
4615 free (save->curr_state);
4617 mark_backtrack_feeds (save->delay_pair->i2, 0);
4619 gcc_assert (next_cycle_replace_deps.is_empty ());
4620 next_cycle_replace_deps = save->next_cycle_deps.copy ();
4621 next_cycle_apply = save->next_cycle_apply.copy ();
4623 free (save);
4625 for (save = backtrack_queue; save; save = save->next)
4627 mark_backtrack_feeds (save->delay_pair->i2, 1);
4631 /* Discard all data associated with the topmost entry in the backtrack
4632 queue. If RESET_TICK is false, we just want to free the data. If true,
4633 we are doing this because we discovered a reason to backtrack. In the
4634 latter case, also reset the INSN_TICK for the shadow insn. */
4635 static void
4636 free_topmost_backtrack_point (bool reset_tick)
4638 struct haifa_saved_data *save = backtrack_queue;
4639 int i;
4641 backtrack_queue = save->next;
4643 if (reset_tick)
4645 struct delay_pair *pair = save->delay_pair;
4646 while (pair)
4648 INSN_TICK (pair->i2) = INVALID_TICK;
4649 INSN_EXACT_TICK (pair->i2) = INVALID_TICK;
4650 pair = pair->next_same_i1;
4652 undo_replacements_for_backtrack (save);
4654 else
4656 save->replacement_deps.release ();
4657 save->replace_apply.release ();
4660 if (targetm.sched.free_sched_context)
4661 targetm.sched.free_sched_context (save->be_saved_data);
4662 if (current_sched_info->restore_state)
4663 free (save->fe_saved_data);
4664 for (i = 0; i <= max_insn_queue_index; i++)
4665 free_INSN_LIST_list (&save->insn_queue[i]);
4666 free (save->insn_queue);
4667 free (save->curr_state);
4668 free (save->ready.vec);
4669 free (save);
4672 /* Free the entire backtrack queue. */
4673 static void
4674 free_backtrack_queue (void)
4676 while (backtrack_queue)
4677 free_topmost_backtrack_point (false);
4680 /* Apply a replacement described by DESC. If IMMEDIATELY is false, we
4681 may have to postpone the replacement until the start of the next cycle,
4682 at which point we will be called again with IMMEDIATELY true. This is
4683 only done for machines which have instruction packets with explicit
4684 parallelism however. */
4685 static void
4686 apply_replacement (dep_t dep, bool immediately)
4688 struct dep_replacement *desc = DEP_REPLACE (dep);
4689 if (!immediately && targetm.sched.exposed_pipeline && reload_completed)
4691 next_cycle_replace_deps.safe_push (dep);
4692 next_cycle_apply.safe_push (1);
4694 else
4696 bool success;
4698 if (QUEUE_INDEX (desc->insn) == QUEUE_SCHEDULED)
4699 return;
4701 if (sched_verbose >= 5)
4702 fprintf (sched_dump, "applying replacement for insn %d\n",
4703 INSN_UID (desc->insn));
4705 success = validate_change (desc->insn, desc->loc, desc->newval, 0);
4706 gcc_assert (success);
4708 update_insn_after_change (desc->insn);
4709 if ((TODO_SPEC (desc->insn) & (HARD_DEP | DEP_POSTPONED)) == 0)
4710 fix_tick_ready (desc->insn);
4712 if (backtrack_queue != NULL)
4714 backtrack_queue->replacement_deps.safe_push (dep);
4715 backtrack_queue->replace_apply.safe_push (1);
4720 /* We have determined that a pattern involved in DEP must be restored.
4721 If IMMEDIATELY is false, we may have to postpone the replacement
4722 until the start of the next cycle, at which point we will be called
4723 again with IMMEDIATELY true. */
4724 static void
4725 restore_pattern (dep_t dep, bool immediately)
4727 rtx_insn *next = DEP_CON (dep);
4728 int tick = INSN_TICK (next);
4730 /* If we already scheduled the insn, the modified version is
4731 correct. */
4732 if (QUEUE_INDEX (next) == QUEUE_SCHEDULED)
4733 return;
4735 if (!immediately && targetm.sched.exposed_pipeline && reload_completed)
4737 next_cycle_replace_deps.safe_push (dep);
4738 next_cycle_apply.safe_push (0);
4739 return;
4743 if (DEP_TYPE (dep) == REG_DEP_CONTROL)
4745 if (sched_verbose >= 5)
4746 fprintf (sched_dump, "restoring pattern for insn %d\n",
4747 INSN_UID (next));
4748 haifa_change_pattern (next, ORIG_PAT (next));
4750 else
4752 struct dep_replacement *desc = DEP_REPLACE (dep);
4753 bool success;
4755 if (sched_verbose >= 5)
4756 fprintf (sched_dump, "restoring pattern for insn %d\n",
4757 INSN_UID (desc->insn));
4758 tick = INSN_TICK (desc->insn);
4760 success = validate_change (desc->insn, desc->loc, desc->orig, 0);
4761 gcc_assert (success);
4762 update_insn_after_change (desc->insn);
4763 if (backtrack_queue != NULL)
4765 backtrack_queue->replacement_deps.safe_push (dep);
4766 backtrack_queue->replace_apply.safe_push (0);
4769 INSN_TICK (next) = tick;
4770 if (TODO_SPEC (next) == DEP_POSTPONED)
4771 return;
4773 if (sd_lists_empty_p (next, SD_LIST_BACK))
4774 TODO_SPEC (next) = 0;
4775 else if (!sd_lists_empty_p (next, SD_LIST_HARD_BACK))
4776 TODO_SPEC (next) = HARD_DEP;
4779 /* Perform pattern replacements that were queued up until the next
4780 cycle. */
4781 static void
4782 perform_replacements_new_cycle (void)
4784 int i;
4785 dep_t dep;
4786 FOR_EACH_VEC_ELT (next_cycle_replace_deps, i, dep)
4788 int apply_p = next_cycle_apply[i];
4789 if (apply_p)
4790 apply_replacement (dep, true);
4791 else
4792 restore_pattern (dep, true);
4794 next_cycle_replace_deps.truncate (0);
4795 next_cycle_apply.truncate (0);
4798 /* Compute INSN_TICK_ESTIMATE for INSN. PROCESSED is a bitmap of
4799 instructions we've previously encountered, a set bit prevents
4800 recursion. BUDGET is a limit on how far ahead we look, it is
4801 reduced on recursive calls. Return true if we produced a good
4802 estimate, or false if we exceeded the budget. */
4803 static bool
4804 estimate_insn_tick (bitmap processed, rtx_insn *insn, int budget)
4806 sd_iterator_def sd_it;
4807 dep_t dep;
4808 int earliest = INSN_TICK (insn);
4810 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
4812 rtx_insn *pro = DEP_PRO (dep);
4813 int t;
4815 if (DEP_STATUS (dep) & DEP_CANCELLED)
4816 continue;
4818 if (QUEUE_INDEX (pro) == QUEUE_SCHEDULED)
4819 gcc_assert (INSN_TICK (pro) + dep_cost (dep) <= INSN_TICK (insn));
4820 else
4822 int cost = dep_cost (dep);
4823 if (cost >= budget)
4824 return false;
4825 if (!bitmap_bit_p (processed, INSN_LUID (pro)))
4827 if (!estimate_insn_tick (processed, pro, budget - cost))
4828 return false;
4830 gcc_assert (INSN_TICK_ESTIMATE (pro) != INVALID_TICK);
4831 t = INSN_TICK_ESTIMATE (pro) + cost;
4832 if (earliest == INVALID_TICK || t > earliest)
4833 earliest = t;
4836 bitmap_set_bit (processed, INSN_LUID (insn));
4837 INSN_TICK_ESTIMATE (insn) = earliest;
4838 return true;
4841 /* Examine the pair of insns in P, and estimate (optimistically, assuming
4842 infinite resources) the cycle in which the delayed shadow can be issued.
4843 Return the number of cycles that must pass before the real insn can be
4844 issued in order to meet this constraint. */
4845 static int
4846 estimate_shadow_tick (struct delay_pair *p)
4848 auto_bitmap processed;
4849 int t;
4850 bool cutoff;
4852 cutoff = !estimate_insn_tick (processed, p->i2,
4853 max_insn_queue_index + pair_delay (p));
4854 if (cutoff)
4855 return max_insn_queue_index;
4856 t = INSN_TICK_ESTIMATE (p->i2) - (clock_var + pair_delay (p) + 1);
4857 if (t > 0)
4858 return t;
4859 return 0;
4862 /* If INSN has no unresolved backwards dependencies, add it to the schedule and
4863 recursively resolve all its forward dependencies. */
4864 static void
4865 resolve_dependencies (rtx_insn *insn)
4867 sd_iterator_def sd_it;
4868 dep_t dep;
4870 /* Don't use sd_lists_empty_p; it ignores debug insns. */
4871 if (DEPS_LIST_FIRST (INSN_HARD_BACK_DEPS (insn)) != NULL
4872 || DEPS_LIST_FIRST (INSN_SPEC_BACK_DEPS (insn)) != NULL)
4873 return;
4875 if (sched_verbose >= 4)
4876 fprintf (sched_dump, ";;\tquickly resolving %d\n", INSN_UID (insn));
4878 if (QUEUE_INDEX (insn) >= 0)
4879 queue_remove (insn);
4881 scheduled_insns.safe_push (insn);
4883 /* Update dependent instructions. */
4884 for (sd_it = sd_iterator_start (insn, SD_LIST_FORW);
4885 sd_iterator_cond (&sd_it, &dep);)
4887 rtx_insn *next = DEP_CON (dep);
4889 if (sched_verbose >= 4)
4890 fprintf (sched_dump, ";;\t\tdep %d against %d\n", INSN_UID (insn),
4891 INSN_UID (next));
4893 /* Resolve the dependence between INSN and NEXT.
4894 sd_resolve_dep () moves current dep to another list thus
4895 advancing the iterator. */
4896 sd_resolve_dep (sd_it);
4898 if (!IS_SPECULATION_BRANCHY_CHECK_P (insn))
4900 resolve_dependencies (next);
4902 else
4903 /* Check always has only one forward dependence (to the first insn in
4904 the recovery block), therefore, this will be executed only once. */
4906 gcc_assert (sd_lists_empty_p (insn, SD_LIST_FORW));
4912 /* Return the head and tail pointers of ebb starting at BEG and ending
4913 at END. */
4914 void
4915 get_ebb_head_tail (basic_block beg, basic_block end,
4916 rtx_insn **headp, rtx_insn **tailp)
4918 rtx_insn *beg_head = BB_HEAD (beg);
4919 rtx_insn * beg_tail = BB_END (beg);
4920 rtx_insn * end_head = BB_HEAD (end);
4921 rtx_insn * end_tail = BB_END (end);
4923 /* Don't include any notes or labels at the beginning of the BEG
4924 basic block, or notes at the end of the END basic blocks. */
4926 if (LABEL_P (beg_head))
4927 beg_head = NEXT_INSN (beg_head);
4929 while (beg_head != beg_tail)
4930 if (NOTE_P (beg_head))
4931 beg_head = NEXT_INSN (beg_head);
4932 else if (DEBUG_INSN_P (beg_head))
4934 rtx_insn * note, *next;
4936 for (note = NEXT_INSN (beg_head);
4937 note != beg_tail;
4938 note = next)
4940 next = NEXT_INSN (note);
4941 if (NOTE_P (note))
4943 if (sched_verbose >= 9)
4944 fprintf (sched_dump, "reorder %i\n", INSN_UID (note));
4946 reorder_insns_nobb (note, note, PREV_INSN (beg_head));
4948 if (BLOCK_FOR_INSN (note) != beg)
4949 df_insn_change_bb (note, beg);
4951 else if (!DEBUG_INSN_P (note))
4952 break;
4955 break;
4957 else
4958 break;
4960 *headp = beg_head;
4962 if (beg == end)
4963 end_head = beg_head;
4964 else if (LABEL_P (end_head))
4965 end_head = NEXT_INSN (end_head);
4967 while (end_head != end_tail)
4968 if (NOTE_P (end_tail))
4969 end_tail = PREV_INSN (end_tail);
4970 else if (DEBUG_INSN_P (end_tail))
4972 rtx_insn * note, *prev;
4974 for (note = PREV_INSN (end_tail);
4975 note != end_head;
4976 note = prev)
4978 prev = PREV_INSN (note);
4979 if (NOTE_P (note))
4981 if (sched_verbose >= 9)
4982 fprintf (sched_dump, "reorder %i\n", INSN_UID (note));
4984 reorder_insns_nobb (note, note, end_tail);
4986 if (end_tail == BB_END (end))
4987 BB_END (end) = note;
4989 if (BLOCK_FOR_INSN (note) != end)
4990 df_insn_change_bb (note, end);
4992 else if (!DEBUG_INSN_P (note))
4993 break;
4996 break;
4998 else
4999 break;
5001 *tailp = end_tail;
5004 /* Return nonzero if there are no real insns in the range [ HEAD, TAIL ]. */
5007 no_real_insns_p (const rtx_insn *head, const rtx_insn *tail)
5009 while (head != NEXT_INSN (tail))
5011 if (!NOTE_P (head) && !LABEL_P (head))
5012 return 0;
5013 head = NEXT_INSN (head);
5015 return 1;
5018 /* Restore-other-notes: NOTE_LIST is the end of a chain of notes
5019 previously found among the insns. Insert them just before HEAD. */
5020 rtx_insn *
5021 restore_other_notes (rtx_insn *head, basic_block head_bb)
5023 if (note_list != 0)
5025 rtx_insn *note_head = note_list;
5027 if (head)
5028 head_bb = BLOCK_FOR_INSN (head);
5029 else
5030 head = NEXT_INSN (bb_note (head_bb));
5032 while (PREV_INSN (note_head))
5034 set_block_for_insn (note_head, head_bb);
5035 note_head = PREV_INSN (note_head);
5037 /* In the above cycle we've missed this note. */
5038 set_block_for_insn (note_head, head_bb);
5040 SET_PREV_INSN (note_head) = PREV_INSN (head);
5041 SET_NEXT_INSN (PREV_INSN (head)) = note_head;
5042 SET_PREV_INSN (head) = note_list;
5043 SET_NEXT_INSN (note_list) = head;
5045 if (BLOCK_FOR_INSN (head) != head_bb)
5046 BB_END (head_bb) = note_list;
5048 head = note_head;
5051 return head;
5054 /* When we know we are going to discard the schedule due to a failed attempt
5055 at modulo scheduling, undo all replacements. */
5056 static void
5057 undo_all_replacements (void)
5059 rtx_insn *insn;
5060 int i;
5062 FOR_EACH_VEC_ELT (scheduled_insns, i, insn)
5064 sd_iterator_def sd_it;
5065 dep_t dep;
5067 /* See if we must undo a replacement. */
5068 for (sd_it = sd_iterator_start (insn, SD_LIST_RES_FORW);
5069 sd_iterator_cond (&sd_it, &dep); sd_iterator_next (&sd_it))
5071 struct dep_replacement *desc = DEP_REPLACE (dep);
5072 if (desc != NULL)
5073 validate_change (desc->insn, desc->loc, desc->orig, 0);
5078 /* Return first non-scheduled insn in the current scheduling block.
5079 This is mostly used for debug-counter purposes. */
5080 static rtx_insn *
5081 first_nonscheduled_insn (void)
5083 rtx_insn *insn = (nonscheduled_insns_begin != NULL_RTX
5084 ? nonscheduled_insns_begin
5085 : current_sched_info->prev_head);
5089 insn = next_nonnote_nondebug_insn (insn);
5091 while (QUEUE_INDEX (insn) == QUEUE_SCHEDULED);
5093 return insn;
5096 /* Move insns that became ready to fire from queue to ready list. */
5098 static void
5099 queue_to_ready (struct ready_list *ready)
5101 rtx_insn *insn;
5102 rtx_insn_list *link;
5103 rtx_insn *skip_insn;
5105 q_ptr = NEXT_Q (q_ptr);
5107 if (dbg_cnt (sched_insn) == false)
5108 /* If debug counter is activated do not requeue the first
5109 nonscheduled insn. */
5110 skip_insn = first_nonscheduled_insn ();
5111 else
5112 skip_insn = NULL;
5114 /* Add all pending insns that can be scheduled without stalls to the
5115 ready list. */
5116 for (link = insn_queue[q_ptr]; link; link = link->next ())
5118 insn = link->insn ();
5119 q_size -= 1;
5121 if (sched_verbose >= 2)
5122 fprintf (sched_dump, ";;\t\tQ-->Ready: insn %s: ",
5123 (*current_sched_info->print_insn) (insn, 0));
5125 /* If the ready list is full, delay the insn for 1 cycle.
5126 See the comment in schedule_block for the rationale. */
5127 if (!reload_completed
5128 && (ready->n_ready - ready->n_debug > MAX_SCHED_READY_INSNS
5129 || (sched_pressure == SCHED_PRESSURE_MODEL
5130 /* Limit pressure recalculations to MAX_SCHED_READY_INSNS
5131 instructions too. */
5132 && model_index (insn) > (model_curr_point
5133 + MAX_SCHED_READY_INSNS)))
5134 && !(sched_pressure == SCHED_PRESSURE_MODEL
5135 && model_curr_point < model_num_insns
5136 /* Always allow the next model instruction to issue. */
5137 && model_index (insn) == model_curr_point)
5138 && !SCHED_GROUP_P (insn)
5139 && insn != skip_insn)
5141 if (sched_verbose >= 2)
5142 fprintf (sched_dump, "keeping in queue, ready full\n");
5143 queue_insn (insn, 1, "ready full");
5145 else
5147 ready_add (ready, insn, false);
5148 if (sched_verbose >= 2)
5149 fprintf (sched_dump, "moving to ready without stalls\n");
5152 free_INSN_LIST_list (&insn_queue[q_ptr]);
5154 /* If there are no ready insns, stall until one is ready and add all
5155 of the pending insns at that point to the ready list. */
5156 if (ready->n_ready == 0)
5158 int stalls;
5160 for (stalls = 1; stalls <= max_insn_queue_index; stalls++)
5162 if ((link = insn_queue[NEXT_Q_AFTER (q_ptr, stalls)]))
5164 for (; link; link = link->next ())
5166 insn = link->insn ();
5167 q_size -= 1;
5169 if (sched_verbose >= 2)
5170 fprintf (sched_dump, ";;\t\tQ-->Ready: insn %s: ",
5171 (*current_sched_info->print_insn) (insn, 0));
5173 ready_add (ready, insn, false);
5174 if (sched_verbose >= 2)
5175 fprintf (sched_dump, "moving to ready with %d stalls\n", stalls);
5177 free_INSN_LIST_list (&insn_queue[NEXT_Q_AFTER (q_ptr, stalls)]);
5179 advance_one_cycle ();
5181 break;
5184 advance_one_cycle ();
5187 q_ptr = NEXT_Q_AFTER (q_ptr, stalls);
5188 clock_var += stalls;
5189 if (sched_verbose >= 2)
5190 fprintf (sched_dump, ";;\tAdvancing clock by %d cycle[s] to %d\n",
5191 stalls, clock_var);
5195 /* Used by early_queue_to_ready. Determines whether it is "ok" to
5196 prematurely move INSN from the queue to the ready list. Currently,
5197 if a target defines the hook 'is_costly_dependence', this function
5198 uses the hook to check whether there exist any dependences which are
5199 considered costly by the target, between INSN and other insns that
5200 have already been scheduled. Dependences are checked up to Y cycles
5201 back, with default Y=1; The flag -fsched-stalled-insns-dep=Y allows
5202 controlling this value.
5203 (Other considerations could be taken into account instead (or in
5204 addition) depending on user flags and target hooks. */
5206 static bool
5207 ok_for_early_queue_removal (rtx_insn *insn)
5209 if (targetm.sched.is_costly_dependence)
5211 int n_cycles;
5212 int i = scheduled_insns.length ();
5213 for (n_cycles = flag_sched_stalled_insns_dep; n_cycles; n_cycles--)
5215 while (i-- > 0)
5217 int cost;
5219 rtx_insn *prev_insn = scheduled_insns[i];
5221 if (!NOTE_P (prev_insn))
5223 dep_t dep;
5225 dep = sd_find_dep_between (prev_insn, insn, true);
5227 if (dep != NULL)
5229 cost = dep_cost (dep);
5231 if (targetm.sched.is_costly_dependence (dep, cost,
5232 flag_sched_stalled_insns_dep - n_cycles))
5233 return false;
5237 if (GET_MODE (prev_insn) == TImode) /* end of dispatch group */
5238 break;
5241 if (i == 0)
5242 break;
5246 return true;
5250 /* Remove insns from the queue, before they become "ready" with respect
5251 to FU latency considerations. */
5253 static int
5254 early_queue_to_ready (state_t state, struct ready_list *ready)
5256 rtx_insn *insn;
5257 rtx_insn_list *link;
5258 rtx_insn_list *next_link;
5259 rtx_insn_list *prev_link;
5260 bool move_to_ready;
5261 int cost;
5262 state_t temp_state = alloca (dfa_state_size);
5263 int stalls;
5264 int insns_removed = 0;
5267 Flag '-fsched-stalled-insns=X' determines the aggressiveness of this
5268 function:
5270 X == 0: There is no limit on how many queued insns can be removed
5271 prematurely. (flag_sched_stalled_insns = -1).
5273 X >= 1: Only X queued insns can be removed prematurely in each
5274 invocation. (flag_sched_stalled_insns = X).
5276 Otherwise: Early queue removal is disabled.
5277 (flag_sched_stalled_insns = 0)
5280 if (! flag_sched_stalled_insns)
5281 return 0;
5283 for (stalls = 0; stalls <= max_insn_queue_index; stalls++)
5285 if ((link = insn_queue[NEXT_Q_AFTER (q_ptr, stalls)]))
5287 if (sched_verbose > 6)
5288 fprintf (sched_dump, ";; look at index %d + %d\n", q_ptr, stalls);
5290 prev_link = 0;
5291 while (link)
5293 next_link = link->next ();
5294 insn = link->insn ();
5295 if (insn && sched_verbose > 6)
5296 print_rtl_single (sched_dump, insn);
5298 memcpy (temp_state, state, dfa_state_size);
5299 if (recog_memoized (insn) < 0)
5300 /* non-negative to indicate that it's not ready
5301 to avoid infinite Q->R->Q->R... */
5302 cost = 0;
5303 else
5304 cost = state_transition (temp_state, insn);
5306 if (sched_verbose >= 6)
5307 fprintf (sched_dump, "transition cost = %d\n", cost);
5309 move_to_ready = false;
5310 if (cost < 0)
5312 move_to_ready = ok_for_early_queue_removal (insn);
5313 if (move_to_ready == true)
5315 /* move from Q to R */
5316 q_size -= 1;
5317 ready_add (ready, insn, false);
5319 if (prev_link)
5320 XEXP (prev_link, 1) = next_link;
5321 else
5322 insn_queue[NEXT_Q_AFTER (q_ptr, stalls)] = next_link;
5324 free_INSN_LIST_node (link);
5326 if (sched_verbose >= 2)
5327 fprintf (sched_dump, ";;\t\tEarly Q-->Ready: insn %s\n",
5328 (*current_sched_info->print_insn) (insn, 0));
5330 insns_removed++;
5331 if (insns_removed == flag_sched_stalled_insns)
5332 /* Remove no more than flag_sched_stalled_insns insns
5333 from Q at a time. */
5334 return insns_removed;
5338 if (move_to_ready == false)
5339 prev_link = link;
5341 link = next_link;
5342 } /* while link */
5343 } /* if link */
5345 } /* for stalls.. */
5347 return insns_removed;
5351 /* Print the ready list for debugging purposes.
5352 If READY_TRY is non-zero then only print insns that max_issue
5353 will consider. */
5354 static void
5355 debug_ready_list_1 (struct ready_list *ready, signed char *ready_try)
5357 rtx_insn **p;
5358 int i;
5360 if (ready->n_ready == 0)
5362 fprintf (sched_dump, "\n");
5363 return;
5366 p = ready_lastpos (ready);
5367 for (i = 0; i < ready->n_ready; i++)
5369 if (ready_try != NULL && ready_try[ready->n_ready - i - 1])
5370 continue;
5372 fprintf (sched_dump, " %s:%d",
5373 (*current_sched_info->print_insn) (p[i], 0),
5374 INSN_LUID (p[i]));
5375 if (sched_pressure != SCHED_PRESSURE_NONE)
5376 fprintf (sched_dump, "(cost=%d",
5377 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (p[i]));
5378 fprintf (sched_dump, ":prio=%d", INSN_PRIORITY (p[i]));
5379 if (INSN_TICK (p[i]) > clock_var)
5380 fprintf (sched_dump, ":delay=%d", INSN_TICK (p[i]) - clock_var);
5381 if (sched_pressure == SCHED_PRESSURE_MODEL)
5382 fprintf (sched_dump, ":idx=%d",
5383 model_index (p[i]));
5384 if (sched_pressure != SCHED_PRESSURE_NONE)
5385 fprintf (sched_dump, ")");
5387 fprintf (sched_dump, "\n");
5390 /* Print the ready list. Callable from debugger. */
5391 static void
5392 debug_ready_list (struct ready_list *ready)
5394 debug_ready_list_1 (ready, NULL);
5397 /* Search INSN for REG_SAVE_NOTE notes and convert them back into insn
5398 NOTEs. This is used for NOTE_INSN_EPILOGUE_BEG, so that sched-ebb
5399 replaces the epilogue note in the correct basic block. */
5400 void
5401 reemit_notes (rtx_insn *insn)
5403 rtx note;
5404 rtx_insn *last = insn;
5406 for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
5408 if (REG_NOTE_KIND (note) == REG_SAVE_NOTE)
5410 enum insn_note note_type = (enum insn_note) INTVAL (XEXP (note, 0));
5412 last = emit_note_before (note_type, last);
5413 remove_note (insn, note);
5418 /* Move INSN. Reemit notes if needed. Update CFG, if needed. */
5419 static void
5420 move_insn (rtx_insn *insn, rtx_insn *last, rtx nt)
5422 if (PREV_INSN (insn) != last)
5424 basic_block bb;
5425 rtx_insn *note;
5426 int jump_p = 0;
5428 bb = BLOCK_FOR_INSN (insn);
5430 /* BB_HEAD is either LABEL or NOTE. */
5431 gcc_assert (BB_HEAD (bb) != insn);
5433 if (BB_END (bb) == insn)
5434 /* If this is last instruction in BB, move end marker one
5435 instruction up. */
5437 /* Jumps are always placed at the end of basic block. */
5438 jump_p = control_flow_insn_p (insn);
5440 gcc_assert (!jump_p
5441 || ((common_sched_info->sched_pass_id == SCHED_RGN_PASS)
5442 && IS_SPECULATION_BRANCHY_CHECK_P (insn))
5443 || (common_sched_info->sched_pass_id
5444 == SCHED_EBB_PASS));
5446 gcc_assert (BLOCK_FOR_INSN (PREV_INSN (insn)) == bb);
5448 BB_END (bb) = PREV_INSN (insn);
5451 gcc_assert (BB_END (bb) != last);
5453 if (jump_p)
5454 /* We move the block note along with jump. */
5456 gcc_assert (nt);
5458 note = NEXT_INSN (insn);
5459 while (NOTE_NOT_BB_P (note) && note != nt)
5460 note = NEXT_INSN (note);
5462 if (note != nt
5463 && (LABEL_P (note)
5464 || BARRIER_P (note)))
5465 note = NEXT_INSN (note);
5467 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
5469 else
5470 note = insn;
5472 SET_NEXT_INSN (PREV_INSN (insn)) = NEXT_INSN (note);
5473 SET_PREV_INSN (NEXT_INSN (note)) = PREV_INSN (insn);
5475 SET_NEXT_INSN (note) = NEXT_INSN (last);
5476 SET_PREV_INSN (NEXT_INSN (last)) = note;
5478 SET_NEXT_INSN (last) = insn;
5479 SET_PREV_INSN (insn) = last;
5481 bb = BLOCK_FOR_INSN (last);
5483 if (jump_p)
5485 fix_jump_move (insn);
5487 if (BLOCK_FOR_INSN (insn) != bb)
5488 move_block_after_check (insn);
5490 gcc_assert (BB_END (bb) == last);
5493 df_insn_change_bb (insn, bb);
5495 /* Update BB_END, if needed. */
5496 if (BB_END (bb) == last)
5497 BB_END (bb) = insn;
5500 SCHED_GROUP_P (insn) = 0;
5503 /* Return true if scheduling INSN will finish current clock cycle. */
5504 static bool
5505 insn_finishes_cycle_p (rtx_insn *insn)
5507 if (SCHED_GROUP_P (insn))
5508 /* After issuing INSN, rest of the sched_group will be forced to issue
5509 in order. Don't make any plans for the rest of cycle. */
5510 return true;
5512 /* Finishing the block will, apparently, finish the cycle. */
5513 if (current_sched_info->insn_finishes_block_p
5514 && current_sched_info->insn_finishes_block_p (insn))
5515 return true;
5517 return false;
5520 /* Helper for autopref_multipass_init. Given a SET in PAT and whether
5521 we're expecting a memory WRITE or not, check that the insn is relevant to
5522 the autoprefetcher modelling code. Return true iff that is the case.
5523 If it is relevant, record the base register of the memory op in BASE and
5524 the offset in OFFSET. */
5526 static bool
5527 analyze_set_insn_for_autopref (rtx pat, bool write, rtx *base, int *offset)
5529 if (GET_CODE (pat) != SET)
5530 return false;
5532 rtx mem = write ? SET_DEST (pat) : SET_SRC (pat);
5533 if (!MEM_P (mem))
5534 return false;
5536 struct address_info info;
5537 decompose_mem_address (&info, mem);
5539 /* TODO: Currently only (base+const) addressing is supported. */
5540 if (info.base == NULL || !REG_P (*info.base)
5541 || (info.disp != NULL && !CONST_INT_P (*info.disp)))
5542 return false;
5544 *base = *info.base;
5545 *offset = info.disp ? INTVAL (*info.disp) : 0;
5546 return true;
5549 /* Functions to model cache auto-prefetcher.
5551 Some of the CPUs have cache auto-prefetcher, which /seems/ to initiate
5552 memory prefetches if it sees instructions with consequitive memory accesses
5553 in the instruction stream. Details of such hardware units are not published,
5554 so we can only guess what exactly is going on there.
5555 In the scheduler, we model abstract auto-prefetcher. If there are memory
5556 insns in the ready list (or the queue) that have same memory base, but
5557 different offsets, then we delay the insns with larger offsets until insns
5558 with smaller offsets get scheduled. If PARAM_SCHED_AUTOPREF_QUEUE_DEPTH
5559 is "1", then we look at the ready list; if it is N>1, then we also look
5560 through N-1 queue entries.
5561 If the param is N>=0, then rank_for_schedule will consider auto-prefetching
5562 among its heuristics.
5563 Param value of "-1" disables modelling of the auto-prefetcher. */
5565 /* Initialize autoprefetcher model data for INSN. */
5566 static void
5567 autopref_multipass_init (const rtx_insn *insn, int write)
5569 autopref_multipass_data_t data = &INSN_AUTOPREF_MULTIPASS_DATA (insn)[write];
5571 gcc_assert (data->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED);
5572 data->base = NULL_RTX;
5573 data->offset = 0;
5574 /* Set insn entry initialized, but not relevant for auto-prefetcher. */
5575 data->status = AUTOPREF_MULTIPASS_DATA_IRRELEVANT;
5577 rtx pat = PATTERN (insn);
5579 /* We have a multi-set insn like a load-multiple or store-multiple.
5580 We care about these as long as all the memory ops inside the PARALLEL
5581 have the same base register. We care about the minimum and maximum
5582 offsets from that base but don't check for the order of those offsets
5583 within the PARALLEL insn itself. */
5584 if (GET_CODE (pat) == PARALLEL)
5586 int n_elems = XVECLEN (pat, 0);
5588 int i, offset;
5589 rtx base, prev_base = NULL_RTX;
5590 int min_offset = INT_MAX;
5592 for (i = 0; i < n_elems; i++)
5594 rtx set = XVECEXP (pat, 0, i);
5595 if (GET_CODE (set) != SET)
5596 return;
5598 if (!analyze_set_insn_for_autopref (set, write, &base, &offset))
5599 return;
5601 /* Ensure that all memory operations in the PARALLEL use the same
5602 base register. */
5603 if (i > 0 && REGNO (base) != REGNO (prev_base))
5604 return;
5605 prev_base = base;
5606 min_offset = MIN (min_offset, offset);
5609 /* If we reached here then we have a valid PARALLEL of multiple memory ops
5610 with prev_base as the base and min_offset containing the offset. */
5611 gcc_assert (prev_base);
5612 data->base = prev_base;
5613 data->offset = min_offset;
5614 data->status = AUTOPREF_MULTIPASS_DATA_NORMAL;
5615 return;
5618 /* Otherwise this is a single set memory operation. */
5619 rtx set = single_set (insn);
5620 if (set == NULL_RTX)
5621 return;
5623 if (!analyze_set_insn_for_autopref (set, write, &data->base,
5624 &data->offset))
5625 return;
5627 /* This insn is relevant for the auto-prefetcher.
5628 The base and offset fields will have been filled in the
5629 analyze_set_insn_for_autopref call above. */
5630 data->status = AUTOPREF_MULTIPASS_DATA_NORMAL;
5633 /* Helper function for rank_for_schedule sorting. */
5634 static int
5635 autopref_rank_for_schedule (const rtx_insn *insn1, const rtx_insn *insn2)
5637 int r = 0;
5638 for (int write = 0; write < 2 && !r; ++write)
5640 autopref_multipass_data_t data1
5641 = &INSN_AUTOPREF_MULTIPASS_DATA (insn1)[write];
5642 autopref_multipass_data_t data2
5643 = &INSN_AUTOPREF_MULTIPASS_DATA (insn2)[write];
5645 if (data1->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED)
5646 autopref_multipass_init (insn1, write);
5648 if (data2->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED)
5649 autopref_multipass_init (insn2, write);
5651 int irrel1 = data1->status == AUTOPREF_MULTIPASS_DATA_IRRELEVANT;
5652 int irrel2 = data2->status == AUTOPREF_MULTIPASS_DATA_IRRELEVANT;
5654 if (!irrel1 && !irrel2)
5655 r = data1->offset - data2->offset;
5656 else
5657 r = irrel2 - irrel1;
5660 return r;
5663 /* True if header of debug dump was printed. */
5664 static bool autopref_multipass_dfa_lookahead_guard_started_dump_p;
5666 /* Helper for autopref_multipass_dfa_lookahead_guard.
5667 Return "1" if INSN1 should be delayed in favor of INSN2. */
5668 static int
5669 autopref_multipass_dfa_lookahead_guard_1 (const rtx_insn *insn1,
5670 const rtx_insn *insn2, int write)
5672 autopref_multipass_data_t data1
5673 = &INSN_AUTOPREF_MULTIPASS_DATA (insn1)[write];
5674 autopref_multipass_data_t data2
5675 = &INSN_AUTOPREF_MULTIPASS_DATA (insn2)[write];
5677 if (data2->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED)
5678 autopref_multipass_init (insn2, write);
5679 if (data2->status == AUTOPREF_MULTIPASS_DATA_IRRELEVANT)
5680 return 0;
5682 if (rtx_equal_p (data1->base, data2->base)
5683 && data1->offset > data2->offset)
5685 if (sched_verbose >= 2)
5687 if (!autopref_multipass_dfa_lookahead_guard_started_dump_p)
5689 fprintf (sched_dump,
5690 ";;\t\tnot trying in max_issue due to autoprefetch "
5691 "model: ");
5692 autopref_multipass_dfa_lookahead_guard_started_dump_p = true;
5695 fprintf (sched_dump, " %d(%d)", INSN_UID (insn1), INSN_UID (insn2));
5698 return 1;
5701 return 0;
5704 /* General note:
5706 We could have also hooked autoprefetcher model into
5707 first_cycle_multipass_backtrack / first_cycle_multipass_issue hooks
5708 to enable intelligent selection of "[r1+0]=r2; [r1+4]=r3" on the same cycle
5709 (e.g., once "[r1+0]=r2" is issued in max_issue(), "[r1+4]=r3" gets
5710 unblocked). We don't bother about this yet because target of interest
5711 (ARM Cortex-A15) can issue only 1 memory operation per cycle. */
5713 /* Implementation of first_cycle_multipass_dfa_lookahead_guard hook.
5714 Return "1" if INSN1 should not be considered in max_issue due to
5715 auto-prefetcher considerations. */
5717 autopref_multipass_dfa_lookahead_guard (rtx_insn *insn1, int ready_index)
5719 int r = 0;
5721 /* Exit early if the param forbids this or if we're not entering here through
5722 normal haifa scheduling. This can happen if selective scheduling is
5723 explicitly enabled. */
5724 if (!insn_queue || PARAM_VALUE (PARAM_SCHED_AUTOPREF_QUEUE_DEPTH) <= 0)
5725 return 0;
5727 if (sched_verbose >= 2 && ready_index == 0)
5728 autopref_multipass_dfa_lookahead_guard_started_dump_p = false;
5730 for (int write = 0; write < 2; ++write)
5732 autopref_multipass_data_t data1
5733 = &INSN_AUTOPREF_MULTIPASS_DATA (insn1)[write];
5735 if (data1->status == AUTOPREF_MULTIPASS_DATA_UNINITIALIZED)
5736 autopref_multipass_init (insn1, write);
5737 if (data1->status == AUTOPREF_MULTIPASS_DATA_IRRELEVANT)
5738 continue;
5740 if (ready_index == 0
5741 && data1->status == AUTOPREF_MULTIPASS_DATA_DONT_DELAY)
5742 /* We allow only a single delay on priviledged instructions.
5743 Doing otherwise would cause infinite loop. */
5745 if (sched_verbose >= 2)
5747 if (!autopref_multipass_dfa_lookahead_guard_started_dump_p)
5749 fprintf (sched_dump,
5750 ";;\t\tnot trying in max_issue due to autoprefetch "
5751 "model: ");
5752 autopref_multipass_dfa_lookahead_guard_started_dump_p = true;
5755 fprintf (sched_dump, " *%d*", INSN_UID (insn1));
5757 continue;
5760 for (int i2 = 0; i2 < ready.n_ready; ++i2)
5762 rtx_insn *insn2 = get_ready_element (i2);
5763 if (insn1 == insn2)
5764 continue;
5765 r = autopref_multipass_dfa_lookahead_guard_1 (insn1, insn2, write);
5766 if (r)
5768 if (ready_index == 0)
5770 r = -1;
5771 data1->status = AUTOPREF_MULTIPASS_DATA_DONT_DELAY;
5773 goto finish;
5777 if (PARAM_VALUE (PARAM_SCHED_AUTOPREF_QUEUE_DEPTH) == 1)
5778 continue;
5780 /* Everything from the current queue slot should have been moved to
5781 the ready list. */
5782 gcc_assert (insn_queue[NEXT_Q_AFTER (q_ptr, 0)] == NULL_RTX);
5784 int n_stalls = PARAM_VALUE (PARAM_SCHED_AUTOPREF_QUEUE_DEPTH) - 1;
5785 if (n_stalls > max_insn_queue_index)
5786 n_stalls = max_insn_queue_index;
5788 for (int stalls = 1; stalls <= n_stalls; ++stalls)
5790 for (rtx_insn_list *link = insn_queue[NEXT_Q_AFTER (q_ptr, stalls)];
5791 link != NULL_RTX;
5792 link = link->next ())
5794 rtx_insn *insn2 = link->insn ();
5795 r = autopref_multipass_dfa_lookahead_guard_1 (insn1, insn2,
5796 write);
5797 if (r)
5799 /* Queue INSN1 until INSN2 can issue. */
5800 r = -stalls;
5801 if (ready_index == 0)
5802 data1->status = AUTOPREF_MULTIPASS_DATA_DONT_DELAY;
5803 goto finish;
5809 finish:
5810 if (sched_verbose >= 2
5811 && autopref_multipass_dfa_lookahead_guard_started_dump_p
5812 && (ready_index == ready.n_ready - 1 || r < 0))
5813 /* This does not /always/ trigger. We don't output EOL if the last
5814 insn is not recognized (INSN_CODE < 0) and lookahead_guard is not
5815 called. We can live with this. */
5816 fprintf (sched_dump, "\n");
5818 return r;
5821 /* Define type for target data used in multipass scheduling. */
5822 #ifndef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DATA_T
5823 # define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DATA_T int
5824 #endif
5825 typedef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DATA_T first_cycle_multipass_data_t;
5827 /* The following structure describe an entry of the stack of choices. */
5828 struct choice_entry
5830 /* Ordinal number of the issued insn in the ready queue. */
5831 int index;
5832 /* The number of the rest insns whose issues we should try. */
5833 int rest;
5834 /* The number of issued essential insns. */
5835 int n;
5836 /* State after issuing the insn. */
5837 state_t state;
5838 /* Target-specific data. */
5839 first_cycle_multipass_data_t target_data;
5842 /* The following array is used to implement a stack of choices used in
5843 function max_issue. */
5844 static struct choice_entry *choice_stack;
5846 /* This holds the value of the target dfa_lookahead hook. */
5847 int dfa_lookahead;
5849 /* The following variable value is maximal number of tries of issuing
5850 insns for the first cycle multipass insn scheduling. We define
5851 this value as constant*(DFA_LOOKAHEAD**ISSUE_RATE). We would not
5852 need this constraint if all real insns (with non-negative codes)
5853 had reservations because in this case the algorithm complexity is
5854 O(DFA_LOOKAHEAD**ISSUE_RATE). Unfortunately, the dfa descriptions
5855 might be incomplete and such insn might occur. For such
5856 descriptions, the complexity of algorithm (without the constraint)
5857 could achieve DFA_LOOKAHEAD ** N , where N is the queue length. */
5858 static int max_lookahead_tries;
5860 /* The following function returns maximal (or close to maximal) number
5861 of insns which can be issued on the same cycle and one of which
5862 insns is insns with the best rank (the first insn in READY). To
5863 make this function tries different samples of ready insns. READY
5864 is current queue `ready'. Global array READY_TRY reflects what
5865 insns are already issued in this try. The function stops immediately,
5866 if it reached the such a solution, that all instruction can be issued.
5867 INDEX will contain index of the best insn in READY. The following
5868 function is used only for first cycle multipass scheduling.
5870 PRIVILEGED_N >= 0
5872 This function expects recognized insns only. All USEs,
5873 CLOBBERs, etc must be filtered elsewhere. */
5875 max_issue (struct ready_list *ready, int privileged_n, state_t state,
5876 bool first_cycle_insn_p, int *index)
5878 int n, i, all, n_ready, best, delay, tries_num;
5879 int more_issue;
5880 struct choice_entry *top;
5881 rtx_insn *insn;
5883 if (sched_fusion)
5884 return 0;
5886 n_ready = ready->n_ready;
5887 gcc_assert (dfa_lookahead >= 1 && privileged_n >= 0
5888 && privileged_n <= n_ready);
5890 /* Init MAX_LOOKAHEAD_TRIES. */
5891 if (max_lookahead_tries == 0)
5893 max_lookahead_tries = 100;
5894 for (i = 0; i < issue_rate; i++)
5895 max_lookahead_tries *= dfa_lookahead;
5898 /* Init max_points. */
5899 more_issue = issue_rate - cycle_issued_insns;
5900 gcc_assert (more_issue >= 0);
5902 /* The number of the issued insns in the best solution. */
5903 best = 0;
5905 top = choice_stack;
5907 /* Set initial state of the search. */
5908 memcpy (top->state, state, dfa_state_size);
5909 top->rest = dfa_lookahead;
5910 top->n = 0;
5911 if (targetm.sched.first_cycle_multipass_begin)
5912 targetm.sched.first_cycle_multipass_begin (&top->target_data,
5913 ready_try, n_ready,
5914 first_cycle_insn_p);
5916 /* Count the number of the insns to search among. */
5917 for (all = i = 0; i < n_ready; i++)
5918 if (!ready_try [i])
5919 all++;
5921 if (sched_verbose >= 2)
5923 fprintf (sched_dump, ";;\t\tmax_issue among %d insns:", all);
5924 debug_ready_list_1 (ready, ready_try);
5927 /* I is the index of the insn to try next. */
5928 i = 0;
5929 tries_num = 0;
5930 for (;;)
5932 if (/* If we've reached a dead end or searched enough of what we have
5933 been asked... */
5934 top->rest == 0
5935 /* or have nothing else to try... */
5936 || i >= n_ready
5937 /* or should not issue more. */
5938 || top->n >= more_issue)
5940 /* ??? (... || i == n_ready). */
5941 gcc_assert (i <= n_ready);
5943 /* We should not issue more than issue_rate instructions. */
5944 gcc_assert (top->n <= more_issue);
5946 if (top == choice_stack)
5947 break;
5949 if (best < top - choice_stack)
5951 if (privileged_n)
5953 n = privileged_n;
5954 /* Try to find issued privileged insn. */
5955 while (n && !ready_try[--n])
5959 if (/* If all insns are equally good... */
5960 privileged_n == 0
5961 /* Or a privileged insn will be issued. */
5962 || ready_try[n])
5963 /* Then we have a solution. */
5965 best = top - choice_stack;
5966 /* This is the index of the insn issued first in this
5967 solution. */
5968 *index = choice_stack [1].index;
5969 if (top->n == more_issue || best == all)
5970 break;
5974 /* Set ready-list index to point to the last insn
5975 ('i++' below will advance it to the next insn). */
5976 i = top->index;
5978 /* Backtrack. */
5979 ready_try [i] = 0;
5981 if (targetm.sched.first_cycle_multipass_backtrack)
5982 targetm.sched.first_cycle_multipass_backtrack (&top->target_data,
5983 ready_try, n_ready);
5985 top--;
5986 memcpy (state, top->state, dfa_state_size);
5988 else if (!ready_try [i])
5990 tries_num++;
5991 if (tries_num > max_lookahead_tries)
5992 break;
5993 insn = ready_element (ready, i);
5994 delay = state_transition (state, insn);
5995 if (delay < 0)
5997 if (state_dead_lock_p (state)
5998 || insn_finishes_cycle_p (insn))
5999 /* We won't issue any more instructions in the next
6000 choice_state. */
6001 top->rest = 0;
6002 else
6003 top->rest--;
6005 n = top->n;
6006 if (memcmp (top->state, state, dfa_state_size) != 0)
6007 n++;
6009 /* Advance to the next choice_entry. */
6010 top++;
6011 /* Initialize it. */
6012 top->rest = dfa_lookahead;
6013 top->index = i;
6014 top->n = n;
6015 memcpy (top->state, state, dfa_state_size);
6016 ready_try [i] = 1;
6018 if (targetm.sched.first_cycle_multipass_issue)
6019 targetm.sched.first_cycle_multipass_issue (&top->target_data,
6020 ready_try, n_ready,
6021 insn,
6022 &((top - 1)
6023 ->target_data));
6025 i = -1;
6029 /* Increase ready-list index. */
6030 i++;
6033 if (targetm.sched.first_cycle_multipass_end)
6034 targetm.sched.first_cycle_multipass_end (best != 0
6035 ? &choice_stack[1].target_data
6036 : NULL);
6038 /* Restore the original state of the DFA. */
6039 memcpy (state, choice_stack->state, dfa_state_size);
6041 return best;
6044 /* The following function chooses insn from READY and modifies
6045 READY. The following function is used only for first
6046 cycle multipass scheduling.
6047 Return:
6048 -1 if cycle should be advanced,
6049 0 if INSN_PTR is set to point to the desirable insn,
6050 1 if choose_ready () should be restarted without advancing the cycle. */
6051 static int
6052 choose_ready (struct ready_list *ready, bool first_cycle_insn_p,
6053 rtx_insn **insn_ptr)
6055 if (dbg_cnt (sched_insn) == false)
6057 if (nonscheduled_insns_begin == NULL_RTX)
6058 nonscheduled_insns_begin = current_sched_info->prev_head;
6060 rtx_insn *insn = first_nonscheduled_insn ();
6062 if (QUEUE_INDEX (insn) == QUEUE_READY)
6063 /* INSN is in the ready_list. */
6065 ready_remove_insn (insn);
6066 *insn_ptr = insn;
6067 return 0;
6070 /* INSN is in the queue. Advance cycle to move it to the ready list. */
6071 gcc_assert (QUEUE_INDEX (insn) >= 0);
6072 return -1;
6075 if (dfa_lookahead <= 0 || SCHED_GROUP_P (ready_element (ready, 0))
6076 || DEBUG_INSN_P (ready_element (ready, 0)))
6078 if (targetm.sched.dispatch (NULL, IS_DISPATCH_ON))
6079 *insn_ptr = ready_remove_first_dispatch (ready);
6080 else
6081 *insn_ptr = ready_remove_first (ready);
6083 return 0;
6085 else
6087 /* Try to choose the best insn. */
6088 int index = 0, i;
6089 rtx_insn *insn;
6091 insn = ready_element (ready, 0);
6092 if (INSN_CODE (insn) < 0)
6094 *insn_ptr = ready_remove_first (ready);
6095 return 0;
6098 /* Filter the search space. */
6099 for (i = 0; i < ready->n_ready; i++)
6101 ready_try[i] = 0;
6103 insn = ready_element (ready, i);
6105 /* If this insn is recognizable we should have already
6106 recognized it earlier.
6107 ??? Not very clear where this is supposed to be done.
6108 See dep_cost_1. */
6109 gcc_checking_assert (INSN_CODE (insn) >= 0
6110 || recog_memoized (insn) < 0);
6111 if (INSN_CODE (insn) < 0)
6113 /* Non-recognized insns at position 0 are handled above. */
6114 gcc_assert (i > 0);
6115 ready_try[i] = 1;
6116 continue;
6119 if (targetm.sched.first_cycle_multipass_dfa_lookahead_guard)
6121 ready_try[i]
6122 = (targetm.sched.first_cycle_multipass_dfa_lookahead_guard
6123 (insn, i));
6125 if (ready_try[i] < 0)
6126 /* Queue instruction for several cycles.
6127 We need to restart choose_ready as we have changed
6128 the ready list. */
6130 change_queue_index (insn, -ready_try[i]);
6131 return 1;
6134 /* Make sure that we didn't end up with 0'th insn filtered out.
6135 Don't be tempted to make life easier for backends and just
6136 requeue 0'th insn if (ready_try[0] == 0) and restart
6137 choose_ready. Backends should be very considerate about
6138 requeueing instructions -- especially the highest priority
6139 one at position 0. */
6140 gcc_assert (ready_try[i] == 0 || i > 0);
6141 if (ready_try[i])
6142 continue;
6145 gcc_assert (ready_try[i] == 0);
6146 /* INSN made it through the scrutiny of filters! */
6149 if (max_issue (ready, 1, curr_state, first_cycle_insn_p, &index) == 0)
6151 *insn_ptr = ready_remove_first (ready);
6152 if (sched_verbose >= 4)
6153 fprintf (sched_dump, ";;\t\tChosen insn (but can't issue) : %s \n",
6154 (*current_sched_info->print_insn) (*insn_ptr, 0));
6155 return 0;
6157 else
6159 if (sched_verbose >= 4)
6160 fprintf (sched_dump, ";;\t\tChosen insn : %s\n",
6161 (*current_sched_info->print_insn)
6162 (ready_element (ready, index), 0));
6164 *insn_ptr = ready_remove (ready, index);
6165 return 0;
6170 /* This function is called when we have successfully scheduled a
6171 block. It uses the schedule stored in the scheduled_insns vector
6172 to rearrange the RTL. PREV_HEAD is used as the anchor to which we
6173 append the scheduled insns; TAIL is the insn after the scheduled
6174 block. TARGET_BB is the argument passed to schedule_block. */
6176 static void
6177 commit_schedule (rtx_insn *prev_head, rtx_insn *tail, basic_block *target_bb)
6179 unsigned int i;
6180 rtx_insn *insn;
6182 last_scheduled_insn = prev_head;
6183 for (i = 0;
6184 scheduled_insns.iterate (i, &insn);
6185 i++)
6187 if (control_flow_insn_p (last_scheduled_insn)
6188 || current_sched_info->advance_target_bb (*target_bb, insn))
6190 *target_bb = current_sched_info->advance_target_bb (*target_bb, 0);
6192 if (sched_verbose)
6194 rtx_insn *x;
6196 x = next_real_insn (last_scheduled_insn);
6197 gcc_assert (x);
6198 dump_new_block_header (1, *target_bb, x, tail);
6201 last_scheduled_insn = bb_note (*target_bb);
6204 if (current_sched_info->begin_move_insn)
6205 (*current_sched_info->begin_move_insn) (insn, last_scheduled_insn);
6206 move_insn (insn, last_scheduled_insn,
6207 current_sched_info->next_tail);
6208 if (!DEBUG_INSN_P (insn))
6209 reemit_notes (insn);
6210 last_scheduled_insn = insn;
6213 scheduled_insns.truncate (0);
6216 /* Examine all insns on the ready list and queue those which can't be
6217 issued in this cycle. TEMP_STATE is temporary scheduler state we
6218 can use as scratch space. If FIRST_CYCLE_INSN_P is true, no insns
6219 have been issued for the current cycle, which means it is valid to
6220 issue an asm statement.
6222 If SHADOWS_ONLY_P is true, we eliminate all real insns and only
6223 leave those for which SHADOW_P is true. If MODULO_EPILOGUE is true,
6224 we only leave insns which have an INSN_EXACT_TICK. */
6226 static void
6227 prune_ready_list (state_t temp_state, bool first_cycle_insn_p,
6228 bool shadows_only_p, bool modulo_epilogue_p)
6230 int i, pass;
6231 bool sched_group_found = false;
6232 int min_cost_group = 0;
6234 if (sched_fusion)
6235 return;
6237 for (i = 0; i < ready.n_ready; i++)
6239 rtx_insn *insn = ready_element (&ready, i);
6240 if (SCHED_GROUP_P (insn))
6242 sched_group_found = true;
6243 break;
6247 /* Make two passes if there's a SCHED_GROUP_P insn; make sure to handle
6248 such an insn first and note its cost. If at least one SCHED_GROUP_P insn
6249 gets queued, then all other insns get queued for one cycle later. */
6250 for (pass = sched_group_found ? 0 : 1; pass < 2; )
6252 int n = ready.n_ready;
6253 for (i = 0; i < n; i++)
6255 rtx_insn *insn = ready_element (&ready, i);
6256 int cost = 0;
6257 const char *reason = "resource conflict";
6259 if (DEBUG_INSN_P (insn))
6260 continue;
6262 if (sched_group_found && !SCHED_GROUP_P (insn)
6263 && ((pass == 0) || (min_cost_group >= 1)))
6265 if (pass == 0)
6266 continue;
6267 cost = min_cost_group;
6268 reason = "not in sched group";
6270 else if (modulo_epilogue_p
6271 && INSN_EXACT_TICK (insn) == INVALID_TICK)
6273 cost = max_insn_queue_index;
6274 reason = "not an epilogue insn";
6276 else if (shadows_only_p && !SHADOW_P (insn))
6278 cost = 1;
6279 reason = "not a shadow";
6281 else if (recog_memoized (insn) < 0)
6283 if (!first_cycle_insn_p
6284 && (GET_CODE (PATTERN (insn)) == ASM_INPUT
6285 || asm_noperands (PATTERN (insn)) >= 0))
6286 cost = 1;
6287 reason = "asm";
6289 else if (sched_pressure != SCHED_PRESSURE_NONE)
6291 if (sched_pressure == SCHED_PRESSURE_MODEL
6292 && INSN_TICK (insn) <= clock_var)
6294 memcpy (temp_state, curr_state, dfa_state_size);
6295 if (state_transition (temp_state, insn) >= 0)
6296 INSN_TICK (insn) = clock_var + 1;
6298 cost = 0;
6300 else
6302 int delay_cost = 0;
6304 if (delay_htab)
6306 struct delay_pair *delay_entry;
6307 delay_entry
6308 = delay_htab->find_with_hash (insn,
6309 htab_hash_pointer (insn));
6310 while (delay_entry && delay_cost == 0)
6312 delay_cost = estimate_shadow_tick (delay_entry);
6313 if (delay_cost > max_insn_queue_index)
6314 delay_cost = max_insn_queue_index;
6315 delay_entry = delay_entry->next_same_i1;
6319 memcpy (temp_state, curr_state, dfa_state_size);
6320 cost = state_transition (temp_state, insn);
6321 if (cost < 0)
6322 cost = 0;
6323 else if (cost == 0)
6324 cost = 1;
6325 if (cost < delay_cost)
6327 cost = delay_cost;
6328 reason = "shadow tick";
6331 if (cost >= 1)
6333 if (SCHED_GROUP_P (insn) && cost > min_cost_group)
6334 min_cost_group = cost;
6335 ready_remove (&ready, i);
6336 /* Normally we'd want to queue INSN for COST cycles. However,
6337 if SCHED_GROUP_P is set, then we must ensure that nothing
6338 else comes between INSN and its predecessor. If there is
6339 some other insn ready to fire on the next cycle, then that
6340 invariant would be broken.
6342 So when SCHED_GROUP_P is set, just queue this insn for a
6343 single cycle. */
6344 queue_insn (insn, SCHED_GROUP_P (insn) ? 1 : cost, reason);
6345 if (i + 1 < n)
6346 break;
6349 if (i == n)
6350 pass++;
6354 /* Called when we detect that the schedule is impossible. We examine the
6355 backtrack queue to find the earliest insn that caused this condition. */
6357 static struct haifa_saved_data *
6358 verify_shadows (void)
6360 struct haifa_saved_data *save, *earliest_fail = NULL;
6361 for (save = backtrack_queue; save; save = save->next)
6363 int t;
6364 struct delay_pair *pair = save->delay_pair;
6365 rtx_insn *i1 = pair->i1;
6367 for (; pair; pair = pair->next_same_i1)
6369 rtx_insn *i2 = pair->i2;
6371 if (QUEUE_INDEX (i2) == QUEUE_SCHEDULED)
6372 continue;
6374 t = INSN_TICK (i1) + pair_delay (pair);
6375 if (t < clock_var)
6377 if (sched_verbose >= 2)
6378 fprintf (sched_dump,
6379 ";;\t\tfailed delay requirements for %d/%d (%d->%d)"
6380 ", not ready\n",
6381 INSN_UID (pair->i1), INSN_UID (pair->i2),
6382 INSN_TICK (pair->i1), INSN_EXACT_TICK (pair->i2));
6383 earliest_fail = save;
6384 break;
6386 if (QUEUE_INDEX (i2) >= 0)
6388 int queued_for = INSN_TICK (i2);
6390 if (t < queued_for)
6392 if (sched_verbose >= 2)
6393 fprintf (sched_dump,
6394 ";;\t\tfailed delay requirements for %d/%d"
6395 " (%d->%d), queued too late\n",
6396 INSN_UID (pair->i1), INSN_UID (pair->i2),
6397 INSN_TICK (pair->i1), INSN_EXACT_TICK (pair->i2));
6398 earliest_fail = save;
6399 break;
6405 return earliest_fail;
6408 /* Print instructions together with useful scheduling information between
6409 HEAD and TAIL (inclusive). */
6410 static void
6411 dump_insn_stream (rtx_insn *head, rtx_insn *tail)
6413 fprintf (sched_dump, ";;\t| insn | prio |\n");
6415 rtx_insn *next_tail = NEXT_INSN (tail);
6416 for (rtx_insn *insn = head; insn != next_tail; insn = NEXT_INSN (insn))
6418 int priority = NOTE_P (insn) ? 0 : INSN_PRIORITY (insn);
6419 const char *pattern = (NOTE_P (insn)
6420 ? "note"
6421 : str_pattern_slim (PATTERN (insn)));
6423 fprintf (sched_dump, ";;\t| %4d | %4d | %-30s ",
6424 INSN_UID (insn), priority, pattern);
6426 if (sched_verbose >= 4)
6428 if (NOTE_P (insn) || LABEL_P (insn) || recog_memoized (insn) < 0)
6429 fprintf (sched_dump, "nothing");
6430 else
6431 print_reservation (sched_dump, insn);
6433 fprintf (sched_dump, "\n");
6437 /* Use forward list scheduling to rearrange insns of block pointed to by
6438 TARGET_BB, possibly bringing insns from subsequent blocks in the same
6439 region. */
6441 bool
6442 schedule_block (basic_block *target_bb, state_t init_state)
6444 int i;
6445 bool success = modulo_ii == 0;
6446 struct sched_block_state ls;
6447 state_t temp_state = NULL; /* It is used for multipass scheduling. */
6448 int sort_p, advance, start_clock_var;
6450 /* Head/tail info for this block. */
6451 rtx_insn *prev_head = current_sched_info->prev_head;
6452 rtx_insn *next_tail = current_sched_info->next_tail;
6453 rtx_insn *head = NEXT_INSN (prev_head);
6454 rtx_insn *tail = PREV_INSN (next_tail);
6456 if ((current_sched_info->flags & DONT_BREAK_DEPENDENCIES) == 0
6457 && sched_pressure != SCHED_PRESSURE_MODEL && !sched_fusion)
6458 find_modifiable_mems (head, tail);
6460 /* We used to have code to avoid getting parameters moved from hard
6461 argument registers into pseudos.
6463 However, it was removed when it proved to be of marginal benefit
6464 and caused problems because schedule_block and compute_forward_dependences
6465 had different notions of what the "head" insn was. */
6467 gcc_assert (head != tail || INSN_P (head));
6469 haifa_recovery_bb_recently_added_p = false;
6471 backtrack_queue = NULL;
6473 /* Debug info. */
6474 if (sched_verbose)
6476 dump_new_block_header (0, *target_bb, head, tail);
6478 if (sched_verbose >= 2)
6480 dump_insn_stream (head, tail);
6481 memset (&rank_for_schedule_stats, 0,
6482 sizeof (rank_for_schedule_stats));
6486 if (init_state == NULL)
6487 state_reset (curr_state);
6488 else
6489 memcpy (curr_state, init_state, dfa_state_size);
6491 /* Clear the ready list. */
6492 ready.first = ready.veclen - 1;
6493 ready.n_ready = 0;
6494 ready.n_debug = 0;
6496 /* It is used for first cycle multipass scheduling. */
6497 temp_state = alloca (dfa_state_size);
6499 if (targetm.sched.init)
6500 targetm.sched.init (sched_dump, sched_verbose, ready.veclen);
6502 /* We start inserting insns after PREV_HEAD. */
6503 last_scheduled_insn = prev_head;
6504 last_nondebug_scheduled_insn = NULL;
6505 nonscheduled_insns_begin = NULL;
6507 gcc_assert ((NOTE_P (last_scheduled_insn)
6508 || DEBUG_INSN_P (last_scheduled_insn))
6509 && BLOCK_FOR_INSN (last_scheduled_insn) == *target_bb);
6511 /* Initialize INSN_QUEUE. Q_SIZE is the total number of insns in the
6512 queue. */
6513 q_ptr = 0;
6514 q_size = 0;
6516 insn_queue = XALLOCAVEC (rtx_insn_list *, max_insn_queue_index + 1);
6517 memset (insn_queue, 0, (max_insn_queue_index + 1) * sizeof (rtx));
6519 /* Start just before the beginning of time. */
6520 clock_var = -1;
6522 /* We need queue and ready lists and clock_var be initialized
6523 in try_ready () (which is called through init_ready_list ()). */
6524 (*current_sched_info->init_ready_list) ();
6526 if (sched_pressure)
6527 sched_pressure_start_bb (*target_bb);
6529 /* The algorithm is O(n^2) in the number of ready insns at any given
6530 time in the worst case. Before reload we are more likely to have
6531 big lists so truncate them to a reasonable size. */
6532 if (!reload_completed
6533 && ready.n_ready - ready.n_debug > MAX_SCHED_READY_INSNS)
6535 ready_sort_debug (&ready);
6536 ready_sort_real (&ready);
6538 /* Find first free-standing insn past MAX_SCHED_READY_INSNS.
6539 If there are debug insns, we know they're first. */
6540 for (i = MAX_SCHED_READY_INSNS + ready.n_debug; i < ready.n_ready; i++)
6541 if (!SCHED_GROUP_P (ready_element (&ready, i)))
6542 break;
6544 if (sched_verbose >= 2)
6546 fprintf (sched_dump,
6547 ";;\t\tReady list on entry: %d insns: ", ready.n_ready);
6548 debug_ready_list (&ready);
6549 fprintf (sched_dump,
6550 ";;\t\t before reload => truncated to %d insns\n", i);
6553 /* Delay all insns past it for 1 cycle. If debug counter is
6554 activated make an exception for the insn right after
6555 nonscheduled_insns_begin. */
6557 rtx_insn *skip_insn;
6559 if (dbg_cnt (sched_insn) == false)
6560 skip_insn = first_nonscheduled_insn ();
6561 else
6562 skip_insn = NULL;
6564 while (i < ready.n_ready)
6566 rtx_insn *insn;
6568 insn = ready_remove (&ready, i);
6570 if (insn != skip_insn)
6571 queue_insn (insn, 1, "list truncated");
6573 if (skip_insn)
6574 ready_add (&ready, skip_insn, true);
6578 /* Now we can restore basic block notes and maintain precise cfg. */
6579 restore_bb_notes (*target_bb);
6581 last_clock_var = -1;
6583 advance = 0;
6585 gcc_assert (scheduled_insns.length () == 0);
6586 sort_p = TRUE;
6587 must_backtrack = false;
6588 modulo_insns_scheduled = 0;
6590 ls.modulo_epilogue = false;
6591 ls.first_cycle_insn_p = true;
6593 /* Loop until all the insns in BB are scheduled. */
6594 while ((*current_sched_info->schedule_more_p) ())
6596 perform_replacements_new_cycle ();
6599 start_clock_var = clock_var;
6601 clock_var++;
6603 advance_one_cycle ();
6605 /* Add to the ready list all pending insns that can be issued now.
6606 If there are no ready insns, increment clock until one
6607 is ready and add all pending insns at that point to the ready
6608 list. */
6609 queue_to_ready (&ready);
6611 gcc_assert (ready.n_ready);
6613 if (sched_verbose >= 2)
6615 fprintf (sched_dump, ";;\t\tReady list after queue_to_ready:");
6616 debug_ready_list (&ready);
6618 advance -= clock_var - start_clock_var;
6620 while (advance > 0);
6622 if (ls.modulo_epilogue)
6624 int stage = clock_var / modulo_ii;
6625 if (stage > modulo_last_stage * 2 + 2)
6627 if (sched_verbose >= 2)
6628 fprintf (sched_dump,
6629 ";;\t\tmodulo scheduled succeeded at II %d\n",
6630 modulo_ii);
6631 success = true;
6632 goto end_schedule;
6635 else if (modulo_ii > 0)
6637 int stage = clock_var / modulo_ii;
6638 if (stage > modulo_max_stages)
6640 if (sched_verbose >= 2)
6641 fprintf (sched_dump,
6642 ";;\t\tfailing schedule due to excessive stages\n");
6643 goto end_schedule;
6645 if (modulo_n_insns == modulo_insns_scheduled
6646 && stage > modulo_last_stage)
6648 if (sched_verbose >= 2)
6649 fprintf (sched_dump,
6650 ";;\t\tfound kernel after %d stages, II %d\n",
6651 stage, modulo_ii);
6652 ls.modulo_epilogue = true;
6656 prune_ready_list (temp_state, true, false, ls.modulo_epilogue);
6657 if (ready.n_ready == 0)
6658 continue;
6659 if (must_backtrack)
6660 goto do_backtrack;
6662 ls.shadows_only_p = false;
6663 cycle_issued_insns = 0;
6664 ls.can_issue_more = issue_rate;
6665 for (;;)
6667 rtx_insn *insn;
6668 int cost;
6669 bool asm_p;
6671 if (sort_p && ready.n_ready > 0)
6673 /* Sort the ready list based on priority. This must be
6674 done every iteration through the loop, as schedule_insn
6675 may have readied additional insns that will not be
6676 sorted correctly. */
6677 ready_sort (&ready);
6679 if (sched_verbose >= 2)
6681 fprintf (sched_dump,
6682 ";;\t\tReady list after ready_sort: ");
6683 debug_ready_list (&ready);
6687 /* We don't want md sched reorder to even see debug isns, so put
6688 them out right away. */
6689 if (ready.n_ready && DEBUG_INSN_P (ready_element (&ready, 0))
6690 && (*current_sched_info->schedule_more_p) ())
6692 while (ready.n_ready && DEBUG_INSN_P (ready_element (&ready, 0)))
6694 rtx_insn *insn = ready_remove_first (&ready);
6695 gcc_assert (DEBUG_INSN_P (insn));
6696 (*current_sched_info->begin_schedule_ready) (insn);
6697 scheduled_insns.safe_push (insn);
6698 last_scheduled_insn = insn;
6699 advance = schedule_insn (insn);
6700 gcc_assert (advance == 0);
6701 if (ready.n_ready > 0)
6702 ready_sort (&ready);
6706 if (ls.first_cycle_insn_p && !ready.n_ready)
6707 break;
6709 resume_after_backtrack:
6710 /* Allow the target to reorder the list, typically for
6711 better instruction bundling. */
6712 if (sort_p
6713 && (ready.n_ready == 0
6714 || !SCHED_GROUP_P (ready_element (&ready, 0))))
6716 if (ls.first_cycle_insn_p && targetm.sched.reorder)
6717 ls.can_issue_more
6718 = targetm.sched.reorder (sched_dump, sched_verbose,
6719 ready_lastpos (&ready),
6720 &ready.n_ready, clock_var);
6721 else if (!ls.first_cycle_insn_p && targetm.sched.reorder2)
6722 ls.can_issue_more
6723 = targetm.sched.reorder2 (sched_dump, sched_verbose,
6724 ready.n_ready
6725 ? ready_lastpos (&ready) : NULL,
6726 &ready.n_ready, clock_var);
6729 restart_choose_ready:
6730 if (sched_verbose >= 2)
6732 fprintf (sched_dump, ";;\tReady list (t = %3d): ",
6733 clock_var);
6734 debug_ready_list (&ready);
6735 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
6736 print_curr_reg_pressure ();
6739 if (ready.n_ready == 0
6740 && ls.can_issue_more
6741 && reload_completed)
6743 /* Allow scheduling insns directly from the queue in case
6744 there's nothing better to do (ready list is empty) but
6745 there are still vacant dispatch slots in the current cycle. */
6746 if (sched_verbose >= 6)
6747 fprintf (sched_dump,";;\t\tSecond chance\n");
6748 memcpy (temp_state, curr_state, dfa_state_size);
6749 if (early_queue_to_ready (temp_state, &ready))
6750 ready_sort (&ready);
6753 if (ready.n_ready == 0
6754 || !ls.can_issue_more
6755 || state_dead_lock_p (curr_state)
6756 || !(*current_sched_info->schedule_more_p) ())
6757 break;
6759 /* Select and remove the insn from the ready list. */
6760 if (sort_p)
6762 int res;
6764 insn = NULL;
6765 res = choose_ready (&ready, ls.first_cycle_insn_p, &insn);
6767 if (res < 0)
6768 /* Finish cycle. */
6769 break;
6770 if (res > 0)
6771 goto restart_choose_ready;
6773 gcc_assert (insn != NULL_RTX);
6775 else
6776 insn = ready_remove_first (&ready);
6778 if (sched_pressure != SCHED_PRESSURE_NONE
6779 && INSN_TICK (insn) > clock_var)
6781 ready_add (&ready, insn, true);
6782 advance = 1;
6783 break;
6786 if (targetm.sched.dfa_new_cycle
6787 && targetm.sched.dfa_new_cycle (sched_dump, sched_verbose,
6788 insn, last_clock_var,
6789 clock_var, &sort_p))
6790 /* SORT_P is used by the target to override sorting
6791 of the ready list. This is needed when the target
6792 has modified its internal structures expecting that
6793 the insn will be issued next. As we need the insn
6794 to have the highest priority (so it will be returned by
6795 the ready_remove_first call above), we invoke
6796 ready_add (&ready, insn, true).
6797 But, still, there is one issue: INSN can be later
6798 discarded by scheduler's front end through
6799 current_sched_info->can_schedule_ready_p, hence, won't
6800 be issued next. */
6802 ready_add (&ready, insn, true);
6803 break;
6806 sort_p = TRUE;
6808 if (current_sched_info->can_schedule_ready_p
6809 && ! (*current_sched_info->can_schedule_ready_p) (insn))
6810 /* We normally get here only if we don't want to move
6811 insn from the split block. */
6813 TODO_SPEC (insn) = DEP_POSTPONED;
6814 goto restart_choose_ready;
6817 if (delay_htab)
6819 /* If this insn is the first part of a delay-slot pair, record a
6820 backtrack point. */
6821 struct delay_pair *delay_entry;
6822 delay_entry
6823 = delay_htab->find_with_hash (insn, htab_hash_pointer (insn));
6824 if (delay_entry)
6826 save_backtrack_point (delay_entry, ls);
6827 if (sched_verbose >= 2)
6828 fprintf (sched_dump, ";;\t\tsaving backtrack point\n");
6832 /* DECISION is made. */
6834 if (modulo_ii > 0 && INSN_UID (insn) < modulo_iter0_max_uid)
6836 modulo_insns_scheduled++;
6837 modulo_last_stage = clock_var / modulo_ii;
6839 if (TODO_SPEC (insn) & SPECULATIVE)
6840 generate_recovery_code (insn);
6842 if (targetm.sched.dispatch (NULL, IS_DISPATCH_ON))
6843 targetm.sched.dispatch_do (insn, ADD_TO_DISPATCH_WINDOW);
6845 /* Update counters, etc in the scheduler's front end. */
6846 (*current_sched_info->begin_schedule_ready) (insn);
6847 scheduled_insns.safe_push (insn);
6848 gcc_assert (NONDEBUG_INSN_P (insn));
6849 last_nondebug_scheduled_insn = last_scheduled_insn = insn;
6851 if (recog_memoized (insn) >= 0)
6853 memcpy (temp_state, curr_state, dfa_state_size);
6854 cost = state_transition (curr_state, insn);
6855 if (sched_pressure != SCHED_PRESSURE_WEIGHTED && !sched_fusion)
6856 gcc_assert (cost < 0);
6857 if (memcmp (temp_state, curr_state, dfa_state_size) != 0)
6858 cycle_issued_insns++;
6859 asm_p = false;
6861 else
6862 asm_p = (GET_CODE (PATTERN (insn)) == ASM_INPUT
6863 || asm_noperands (PATTERN (insn)) >= 0);
6865 if (targetm.sched.variable_issue)
6866 ls.can_issue_more =
6867 targetm.sched.variable_issue (sched_dump, sched_verbose,
6868 insn, ls.can_issue_more);
6869 /* A naked CLOBBER or USE generates no instruction, so do
6870 not count them against the issue rate. */
6871 else if (GET_CODE (PATTERN (insn)) != USE
6872 && GET_CODE (PATTERN (insn)) != CLOBBER)
6873 ls.can_issue_more--;
6874 advance = schedule_insn (insn);
6876 if (SHADOW_P (insn))
6877 ls.shadows_only_p = true;
6879 /* After issuing an asm insn we should start a new cycle. */
6880 if (advance == 0 && asm_p)
6881 advance = 1;
6883 if (must_backtrack)
6884 break;
6886 if (advance != 0)
6887 break;
6889 ls.first_cycle_insn_p = false;
6890 if (ready.n_ready > 0)
6891 prune_ready_list (temp_state, false, ls.shadows_only_p,
6892 ls.modulo_epilogue);
6895 do_backtrack:
6896 if (!must_backtrack)
6897 for (i = 0; i < ready.n_ready; i++)
6899 rtx_insn *insn = ready_element (&ready, i);
6900 if (INSN_EXACT_TICK (insn) == clock_var)
6902 must_backtrack = true;
6903 clock_var++;
6904 break;
6907 if (must_backtrack && modulo_ii > 0)
6909 if (modulo_backtracks_left == 0)
6910 goto end_schedule;
6911 modulo_backtracks_left--;
6913 while (must_backtrack)
6915 struct haifa_saved_data *failed;
6916 rtx_insn *failed_insn;
6918 must_backtrack = false;
6919 failed = verify_shadows ();
6920 gcc_assert (failed);
6922 failed_insn = failed->delay_pair->i1;
6923 /* Clear these queues. */
6924 perform_replacements_new_cycle ();
6925 toggle_cancelled_flags (false);
6926 unschedule_insns_until (failed_insn);
6927 while (failed != backtrack_queue)
6928 free_topmost_backtrack_point (true);
6929 restore_last_backtrack_point (&ls);
6930 if (sched_verbose >= 2)
6931 fprintf (sched_dump, ";;\t\trewind to cycle %d\n", clock_var);
6932 /* Delay by at least a cycle. This could cause additional
6933 backtracking. */
6934 queue_insn (failed_insn, 1, "backtracked");
6935 advance = 0;
6936 if (must_backtrack)
6937 continue;
6938 if (ready.n_ready > 0)
6939 goto resume_after_backtrack;
6940 else
6942 if (clock_var == 0 && ls.first_cycle_insn_p)
6943 goto end_schedule;
6944 advance = 1;
6945 break;
6948 ls.first_cycle_insn_p = true;
6950 if (ls.modulo_epilogue)
6951 success = true;
6952 end_schedule:
6953 if (!ls.first_cycle_insn_p || advance)
6954 advance_one_cycle ();
6955 perform_replacements_new_cycle ();
6956 if (modulo_ii > 0)
6958 /* Once again, debug insn suckiness: they can be on the ready list
6959 even if they have unresolved dependencies. To make our view
6960 of the world consistent, remove such "ready" insns. */
6961 restart_debug_insn_loop:
6962 for (i = ready.n_ready - 1; i >= 0; i--)
6964 rtx_insn *x;
6966 x = ready_element (&ready, i);
6967 if (DEPS_LIST_FIRST (INSN_HARD_BACK_DEPS (x)) != NULL
6968 || DEPS_LIST_FIRST (INSN_SPEC_BACK_DEPS (x)) != NULL)
6970 ready_remove (&ready, i);
6971 goto restart_debug_insn_loop;
6974 for (i = ready.n_ready - 1; i >= 0; i--)
6976 rtx_insn *x;
6978 x = ready_element (&ready, i);
6979 resolve_dependencies (x);
6981 for (i = 0; i <= max_insn_queue_index; i++)
6983 rtx_insn_list *link;
6984 while ((link = insn_queue[i]) != NULL)
6986 rtx_insn *x = link->insn ();
6987 insn_queue[i] = link->next ();
6988 QUEUE_INDEX (x) = QUEUE_NOWHERE;
6989 free_INSN_LIST_node (link);
6990 resolve_dependencies (x);
6995 if (!success)
6996 undo_all_replacements ();
6998 /* Debug info. */
6999 if (sched_verbose)
7001 fprintf (sched_dump, ";;\tReady list (final): ");
7002 debug_ready_list (&ready);
7005 if (modulo_ii == 0 && current_sched_info->queue_must_finish_empty)
7006 /* Sanity check -- queue must be empty now. Meaningless if region has
7007 multiple bbs. */
7008 gcc_assert (!q_size && !ready.n_ready && !ready.n_debug);
7009 else if (modulo_ii == 0)
7011 /* We must maintain QUEUE_INDEX between blocks in region. */
7012 for (i = ready.n_ready - 1; i >= 0; i--)
7014 rtx_insn *x;
7016 x = ready_element (&ready, i);
7017 QUEUE_INDEX (x) = QUEUE_NOWHERE;
7018 TODO_SPEC (x) = HARD_DEP;
7021 if (q_size)
7022 for (i = 0; i <= max_insn_queue_index; i++)
7024 rtx_insn_list *link;
7025 for (link = insn_queue[i]; link; link = link->next ())
7027 rtx_insn *x;
7029 x = link->insn ();
7030 QUEUE_INDEX (x) = QUEUE_NOWHERE;
7031 TODO_SPEC (x) = HARD_DEP;
7033 free_INSN_LIST_list (&insn_queue[i]);
7037 if (sched_pressure == SCHED_PRESSURE_MODEL)
7038 model_end_schedule ();
7040 if (success)
7042 commit_schedule (prev_head, tail, target_bb);
7043 if (sched_verbose)
7044 fprintf (sched_dump, ";; total time = %d\n", clock_var);
7046 else
7047 last_scheduled_insn = tail;
7049 scheduled_insns.truncate (0);
7051 if (!current_sched_info->queue_must_finish_empty
7052 || haifa_recovery_bb_recently_added_p)
7054 /* INSN_TICK (minimum clock tick at which the insn becomes
7055 ready) may be not correct for the insn in the subsequent
7056 blocks of the region. We should use a correct value of
7057 `clock_var' or modify INSN_TICK. It is better to keep
7058 clock_var value equal to 0 at the start of a basic block.
7059 Therefore we modify INSN_TICK here. */
7060 fix_inter_tick (NEXT_INSN (prev_head), last_scheduled_insn);
7063 if (targetm.sched.finish)
7065 targetm.sched.finish (sched_dump, sched_verbose);
7066 /* Target might have added some instructions to the scheduled block
7067 in its md_finish () hook. These new insns don't have any data
7068 initialized and to identify them we extend h_i_d so that they'll
7069 get zero luids. */
7070 sched_extend_luids ();
7073 /* Update head/tail boundaries. */
7074 head = NEXT_INSN (prev_head);
7075 tail = last_scheduled_insn;
7077 if (sched_verbose)
7079 fprintf (sched_dump, ";; new head = %d\n;; new tail = %d\n",
7080 INSN_UID (head), INSN_UID (tail));
7082 if (sched_verbose >= 2)
7084 dump_insn_stream (head, tail);
7085 print_rank_for_schedule_stats (";; TOTAL ", &rank_for_schedule_stats,
7086 NULL);
7089 fprintf (sched_dump, "\n");
7092 head = restore_other_notes (head, NULL);
7094 current_sched_info->head = head;
7095 current_sched_info->tail = tail;
7097 free_backtrack_queue ();
7099 return success;
7102 /* Set_priorities: compute priority of each insn in the block. */
7105 set_priorities (rtx_insn *head, rtx_insn *tail)
7107 rtx_insn *insn;
7108 int n_insn;
7109 int sched_max_insns_priority =
7110 current_sched_info->sched_max_insns_priority;
7111 rtx_insn *prev_head;
7113 if (head == tail && ! INSN_P (head))
7114 gcc_unreachable ();
7116 n_insn = 0;
7118 prev_head = PREV_INSN (head);
7119 for (insn = tail; insn != prev_head; insn = PREV_INSN (insn))
7121 if (!INSN_P (insn))
7122 continue;
7124 n_insn++;
7125 (void) priority (insn);
7127 gcc_assert (INSN_PRIORITY_KNOWN (insn));
7129 sched_max_insns_priority = MAX (sched_max_insns_priority,
7130 INSN_PRIORITY (insn));
7133 current_sched_info->sched_max_insns_priority = sched_max_insns_priority;
7135 return n_insn;
7138 /* Set sched_dump and sched_verbose for the desired debugging output. */
7139 void
7140 setup_sched_dump (void)
7142 sched_verbose = sched_verbose_param;
7143 sched_dump = dump_file;
7144 if (!dump_file)
7145 sched_verbose = 0;
7148 /* Allocate data for register pressure sensitive scheduling. */
7149 static void
7150 alloc_global_sched_pressure_data (void)
7152 if (sched_pressure != SCHED_PRESSURE_NONE)
7154 int i, max_regno = max_reg_num ();
7156 if (sched_dump != NULL)
7157 /* We need info about pseudos for rtl dumps about pseudo
7158 classes and costs. */
7159 regstat_init_n_sets_and_refs ();
7160 ira_set_pseudo_classes (true, sched_verbose ? sched_dump : NULL);
7161 sched_regno_pressure_class
7162 = (enum reg_class *) xmalloc (max_regno * sizeof (enum reg_class));
7163 for (i = 0; i < max_regno; i++)
7164 sched_regno_pressure_class[i]
7165 = (i < FIRST_PSEUDO_REGISTER
7166 ? ira_pressure_class_translate[REGNO_REG_CLASS (i)]
7167 : ira_pressure_class_translate[reg_allocno_class (i)]);
7168 curr_reg_live = BITMAP_ALLOC (NULL);
7169 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
7171 saved_reg_live = BITMAP_ALLOC (NULL);
7172 region_ref_regs = BITMAP_ALLOC (NULL);
7174 if (sched_pressure == SCHED_PRESSURE_MODEL)
7175 tmp_bitmap = BITMAP_ALLOC (NULL);
7177 /* Calculate number of CALL_SAVED_REGS and FIXED_REGS in register classes
7178 that we calculate register pressure for. */
7179 for (int c = 0; c < ira_pressure_classes_num; ++c)
7181 enum reg_class cl = ira_pressure_classes[c];
7183 call_saved_regs_num[cl] = 0;
7184 fixed_regs_num[cl] = 0;
7186 for (int i = 0; i < ira_class_hard_regs_num[cl]; ++i)
7187 if (!call_used_regs[ira_class_hard_regs[cl][i]])
7188 ++call_saved_regs_num[cl];
7189 else if (fixed_regs[ira_class_hard_regs[cl][i]])
7190 ++fixed_regs_num[cl];
7195 /* Free data for register pressure sensitive scheduling. Also called
7196 from schedule_region when stopping sched-pressure early. */
7197 void
7198 free_global_sched_pressure_data (void)
7200 if (sched_pressure != SCHED_PRESSURE_NONE)
7202 if (regstat_n_sets_and_refs != NULL)
7203 regstat_free_n_sets_and_refs ();
7204 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
7206 BITMAP_FREE (region_ref_regs);
7207 BITMAP_FREE (saved_reg_live);
7209 if (sched_pressure == SCHED_PRESSURE_MODEL)
7210 BITMAP_FREE (tmp_bitmap);
7211 BITMAP_FREE (curr_reg_live);
7212 free (sched_regno_pressure_class);
7216 /* Initialize some global state for the scheduler. This function works
7217 with the common data shared between all the schedulers. It is called
7218 from the scheduler specific initialization routine. */
7220 void
7221 sched_init (void)
7223 /* Disable speculative loads in their presence if cc0 defined. */
7224 if (HAVE_cc0)
7225 flag_schedule_speculative_load = 0;
7227 if (targetm.sched.dispatch (NULL, IS_DISPATCH_ON))
7228 targetm.sched.dispatch_do (NULL, DISPATCH_INIT);
7230 if (live_range_shrinkage_p)
7231 sched_pressure = SCHED_PRESSURE_WEIGHTED;
7232 else if (flag_sched_pressure
7233 && !reload_completed
7234 && common_sched_info->sched_pass_id == SCHED_RGN_PASS)
7235 sched_pressure = ((enum sched_pressure_algorithm)
7236 PARAM_VALUE (PARAM_SCHED_PRESSURE_ALGORITHM));
7237 else
7238 sched_pressure = SCHED_PRESSURE_NONE;
7240 if (sched_pressure != SCHED_PRESSURE_NONE)
7241 ira_setup_eliminable_regset ();
7243 /* Initialize SPEC_INFO. */
7244 if (targetm.sched.set_sched_flags)
7246 spec_info = &spec_info_var;
7247 targetm.sched.set_sched_flags (spec_info);
7249 if (spec_info->mask != 0)
7251 spec_info->data_weakness_cutoff =
7252 (PARAM_VALUE (PARAM_SCHED_SPEC_PROB_CUTOFF) * MAX_DEP_WEAK) / 100;
7253 spec_info->control_weakness_cutoff =
7254 (PARAM_VALUE (PARAM_SCHED_SPEC_PROB_CUTOFF)
7255 * REG_BR_PROB_BASE) / 100;
7257 else
7258 /* So we won't read anything accidentally. */
7259 spec_info = NULL;
7262 else
7263 /* So we won't read anything accidentally. */
7264 spec_info = 0;
7266 /* Initialize issue_rate. */
7267 if (targetm.sched.issue_rate)
7268 issue_rate = targetm.sched.issue_rate ();
7269 else
7270 issue_rate = 1;
7272 if (targetm.sched.first_cycle_multipass_dfa_lookahead
7273 /* Don't use max_issue with reg_pressure scheduling. Multipass
7274 scheduling and reg_pressure scheduling undo each other's decisions. */
7275 && sched_pressure == SCHED_PRESSURE_NONE)
7276 dfa_lookahead = targetm.sched.first_cycle_multipass_dfa_lookahead ();
7277 else
7278 dfa_lookahead = 0;
7280 /* Set to "0" so that we recalculate. */
7281 max_lookahead_tries = 0;
7283 if (targetm.sched.init_dfa_pre_cycle_insn)
7284 targetm.sched.init_dfa_pre_cycle_insn ();
7286 if (targetm.sched.init_dfa_post_cycle_insn)
7287 targetm.sched.init_dfa_post_cycle_insn ();
7289 dfa_start ();
7290 dfa_state_size = state_size ();
7292 init_alias_analysis ();
7294 if (!sched_no_dce)
7295 df_set_flags (DF_LR_RUN_DCE);
7296 df_note_add_problem ();
7298 /* More problems needed for interloop dep calculation in SMS. */
7299 if (common_sched_info->sched_pass_id == SCHED_SMS_PASS)
7301 df_rd_add_problem ();
7302 df_chain_add_problem (DF_DU_CHAIN + DF_UD_CHAIN);
7305 df_analyze ();
7307 /* Do not run DCE after reload, as this can kill nops inserted
7308 by bundling. */
7309 if (reload_completed)
7310 df_clear_flags (DF_LR_RUN_DCE);
7312 regstat_compute_calls_crossed ();
7314 if (targetm.sched.init_global)
7315 targetm.sched.init_global (sched_dump, sched_verbose, get_max_uid () + 1);
7317 alloc_global_sched_pressure_data ();
7319 curr_state = xmalloc (dfa_state_size);
7322 static void haifa_init_only_bb (basic_block, basic_block);
7324 /* Initialize data structures specific to the Haifa scheduler. */
7325 void
7326 haifa_sched_init (void)
7328 setup_sched_dump ();
7329 sched_init ();
7331 scheduled_insns.create (0);
7333 if (spec_info != NULL)
7335 sched_deps_info->use_deps_list = 1;
7336 sched_deps_info->generate_spec_deps = 1;
7339 /* Initialize luids, dependency caches, target and h_i_d for the
7340 whole function. */
7342 sched_init_bbs ();
7344 auto_vec<basic_block> bbs (n_basic_blocks_for_fn (cfun));
7345 basic_block bb;
7346 FOR_EACH_BB_FN (bb, cfun)
7347 bbs.quick_push (bb);
7348 sched_init_luids (bbs);
7349 sched_deps_init (true);
7350 sched_extend_target ();
7351 haifa_init_h_i_d (bbs);
7354 sched_init_only_bb = haifa_init_only_bb;
7355 sched_split_block = sched_split_block_1;
7356 sched_create_empty_bb = sched_create_empty_bb_1;
7357 haifa_recovery_bb_ever_added_p = false;
7359 nr_begin_data = nr_begin_control = nr_be_in_data = nr_be_in_control = 0;
7360 before_recovery = 0;
7361 after_recovery = 0;
7363 modulo_ii = 0;
7366 /* Finish work with the data specific to the Haifa scheduler. */
7367 void
7368 haifa_sched_finish (void)
7370 sched_create_empty_bb = NULL;
7371 sched_split_block = NULL;
7372 sched_init_only_bb = NULL;
7374 if (spec_info && spec_info->dump)
7376 char c = reload_completed ? 'a' : 'b';
7378 fprintf (spec_info->dump,
7379 ";; %s:\n", current_function_name ());
7381 fprintf (spec_info->dump,
7382 ";; Procedure %cr-begin-data-spec motions == %d\n",
7383 c, nr_begin_data);
7384 fprintf (spec_info->dump,
7385 ";; Procedure %cr-be-in-data-spec motions == %d\n",
7386 c, nr_be_in_data);
7387 fprintf (spec_info->dump,
7388 ";; Procedure %cr-begin-control-spec motions == %d\n",
7389 c, nr_begin_control);
7390 fprintf (spec_info->dump,
7391 ";; Procedure %cr-be-in-control-spec motions == %d\n",
7392 c, nr_be_in_control);
7395 scheduled_insns.release ();
7397 /* Finalize h_i_d, dependency caches, and luids for the whole
7398 function. Target will be finalized in md_global_finish (). */
7399 sched_deps_finish ();
7400 sched_finish_luids ();
7401 current_sched_info = NULL;
7402 insn_queue = NULL;
7403 sched_finish ();
7406 /* Free global data used during insn scheduling. This function works with
7407 the common data shared between the schedulers. */
7409 void
7410 sched_finish (void)
7412 haifa_finish_h_i_d ();
7413 free_global_sched_pressure_data ();
7414 free (curr_state);
7416 if (targetm.sched.finish_global)
7417 targetm.sched.finish_global (sched_dump, sched_verbose);
7419 end_alias_analysis ();
7421 regstat_free_calls_crossed ();
7423 dfa_finish ();
7426 /* Free all delay_pair structures that were recorded. */
7427 void
7428 free_delay_pairs (void)
7430 if (delay_htab)
7432 delay_htab->empty ();
7433 delay_htab_i2->empty ();
7437 /* Fix INSN_TICKs of the instructions in the current block as well as
7438 INSN_TICKs of their dependents.
7439 HEAD and TAIL are the begin and the end of the current scheduled block. */
7440 static void
7441 fix_inter_tick (rtx_insn *head, rtx_insn *tail)
7443 /* Set of instructions with corrected INSN_TICK. */
7444 auto_bitmap processed;
7445 /* ??? It is doubtful if we should assume that cycle advance happens on
7446 basic block boundaries. Basically insns that are unconditionally ready
7447 on the start of the block are more preferable then those which have
7448 a one cycle dependency over insn from the previous block. */
7449 int next_clock = clock_var + 1;
7451 /* Iterates over scheduled instructions and fix their INSN_TICKs and
7452 INSN_TICKs of dependent instructions, so that INSN_TICKs are consistent
7453 across different blocks. */
7454 for (tail = NEXT_INSN (tail); head != tail; head = NEXT_INSN (head))
7456 if (INSN_P (head))
7458 int tick;
7459 sd_iterator_def sd_it;
7460 dep_t dep;
7462 tick = INSN_TICK (head);
7463 gcc_assert (tick >= MIN_TICK);
7465 /* Fix INSN_TICK of instruction from just scheduled block. */
7466 if (bitmap_set_bit (processed, INSN_LUID (head)))
7468 tick -= next_clock;
7470 if (tick < MIN_TICK)
7471 tick = MIN_TICK;
7473 INSN_TICK (head) = tick;
7476 if (DEBUG_INSN_P (head))
7477 continue;
7479 FOR_EACH_DEP (head, SD_LIST_RES_FORW, sd_it, dep)
7481 rtx_insn *next;
7483 next = DEP_CON (dep);
7484 tick = INSN_TICK (next);
7486 if (tick != INVALID_TICK
7487 /* If NEXT has its INSN_TICK calculated, fix it.
7488 If not - it will be properly calculated from
7489 scratch later in fix_tick_ready. */
7490 && bitmap_set_bit (processed, INSN_LUID (next)))
7492 tick -= next_clock;
7494 if (tick < MIN_TICK)
7495 tick = MIN_TICK;
7497 if (tick > INTER_TICK (next))
7498 INTER_TICK (next) = tick;
7499 else
7500 tick = INTER_TICK (next);
7502 INSN_TICK (next) = tick;
7509 /* Check if NEXT is ready to be added to the ready or queue list.
7510 If "yes", add it to the proper list.
7511 Returns:
7512 -1 - is not ready yet,
7513 0 - added to the ready list,
7514 0 < N - queued for N cycles. */
7516 try_ready (rtx_insn *next)
7518 ds_t old_ts, new_ts;
7520 old_ts = TODO_SPEC (next);
7522 gcc_assert (!(old_ts & ~(SPECULATIVE | HARD_DEP | DEP_CONTROL | DEP_POSTPONED))
7523 && (old_ts == HARD_DEP
7524 || old_ts == DEP_POSTPONED
7525 || (old_ts & SPECULATIVE)
7526 || old_ts == DEP_CONTROL));
7528 new_ts = recompute_todo_spec (next, false);
7530 if (new_ts & (HARD_DEP | DEP_POSTPONED))
7531 gcc_assert (new_ts == old_ts
7532 && QUEUE_INDEX (next) == QUEUE_NOWHERE);
7533 else if (current_sched_info->new_ready)
7534 new_ts = current_sched_info->new_ready (next, new_ts);
7536 /* * if !(old_ts & SPECULATIVE) (e.g. HARD_DEP or 0), then insn might
7537 have its original pattern or changed (speculative) one. This is due
7538 to changing ebb in region scheduling.
7539 * But if (old_ts & SPECULATIVE), then we are pretty sure that insn
7540 has speculative pattern.
7542 We can't assert (!(new_ts & HARD_DEP) || new_ts == old_ts) here because
7543 control-speculative NEXT could have been discarded by sched-rgn.c
7544 (the same case as when discarded by can_schedule_ready_p ()). */
7546 if ((new_ts & SPECULATIVE)
7547 /* If (old_ts == new_ts), then (old_ts & SPECULATIVE) and we don't
7548 need to change anything. */
7549 && new_ts != old_ts)
7551 int res;
7552 rtx new_pat;
7554 gcc_assert ((new_ts & SPECULATIVE) && !(new_ts & ~SPECULATIVE));
7556 res = haifa_speculate_insn (next, new_ts, &new_pat);
7558 switch (res)
7560 case -1:
7561 /* It would be nice to change DEP_STATUS of all dependences,
7562 which have ((DEP_STATUS & SPECULATIVE) == new_ts) to HARD_DEP,
7563 so we won't reanalyze anything. */
7564 new_ts = HARD_DEP;
7565 break;
7567 case 0:
7568 /* We follow the rule, that every speculative insn
7569 has non-null ORIG_PAT. */
7570 if (!ORIG_PAT (next))
7571 ORIG_PAT (next) = PATTERN (next);
7572 break;
7574 case 1:
7575 if (!ORIG_PAT (next))
7576 /* If we gonna to overwrite the original pattern of insn,
7577 save it. */
7578 ORIG_PAT (next) = PATTERN (next);
7580 res = haifa_change_pattern (next, new_pat);
7581 gcc_assert (res);
7582 break;
7584 default:
7585 gcc_unreachable ();
7589 /* We need to restore pattern only if (new_ts == 0), because otherwise it is
7590 either correct (new_ts & SPECULATIVE),
7591 or we simply don't care (new_ts & HARD_DEP). */
7593 gcc_assert (!ORIG_PAT (next)
7594 || !IS_SPECULATION_BRANCHY_CHECK_P (next));
7596 TODO_SPEC (next) = new_ts;
7598 if (new_ts & (HARD_DEP | DEP_POSTPONED))
7600 /* We can't assert (QUEUE_INDEX (next) == QUEUE_NOWHERE) here because
7601 control-speculative NEXT could have been discarded by sched-rgn.c
7602 (the same case as when discarded by can_schedule_ready_p ()). */
7603 /*gcc_assert (QUEUE_INDEX (next) == QUEUE_NOWHERE);*/
7605 change_queue_index (next, QUEUE_NOWHERE);
7607 return -1;
7609 else if (!(new_ts & BEGIN_SPEC)
7610 && ORIG_PAT (next) && PREDICATED_PAT (next) == NULL_RTX
7611 && !IS_SPECULATION_CHECK_P (next))
7612 /* We should change pattern of every previously speculative
7613 instruction - and we determine if NEXT was speculative by using
7614 ORIG_PAT field. Except one case - speculation checks have ORIG_PAT
7615 pat too, so skip them. */
7617 bool success = haifa_change_pattern (next, ORIG_PAT (next));
7618 gcc_assert (success);
7619 ORIG_PAT (next) = 0;
7622 if (sched_verbose >= 2)
7624 fprintf (sched_dump, ";;\t\tdependencies resolved: insn %s",
7625 (*current_sched_info->print_insn) (next, 0));
7627 if (spec_info && spec_info->dump)
7629 if (new_ts & BEGIN_DATA)
7630 fprintf (spec_info->dump, "; data-spec;");
7631 if (new_ts & BEGIN_CONTROL)
7632 fprintf (spec_info->dump, "; control-spec;");
7633 if (new_ts & BE_IN_CONTROL)
7634 fprintf (spec_info->dump, "; in-control-spec;");
7636 if (TODO_SPEC (next) & DEP_CONTROL)
7637 fprintf (sched_dump, " predicated");
7638 fprintf (sched_dump, "\n");
7641 adjust_priority (next);
7643 return fix_tick_ready (next);
7646 /* Calculate INSN_TICK of NEXT and add it to either ready or queue list. */
7647 static int
7648 fix_tick_ready (rtx_insn *next)
7650 int tick, delay;
7652 if (!DEBUG_INSN_P (next) && !sd_lists_empty_p (next, SD_LIST_RES_BACK))
7654 int full_p;
7655 sd_iterator_def sd_it;
7656 dep_t dep;
7658 tick = INSN_TICK (next);
7659 /* if tick is not equal to INVALID_TICK, then update
7660 INSN_TICK of NEXT with the most recent resolved dependence
7661 cost. Otherwise, recalculate from scratch. */
7662 full_p = (tick == INVALID_TICK);
7664 FOR_EACH_DEP (next, SD_LIST_RES_BACK, sd_it, dep)
7666 rtx_insn *pro = DEP_PRO (dep);
7667 int tick1;
7669 gcc_assert (INSN_TICK (pro) >= MIN_TICK);
7671 tick1 = INSN_TICK (pro) + dep_cost (dep);
7672 if (tick1 > tick)
7673 tick = tick1;
7675 if (!full_p)
7676 break;
7679 else
7680 tick = -1;
7682 INSN_TICK (next) = tick;
7684 delay = tick - clock_var;
7685 if (delay <= 0 || sched_pressure != SCHED_PRESSURE_NONE || sched_fusion)
7686 delay = QUEUE_READY;
7688 change_queue_index (next, delay);
7690 return delay;
7693 /* Move NEXT to the proper queue list with (DELAY >= 1),
7694 or add it to the ready list (DELAY == QUEUE_READY),
7695 or remove it from ready and queue lists at all (DELAY == QUEUE_NOWHERE). */
7696 static void
7697 change_queue_index (rtx_insn *next, int delay)
7699 int i = QUEUE_INDEX (next);
7701 gcc_assert (QUEUE_NOWHERE <= delay && delay <= max_insn_queue_index
7702 && delay != 0);
7703 gcc_assert (i != QUEUE_SCHEDULED);
7705 if ((delay > 0 && NEXT_Q_AFTER (q_ptr, delay) == i)
7706 || (delay < 0 && delay == i))
7707 /* We have nothing to do. */
7708 return;
7710 /* Remove NEXT from wherever it is now. */
7711 if (i == QUEUE_READY)
7712 ready_remove_insn (next);
7713 else if (i >= 0)
7714 queue_remove (next);
7716 /* Add it to the proper place. */
7717 if (delay == QUEUE_READY)
7718 ready_add (readyp, next, false);
7719 else if (delay >= 1)
7720 queue_insn (next, delay, "change queue index");
7722 if (sched_verbose >= 2)
7724 fprintf (sched_dump, ";;\t\ttick updated: insn %s",
7725 (*current_sched_info->print_insn) (next, 0));
7727 if (delay == QUEUE_READY)
7728 fprintf (sched_dump, " into ready\n");
7729 else if (delay >= 1)
7730 fprintf (sched_dump, " into queue with cost=%d\n", delay);
7731 else
7732 fprintf (sched_dump, " removed from ready or queue lists\n");
7736 static int sched_ready_n_insns = -1;
7738 /* Initialize per region data structures. */
7739 void
7740 sched_extend_ready_list (int new_sched_ready_n_insns)
7742 int i;
7744 if (sched_ready_n_insns == -1)
7745 /* At the first call we need to initialize one more choice_stack
7746 entry. */
7748 i = 0;
7749 sched_ready_n_insns = 0;
7750 scheduled_insns.reserve (new_sched_ready_n_insns);
7752 else
7753 i = sched_ready_n_insns + 1;
7755 ready.veclen = new_sched_ready_n_insns + issue_rate;
7756 ready.vec = XRESIZEVEC (rtx_insn *, ready.vec, ready.veclen);
7758 gcc_assert (new_sched_ready_n_insns >= sched_ready_n_insns);
7760 ready_try = (signed char *) xrecalloc (ready_try, new_sched_ready_n_insns,
7761 sched_ready_n_insns,
7762 sizeof (*ready_try));
7764 /* We allocate +1 element to save initial state in the choice_stack[0]
7765 entry. */
7766 choice_stack = XRESIZEVEC (struct choice_entry, choice_stack,
7767 new_sched_ready_n_insns + 1);
7769 for (; i <= new_sched_ready_n_insns; i++)
7771 choice_stack[i].state = xmalloc (dfa_state_size);
7773 if (targetm.sched.first_cycle_multipass_init)
7774 targetm.sched.first_cycle_multipass_init (&(choice_stack[i]
7775 .target_data));
7778 sched_ready_n_insns = new_sched_ready_n_insns;
7781 /* Free per region data structures. */
7782 void
7783 sched_finish_ready_list (void)
7785 int i;
7787 free (ready.vec);
7788 ready.vec = NULL;
7789 ready.veclen = 0;
7791 free (ready_try);
7792 ready_try = NULL;
7794 for (i = 0; i <= sched_ready_n_insns; i++)
7796 if (targetm.sched.first_cycle_multipass_fini)
7797 targetm.sched.first_cycle_multipass_fini (&(choice_stack[i]
7798 .target_data));
7800 free (choice_stack [i].state);
7802 free (choice_stack);
7803 choice_stack = NULL;
7805 sched_ready_n_insns = -1;
7808 static int
7809 haifa_luid_for_non_insn (rtx x)
7811 gcc_assert (NOTE_P (x) || LABEL_P (x));
7813 return 0;
7816 /* Generates recovery code for INSN. */
7817 static void
7818 generate_recovery_code (rtx_insn *insn)
7820 if (TODO_SPEC (insn) & BEGIN_SPEC)
7821 begin_speculative_block (insn);
7823 /* Here we have insn with no dependencies to
7824 instructions other then CHECK_SPEC ones. */
7826 if (TODO_SPEC (insn) & BE_IN_SPEC)
7827 add_to_speculative_block (insn);
7830 /* Helper function.
7831 Tries to add speculative dependencies of type FS between instructions
7832 in deps_list L and TWIN. */
7833 static void
7834 process_insn_forw_deps_be_in_spec (rtx_insn *insn, rtx_insn *twin, ds_t fs)
7836 sd_iterator_def sd_it;
7837 dep_t dep;
7839 FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep)
7841 ds_t ds;
7842 rtx_insn *consumer;
7844 consumer = DEP_CON (dep);
7846 ds = DEP_STATUS (dep);
7848 if (/* If we want to create speculative dep. */
7850 /* And we can do that because this is a true dep. */
7851 && (ds & DEP_TYPES) == DEP_TRUE)
7853 gcc_assert (!(ds & BE_IN_SPEC));
7855 if (/* If this dep can be overcome with 'begin speculation'. */
7856 ds & BEGIN_SPEC)
7857 /* Then we have a choice: keep the dep 'begin speculative'
7858 or transform it into 'be in speculative'. */
7860 if (/* In try_ready we assert that if insn once became ready
7861 it can be removed from the ready (or queue) list only
7862 due to backend decision. Hence we can't let the
7863 probability of the speculative dep to decrease. */
7864 ds_weak (ds) <= ds_weak (fs))
7866 ds_t new_ds;
7868 new_ds = (ds & ~BEGIN_SPEC) | fs;
7870 if (/* consumer can 'be in speculative'. */
7871 sched_insn_is_legitimate_for_speculation_p (consumer,
7872 new_ds))
7873 /* Transform it to be in speculative. */
7874 ds = new_ds;
7877 else
7878 /* Mark the dep as 'be in speculative'. */
7879 ds |= fs;
7883 dep_def _new_dep, *new_dep = &_new_dep;
7885 init_dep_1 (new_dep, twin, consumer, DEP_TYPE (dep), ds);
7886 sd_add_dep (new_dep, false);
7891 /* Generates recovery code for BEGIN speculative INSN. */
7892 static void
7893 begin_speculative_block (rtx_insn *insn)
7895 if (TODO_SPEC (insn) & BEGIN_DATA)
7896 nr_begin_data++;
7897 if (TODO_SPEC (insn) & BEGIN_CONTROL)
7898 nr_begin_control++;
7900 create_check_block_twin (insn, false);
7902 TODO_SPEC (insn) &= ~BEGIN_SPEC;
7905 static void haifa_init_insn (rtx_insn *);
7907 /* Generates recovery code for BE_IN speculative INSN. */
7908 static void
7909 add_to_speculative_block (rtx_insn *insn)
7911 ds_t ts;
7912 sd_iterator_def sd_it;
7913 dep_t dep;
7914 auto_vec<rtx_insn *, 10> twins;
7916 ts = TODO_SPEC (insn);
7917 gcc_assert (!(ts & ~BE_IN_SPEC));
7919 if (ts & BE_IN_DATA)
7920 nr_be_in_data++;
7921 if (ts & BE_IN_CONTROL)
7922 nr_be_in_control++;
7924 TODO_SPEC (insn) &= ~BE_IN_SPEC;
7925 gcc_assert (!TODO_SPEC (insn));
7927 DONE_SPEC (insn) |= ts;
7929 /* First we convert all simple checks to branchy. */
7930 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
7931 sd_iterator_cond (&sd_it, &dep);)
7933 rtx_insn *check = DEP_PRO (dep);
7935 if (IS_SPECULATION_SIMPLE_CHECK_P (check))
7937 create_check_block_twin (check, true);
7939 /* Restart search. */
7940 sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
7942 else
7943 /* Continue search. */
7944 sd_iterator_next (&sd_it);
7947 auto_vec<rtx_insn *> priorities_roots;
7948 clear_priorities (insn, &priorities_roots);
7950 while (1)
7952 rtx_insn *check, *twin;
7953 basic_block rec;
7955 /* Get the first backward dependency of INSN. */
7956 sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
7957 if (!sd_iterator_cond (&sd_it, &dep))
7958 /* INSN has no backward dependencies left. */
7959 break;
7961 gcc_assert ((DEP_STATUS (dep) & BEGIN_SPEC) == 0
7962 && (DEP_STATUS (dep) & BE_IN_SPEC) != 0
7963 && (DEP_STATUS (dep) & DEP_TYPES) == DEP_TRUE);
7965 check = DEP_PRO (dep);
7967 gcc_assert (!IS_SPECULATION_CHECK_P (check) && !ORIG_PAT (check)
7968 && QUEUE_INDEX (check) == QUEUE_NOWHERE);
7970 rec = BLOCK_FOR_INSN (check);
7972 twin = emit_insn_before (copy_insn (PATTERN (insn)), BB_END (rec));
7973 haifa_init_insn (twin);
7975 sd_copy_back_deps (twin, insn, true);
7977 if (sched_verbose && spec_info->dump)
7978 /* INSN_BB (insn) isn't determined for twin insns yet.
7979 So we can't use current_sched_info->print_insn. */
7980 fprintf (spec_info->dump, ";;\t\tGenerated twin insn : %d/rec%d\n",
7981 INSN_UID (twin), rec->index);
7983 twins.safe_push (twin);
7985 /* Add dependences between TWIN and all appropriate
7986 instructions from REC. */
7987 FOR_EACH_DEP (insn, SD_LIST_SPEC_BACK, sd_it, dep)
7989 rtx_insn *pro = DEP_PRO (dep);
7991 gcc_assert (DEP_TYPE (dep) == REG_DEP_TRUE);
7993 /* INSN might have dependencies from the instructions from
7994 several recovery blocks. At this iteration we process those
7995 producers that reside in REC. */
7996 if (BLOCK_FOR_INSN (pro) == rec)
7998 dep_def _new_dep, *new_dep = &_new_dep;
8000 init_dep (new_dep, pro, twin, REG_DEP_TRUE);
8001 sd_add_dep (new_dep, false);
8005 process_insn_forw_deps_be_in_spec (insn, twin, ts);
8007 /* Remove all dependencies between INSN and insns in REC. */
8008 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
8009 sd_iterator_cond (&sd_it, &dep);)
8011 rtx_insn *pro = DEP_PRO (dep);
8013 if (BLOCK_FOR_INSN (pro) == rec)
8014 sd_delete_dep (sd_it);
8015 else
8016 sd_iterator_next (&sd_it);
8020 /* We couldn't have added the dependencies between INSN and TWINS earlier
8021 because that would make TWINS appear in the INSN_BACK_DEPS (INSN). */
8022 unsigned int i;
8023 rtx_insn *twin;
8024 FOR_EACH_VEC_ELT_REVERSE (twins, i, twin)
8026 dep_def _new_dep, *new_dep = &_new_dep;
8028 init_dep (new_dep, insn, twin, REG_DEP_OUTPUT);
8029 sd_add_dep (new_dep, false);
8032 calc_priorities (priorities_roots);
8035 /* Extends and fills with zeros (only the new part) array pointed to by P. */
8036 void *
8037 xrecalloc (void *p, size_t new_nmemb, size_t old_nmemb, size_t size)
8039 gcc_assert (new_nmemb >= old_nmemb);
8040 p = XRESIZEVAR (void, p, new_nmemb * size);
8041 memset (((char *) p) + old_nmemb * size, 0, (new_nmemb - old_nmemb) * size);
8042 return p;
8045 /* Helper function.
8046 Find fallthru edge from PRED. */
8047 edge
8048 find_fallthru_edge_from (basic_block pred)
8050 edge e;
8051 basic_block succ;
8053 succ = pred->next_bb;
8054 gcc_assert (succ->prev_bb == pred);
8056 if (EDGE_COUNT (pred->succs) <= EDGE_COUNT (succ->preds))
8058 e = find_fallthru_edge (pred->succs);
8060 if (e)
8062 gcc_assert (e->dest == succ);
8063 return e;
8066 else
8068 e = find_fallthru_edge (succ->preds);
8070 if (e)
8072 gcc_assert (e->src == pred);
8073 return e;
8077 return NULL;
8080 /* Extend per basic block data structures. */
8081 static void
8082 sched_extend_bb (void)
8084 /* The following is done to keep current_sched_info->next_tail non null. */
8085 rtx_insn *end = BB_END (EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb);
8086 rtx_insn *insn = DEBUG_INSN_P (end) ? prev_nondebug_insn (end) : end;
8087 if (NEXT_INSN (end) == 0
8088 || (!NOTE_P (insn)
8089 && !LABEL_P (insn)
8090 /* Don't emit a NOTE if it would end up before a BARRIER. */
8091 && !BARRIER_P (next_nondebug_insn (end))))
8093 rtx_note *note = emit_note_after (NOTE_INSN_DELETED, end);
8094 /* Make note appear outside BB. */
8095 set_block_for_insn (note, NULL);
8096 BB_END (EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb) = end;
8100 /* Init per basic block data structures. */
8101 void
8102 sched_init_bbs (void)
8104 sched_extend_bb ();
8107 /* Initialize BEFORE_RECOVERY variable. */
8108 static void
8109 init_before_recovery (basic_block *before_recovery_ptr)
8111 basic_block last;
8112 edge e;
8114 last = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
8115 e = find_fallthru_edge_from (last);
8117 if (e)
8119 /* We create two basic blocks:
8120 1. Single instruction block is inserted right after E->SRC
8121 and has jump to
8122 2. Empty block right before EXIT_BLOCK.
8123 Between these two blocks recovery blocks will be emitted. */
8125 basic_block single, empty;
8127 /* If the fallthrough edge to exit we've found is from the block we've
8128 created before, don't do anything more. */
8129 if (last == after_recovery)
8130 return;
8132 adding_bb_to_current_region_p = false;
8134 single = sched_create_empty_bb (last);
8135 empty = sched_create_empty_bb (single);
8137 /* Add new blocks to the root loop. */
8138 if (current_loops != NULL)
8140 add_bb_to_loop (single, (*current_loops->larray)[0]);
8141 add_bb_to_loop (empty, (*current_loops->larray)[0]);
8144 single->count = last->count;
8145 empty->count = last->count;
8146 BB_COPY_PARTITION (single, last);
8147 BB_COPY_PARTITION (empty, last);
8149 redirect_edge_succ (e, single);
8150 make_single_succ_edge (single, empty, 0);
8151 make_single_succ_edge (empty, EXIT_BLOCK_PTR_FOR_FN (cfun),
8152 EDGE_FALLTHRU);
8154 rtx_code_label *label = block_label (empty);
8155 rtx_jump_insn *x = emit_jump_insn_after (targetm.gen_jump (label),
8156 BB_END (single));
8157 JUMP_LABEL (x) = label;
8158 LABEL_NUSES (label)++;
8159 haifa_init_insn (x);
8161 emit_barrier_after (x);
8163 sched_init_only_bb (empty, NULL);
8164 sched_init_only_bb (single, NULL);
8165 sched_extend_bb ();
8167 adding_bb_to_current_region_p = true;
8168 before_recovery = single;
8169 after_recovery = empty;
8171 if (before_recovery_ptr)
8172 *before_recovery_ptr = before_recovery;
8174 if (sched_verbose >= 2 && spec_info->dump)
8175 fprintf (spec_info->dump,
8176 ";;\t\tFixed fallthru to EXIT : %d->>%d->%d->>EXIT\n",
8177 last->index, single->index, empty->index);
8179 else
8180 before_recovery = last;
8183 /* Returns new recovery block. */
8184 basic_block
8185 sched_create_recovery_block (basic_block *before_recovery_ptr)
8187 rtx_insn *barrier;
8188 basic_block rec;
8190 haifa_recovery_bb_recently_added_p = true;
8191 haifa_recovery_bb_ever_added_p = true;
8193 init_before_recovery (before_recovery_ptr);
8195 barrier = get_last_bb_insn (before_recovery);
8196 gcc_assert (BARRIER_P (barrier));
8198 rtx_insn *label = emit_label_after (gen_label_rtx (), barrier);
8200 rec = create_basic_block (label, label, before_recovery);
8202 /* A recovery block always ends with an unconditional jump. */
8203 emit_barrier_after (BB_END (rec));
8205 if (BB_PARTITION (before_recovery) != BB_UNPARTITIONED)
8206 BB_SET_PARTITION (rec, BB_COLD_PARTITION);
8208 if (sched_verbose && spec_info->dump)
8209 fprintf (spec_info->dump, ";;\t\tGenerated recovery block rec%d\n",
8210 rec->index);
8212 return rec;
8215 /* Create edges: FIRST_BB -> REC; FIRST_BB -> SECOND_BB; REC -> SECOND_BB
8216 and emit necessary jumps. */
8217 void
8218 sched_create_recovery_edges (basic_block first_bb, basic_block rec,
8219 basic_block second_bb)
8221 int edge_flags;
8223 /* This is fixing of incoming edge. */
8224 /* ??? Which other flags should be specified? */
8225 if (BB_PARTITION (first_bb) != BB_PARTITION (rec))
8226 /* Partition type is the same, if it is "unpartitioned". */
8227 edge_flags = EDGE_CROSSING;
8228 else
8229 edge_flags = 0;
8231 edge e2 = single_succ_edge (first_bb);
8232 edge e = make_edge (first_bb, rec, edge_flags);
8234 /* TODO: The actual probability can be determined and is computed as
8235 'todo_spec' variable in create_check_block_twin and
8236 in sel-sched.c `check_ds' in create_speculation_check. */
8237 e->probability = profile_probability::very_unlikely ();
8238 rec->count = e->count ();
8239 e2->probability = e->probability.invert ();
8241 rtx_code_label *label = block_label (second_bb);
8242 rtx_jump_insn *jump = emit_jump_insn_after (targetm.gen_jump (label),
8243 BB_END (rec));
8244 JUMP_LABEL (jump) = label;
8245 LABEL_NUSES (label)++;
8247 if (BB_PARTITION (second_bb) != BB_PARTITION (rec))
8248 /* Partition type is the same, if it is "unpartitioned". */
8250 /* Rewritten from cfgrtl.c. */
8251 if (crtl->has_bb_partition && targetm_common.have_named_sections)
8253 /* We don't need the same note for the check because
8254 any_condjump_p (check) == true. */
8255 CROSSING_JUMP_P (jump) = 1;
8257 edge_flags = EDGE_CROSSING;
8259 else
8260 edge_flags = 0;
8262 make_single_succ_edge (rec, second_bb, edge_flags);
8263 if (dom_info_available_p (CDI_DOMINATORS))
8264 set_immediate_dominator (CDI_DOMINATORS, rec, first_bb);
8267 /* This function creates recovery code for INSN. If MUTATE_P is nonzero,
8268 INSN is a simple check, that should be converted to branchy one. */
8269 static void
8270 create_check_block_twin (rtx_insn *insn, bool mutate_p)
8272 basic_block rec;
8273 rtx_insn *label, *check, *twin;
8274 rtx check_pat;
8275 ds_t fs;
8276 sd_iterator_def sd_it;
8277 dep_t dep;
8278 dep_def _new_dep, *new_dep = &_new_dep;
8279 ds_t todo_spec;
8281 gcc_assert (ORIG_PAT (insn) != NULL_RTX);
8283 if (!mutate_p)
8284 todo_spec = TODO_SPEC (insn);
8285 else
8287 gcc_assert (IS_SPECULATION_SIMPLE_CHECK_P (insn)
8288 && (TODO_SPEC (insn) & SPECULATIVE) == 0);
8290 todo_spec = CHECK_SPEC (insn);
8293 todo_spec &= SPECULATIVE;
8295 /* Create recovery block. */
8296 if (mutate_p || targetm.sched.needs_block_p (todo_spec))
8298 rec = sched_create_recovery_block (NULL);
8299 label = BB_HEAD (rec);
8301 else
8303 rec = EXIT_BLOCK_PTR_FOR_FN (cfun);
8304 label = NULL;
8307 /* Emit CHECK. */
8308 check_pat = targetm.sched.gen_spec_check (insn, label, todo_spec);
8310 if (rec != EXIT_BLOCK_PTR_FOR_FN (cfun))
8312 /* To have mem_reg alive at the beginning of second_bb,
8313 we emit check BEFORE insn, so insn after splitting
8314 insn will be at the beginning of second_bb, which will
8315 provide us with the correct life information. */
8316 check = emit_jump_insn_before (check_pat, insn);
8317 JUMP_LABEL (check) = label;
8318 LABEL_NUSES (label)++;
8320 else
8321 check = emit_insn_before (check_pat, insn);
8323 /* Extend data structures. */
8324 haifa_init_insn (check);
8326 /* CHECK is being added to current region. Extend ready list. */
8327 gcc_assert (sched_ready_n_insns != -1);
8328 sched_extend_ready_list (sched_ready_n_insns + 1);
8330 if (current_sched_info->add_remove_insn)
8331 current_sched_info->add_remove_insn (insn, 0);
8333 RECOVERY_BLOCK (check) = rec;
8335 if (sched_verbose && spec_info->dump)
8336 fprintf (spec_info->dump, ";;\t\tGenerated check insn : %s\n",
8337 (*current_sched_info->print_insn) (check, 0));
8339 gcc_assert (ORIG_PAT (insn));
8341 /* Initialize TWIN (twin is a duplicate of original instruction
8342 in the recovery block). */
8343 if (rec != EXIT_BLOCK_PTR_FOR_FN (cfun))
8345 sd_iterator_def sd_it;
8346 dep_t dep;
8348 FOR_EACH_DEP (insn, SD_LIST_RES_BACK, sd_it, dep)
8349 if ((DEP_STATUS (dep) & DEP_OUTPUT) != 0)
8351 struct _dep _dep2, *dep2 = &_dep2;
8353 init_dep (dep2, DEP_PRO (dep), check, REG_DEP_TRUE);
8355 sd_add_dep (dep2, true);
8358 twin = emit_insn_after (ORIG_PAT (insn), BB_END (rec));
8359 haifa_init_insn (twin);
8361 if (sched_verbose && spec_info->dump)
8362 /* INSN_BB (insn) isn't determined for twin insns yet.
8363 So we can't use current_sched_info->print_insn. */
8364 fprintf (spec_info->dump, ";;\t\tGenerated twin insn : %d/rec%d\n",
8365 INSN_UID (twin), rec->index);
8367 else
8369 ORIG_PAT (check) = ORIG_PAT (insn);
8370 HAS_INTERNAL_DEP (check) = 1;
8371 twin = check;
8372 /* ??? We probably should change all OUTPUT dependencies to
8373 (TRUE | OUTPUT). */
8376 /* Copy all resolved back dependencies of INSN to TWIN. This will
8377 provide correct value for INSN_TICK (TWIN). */
8378 sd_copy_back_deps (twin, insn, true);
8380 if (rec != EXIT_BLOCK_PTR_FOR_FN (cfun))
8381 /* In case of branchy check, fix CFG. */
8383 basic_block first_bb, second_bb;
8384 rtx_insn *jump;
8386 first_bb = BLOCK_FOR_INSN (check);
8387 second_bb = sched_split_block (first_bb, check);
8389 sched_create_recovery_edges (first_bb, rec, second_bb);
8391 sched_init_only_bb (second_bb, first_bb);
8392 sched_init_only_bb (rec, EXIT_BLOCK_PTR_FOR_FN (cfun));
8394 jump = BB_END (rec);
8395 haifa_init_insn (jump);
8398 /* Move backward dependences from INSN to CHECK and
8399 move forward dependences from INSN to TWIN. */
8401 /* First, create dependencies between INSN's producers and CHECK & TWIN. */
8402 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
8404 rtx_insn *pro = DEP_PRO (dep);
8405 ds_t ds;
8407 /* If BEGIN_DATA: [insn ~~TRUE~~> producer]:
8408 check --TRUE--> producer ??? or ANTI ???
8409 twin --TRUE--> producer
8410 twin --ANTI--> check
8412 If BEGIN_CONTROL: [insn ~~ANTI~~> producer]:
8413 check --ANTI--> producer
8414 twin --ANTI--> producer
8415 twin --ANTI--> check
8417 If BE_IN_SPEC: [insn ~~TRUE~~> producer]:
8418 check ~~TRUE~~> producer
8419 twin ~~TRUE~~> producer
8420 twin --ANTI--> check */
8422 ds = DEP_STATUS (dep);
8424 if (ds & BEGIN_SPEC)
8426 gcc_assert (!mutate_p);
8427 ds &= ~BEGIN_SPEC;
8430 init_dep_1 (new_dep, pro, check, DEP_TYPE (dep), ds);
8431 sd_add_dep (new_dep, false);
8433 if (rec != EXIT_BLOCK_PTR_FOR_FN (cfun))
8435 DEP_CON (new_dep) = twin;
8436 sd_add_dep (new_dep, false);
8440 /* Second, remove backward dependencies of INSN. */
8441 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
8442 sd_iterator_cond (&sd_it, &dep);)
8444 if ((DEP_STATUS (dep) & BEGIN_SPEC)
8445 || mutate_p)
8446 /* We can delete this dep because we overcome it with
8447 BEGIN_SPECULATION. */
8448 sd_delete_dep (sd_it);
8449 else
8450 sd_iterator_next (&sd_it);
8453 /* Future Speculations. Determine what BE_IN speculations will be like. */
8454 fs = 0;
8456 /* Fields (DONE_SPEC (x) & BEGIN_SPEC) and CHECK_SPEC (x) are set only
8457 here. */
8459 gcc_assert (!DONE_SPEC (insn));
8461 if (!mutate_p)
8463 ds_t ts = TODO_SPEC (insn);
8465 DONE_SPEC (insn) = ts & BEGIN_SPEC;
8466 CHECK_SPEC (check) = ts & BEGIN_SPEC;
8468 /* Luckiness of future speculations solely depends upon initial
8469 BEGIN speculation. */
8470 if (ts & BEGIN_DATA)
8471 fs = set_dep_weak (fs, BE_IN_DATA, get_dep_weak (ts, BEGIN_DATA));
8472 if (ts & BEGIN_CONTROL)
8473 fs = set_dep_weak (fs, BE_IN_CONTROL,
8474 get_dep_weak (ts, BEGIN_CONTROL));
8476 else
8477 CHECK_SPEC (check) = CHECK_SPEC (insn);
8479 /* Future speculations: call the helper. */
8480 process_insn_forw_deps_be_in_spec (insn, twin, fs);
8482 if (rec != EXIT_BLOCK_PTR_FOR_FN (cfun))
8484 /* Which types of dependencies should we use here is,
8485 generally, machine-dependent question... But, for now,
8486 it is not. */
8488 if (!mutate_p)
8490 init_dep (new_dep, insn, check, REG_DEP_TRUE);
8491 sd_add_dep (new_dep, false);
8493 init_dep (new_dep, insn, twin, REG_DEP_OUTPUT);
8494 sd_add_dep (new_dep, false);
8496 else
8498 if (spec_info->dump)
8499 fprintf (spec_info->dump, ";;\t\tRemoved simple check : %s\n",
8500 (*current_sched_info->print_insn) (insn, 0));
8502 /* Remove all dependencies of the INSN. */
8504 sd_it = sd_iterator_start (insn, (SD_LIST_FORW
8505 | SD_LIST_BACK
8506 | SD_LIST_RES_BACK));
8507 while (sd_iterator_cond (&sd_it, &dep))
8508 sd_delete_dep (sd_it);
8511 /* If former check (INSN) already was moved to the ready (or queue)
8512 list, add new check (CHECK) there too. */
8513 if (QUEUE_INDEX (insn) != QUEUE_NOWHERE)
8514 try_ready (check);
8516 /* Remove old check from instruction stream and free its
8517 data. */
8518 sched_remove_insn (insn);
8521 init_dep (new_dep, check, twin, REG_DEP_ANTI);
8522 sd_add_dep (new_dep, false);
8524 else
8526 init_dep_1 (new_dep, insn, check, REG_DEP_TRUE, DEP_TRUE | DEP_OUTPUT);
8527 sd_add_dep (new_dep, false);
8530 if (!mutate_p)
8531 /* Fix priorities. If MUTATE_P is nonzero, this is not necessary,
8532 because it'll be done later in add_to_speculative_block. */
8534 auto_vec<rtx_insn *> priorities_roots;
8536 clear_priorities (twin, &priorities_roots);
8537 calc_priorities (priorities_roots);
8541 /* Removes dependency between instructions in the recovery block REC
8542 and usual region instructions. It keeps inner dependences so it
8543 won't be necessary to recompute them. */
8544 static void
8545 fix_recovery_deps (basic_block rec)
8547 rtx_insn *note, *insn, *jump;
8548 auto_vec<rtx_insn *, 10> ready_list;
8549 auto_bitmap in_ready;
8551 /* NOTE - a basic block note. */
8552 note = NEXT_INSN (BB_HEAD (rec));
8553 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
8554 insn = BB_END (rec);
8555 gcc_assert (JUMP_P (insn));
8556 insn = PREV_INSN (insn);
8560 sd_iterator_def sd_it;
8561 dep_t dep;
8563 for (sd_it = sd_iterator_start (insn, SD_LIST_FORW);
8564 sd_iterator_cond (&sd_it, &dep);)
8566 rtx_insn *consumer = DEP_CON (dep);
8568 if (BLOCK_FOR_INSN (consumer) != rec)
8570 sd_delete_dep (sd_it);
8572 if (bitmap_set_bit (in_ready, INSN_LUID (consumer)))
8573 ready_list.safe_push (consumer);
8575 else
8577 gcc_assert ((DEP_STATUS (dep) & DEP_TYPES) == DEP_TRUE);
8579 sd_iterator_next (&sd_it);
8583 insn = PREV_INSN (insn);
8585 while (insn != note);
8587 /* Try to add instructions to the ready or queue list. */
8588 unsigned int i;
8589 rtx_insn *temp;
8590 FOR_EACH_VEC_ELT_REVERSE (ready_list, i, temp)
8591 try_ready (temp);
8593 /* Fixing jump's dependences. */
8594 insn = BB_HEAD (rec);
8595 jump = BB_END (rec);
8597 gcc_assert (LABEL_P (insn));
8598 insn = NEXT_INSN (insn);
8600 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (insn));
8601 add_jump_dependencies (insn, jump);
8604 /* Change pattern of INSN to NEW_PAT. Invalidate cached haifa
8605 instruction data. */
8606 static bool
8607 haifa_change_pattern (rtx_insn *insn, rtx new_pat)
8609 int t;
8611 t = validate_change (insn, &PATTERN (insn), new_pat, 0);
8612 if (!t)
8613 return false;
8615 update_insn_after_change (insn);
8616 return true;
8619 /* -1 - can't speculate,
8620 0 - for speculation with REQUEST mode it is OK to use
8621 current instruction pattern,
8622 1 - need to change pattern for *NEW_PAT to be speculative. */
8624 sched_speculate_insn (rtx_insn *insn, ds_t request, rtx *new_pat)
8626 gcc_assert (current_sched_info->flags & DO_SPECULATION
8627 && (request & SPECULATIVE)
8628 && sched_insn_is_legitimate_for_speculation_p (insn, request));
8630 if ((request & spec_info->mask) != request)
8631 return -1;
8633 if (request & BE_IN_SPEC
8634 && !(request & BEGIN_SPEC))
8635 return 0;
8637 return targetm.sched.speculate_insn (insn, request, new_pat);
8640 static int
8641 haifa_speculate_insn (rtx_insn *insn, ds_t request, rtx *new_pat)
8643 gcc_assert (sched_deps_info->generate_spec_deps
8644 && !IS_SPECULATION_CHECK_P (insn));
8646 if (HAS_INTERNAL_DEP (insn)
8647 || SCHED_GROUP_P (insn))
8648 return -1;
8650 return sched_speculate_insn (insn, request, new_pat);
8653 /* Print some information about block BB, which starts with HEAD and
8654 ends with TAIL, before scheduling it.
8655 I is zero, if scheduler is about to start with the fresh ebb. */
8656 static void
8657 dump_new_block_header (int i, basic_block bb, rtx_insn *head, rtx_insn *tail)
8659 if (!i)
8660 fprintf (sched_dump,
8661 ";; ======================================================\n");
8662 else
8663 fprintf (sched_dump,
8664 ";; =====================ADVANCING TO=====================\n");
8665 fprintf (sched_dump,
8666 ";; -- basic block %d from %d to %d -- %s reload\n",
8667 bb->index, INSN_UID (head), INSN_UID (tail),
8668 (reload_completed ? "after" : "before"));
8669 fprintf (sched_dump,
8670 ";; ======================================================\n");
8671 fprintf (sched_dump, "\n");
8674 /* Unlink basic block notes and labels and saves them, so they
8675 can be easily restored. We unlink basic block notes in EBB to
8676 provide back-compatibility with the previous code, as target backends
8677 assume, that there'll be only instructions between
8678 current_sched_info->{head and tail}. We restore these notes as soon
8679 as we can.
8680 FIRST (LAST) is the first (last) basic block in the ebb.
8681 NB: In usual case (FIRST == LAST) nothing is really done. */
8682 void
8683 unlink_bb_notes (basic_block first, basic_block last)
8685 /* We DON'T unlink basic block notes of the first block in the ebb. */
8686 if (first == last)
8687 return;
8689 bb_header = XNEWVEC (rtx_insn *, last_basic_block_for_fn (cfun));
8691 /* Make a sentinel. */
8692 if (last->next_bb != EXIT_BLOCK_PTR_FOR_FN (cfun))
8693 bb_header[last->next_bb->index] = 0;
8695 first = first->next_bb;
8698 rtx_insn *prev, *label, *note, *next;
8700 label = BB_HEAD (last);
8701 if (LABEL_P (label))
8702 note = NEXT_INSN (label);
8703 else
8704 note = label;
8705 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
8707 prev = PREV_INSN (label);
8708 next = NEXT_INSN (note);
8709 gcc_assert (prev && next);
8711 SET_NEXT_INSN (prev) = next;
8712 SET_PREV_INSN (next) = prev;
8714 bb_header[last->index] = label;
8716 if (last == first)
8717 break;
8719 last = last->prev_bb;
8721 while (1);
8724 /* Restore basic block notes.
8725 FIRST is the first basic block in the ebb. */
8726 static void
8727 restore_bb_notes (basic_block first)
8729 if (!bb_header)
8730 return;
8732 /* We DON'T unlink basic block notes of the first block in the ebb. */
8733 first = first->next_bb;
8734 /* Remember: FIRST is actually a second basic block in the ebb. */
8736 while (first != EXIT_BLOCK_PTR_FOR_FN (cfun)
8737 && bb_header[first->index])
8739 rtx_insn *prev, *label, *note, *next;
8741 label = bb_header[first->index];
8742 prev = PREV_INSN (label);
8743 next = NEXT_INSN (prev);
8745 if (LABEL_P (label))
8746 note = NEXT_INSN (label);
8747 else
8748 note = label;
8749 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
8751 bb_header[first->index] = 0;
8753 SET_NEXT_INSN (prev) = label;
8754 SET_NEXT_INSN (note) = next;
8755 SET_PREV_INSN (next) = note;
8757 first = first->next_bb;
8760 free (bb_header);
8761 bb_header = 0;
8764 /* Helper function.
8765 Fix CFG after both in- and inter-block movement of
8766 control_flow_insn_p JUMP. */
8767 static void
8768 fix_jump_move (rtx_insn *jump)
8770 basic_block bb, jump_bb, jump_bb_next;
8772 bb = BLOCK_FOR_INSN (PREV_INSN (jump));
8773 jump_bb = BLOCK_FOR_INSN (jump);
8774 jump_bb_next = jump_bb->next_bb;
8776 gcc_assert (common_sched_info->sched_pass_id == SCHED_EBB_PASS
8777 || IS_SPECULATION_BRANCHY_CHECK_P (jump));
8779 if (!NOTE_INSN_BASIC_BLOCK_P (BB_END (jump_bb_next)))
8780 /* if jump_bb_next is not empty. */
8781 BB_END (jump_bb) = BB_END (jump_bb_next);
8783 if (BB_END (bb) != PREV_INSN (jump))
8784 /* Then there are instruction after jump that should be placed
8785 to jump_bb_next. */
8786 BB_END (jump_bb_next) = BB_END (bb);
8787 else
8788 /* Otherwise jump_bb_next is empty. */
8789 BB_END (jump_bb_next) = NEXT_INSN (BB_HEAD (jump_bb_next));
8791 /* To make assertion in move_insn happy. */
8792 BB_END (bb) = PREV_INSN (jump);
8794 update_bb_for_insn (jump_bb_next);
8797 /* Fix CFG after interblock movement of control_flow_insn_p JUMP. */
8798 static void
8799 move_block_after_check (rtx_insn *jump)
8801 basic_block bb, jump_bb, jump_bb_next;
8802 vec<edge, va_gc> *t;
8804 bb = BLOCK_FOR_INSN (PREV_INSN (jump));
8805 jump_bb = BLOCK_FOR_INSN (jump);
8806 jump_bb_next = jump_bb->next_bb;
8808 update_bb_for_insn (jump_bb);
8810 gcc_assert (IS_SPECULATION_CHECK_P (jump)
8811 || IS_SPECULATION_CHECK_P (BB_END (jump_bb_next)));
8813 unlink_block (jump_bb_next);
8814 link_block (jump_bb_next, bb);
8816 t = bb->succs;
8817 bb->succs = 0;
8818 move_succs (&(jump_bb->succs), bb);
8819 move_succs (&(jump_bb_next->succs), jump_bb);
8820 move_succs (&t, jump_bb_next);
8822 df_mark_solutions_dirty ();
8824 common_sched_info->fix_recovery_cfg
8825 (bb->index, jump_bb->index, jump_bb_next->index);
8828 /* Helper function for move_block_after_check.
8829 This functions attaches edge vector pointed to by SUCCSP to
8830 block TO. */
8831 static void
8832 move_succs (vec<edge, va_gc> **succsp, basic_block to)
8834 edge e;
8835 edge_iterator ei;
8837 gcc_assert (to->succs == 0);
8839 to->succs = *succsp;
8841 FOR_EACH_EDGE (e, ei, to->succs)
8842 e->src = to;
8844 *succsp = 0;
8847 /* Remove INSN from the instruction stream.
8848 INSN should have any dependencies. */
8849 static void
8850 sched_remove_insn (rtx_insn *insn)
8852 sd_finish_insn (insn);
8854 change_queue_index (insn, QUEUE_NOWHERE);
8855 current_sched_info->add_remove_insn (insn, 1);
8856 delete_insn (insn);
8859 /* Clear priorities of all instructions, that are forward dependent on INSN.
8860 Store in vector pointed to by ROOTS_PTR insns on which priority () should
8861 be invoked to initialize all cleared priorities. */
8862 static void
8863 clear_priorities (rtx_insn *insn, rtx_vec_t *roots_ptr)
8865 sd_iterator_def sd_it;
8866 dep_t dep;
8867 bool insn_is_root_p = true;
8869 gcc_assert (QUEUE_INDEX (insn) != QUEUE_SCHEDULED);
8871 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
8873 rtx_insn *pro = DEP_PRO (dep);
8875 if (INSN_PRIORITY_STATUS (pro) >= 0
8876 && QUEUE_INDEX (insn) != QUEUE_SCHEDULED)
8878 /* If DEP doesn't contribute to priority then INSN itself should
8879 be added to priority roots. */
8880 if (contributes_to_priority_p (dep))
8881 insn_is_root_p = false;
8883 INSN_PRIORITY_STATUS (pro) = -1;
8884 clear_priorities (pro, roots_ptr);
8888 if (insn_is_root_p)
8889 roots_ptr->safe_push (insn);
8892 /* Recompute priorities of instructions, whose priorities might have been
8893 changed. ROOTS is a vector of instructions whose priority computation will
8894 trigger initialization of all cleared priorities. */
8895 static void
8896 calc_priorities (rtx_vec_t roots)
8898 int i;
8899 rtx_insn *insn;
8901 FOR_EACH_VEC_ELT (roots, i, insn)
8902 priority (insn);
8906 /* Add dependences between JUMP and other instructions in the recovery
8907 block. INSN is the first insn the recovery block. */
8908 static void
8909 add_jump_dependencies (rtx_insn *insn, rtx_insn *jump)
8913 insn = NEXT_INSN (insn);
8914 if (insn == jump)
8915 break;
8917 if (dep_list_size (insn, SD_LIST_FORW) == 0)
8919 dep_def _new_dep, *new_dep = &_new_dep;
8921 init_dep (new_dep, insn, jump, REG_DEP_ANTI);
8922 sd_add_dep (new_dep, false);
8925 while (1);
8927 gcc_assert (!sd_lists_empty_p (jump, SD_LIST_BACK));
8930 /* Extend data structures for logical insn UID. */
8931 void
8932 sched_extend_luids (void)
8934 int new_luids_max_uid = get_max_uid () + 1;
8936 sched_luids.safe_grow_cleared (new_luids_max_uid);
8939 /* Initialize LUID for INSN. */
8940 void
8941 sched_init_insn_luid (rtx_insn *insn)
8943 int i = INSN_P (insn) ? 1 : common_sched_info->luid_for_non_insn (insn);
8944 int luid;
8946 if (i >= 0)
8948 luid = sched_max_luid;
8949 sched_max_luid += i;
8951 else
8952 luid = -1;
8954 SET_INSN_LUID (insn, luid);
8957 /* Initialize luids for BBS.
8958 The hook common_sched_info->luid_for_non_insn () is used to determine
8959 if notes, labels, etc. need luids. */
8960 void
8961 sched_init_luids (bb_vec_t bbs)
8963 int i;
8964 basic_block bb;
8966 sched_extend_luids ();
8967 FOR_EACH_VEC_ELT (bbs, i, bb)
8969 rtx_insn *insn;
8971 FOR_BB_INSNS (bb, insn)
8972 sched_init_insn_luid (insn);
8976 /* Free LUIDs. */
8977 void
8978 sched_finish_luids (void)
8980 sched_luids.release ();
8981 sched_max_luid = 1;
8984 /* Return logical uid of INSN. Helpful while debugging. */
8986 insn_luid (rtx_insn *insn)
8988 return INSN_LUID (insn);
8991 /* Extend per insn data in the target. */
8992 void
8993 sched_extend_target (void)
8995 if (targetm.sched.h_i_d_extended)
8996 targetm.sched.h_i_d_extended ();
8999 /* Extend global scheduler structures (those, that live across calls to
9000 schedule_block) to include information about just emitted INSN. */
9001 static void
9002 extend_h_i_d (void)
9004 int reserve = (get_max_uid () + 1 - h_i_d.length ());
9005 if (reserve > 0
9006 && ! h_i_d.space (reserve))
9008 h_i_d.safe_grow_cleared (3 * get_max_uid () / 2);
9009 sched_extend_target ();
9013 /* Initialize h_i_d entry of the INSN with default values.
9014 Values, that are not explicitly initialized here, hold zero. */
9015 static void
9016 init_h_i_d (rtx_insn *insn)
9018 if (INSN_LUID (insn) > 0)
9020 INSN_COST (insn) = -1;
9021 QUEUE_INDEX (insn) = QUEUE_NOWHERE;
9022 INSN_TICK (insn) = INVALID_TICK;
9023 INSN_EXACT_TICK (insn) = INVALID_TICK;
9024 INTER_TICK (insn) = INVALID_TICK;
9025 TODO_SPEC (insn) = HARD_DEP;
9026 INSN_AUTOPREF_MULTIPASS_DATA (insn)[0].status
9027 = AUTOPREF_MULTIPASS_DATA_UNINITIALIZED;
9028 INSN_AUTOPREF_MULTIPASS_DATA (insn)[1].status
9029 = AUTOPREF_MULTIPASS_DATA_UNINITIALIZED;
9033 /* Initialize haifa_insn_data for BBS. */
9034 void
9035 haifa_init_h_i_d (bb_vec_t bbs)
9037 int i;
9038 basic_block bb;
9040 extend_h_i_d ();
9041 FOR_EACH_VEC_ELT (bbs, i, bb)
9043 rtx_insn *insn;
9045 FOR_BB_INSNS (bb, insn)
9046 init_h_i_d (insn);
9050 /* Finalize haifa_insn_data. */
9051 void
9052 haifa_finish_h_i_d (void)
9054 int i;
9055 haifa_insn_data_t data;
9056 reg_use_data *use, *next_use;
9057 reg_set_data *set, *next_set;
9059 FOR_EACH_VEC_ELT (h_i_d, i, data)
9061 free (data->max_reg_pressure);
9062 free (data->reg_pressure);
9063 for (use = data->reg_use_list; use != NULL; use = next_use)
9065 next_use = use->next_insn_use;
9066 free (use);
9068 for (set = data->reg_set_list; set != NULL; set = next_set)
9070 next_set = set->next_insn_set;
9071 free (set);
9075 h_i_d.release ();
9078 /* Init data for the new insn INSN. */
9079 static void
9080 haifa_init_insn (rtx_insn *insn)
9082 gcc_assert (insn != NULL);
9084 sched_extend_luids ();
9085 sched_init_insn_luid (insn);
9086 sched_extend_target ();
9087 sched_deps_init (false);
9088 extend_h_i_d ();
9089 init_h_i_d (insn);
9091 if (adding_bb_to_current_region_p)
9093 sd_init_insn (insn);
9095 /* Extend dependency caches by one element. */
9096 extend_dependency_caches (1, false);
9098 if (sched_pressure != SCHED_PRESSURE_NONE)
9099 init_insn_reg_pressure_info (insn);
9102 /* Init data for the new basic block BB which comes after AFTER. */
9103 static void
9104 haifa_init_only_bb (basic_block bb, basic_block after)
9106 gcc_assert (bb != NULL);
9108 sched_init_bbs ();
9110 if (common_sched_info->add_block)
9111 /* This changes only data structures of the front-end. */
9112 common_sched_info->add_block (bb, after);
9115 /* A generic version of sched_split_block (). */
9116 basic_block
9117 sched_split_block_1 (basic_block first_bb, rtx after)
9119 edge e;
9121 e = split_block (first_bb, after);
9122 gcc_assert (e->src == first_bb);
9124 /* sched_split_block emits note if *check == BB_END. Probably it
9125 is better to rip that note off. */
9127 return e->dest;
9130 /* A generic version of sched_create_empty_bb (). */
9131 basic_block
9132 sched_create_empty_bb_1 (basic_block after)
9134 return create_empty_bb (after);
9137 /* Insert PAT as an INSN into the schedule and update the necessary data
9138 structures to account for it. */
9139 rtx_insn *
9140 sched_emit_insn (rtx pat)
9142 rtx_insn *insn = emit_insn_before (pat, first_nonscheduled_insn ());
9143 haifa_init_insn (insn);
9145 if (current_sched_info->add_remove_insn)
9146 current_sched_info->add_remove_insn (insn, 0);
9148 (*current_sched_info->begin_schedule_ready) (insn);
9149 scheduled_insns.safe_push (insn);
9151 last_scheduled_insn = insn;
9152 return insn;
9155 /* This function returns a candidate satisfying dispatch constraints from
9156 the ready list. */
9158 static rtx_insn *
9159 ready_remove_first_dispatch (struct ready_list *ready)
9161 int i;
9162 rtx_insn *insn = ready_element (ready, 0);
9164 if (ready->n_ready == 1
9165 || !INSN_P (insn)
9166 || INSN_CODE (insn) < 0
9167 || !active_insn_p (insn)
9168 || targetm.sched.dispatch (insn, FITS_DISPATCH_WINDOW))
9169 return ready_remove_first (ready);
9171 for (i = 1; i < ready->n_ready; i++)
9173 insn = ready_element (ready, i);
9175 if (!INSN_P (insn)
9176 || INSN_CODE (insn) < 0
9177 || !active_insn_p (insn))
9178 continue;
9180 if (targetm.sched.dispatch (insn, FITS_DISPATCH_WINDOW))
9182 /* Return ith element of ready. */
9183 insn = ready_remove (ready, i);
9184 return insn;
9188 if (targetm.sched.dispatch (NULL, DISPATCH_VIOLATION))
9189 return ready_remove_first (ready);
9191 for (i = 1; i < ready->n_ready; i++)
9193 insn = ready_element (ready, i);
9195 if (!INSN_P (insn)
9196 || INSN_CODE (insn) < 0
9197 || !active_insn_p (insn))
9198 continue;
9200 /* Return i-th element of ready. */
9201 if (targetm.sched.dispatch (insn, IS_CMP))
9202 return ready_remove (ready, i);
9205 return ready_remove_first (ready);
9208 /* Get number of ready insn in the ready list. */
9211 number_in_ready (void)
9213 return ready.n_ready;
9216 /* Get number of ready's in the ready list. */
9218 rtx_insn *
9219 get_ready_element (int i)
9221 return ready_element (&ready, i);
9224 #endif /* INSN_SCHEDULING */