* gcc.dg/c11-complex-1.c: Use dg-add-options ieee.
[official-gcc.git] / gcc / haifa-sched.c
blobbeddc116ffb9d82983b64153286c7984554e0162
1 /* Instruction scheduling pass.
2 Copyright (C) 1992-2013 Free Software Foundation, Inc.
3 Contributed by Michael Tiemann (tiemann@cygnus.com) Enhanced by,
4 and currently maintained by, Jim Wilson (wilson@cygnus.com)
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 is found 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_backward_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 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 "tm.h"
129 #include "diagnostic-core.h"
130 #include "hard-reg-set.h"
131 #include "rtl.h"
132 #include "tm_p.h"
133 #include "regs.h"
134 #include "function.h"
135 #include "flags.h"
136 #include "insn-config.h"
137 #include "insn-attr.h"
138 #include "except.h"
139 #include "recog.h"
140 #include "sched-int.h"
141 #include "target.h"
142 #include "common/common-target.h"
143 #include "params.h"
144 #include "dbgcnt.h"
145 #include "cfgloop.h"
146 #include "ira.h"
147 #include "emit-rtl.h" /* FIXME: Can go away once crtl is moved to rtl.h. */
148 #include "hash-table.h"
149 #include "dumpfile.h"
151 #ifdef INSN_SCHEDULING
153 /* True if we do register pressure relief through live-range
154 shrinkage. */
155 static bool live_range_shrinkage_p;
157 /* Switch on live range shrinkage. */
158 void
159 initialize_live_range_shrinkage (void)
161 live_range_shrinkage_p = true;
164 /* Switch off live range shrinkage. */
165 void
166 finish_live_range_shrinkage (void)
168 live_range_shrinkage_p = false;
171 /* issue_rate is the number of insns that can be scheduled in the same
172 machine cycle. It can be defined in the config/mach/mach.h file,
173 otherwise we set it to 1. */
175 int issue_rate;
177 /* This can be set to true by a backend if the scheduler should not
178 enable a DCE pass. */
179 bool sched_no_dce;
181 /* The current initiation interval used when modulo scheduling. */
182 static int modulo_ii;
184 /* The maximum number of stages we are prepared to handle. */
185 static int modulo_max_stages;
187 /* The number of insns that exist in each iteration of the loop. We use this
188 to detect when we've scheduled all insns from the first iteration. */
189 static int modulo_n_insns;
191 /* The current count of insns in the first iteration of the loop that have
192 already been scheduled. */
193 static int modulo_insns_scheduled;
195 /* The maximum uid of insns from the first iteration of the loop. */
196 static int modulo_iter0_max_uid;
198 /* The number of times we should attempt to backtrack when modulo scheduling.
199 Decreased each time we have to backtrack. */
200 static int modulo_backtracks_left;
202 /* The stage in which the last insn from the original loop was
203 scheduled. */
204 static int modulo_last_stage;
206 /* sched-verbose controls the amount of debugging output the
207 scheduler prints. It is controlled by -fsched-verbose=N:
208 N>0 and no -DSR : the output is directed to stderr.
209 N>=10 will direct the printouts to stderr (regardless of -dSR).
210 N=1: same as -dSR.
211 N=2: bb's probabilities, detailed ready list info, unit/insn info.
212 N=3: rtl at abort point, control-flow, regions info.
213 N=5: dependences info. */
215 int sched_verbose = 0;
217 /* Debugging file. All printouts are sent to dump, which is always set,
218 either to stderr, or to the dump listing file (-dRS). */
219 FILE *sched_dump = 0;
221 /* This is a placeholder for the scheduler parameters common
222 to all schedulers. */
223 struct common_sched_info_def *common_sched_info;
225 #define INSN_TICK(INSN) (HID (INSN)->tick)
226 #define INSN_EXACT_TICK(INSN) (HID (INSN)->exact_tick)
227 #define INSN_TICK_ESTIMATE(INSN) (HID (INSN)->tick_estimate)
228 #define INTER_TICK(INSN) (HID (INSN)->inter_tick)
229 #define FEEDS_BACKTRACK_INSN(INSN) (HID (INSN)->feeds_backtrack_insn)
230 #define SHADOW_P(INSN) (HID (INSN)->shadow_p)
231 #define MUST_RECOMPUTE_SPEC_P(INSN) (HID (INSN)->must_recompute_spec)
232 /* Cached cost of the instruction. Use insn_cost to get cost of the
233 insn. -1 here means that the field is not initialized. */
234 #define INSN_COST(INSN) (HID (INSN)->cost)
236 /* If INSN_TICK of an instruction is equal to INVALID_TICK,
237 then it should be recalculated from scratch. */
238 #define INVALID_TICK (-(max_insn_queue_index + 1))
239 /* The minimal value of the INSN_TICK of an instruction. */
240 #define MIN_TICK (-max_insn_queue_index)
242 /* List of important notes we must keep around. This is a pointer to the
243 last element in the list. */
244 rtx note_list;
246 static struct spec_info_def spec_info_var;
247 /* Description of the speculative part of the scheduling.
248 If NULL - no speculation. */
249 spec_info_t spec_info = NULL;
251 /* True, if recovery block was added during scheduling of current block.
252 Used to determine, if we need to fix INSN_TICKs. */
253 static bool haifa_recovery_bb_recently_added_p;
255 /* True, if recovery block was added during this scheduling pass.
256 Used to determine if we should have empty memory pools of dependencies
257 after finishing current region. */
258 bool haifa_recovery_bb_ever_added_p;
260 /* Counters of different types of speculative instructions. */
261 static int nr_begin_data, nr_be_in_data, nr_begin_control, nr_be_in_control;
263 /* Array used in {unlink, restore}_bb_notes. */
264 static rtx *bb_header = 0;
266 /* Basic block after which recovery blocks will be created. */
267 static basic_block before_recovery;
269 /* Basic block just before the EXIT_BLOCK and after recovery, if we have
270 created it. */
271 basic_block after_recovery;
273 /* FALSE if we add bb to another region, so we don't need to initialize it. */
274 bool adding_bb_to_current_region_p = true;
276 /* Queues, etc. */
278 /* An instruction is ready to be scheduled when all insns preceding it
279 have already been scheduled. It is important to ensure that all
280 insns which use its result will not be executed until its result
281 has been computed. An insn is maintained in one of four structures:
283 (P) the "Pending" set of insns which cannot be scheduled until
284 their dependencies have been satisfied.
285 (Q) the "Queued" set of insns that can be scheduled when sufficient
286 time has passed.
287 (R) the "Ready" list of unscheduled, uncommitted insns.
288 (S) the "Scheduled" list of insns.
290 Initially, all insns are either "Pending" or "Ready" depending on
291 whether their dependencies are satisfied.
293 Insns move from the "Ready" list to the "Scheduled" list as they
294 are committed to the schedule. As this occurs, the insns in the
295 "Pending" list have their dependencies satisfied and move to either
296 the "Ready" list or the "Queued" set depending on whether
297 sufficient time has passed to make them ready. As time passes,
298 insns move from the "Queued" set to the "Ready" list.
300 The "Pending" list (P) are the insns in the INSN_FORW_DEPS of the
301 unscheduled insns, i.e., those that are ready, queued, and pending.
302 The "Queued" set (Q) is implemented by the variable `insn_queue'.
303 The "Ready" list (R) is implemented by the variables `ready' and
304 `n_ready'.
305 The "Scheduled" list (S) is the new insn chain built by this pass.
307 The transition (R->S) is implemented in the scheduling loop in
308 `schedule_block' when the best insn to schedule is chosen.
309 The transitions (P->R and P->Q) are implemented in `schedule_insn' as
310 insns move from the ready list to the scheduled list.
311 The transition (Q->R) is implemented in 'queue_to_insn' as time
312 passes or stalls are introduced. */
314 /* Implement a circular buffer to delay instructions until sufficient
315 time has passed. For the new pipeline description interface,
316 MAX_INSN_QUEUE_INDEX is a power of two minus one which is not less
317 than maximal time of instruction execution computed by genattr.c on
318 the base maximal time of functional unit reservations and getting a
319 result. This is the longest time an insn may be queued. */
321 static rtx *insn_queue;
322 static int q_ptr = 0;
323 static int q_size = 0;
324 #define NEXT_Q(X) (((X)+1) & max_insn_queue_index)
325 #define NEXT_Q_AFTER(X, C) (((X)+C) & max_insn_queue_index)
327 #define QUEUE_SCHEDULED (-3)
328 #define QUEUE_NOWHERE (-2)
329 #define QUEUE_READY (-1)
330 /* QUEUE_SCHEDULED - INSN is scheduled.
331 QUEUE_NOWHERE - INSN isn't scheduled yet and is neither in
332 queue or ready list.
333 QUEUE_READY - INSN is in ready list.
334 N >= 0 - INSN queued for X [where NEXT_Q_AFTER (q_ptr, X) == N] cycles. */
336 #define QUEUE_INDEX(INSN) (HID (INSN)->queue_index)
338 /* The following variable value refers for all current and future
339 reservations of the processor units. */
340 state_t curr_state;
342 /* The following variable value is size of memory representing all
343 current and future reservations of the processor units. */
344 size_t dfa_state_size;
346 /* The following array is used to find the best insn from ready when
347 the automaton pipeline interface is used. */
348 char *ready_try = NULL;
350 /* The ready list. */
351 struct ready_list ready = {NULL, 0, 0, 0, 0};
353 /* The pointer to the ready list (to be removed). */
354 static struct ready_list *readyp = &ready;
356 /* Scheduling clock. */
357 static int clock_var;
359 /* Clock at which the previous instruction was issued. */
360 static int last_clock_var;
362 /* Set to true if, when queuing a shadow insn, we discover that it would be
363 scheduled too late. */
364 static bool must_backtrack;
366 /* The following variable value is number of essential insns issued on
367 the current cycle. An insn is essential one if it changes the
368 processors state. */
369 int cycle_issued_insns;
371 /* This records the actual schedule. It is built up during the main phase
372 of schedule_block, and afterwards used to reorder the insns in the RTL. */
373 static vec<rtx> scheduled_insns;
375 static int may_trap_exp (const_rtx, int);
377 /* Nonzero iff the address is comprised from at most 1 register. */
378 #define CONST_BASED_ADDRESS_P(x) \
379 (REG_P (x) \
380 || ((GET_CODE (x) == PLUS || GET_CODE (x) == MINUS \
381 || (GET_CODE (x) == LO_SUM)) \
382 && (CONSTANT_P (XEXP (x, 0)) \
383 || CONSTANT_P (XEXP (x, 1)))))
385 /* Returns a class that insn with GET_DEST(insn)=x may belong to,
386 as found by analyzing insn's expression. */
389 static int haifa_luid_for_non_insn (rtx x);
391 /* Haifa version of sched_info hooks common to all headers. */
392 const struct common_sched_info_def haifa_common_sched_info =
394 NULL, /* fix_recovery_cfg */
395 NULL, /* add_block */
396 NULL, /* estimate_number_of_insns */
397 haifa_luid_for_non_insn, /* luid_for_non_insn */
398 SCHED_PASS_UNKNOWN /* sched_pass_id */
401 /* Mapping from instruction UID to its Logical UID. */
402 vec<int> sched_luids = vNULL;
404 /* Next LUID to assign to an instruction. */
405 int sched_max_luid = 1;
407 /* Haifa Instruction Data. */
408 vec<haifa_insn_data_def> h_i_d = vNULL;
410 void (* sched_init_only_bb) (basic_block, basic_block);
412 /* Split block function. Different schedulers might use different functions
413 to handle their internal data consistent. */
414 basic_block (* sched_split_block) (basic_block, rtx);
416 /* Create empty basic block after the specified block. */
417 basic_block (* sched_create_empty_bb) (basic_block);
419 /* Return the number of cycles until INSN is expected to be ready.
420 Return zero if it already is. */
421 static int
422 insn_delay (rtx insn)
424 return MAX (INSN_TICK (insn) - clock_var, 0);
427 static int
428 may_trap_exp (const_rtx x, int is_store)
430 enum rtx_code code;
432 if (x == 0)
433 return TRAP_FREE;
434 code = GET_CODE (x);
435 if (is_store)
437 if (code == MEM && may_trap_p (x))
438 return TRAP_RISKY;
439 else
440 return TRAP_FREE;
442 if (code == MEM)
444 /* The insn uses memory: a volatile load. */
445 if (MEM_VOLATILE_P (x))
446 return IRISKY;
447 /* An exception-free load. */
448 if (!may_trap_p (x))
449 return IFREE;
450 /* A load with 1 base register, to be further checked. */
451 if (CONST_BASED_ADDRESS_P (XEXP (x, 0)))
452 return PFREE_CANDIDATE;
453 /* No info on the load, to be further checked. */
454 return PRISKY_CANDIDATE;
456 else
458 const char *fmt;
459 int i, insn_class = TRAP_FREE;
461 /* Neither store nor load, check if it may cause a trap. */
462 if (may_trap_p (x))
463 return TRAP_RISKY;
464 /* Recursive step: walk the insn... */
465 fmt = GET_RTX_FORMAT (code);
466 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
468 if (fmt[i] == 'e')
470 int tmp_class = may_trap_exp (XEXP (x, i), is_store);
471 insn_class = WORST_CLASS (insn_class, tmp_class);
473 else if (fmt[i] == 'E')
475 int j;
476 for (j = 0; j < XVECLEN (x, i); j++)
478 int tmp_class = may_trap_exp (XVECEXP (x, i, j), is_store);
479 insn_class = WORST_CLASS (insn_class, tmp_class);
480 if (insn_class == TRAP_RISKY || insn_class == IRISKY)
481 break;
484 if (insn_class == TRAP_RISKY || insn_class == IRISKY)
485 break;
487 return insn_class;
491 /* Classifies rtx X of an insn for the purpose of verifying that X can be
492 executed speculatively (and consequently the insn can be moved
493 speculatively), by examining X, returning:
494 TRAP_RISKY: store, or risky non-load insn (e.g. division by variable).
495 TRAP_FREE: non-load insn.
496 IFREE: load from a globally safe location.
497 IRISKY: volatile load.
498 PFREE_CANDIDATE, PRISKY_CANDIDATE: load that need to be checked for
499 being either PFREE or PRISKY. */
501 static int
502 haifa_classify_rtx (const_rtx x)
504 int tmp_class = TRAP_FREE;
505 int insn_class = TRAP_FREE;
506 enum rtx_code code;
508 if (GET_CODE (x) == PARALLEL)
510 int i, len = XVECLEN (x, 0);
512 for (i = len - 1; i >= 0; i--)
514 tmp_class = haifa_classify_rtx (XVECEXP (x, 0, i));
515 insn_class = WORST_CLASS (insn_class, tmp_class);
516 if (insn_class == TRAP_RISKY || insn_class == IRISKY)
517 break;
520 else
522 code = GET_CODE (x);
523 switch (code)
525 case CLOBBER:
526 /* Test if it is a 'store'. */
527 tmp_class = may_trap_exp (XEXP (x, 0), 1);
528 break;
529 case SET:
530 /* Test if it is a store. */
531 tmp_class = may_trap_exp (SET_DEST (x), 1);
532 if (tmp_class == TRAP_RISKY)
533 break;
534 /* Test if it is a load. */
535 tmp_class =
536 WORST_CLASS (tmp_class,
537 may_trap_exp (SET_SRC (x), 0));
538 break;
539 case COND_EXEC:
540 tmp_class = haifa_classify_rtx (COND_EXEC_CODE (x));
541 if (tmp_class == TRAP_RISKY)
542 break;
543 tmp_class = WORST_CLASS (tmp_class,
544 may_trap_exp (COND_EXEC_TEST (x), 0));
545 break;
546 case TRAP_IF:
547 tmp_class = TRAP_RISKY;
548 break;
549 default:;
551 insn_class = tmp_class;
554 return insn_class;
558 haifa_classify_insn (const_rtx insn)
560 return haifa_classify_rtx (PATTERN (insn));
563 /* After the scheduler initialization function has been called, this function
564 can be called to enable modulo scheduling. II is the initiation interval
565 we should use, it affects the delays for delay_pairs that were recorded as
566 separated by a given number of stages.
568 MAX_STAGES provides us with a limit
569 after which we give up scheduling; the caller must have unrolled at least
570 as many copies of the loop body and recorded delay_pairs for them.
572 INSNS is the number of real (non-debug) insns in one iteration of
573 the loop. MAX_UID can be used to test whether an insn belongs to
574 the first iteration of the loop; all of them have a uid lower than
575 MAX_UID. */
576 void
577 set_modulo_params (int ii, int max_stages, int insns, int max_uid)
579 modulo_ii = ii;
580 modulo_max_stages = max_stages;
581 modulo_n_insns = insns;
582 modulo_iter0_max_uid = max_uid;
583 modulo_backtracks_left = PARAM_VALUE (PARAM_MAX_MODULO_BACKTRACK_ATTEMPTS);
586 /* A structure to record a pair of insns where the first one is a real
587 insn that has delay slots, and the second is its delayed shadow.
588 I1 is scheduled normally and will emit an assembly instruction,
589 while I2 describes the side effect that takes place at the
590 transition between cycles CYCLES and (CYCLES + 1) after I1. */
591 struct delay_pair
593 struct delay_pair *next_same_i1;
594 rtx i1, i2;
595 int cycles;
596 /* When doing modulo scheduling, we a delay_pair can also be used to
597 show that I1 and I2 are the same insn in a different stage. If that
598 is the case, STAGES will be nonzero. */
599 int stages;
602 /* Helpers for delay hashing. */
604 struct delay_i1_hasher : typed_noop_remove <delay_pair>
606 typedef delay_pair value_type;
607 typedef void compare_type;
608 static inline hashval_t hash (const value_type *);
609 static inline bool equal (const value_type *, const compare_type *);
612 /* Returns a hash value for X, based on hashing just I1. */
614 inline hashval_t
615 delay_i1_hasher::hash (const value_type *x)
617 return htab_hash_pointer (x->i1);
620 /* Return true if I1 of pair X is the same as that of pair Y. */
622 inline bool
623 delay_i1_hasher::equal (const value_type *x, const compare_type *y)
625 return x->i1 == y;
628 struct delay_i2_hasher : typed_free_remove <delay_pair>
630 typedef delay_pair value_type;
631 typedef void compare_type;
632 static inline hashval_t hash (const value_type *);
633 static inline bool equal (const value_type *, const compare_type *);
636 /* Returns a hash value for X, based on hashing just I2. */
638 inline hashval_t
639 delay_i2_hasher::hash (const value_type *x)
641 return htab_hash_pointer (x->i2);
644 /* Return true if I2 of pair X is the same as that of pair Y. */
646 inline bool
647 delay_i2_hasher::equal (const value_type *x, const compare_type *y)
649 return x->i2 == y;
652 /* Two hash tables to record delay_pairs, one indexed by I1 and the other
653 indexed by I2. */
654 static hash_table <delay_i1_hasher> delay_htab;
655 static hash_table <delay_i2_hasher> delay_htab_i2;
657 /* Called through htab_traverse. Walk the hashtable using I2 as
658 index, and delete all elements involving an UID higher than
659 that pointed to by *DATA. */
661 haifa_htab_i2_traverse (delay_pair **slot, int *data)
663 int maxuid = *data;
664 struct delay_pair *p = *slot;
665 if (INSN_UID (p->i2) >= maxuid || INSN_UID (p->i1) >= maxuid)
667 delay_htab_i2.clear_slot (slot);
669 return 1;
672 /* Called through htab_traverse. Walk the hashtable using I2 as
673 index, and delete all elements involving an UID higher than
674 that pointed to by *DATA. */
676 haifa_htab_i1_traverse (delay_pair **pslot, int *data)
678 int maxuid = *data;
679 struct delay_pair *p, *first, **pprev;
681 if (INSN_UID ((*pslot)->i1) >= maxuid)
683 delay_htab.clear_slot (pslot);
684 return 1;
686 pprev = &first;
687 for (p = *pslot; p; p = p->next_same_i1)
689 if (INSN_UID (p->i2) < maxuid)
691 *pprev = p;
692 pprev = &p->next_same_i1;
695 *pprev = NULL;
696 if (first == NULL)
697 delay_htab.clear_slot (pslot);
698 else
699 *pslot = first;
700 return 1;
703 /* Discard all delay pairs which involve an insn with an UID higher
704 than MAX_UID. */
705 void
706 discard_delay_pairs_above (int max_uid)
708 delay_htab.traverse <int *, haifa_htab_i1_traverse> (&max_uid);
709 delay_htab_i2.traverse <int *, haifa_htab_i2_traverse> (&max_uid);
712 /* This function can be called by a port just before it starts the final
713 scheduling pass. It records the fact that an instruction with delay
714 slots has been split into two insns, I1 and I2. The first one will be
715 scheduled normally and initiates the operation. The second one is a
716 shadow which must follow a specific number of cycles after I1; its only
717 purpose is to show the side effect that occurs at that cycle in the RTL.
718 If a JUMP_INSN or a CALL_INSN has been split, I1 should be a normal INSN,
719 while I2 retains the original insn type.
721 There are two ways in which the number of cycles can be specified,
722 involving the CYCLES and STAGES arguments to this function. If STAGES
723 is zero, we just use the value of CYCLES. Otherwise, STAGES is a factor
724 which is multiplied by MODULO_II to give the number of cycles. This is
725 only useful if the caller also calls set_modulo_params to enable modulo
726 scheduling. */
728 void
729 record_delay_slot_pair (rtx i1, rtx i2, int cycles, int stages)
731 struct delay_pair *p = XNEW (struct delay_pair);
732 struct delay_pair **slot;
734 p->i1 = i1;
735 p->i2 = i2;
736 p->cycles = cycles;
737 p->stages = stages;
739 if (!delay_htab.is_created ())
741 delay_htab.create (10);
742 delay_htab_i2.create (10);
744 slot = delay_htab.find_slot_with_hash (i1, htab_hash_pointer (i1), INSERT);
745 p->next_same_i1 = *slot;
746 *slot = p;
747 slot = delay_htab_i2.find_slot_with_hash (i2, htab_hash_pointer (i2), INSERT);
748 *slot = p;
751 /* Examine the delay pair hashtable to see if INSN is a shadow for another,
752 and return the other insn if so. Return NULL otherwise. */
754 real_insn_for_shadow (rtx insn)
756 struct delay_pair *pair;
758 if (!delay_htab.is_created ())
759 return NULL_RTX;
761 pair = delay_htab_i2.find_with_hash (insn, htab_hash_pointer (insn));
762 if (!pair || pair->stages > 0)
763 return NULL_RTX;
764 return pair->i1;
767 /* For a pair P of insns, return the fixed distance in cycles from the first
768 insn after which the second must be scheduled. */
769 static int
770 pair_delay (struct delay_pair *p)
772 if (p->stages == 0)
773 return p->cycles;
774 else
775 return p->stages * modulo_ii;
778 /* Given an insn INSN, add a dependence on its delayed shadow if it
779 has one. Also try to find situations where shadows depend on each other
780 and add dependencies to the real insns to limit the amount of backtracking
781 needed. */
782 void
783 add_delay_dependencies (rtx insn)
785 struct delay_pair *pair;
786 sd_iterator_def sd_it;
787 dep_t dep;
789 if (!delay_htab.is_created ())
790 return;
792 pair = delay_htab_i2.find_with_hash (insn, htab_hash_pointer (insn));
793 if (!pair)
794 return;
795 add_dependence (insn, pair->i1, REG_DEP_ANTI);
796 if (pair->stages)
797 return;
799 FOR_EACH_DEP (pair->i2, SD_LIST_BACK, sd_it, dep)
801 rtx pro = DEP_PRO (dep);
802 struct delay_pair *other_pair
803 = delay_htab_i2.find_with_hash (pro, htab_hash_pointer (pro));
804 if (!other_pair || other_pair->stages)
805 continue;
806 if (pair_delay (other_pair) >= pair_delay (pair))
808 if (sched_verbose >= 4)
810 fprintf (sched_dump, ";;\tadding dependence %d <- %d\n",
811 INSN_UID (other_pair->i1),
812 INSN_UID (pair->i1));
813 fprintf (sched_dump, ";;\tpair1 %d <- %d, cost %d\n",
814 INSN_UID (pair->i1),
815 INSN_UID (pair->i2),
816 pair_delay (pair));
817 fprintf (sched_dump, ";;\tpair2 %d <- %d, cost %d\n",
818 INSN_UID (other_pair->i1),
819 INSN_UID (other_pair->i2),
820 pair_delay (other_pair));
822 add_dependence (pair->i1, other_pair->i1, REG_DEP_ANTI);
827 /* Forward declarations. */
829 static int priority (rtx);
830 static int rank_for_schedule (const void *, const void *);
831 static void swap_sort (rtx *, int);
832 static void queue_insn (rtx, int, const char *);
833 static int schedule_insn (rtx);
834 static void adjust_priority (rtx);
835 static void advance_one_cycle (void);
836 static void extend_h_i_d (void);
839 /* Notes handling mechanism:
840 =========================
841 Generally, NOTES are saved before scheduling and restored after scheduling.
842 The scheduler distinguishes between two types of notes:
844 (1) LOOP_BEGIN, LOOP_END, SETJMP, EHREGION_BEG, EHREGION_END notes:
845 Before scheduling a region, a pointer to the note is added to the insn
846 that follows or precedes it. (This happens as part of the data dependence
847 computation). After scheduling an insn, the pointer contained in it is
848 used for regenerating the corresponding note (in reemit_notes).
850 (2) All other notes (e.g. INSN_DELETED): Before scheduling a block,
851 these notes are put in a list (in rm_other_notes() and
852 unlink_other_notes ()). After scheduling the block, these notes are
853 inserted at the beginning of the block (in schedule_block()). */
855 static void ready_add (struct ready_list *, rtx, bool);
856 static rtx ready_remove_first (struct ready_list *);
857 static rtx ready_remove_first_dispatch (struct ready_list *ready);
859 static void queue_to_ready (struct ready_list *);
860 static int early_queue_to_ready (state_t, struct ready_list *);
862 static void debug_ready_list (struct ready_list *);
864 /* The following functions are used to implement multi-pass scheduling
865 on the first cycle. */
866 static rtx ready_remove (struct ready_list *, int);
867 static void ready_remove_insn (rtx);
869 static void fix_inter_tick (rtx, rtx);
870 static int fix_tick_ready (rtx);
871 static void change_queue_index (rtx, int);
873 /* The following functions are used to implement scheduling of data/control
874 speculative instructions. */
876 static void extend_h_i_d (void);
877 static void init_h_i_d (rtx);
878 static int haifa_speculate_insn (rtx, ds_t, rtx *);
879 static void generate_recovery_code (rtx);
880 static void process_insn_forw_deps_be_in_spec (rtx, rtx, ds_t);
881 static void begin_speculative_block (rtx);
882 static void add_to_speculative_block (rtx);
883 static void init_before_recovery (basic_block *);
884 static void create_check_block_twin (rtx, bool);
885 static void fix_recovery_deps (basic_block);
886 static bool haifa_change_pattern (rtx, rtx);
887 static void dump_new_block_header (int, basic_block, rtx, rtx);
888 static void restore_bb_notes (basic_block);
889 static void fix_jump_move (rtx);
890 static void move_block_after_check (rtx);
891 static void move_succs (vec<edge, va_gc> **, basic_block);
892 static void sched_remove_insn (rtx);
893 static void clear_priorities (rtx, rtx_vec_t *);
894 static void calc_priorities (rtx_vec_t);
895 static void add_jump_dependencies (rtx, rtx);
897 #endif /* INSN_SCHEDULING */
899 /* Point to state used for the current scheduling pass. */
900 struct haifa_sched_info *current_sched_info;
902 #ifndef INSN_SCHEDULING
903 void
904 schedule_insns (void)
907 #else
909 /* Do register pressure sensitive insn scheduling if the flag is set
910 up. */
911 enum sched_pressure_algorithm sched_pressure;
913 /* Map regno -> its pressure class. The map defined only when
914 SCHED_PRESSURE != SCHED_PRESSURE_NONE. */
915 enum reg_class *sched_regno_pressure_class;
917 /* The current register pressure. Only elements corresponding pressure
918 classes are defined. */
919 static int curr_reg_pressure[N_REG_CLASSES];
921 /* Saved value of the previous array. */
922 static int saved_reg_pressure[N_REG_CLASSES];
924 /* Register living at given scheduling point. */
925 static bitmap curr_reg_live;
927 /* Saved value of the previous array. */
928 static bitmap saved_reg_live;
930 /* Registers mentioned in the current region. */
931 static bitmap region_ref_regs;
933 /* Initiate register pressure relative info for scheduling the current
934 region. Currently it is only clearing register mentioned in the
935 current region. */
936 void
937 sched_init_region_reg_pressure_info (void)
939 bitmap_clear (region_ref_regs);
942 /* PRESSURE[CL] describes the pressure on register class CL. Update it
943 for the birth (if BIRTH_P) or death (if !BIRTH_P) of register REGNO.
944 LIVE tracks the set of live registers; if it is null, assume that
945 every birth or death is genuine. */
946 static inline void
947 mark_regno_birth_or_death (bitmap live, int *pressure, int regno, bool birth_p)
949 enum reg_class pressure_class;
951 pressure_class = sched_regno_pressure_class[regno];
952 if (regno >= FIRST_PSEUDO_REGISTER)
954 if (pressure_class != NO_REGS)
956 if (birth_p)
958 if (!live || bitmap_set_bit (live, regno))
959 pressure[pressure_class]
960 += (ira_reg_class_max_nregs
961 [pressure_class][PSEUDO_REGNO_MODE (regno)]);
963 else
965 if (!live || bitmap_clear_bit (live, regno))
966 pressure[pressure_class]
967 -= (ira_reg_class_max_nregs
968 [pressure_class][PSEUDO_REGNO_MODE (regno)]);
972 else if (pressure_class != NO_REGS
973 && ! TEST_HARD_REG_BIT (ira_no_alloc_regs, regno))
975 if (birth_p)
977 if (!live || bitmap_set_bit (live, regno))
978 pressure[pressure_class]++;
980 else
982 if (!live || bitmap_clear_bit (live, regno))
983 pressure[pressure_class]--;
988 /* Initiate current register pressure related info from living
989 registers given by LIVE. */
990 static void
991 initiate_reg_pressure_info (bitmap live)
993 int i;
994 unsigned int j;
995 bitmap_iterator bi;
997 for (i = 0; i < ira_pressure_classes_num; i++)
998 curr_reg_pressure[ira_pressure_classes[i]] = 0;
999 bitmap_clear (curr_reg_live);
1000 EXECUTE_IF_SET_IN_BITMAP (live, 0, j, bi)
1001 if (sched_pressure == SCHED_PRESSURE_MODEL
1002 || current_nr_blocks == 1
1003 || bitmap_bit_p (region_ref_regs, j))
1004 mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure, j, true);
1007 /* Mark registers in X as mentioned in the current region. */
1008 static void
1009 setup_ref_regs (rtx x)
1011 int i, j, regno;
1012 const RTX_CODE code = GET_CODE (x);
1013 const char *fmt;
1015 if (REG_P (x))
1017 regno = REGNO (x);
1018 if (HARD_REGISTER_NUM_P (regno))
1019 bitmap_set_range (region_ref_regs, regno,
1020 hard_regno_nregs[regno][GET_MODE (x)]);
1021 else
1022 bitmap_set_bit (region_ref_regs, REGNO (x));
1023 return;
1025 fmt = GET_RTX_FORMAT (code);
1026 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1027 if (fmt[i] == 'e')
1028 setup_ref_regs (XEXP (x, i));
1029 else if (fmt[i] == 'E')
1031 for (j = 0; j < XVECLEN (x, i); j++)
1032 setup_ref_regs (XVECEXP (x, i, j));
1036 /* Initiate current register pressure related info at the start of
1037 basic block BB. */
1038 static void
1039 initiate_bb_reg_pressure_info (basic_block bb)
1041 unsigned int i ATTRIBUTE_UNUSED;
1042 rtx insn;
1044 if (current_nr_blocks > 1)
1045 FOR_BB_INSNS (bb, insn)
1046 if (NONDEBUG_INSN_P (insn))
1047 setup_ref_regs (PATTERN (insn));
1048 initiate_reg_pressure_info (df_get_live_in (bb));
1049 #ifdef EH_RETURN_DATA_REGNO
1050 if (bb_has_eh_pred (bb))
1051 for (i = 0; ; ++i)
1053 unsigned int regno = EH_RETURN_DATA_REGNO (i);
1055 if (regno == INVALID_REGNUM)
1056 break;
1057 if (! bitmap_bit_p (df_get_live_in (bb), regno))
1058 mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure,
1059 regno, true);
1061 #endif
1064 /* Save current register pressure related info. */
1065 static void
1066 save_reg_pressure (void)
1068 int i;
1070 for (i = 0; i < ira_pressure_classes_num; i++)
1071 saved_reg_pressure[ira_pressure_classes[i]]
1072 = curr_reg_pressure[ira_pressure_classes[i]];
1073 bitmap_copy (saved_reg_live, curr_reg_live);
1076 /* Restore saved register pressure related info. */
1077 static void
1078 restore_reg_pressure (void)
1080 int i;
1082 for (i = 0; i < ira_pressure_classes_num; i++)
1083 curr_reg_pressure[ira_pressure_classes[i]]
1084 = saved_reg_pressure[ira_pressure_classes[i]];
1085 bitmap_copy (curr_reg_live, saved_reg_live);
1088 /* Return TRUE if the register is dying after its USE. */
1089 static bool
1090 dying_use_p (struct reg_use_data *use)
1092 struct reg_use_data *next;
1094 for (next = use->next_regno_use; next != use; next = next->next_regno_use)
1095 if (NONDEBUG_INSN_P (next->insn)
1096 && QUEUE_INDEX (next->insn) != QUEUE_SCHEDULED)
1097 return false;
1098 return true;
1101 /* Print info about the current register pressure and its excess for
1102 each pressure class. */
1103 static void
1104 print_curr_reg_pressure (void)
1106 int i;
1107 enum reg_class cl;
1109 fprintf (sched_dump, ";;\t");
1110 for (i = 0; i < ira_pressure_classes_num; i++)
1112 cl = ira_pressure_classes[i];
1113 gcc_assert (curr_reg_pressure[cl] >= 0);
1114 fprintf (sched_dump, " %s:%d(%d)", reg_class_names[cl],
1115 curr_reg_pressure[cl],
1116 curr_reg_pressure[cl] - ira_class_hard_regs_num[cl]);
1118 fprintf (sched_dump, "\n");
1121 /* Determine if INSN has a condition that is clobbered if a register
1122 in SET_REGS is modified. */
1123 static bool
1124 cond_clobbered_p (rtx insn, HARD_REG_SET set_regs)
1126 rtx pat = PATTERN (insn);
1127 gcc_assert (GET_CODE (pat) == COND_EXEC);
1128 if (TEST_HARD_REG_BIT (set_regs, REGNO (XEXP (COND_EXEC_TEST (pat), 0))))
1130 sd_iterator_def sd_it;
1131 dep_t dep;
1132 haifa_change_pattern (insn, ORIG_PAT (insn));
1133 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
1134 DEP_STATUS (dep) &= ~DEP_CANCELLED;
1135 TODO_SPEC (insn) = HARD_DEP;
1136 if (sched_verbose >= 2)
1137 fprintf (sched_dump,
1138 ";;\t\tdequeue insn %s because of clobbered condition\n",
1139 (*current_sched_info->print_insn) (insn, 0));
1140 return true;
1143 return false;
1146 /* This function should be called after modifying the pattern of INSN,
1147 to update scheduler data structures as needed. */
1148 static void
1149 update_insn_after_change (rtx insn)
1151 sd_iterator_def sd_it;
1152 dep_t dep;
1154 dfa_clear_single_insn_cache (insn);
1156 sd_it = sd_iterator_start (insn,
1157 SD_LIST_FORW | SD_LIST_BACK | SD_LIST_RES_BACK);
1158 while (sd_iterator_cond (&sd_it, &dep))
1160 DEP_COST (dep) = UNKNOWN_DEP_COST;
1161 sd_iterator_next (&sd_it);
1164 /* Invalidate INSN_COST, so it'll be recalculated. */
1165 INSN_COST (insn) = -1;
1166 /* Invalidate INSN_TICK, so it'll be recalculated. */
1167 INSN_TICK (insn) = INVALID_TICK;
1171 /* Two VECs, one to hold dependencies for which pattern replacements
1172 need to be applied or restored at the start of the next cycle, and
1173 another to hold an integer that is either one, to apply the
1174 corresponding replacement, or zero to restore it. */
1175 static vec<dep_t> next_cycle_replace_deps;
1176 static vec<int> next_cycle_apply;
1178 static void apply_replacement (dep_t, bool);
1179 static void restore_pattern (dep_t, bool);
1181 /* Look at the remaining dependencies for insn NEXT, and compute and return
1182 the TODO_SPEC value we should use for it. This is called after one of
1183 NEXT's dependencies has been resolved.
1184 We also perform pattern replacements for predication, and for broken
1185 replacement dependencies. The latter is only done if FOR_BACKTRACK is
1186 false. */
1188 static ds_t
1189 recompute_todo_spec (rtx next, bool for_backtrack)
1191 ds_t new_ds;
1192 sd_iterator_def sd_it;
1193 dep_t dep, modify_dep = NULL;
1194 int n_spec = 0;
1195 int n_control = 0;
1196 int n_replace = 0;
1197 bool first_p = true;
1199 if (sd_lists_empty_p (next, SD_LIST_BACK))
1200 /* NEXT has all its dependencies resolved. */
1201 return 0;
1203 if (!sd_lists_empty_p (next, SD_LIST_HARD_BACK))
1204 return HARD_DEP;
1206 /* Now we've got NEXT with speculative deps only.
1207 1. Look at the deps to see what we have to do.
1208 2. Check if we can do 'todo'. */
1209 new_ds = 0;
1211 FOR_EACH_DEP (next, SD_LIST_BACK, sd_it, dep)
1213 rtx pro = DEP_PRO (dep);
1214 ds_t ds = DEP_STATUS (dep) & SPECULATIVE;
1216 if (DEBUG_INSN_P (pro) && !DEBUG_INSN_P (next))
1217 continue;
1219 if (ds)
1221 n_spec++;
1222 if (first_p)
1224 first_p = false;
1226 new_ds = ds;
1228 else
1229 new_ds = ds_merge (new_ds, ds);
1231 else if (DEP_TYPE (dep) == REG_DEP_CONTROL)
1233 if (QUEUE_INDEX (pro) != QUEUE_SCHEDULED)
1235 n_control++;
1236 modify_dep = dep;
1238 DEP_STATUS (dep) &= ~DEP_CANCELLED;
1240 else if (DEP_REPLACE (dep) != NULL)
1242 if (QUEUE_INDEX (pro) != QUEUE_SCHEDULED)
1244 n_replace++;
1245 modify_dep = dep;
1247 DEP_STATUS (dep) &= ~DEP_CANCELLED;
1251 if (n_replace > 0 && n_control == 0 && n_spec == 0)
1253 if (!dbg_cnt (sched_breakdep))
1254 return HARD_DEP;
1255 FOR_EACH_DEP (next, SD_LIST_BACK, sd_it, dep)
1257 struct dep_replacement *desc = DEP_REPLACE (dep);
1258 if (desc != NULL)
1260 if (desc->insn == next && !for_backtrack)
1262 gcc_assert (n_replace == 1);
1263 apply_replacement (dep, true);
1265 DEP_STATUS (dep) |= DEP_CANCELLED;
1268 return 0;
1271 else if (n_control == 1 && n_replace == 0 && n_spec == 0)
1273 rtx pro, other, new_pat;
1274 rtx cond = NULL_RTX;
1275 bool success;
1276 rtx prev = NULL_RTX;
1277 int i;
1278 unsigned regno;
1280 if ((current_sched_info->flags & DO_PREDICATION) == 0
1281 || (ORIG_PAT (next) != NULL_RTX
1282 && PREDICATED_PAT (next) == NULL_RTX))
1283 return HARD_DEP;
1285 pro = DEP_PRO (modify_dep);
1286 other = real_insn_for_shadow (pro);
1287 if (other != NULL_RTX)
1288 pro = other;
1290 cond = sched_get_reverse_condition_uncached (pro);
1291 regno = REGNO (XEXP (cond, 0));
1293 /* Find the last scheduled insn that modifies the condition register.
1294 We can stop looking once we find the insn we depend on through the
1295 REG_DEP_CONTROL; if the condition register isn't modified after it,
1296 we know that it still has the right value. */
1297 if (QUEUE_INDEX (pro) == QUEUE_SCHEDULED)
1298 FOR_EACH_VEC_ELT_REVERSE (scheduled_insns, i, prev)
1300 HARD_REG_SET t;
1302 find_all_hard_reg_sets (prev, &t);
1303 if (TEST_HARD_REG_BIT (t, regno))
1304 return HARD_DEP;
1305 if (prev == pro)
1306 break;
1308 if (ORIG_PAT (next) == NULL_RTX)
1310 ORIG_PAT (next) = PATTERN (next);
1312 new_pat = gen_rtx_COND_EXEC (VOIDmode, cond, PATTERN (next));
1313 success = haifa_change_pattern (next, new_pat);
1314 if (!success)
1315 return HARD_DEP;
1316 PREDICATED_PAT (next) = new_pat;
1318 else if (PATTERN (next) != PREDICATED_PAT (next))
1320 bool success = haifa_change_pattern (next,
1321 PREDICATED_PAT (next));
1322 gcc_assert (success);
1324 DEP_STATUS (modify_dep) |= DEP_CANCELLED;
1325 return DEP_CONTROL;
1328 if (PREDICATED_PAT (next) != NULL_RTX)
1330 int tick = INSN_TICK (next);
1331 bool success = haifa_change_pattern (next,
1332 ORIG_PAT (next));
1333 INSN_TICK (next) = tick;
1334 gcc_assert (success);
1337 /* We can't handle the case where there are both speculative and control
1338 dependencies, so we return HARD_DEP in such a case. Also fail if
1339 we have speculative dependencies with not enough points, or more than
1340 one control dependency. */
1341 if ((n_spec > 0 && (n_control > 0 || n_replace > 0))
1342 || (n_spec > 0
1343 /* Too few points? */
1344 && ds_weak (new_ds) < spec_info->data_weakness_cutoff)
1345 || n_control > 0
1346 || n_replace > 0)
1347 return HARD_DEP;
1349 return new_ds;
1352 /* Pointer to the last instruction scheduled. */
1353 static rtx last_scheduled_insn;
1355 /* Pointer to the last nondebug instruction scheduled within the
1356 block, or the prev_head of the scheduling block. Used by
1357 rank_for_schedule, so that insns independent of the last scheduled
1358 insn will be preferred over dependent instructions. */
1359 static rtx last_nondebug_scheduled_insn;
1361 /* Pointer that iterates through the list of unscheduled insns if we
1362 have a dbg_cnt enabled. It always points at an insn prior to the
1363 first unscheduled one. */
1364 static rtx nonscheduled_insns_begin;
1366 /* Compute cost of executing INSN.
1367 This is the number of cycles between instruction issue and
1368 instruction results. */
1370 insn_cost (rtx insn)
1372 int cost;
1374 if (sel_sched_p ())
1376 if (recog_memoized (insn) < 0)
1377 return 0;
1379 cost = insn_default_latency (insn);
1380 if (cost < 0)
1381 cost = 0;
1383 return cost;
1386 cost = INSN_COST (insn);
1388 if (cost < 0)
1390 /* A USE insn, or something else we don't need to
1391 understand. We can't pass these directly to
1392 result_ready_cost or insn_default_latency because it will
1393 trigger a fatal error for unrecognizable insns. */
1394 if (recog_memoized (insn) < 0)
1396 INSN_COST (insn) = 0;
1397 return 0;
1399 else
1401 cost = insn_default_latency (insn);
1402 if (cost < 0)
1403 cost = 0;
1405 INSN_COST (insn) = cost;
1409 return cost;
1412 /* Compute cost of dependence LINK.
1413 This is the number of cycles between instruction issue and
1414 instruction results.
1415 ??? We also use this function to call recog_memoized on all insns. */
1417 dep_cost_1 (dep_t link, dw_t dw)
1419 rtx insn = DEP_PRO (link);
1420 rtx used = DEP_CON (link);
1421 int cost;
1423 if (DEP_COST (link) != UNKNOWN_DEP_COST)
1424 return DEP_COST (link);
1426 if (delay_htab.is_created ())
1428 struct delay_pair *delay_entry;
1429 delay_entry
1430 = delay_htab_i2.find_with_hash (used, htab_hash_pointer (used));
1431 if (delay_entry)
1433 if (delay_entry->i1 == insn)
1435 DEP_COST (link) = pair_delay (delay_entry);
1436 return DEP_COST (link);
1441 /* A USE insn should never require the value used to be computed.
1442 This allows the computation of a function's result and parameter
1443 values to overlap the return and call. We don't care about the
1444 dependence cost when only decreasing register pressure. */
1445 if (recog_memoized (used) < 0)
1447 cost = 0;
1448 recog_memoized (insn);
1450 else
1452 enum reg_note dep_type = DEP_TYPE (link);
1454 cost = insn_cost (insn);
1456 if (INSN_CODE (insn) >= 0)
1458 if (dep_type == REG_DEP_ANTI)
1459 cost = 0;
1460 else if (dep_type == REG_DEP_OUTPUT)
1462 cost = (insn_default_latency (insn)
1463 - insn_default_latency (used));
1464 if (cost <= 0)
1465 cost = 1;
1467 else if (bypass_p (insn))
1468 cost = insn_latency (insn, used);
1472 if (targetm.sched.adjust_cost_2)
1473 cost = targetm.sched.adjust_cost_2 (used, (int) dep_type, insn, cost,
1474 dw);
1475 else if (targetm.sched.adjust_cost != NULL)
1477 /* This variable is used for backward compatibility with the
1478 targets. */
1479 rtx dep_cost_rtx_link = alloc_INSN_LIST (NULL_RTX, NULL_RTX);
1481 /* Make it self-cycled, so that if some tries to walk over this
1482 incomplete list he/she will be caught in an endless loop. */
1483 XEXP (dep_cost_rtx_link, 1) = dep_cost_rtx_link;
1485 /* Targets use only REG_NOTE_KIND of the link. */
1486 PUT_REG_NOTE_KIND (dep_cost_rtx_link, DEP_TYPE (link));
1488 cost = targetm.sched.adjust_cost (used, dep_cost_rtx_link,
1489 insn, cost);
1491 free_INSN_LIST_node (dep_cost_rtx_link);
1494 if (cost < 0)
1495 cost = 0;
1498 DEP_COST (link) = cost;
1499 return cost;
1502 /* Compute cost of dependence LINK.
1503 This is the number of cycles between instruction issue and
1504 instruction results. */
1506 dep_cost (dep_t link)
1508 return dep_cost_1 (link, 0);
1511 /* Use this sel-sched.c friendly function in reorder2 instead of increasing
1512 INSN_PRIORITY explicitly. */
1513 void
1514 increase_insn_priority (rtx insn, int amount)
1516 if (!sel_sched_p ())
1518 /* We're dealing with haifa-sched.c INSN_PRIORITY. */
1519 if (INSN_PRIORITY_KNOWN (insn))
1520 INSN_PRIORITY (insn) += amount;
1522 else
1524 /* In sel-sched.c INSN_PRIORITY is not kept up to date.
1525 Use EXPR_PRIORITY instead. */
1526 sel_add_to_insn_priority (insn, amount);
1530 /* Return 'true' if DEP should be included in priority calculations. */
1531 static bool
1532 contributes_to_priority_p (dep_t dep)
1534 if (DEBUG_INSN_P (DEP_CON (dep))
1535 || DEBUG_INSN_P (DEP_PRO (dep)))
1536 return false;
1538 /* Critical path is meaningful in block boundaries only. */
1539 if (!current_sched_info->contributes_to_priority (DEP_CON (dep),
1540 DEP_PRO (dep)))
1541 return false;
1543 if (DEP_REPLACE (dep) != NULL)
1544 return false;
1546 /* If flag COUNT_SPEC_IN_CRITICAL_PATH is set,
1547 then speculative instructions will less likely be
1548 scheduled. That is because the priority of
1549 their producers will increase, and, thus, the
1550 producers will more likely be scheduled, thus,
1551 resolving the dependence. */
1552 if (sched_deps_info->generate_spec_deps
1553 && !(spec_info->flags & COUNT_SPEC_IN_CRITICAL_PATH)
1554 && (DEP_STATUS (dep) & SPECULATIVE))
1555 return false;
1557 return true;
1560 /* Compute the number of nondebug deps in list LIST for INSN. */
1562 static int
1563 dep_list_size (rtx insn, sd_list_types_def list)
1565 sd_iterator_def sd_it;
1566 dep_t dep;
1567 int dbgcount = 0, nodbgcount = 0;
1569 if (!MAY_HAVE_DEBUG_INSNS)
1570 return sd_lists_size (insn, list);
1572 FOR_EACH_DEP (insn, list, sd_it, dep)
1574 if (DEBUG_INSN_P (DEP_CON (dep)))
1575 dbgcount++;
1576 else if (!DEBUG_INSN_P (DEP_PRO (dep)))
1577 nodbgcount++;
1580 gcc_assert (dbgcount + nodbgcount == sd_lists_size (insn, list));
1582 return nodbgcount;
1585 /* Compute the priority number for INSN. */
1586 static int
1587 priority (rtx insn)
1589 if (! INSN_P (insn))
1590 return 0;
1592 /* We should not be interested in priority of an already scheduled insn. */
1593 gcc_assert (QUEUE_INDEX (insn) != QUEUE_SCHEDULED);
1595 if (!INSN_PRIORITY_KNOWN (insn))
1597 int this_priority = -1;
1599 if (dep_list_size (insn, SD_LIST_FORW) == 0)
1600 /* ??? We should set INSN_PRIORITY to insn_cost when and insn has
1601 some forward deps but all of them are ignored by
1602 contributes_to_priority hook. At the moment we set priority of
1603 such insn to 0. */
1604 this_priority = insn_cost (insn);
1605 else
1607 rtx prev_first, twin;
1608 basic_block rec;
1610 /* For recovery check instructions we calculate priority slightly
1611 different than that of normal instructions. Instead of walking
1612 through INSN_FORW_DEPS (check) list, we walk through
1613 INSN_FORW_DEPS list of each instruction in the corresponding
1614 recovery block. */
1616 /* Selective scheduling does not define RECOVERY_BLOCK macro. */
1617 rec = sel_sched_p () ? NULL : RECOVERY_BLOCK (insn);
1618 if (!rec || rec == EXIT_BLOCK_PTR)
1620 prev_first = PREV_INSN (insn);
1621 twin = insn;
1623 else
1625 prev_first = NEXT_INSN (BB_HEAD (rec));
1626 twin = PREV_INSN (BB_END (rec));
1631 sd_iterator_def sd_it;
1632 dep_t dep;
1634 FOR_EACH_DEP (twin, SD_LIST_FORW, sd_it, dep)
1636 rtx next;
1637 int next_priority;
1639 next = DEP_CON (dep);
1641 if (BLOCK_FOR_INSN (next) != rec)
1643 int cost;
1645 if (!contributes_to_priority_p (dep))
1646 continue;
1648 if (twin == insn)
1649 cost = dep_cost (dep);
1650 else
1652 struct _dep _dep1, *dep1 = &_dep1;
1654 init_dep (dep1, insn, next, REG_DEP_ANTI);
1656 cost = dep_cost (dep1);
1659 next_priority = cost + priority (next);
1661 if (next_priority > this_priority)
1662 this_priority = next_priority;
1666 twin = PREV_INSN (twin);
1668 while (twin != prev_first);
1671 if (this_priority < 0)
1673 gcc_assert (this_priority == -1);
1675 this_priority = insn_cost (insn);
1678 INSN_PRIORITY (insn) = this_priority;
1679 INSN_PRIORITY_STATUS (insn) = 1;
1682 return INSN_PRIORITY (insn);
1685 /* Macros and functions for keeping the priority queue sorted, and
1686 dealing with queuing and dequeuing of instructions. */
1688 #define SCHED_SORT(READY, N_READY) \
1689 do { if ((N_READY) == 2) \
1690 swap_sort (READY, N_READY); \
1691 else if ((N_READY) > 2) \
1692 qsort (READY, N_READY, sizeof (rtx), rank_for_schedule); } \
1693 while (0)
1695 /* For each pressure class CL, set DEATH[CL] to the number of registers
1696 in that class that die in INSN. */
1698 static void
1699 calculate_reg_deaths (rtx insn, int *death)
1701 int i;
1702 struct reg_use_data *use;
1704 for (i = 0; i < ira_pressure_classes_num; i++)
1705 death[ira_pressure_classes[i]] = 0;
1706 for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
1707 if (dying_use_p (use))
1708 mark_regno_birth_or_death (0, death, use->regno, true);
1711 /* Setup info about the current register pressure impact of scheduling
1712 INSN at the current scheduling point. */
1713 static void
1714 setup_insn_reg_pressure_info (rtx insn)
1716 int i, change, before, after, hard_regno;
1717 int excess_cost_change;
1718 enum machine_mode mode;
1719 enum reg_class cl;
1720 struct reg_pressure_data *pressure_info;
1721 int *max_reg_pressure;
1722 static int death[N_REG_CLASSES];
1724 gcc_checking_assert (!DEBUG_INSN_P (insn));
1726 excess_cost_change = 0;
1727 calculate_reg_deaths (insn, death);
1728 pressure_info = INSN_REG_PRESSURE (insn);
1729 max_reg_pressure = INSN_MAX_REG_PRESSURE (insn);
1730 gcc_assert (pressure_info != NULL && max_reg_pressure != NULL);
1731 for (i = 0; i < ira_pressure_classes_num; i++)
1733 cl = ira_pressure_classes[i];
1734 gcc_assert (curr_reg_pressure[cl] >= 0);
1735 change = (int) pressure_info[i].set_increase - death[cl];
1736 before = MAX (0, max_reg_pressure[i] - ira_class_hard_regs_num[cl]);
1737 after = MAX (0, max_reg_pressure[i] + change
1738 - ira_class_hard_regs_num[cl]);
1739 hard_regno = ira_class_hard_regs[cl][0];
1740 gcc_assert (hard_regno >= 0);
1741 mode = reg_raw_mode[hard_regno];
1742 excess_cost_change += ((after - before)
1743 * (ira_memory_move_cost[mode][cl][0]
1744 + ira_memory_move_cost[mode][cl][1]));
1746 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insn) = excess_cost_change;
1749 /* This is the first page of code related to SCHED_PRESSURE_MODEL.
1750 It tries to make the scheduler take register pressure into account
1751 without introducing too many unnecessary stalls. It hooks into the
1752 main scheduling algorithm at several points:
1754 - Before scheduling starts, model_start_schedule constructs a
1755 "model schedule" for the current block. This model schedule is
1756 chosen solely to keep register pressure down. It does not take the
1757 target's pipeline or the original instruction order into account,
1758 except as a tie-breaker. It also doesn't work to a particular
1759 pressure limit.
1761 This model schedule gives us an idea of what pressure can be
1762 achieved for the block and gives us an example of a schedule that
1763 keeps to that pressure. It also makes the final schedule less
1764 dependent on the original instruction order. This is important
1765 because the original order can either be "wide" (many values live
1766 at once, such as in user-scheduled code) or "narrow" (few values
1767 live at once, such as after loop unrolling, where several
1768 iterations are executed sequentially).
1770 We do not apply this model schedule to the rtx stream. We simply
1771 record it in model_schedule. We also compute the maximum pressure,
1772 MP, that was seen during this schedule.
1774 - Instructions are added to the ready queue even if they require
1775 a stall. The length of the stall is instead computed as:
1777 MAX (INSN_TICK (INSN) - clock_var, 0)
1779 (= insn_delay). This allows rank_for_schedule to choose between
1780 introducing a deliberate stall or increasing pressure.
1782 - Before sorting the ready queue, model_set_excess_costs assigns
1783 a pressure-based cost to each ready instruction in the queue.
1784 This is the instruction's INSN_REG_PRESSURE_EXCESS_COST_CHANGE
1785 (ECC for short) and is effectively measured in cycles.
1787 - rank_for_schedule ranks instructions based on:
1789 ECC (insn) + insn_delay (insn)
1791 then as:
1793 insn_delay (insn)
1795 So, for example, an instruction X1 with an ECC of 1 that can issue
1796 now will win over an instruction X0 with an ECC of zero that would
1797 introduce a stall of one cycle. However, an instruction X2 with an
1798 ECC of 2 that can issue now will lose to both X0 and X1.
1800 - When an instruction is scheduled, model_recompute updates the model
1801 schedule with the new pressures (some of which might now exceed the
1802 original maximum pressure MP). model_update_limit_points then searches
1803 for the new point of maximum pressure, if not already known. */
1805 /* Used to separate high-verbosity debug information for SCHED_PRESSURE_MODEL
1806 from surrounding debug information. */
1807 #define MODEL_BAR \
1808 ";;\t\t+------------------------------------------------------\n"
1810 /* Information about the pressure on a particular register class at a
1811 particular point of the model schedule. */
1812 struct model_pressure_data {
1813 /* The pressure at this point of the model schedule, or -1 if the
1814 point is associated with an instruction that has already been
1815 scheduled. */
1816 int ref_pressure;
1818 /* The maximum pressure during or after this point of the model schedule. */
1819 int max_pressure;
1822 /* Per-instruction information that is used while building the model
1823 schedule. Here, "schedule" refers to the model schedule rather
1824 than the main schedule. */
1825 struct model_insn_info {
1826 /* The instruction itself. */
1827 rtx insn;
1829 /* If this instruction is in model_worklist, these fields link to the
1830 previous (higher-priority) and next (lower-priority) instructions
1831 in the list. */
1832 struct model_insn_info *prev;
1833 struct model_insn_info *next;
1835 /* While constructing the schedule, QUEUE_INDEX describes whether an
1836 instruction has already been added to the schedule (QUEUE_SCHEDULED),
1837 is in model_worklist (QUEUE_READY), or neither (QUEUE_NOWHERE).
1838 old_queue records the value that QUEUE_INDEX had before scheduling
1839 started, so that we can restore it once the schedule is complete. */
1840 int old_queue;
1842 /* The relative importance of an unscheduled instruction. Higher
1843 values indicate greater importance. */
1844 unsigned int model_priority;
1846 /* The length of the longest path of satisfied true dependencies
1847 that leads to this instruction. */
1848 unsigned int depth;
1850 /* The length of the longest path of dependencies of any kind
1851 that leads from this instruction. */
1852 unsigned int alap;
1854 /* The number of predecessor nodes that must still be scheduled. */
1855 int unscheduled_preds;
1858 /* Information about the pressure limit for a particular register class.
1859 This structure is used when applying a model schedule to the main
1860 schedule. */
1861 struct model_pressure_limit {
1862 /* The maximum register pressure seen in the original model schedule. */
1863 int orig_pressure;
1865 /* The maximum register pressure seen in the current model schedule
1866 (which excludes instructions that have already been scheduled). */
1867 int pressure;
1869 /* The point of the current model schedule at which PRESSURE is first
1870 reached. It is set to -1 if the value needs to be recomputed. */
1871 int point;
1874 /* Describes a particular way of measuring register pressure. */
1875 struct model_pressure_group {
1876 /* Index PCI describes the maximum pressure on ira_pressure_classes[PCI]. */
1877 struct model_pressure_limit limits[N_REG_CLASSES];
1879 /* Index (POINT * ira_num_pressure_classes + PCI) describes the pressure
1880 on register class ira_pressure_classes[PCI] at point POINT of the
1881 current model schedule. A POINT of model_num_insns describes the
1882 pressure at the end of the schedule. */
1883 struct model_pressure_data *model;
1886 /* Index POINT gives the instruction at point POINT of the model schedule.
1887 This array doesn't change during main scheduling. */
1888 static vec<rtx> model_schedule;
1890 /* The list of instructions in the model worklist, sorted in order of
1891 decreasing priority. */
1892 static struct model_insn_info *model_worklist;
1894 /* Index I describes the instruction with INSN_LUID I. */
1895 static struct model_insn_info *model_insns;
1897 /* The number of instructions in the model schedule. */
1898 static int model_num_insns;
1900 /* The index of the first instruction in model_schedule that hasn't yet been
1901 added to the main schedule, or model_num_insns if all of them have. */
1902 static int model_curr_point;
1904 /* Describes the pressure before each instruction in the model schedule. */
1905 static struct model_pressure_group model_before_pressure;
1907 /* The first unused model_priority value (as used in model_insn_info). */
1908 static unsigned int model_next_priority;
1911 /* The model_pressure_data for ira_pressure_classes[PCI] in GROUP
1912 at point POINT of the model schedule. */
1913 #define MODEL_PRESSURE_DATA(GROUP, POINT, PCI) \
1914 (&(GROUP)->model[(POINT) * ira_pressure_classes_num + (PCI)])
1916 /* The maximum pressure on ira_pressure_classes[PCI] in GROUP at or
1917 after point POINT of the model schedule. */
1918 #define MODEL_MAX_PRESSURE(GROUP, POINT, PCI) \
1919 (MODEL_PRESSURE_DATA (GROUP, POINT, PCI)->max_pressure)
1921 /* The pressure on ira_pressure_classes[PCI] in GROUP at point POINT
1922 of the model schedule. */
1923 #define MODEL_REF_PRESSURE(GROUP, POINT, PCI) \
1924 (MODEL_PRESSURE_DATA (GROUP, POINT, PCI)->ref_pressure)
1926 /* Information about INSN that is used when creating the model schedule. */
1927 #define MODEL_INSN_INFO(INSN) \
1928 (&model_insns[INSN_LUID (INSN)])
1930 /* The instruction at point POINT of the model schedule. */
1931 #define MODEL_INSN(POINT) \
1932 (model_schedule[POINT])
1935 /* Return INSN's index in the model schedule, or model_num_insns if it
1936 doesn't belong to that schedule. */
1938 static int
1939 model_index (rtx insn)
1941 if (INSN_MODEL_INDEX (insn) == 0)
1942 return model_num_insns;
1943 return INSN_MODEL_INDEX (insn) - 1;
1946 /* Make sure that GROUP->limits is up-to-date for the current point
1947 of the model schedule. */
1949 static void
1950 model_update_limit_points_in_group (struct model_pressure_group *group)
1952 int pci, max_pressure, point;
1954 for (pci = 0; pci < ira_pressure_classes_num; pci++)
1956 /* We may have passed the final point at which the pressure in
1957 group->limits[pci].pressure was reached. Update the limit if so. */
1958 max_pressure = MODEL_MAX_PRESSURE (group, model_curr_point, pci);
1959 group->limits[pci].pressure = max_pressure;
1961 /* Find the point at which MAX_PRESSURE is first reached. We need
1962 to search in three cases:
1964 - We've already moved past the previous pressure point.
1965 In this case we search forward from model_curr_point.
1967 - We scheduled the previous point of maximum pressure ahead of
1968 its position in the model schedule, but doing so didn't bring
1969 the pressure point earlier. In this case we search forward
1970 from that previous pressure point.
1972 - Scheduling an instruction early caused the maximum pressure
1973 to decrease. In this case we will have set the pressure
1974 point to -1, and we search forward from model_curr_point. */
1975 point = MAX (group->limits[pci].point, model_curr_point);
1976 while (point < model_num_insns
1977 && MODEL_REF_PRESSURE (group, point, pci) < max_pressure)
1978 point++;
1979 group->limits[pci].point = point;
1981 gcc_assert (MODEL_REF_PRESSURE (group, point, pci) == max_pressure);
1982 gcc_assert (MODEL_MAX_PRESSURE (group, point, pci) == max_pressure);
1986 /* Make sure that all register-pressure limits are up-to-date for the
1987 current position in the model schedule. */
1989 static void
1990 model_update_limit_points (void)
1992 model_update_limit_points_in_group (&model_before_pressure);
1995 /* Return the model_index of the last unscheduled use in chain USE
1996 outside of USE's instruction. Return -1 if there are no other uses,
1997 or model_num_insns if the register is live at the end of the block. */
1999 static int
2000 model_last_use_except (struct reg_use_data *use)
2002 struct reg_use_data *next;
2003 int last, index;
2005 last = -1;
2006 for (next = use->next_regno_use; next != use; next = next->next_regno_use)
2007 if (NONDEBUG_INSN_P (next->insn)
2008 && QUEUE_INDEX (next->insn) != QUEUE_SCHEDULED)
2010 index = model_index (next->insn);
2011 if (index == model_num_insns)
2012 return model_num_insns;
2013 if (last < index)
2014 last = index;
2016 return last;
2019 /* An instruction with model_index POINT has just been scheduled, and it
2020 adds DELTA to the pressure on ira_pressure_classes[PCI] after POINT - 1.
2021 Update MODEL_REF_PRESSURE (GROUP, POINT, PCI) and
2022 MODEL_MAX_PRESSURE (GROUP, POINT, PCI) accordingly. */
2024 static void
2025 model_start_update_pressure (struct model_pressure_group *group,
2026 int point, int pci, int delta)
2028 int next_max_pressure;
2030 if (point == model_num_insns)
2032 /* The instruction wasn't part of the model schedule; it was moved
2033 from a different block. Update the pressure for the end of
2034 the model schedule. */
2035 MODEL_REF_PRESSURE (group, point, pci) += delta;
2036 MODEL_MAX_PRESSURE (group, point, pci) += delta;
2038 else
2040 /* Record that this instruction has been scheduled. Nothing now
2041 changes between POINT and POINT + 1, so get the maximum pressure
2042 from the latter. If the maximum pressure decreases, the new
2043 pressure point may be before POINT. */
2044 MODEL_REF_PRESSURE (group, point, pci) = -1;
2045 next_max_pressure = MODEL_MAX_PRESSURE (group, point + 1, pci);
2046 if (MODEL_MAX_PRESSURE (group, point, pci) > next_max_pressure)
2048 MODEL_MAX_PRESSURE (group, point, pci) = next_max_pressure;
2049 if (group->limits[pci].point == point)
2050 group->limits[pci].point = -1;
2055 /* Record that scheduling a later instruction has changed the pressure
2056 at point POINT of the model schedule by DELTA (which might be 0).
2057 Update GROUP accordingly. Return nonzero if these changes might
2058 trigger changes to previous points as well. */
2060 static int
2061 model_update_pressure (struct model_pressure_group *group,
2062 int point, int pci, int delta)
2064 int ref_pressure, max_pressure, next_max_pressure;
2066 /* If POINT hasn't yet been scheduled, update its pressure. */
2067 ref_pressure = MODEL_REF_PRESSURE (group, point, pci);
2068 if (ref_pressure >= 0 && delta != 0)
2070 ref_pressure += delta;
2071 MODEL_REF_PRESSURE (group, point, pci) = ref_pressure;
2073 /* Check whether the maximum pressure in the overall schedule
2074 has increased. (This means that the MODEL_MAX_PRESSURE of
2075 every point <= POINT will need to increae too; see below.) */
2076 if (group->limits[pci].pressure < ref_pressure)
2077 group->limits[pci].pressure = ref_pressure;
2079 /* If we are at maximum pressure, and the maximum pressure
2080 point was previously unknown or later than POINT,
2081 bring it forward. */
2082 if (group->limits[pci].pressure == ref_pressure
2083 && !IN_RANGE (group->limits[pci].point, 0, point))
2084 group->limits[pci].point = point;
2086 /* If POINT used to be the point of maximum pressure, but isn't
2087 any longer, we need to recalculate it using a forward walk. */
2088 if (group->limits[pci].pressure > ref_pressure
2089 && group->limits[pci].point == point)
2090 group->limits[pci].point = -1;
2093 /* Update the maximum pressure at POINT. Changes here might also
2094 affect the maximum pressure at POINT - 1. */
2095 next_max_pressure = MODEL_MAX_PRESSURE (group, point + 1, pci);
2096 max_pressure = MAX (ref_pressure, next_max_pressure);
2097 if (MODEL_MAX_PRESSURE (group, point, pci) != max_pressure)
2099 MODEL_MAX_PRESSURE (group, point, pci) = max_pressure;
2100 return 1;
2102 return 0;
2105 /* INSN has just been scheduled. Update the model schedule accordingly. */
2107 static void
2108 model_recompute (rtx insn)
2110 struct {
2111 int last_use;
2112 int regno;
2113 } uses[FIRST_PSEUDO_REGISTER + MAX_RECOG_OPERANDS];
2114 struct reg_use_data *use;
2115 struct reg_pressure_data *reg_pressure;
2116 int delta[N_REG_CLASSES];
2117 int pci, point, mix, new_last, cl, ref_pressure, queue;
2118 unsigned int i, num_uses, num_pending_births;
2119 bool print_p;
2121 /* The destinations of INSN were previously live from POINT onwards, but are
2122 now live from model_curr_point onwards. Set up DELTA accordingly. */
2123 point = model_index (insn);
2124 reg_pressure = INSN_REG_PRESSURE (insn);
2125 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2127 cl = ira_pressure_classes[pci];
2128 delta[cl] = reg_pressure[pci].set_increase;
2131 /* Record which registers previously died at POINT, but which now die
2132 before POINT. Adjust DELTA so that it represents the effect of
2133 this change after POINT - 1. Set NUM_PENDING_BIRTHS to the number of
2134 registers that will be born in the range [model_curr_point, POINT). */
2135 num_uses = 0;
2136 num_pending_births = 0;
2137 for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
2139 new_last = model_last_use_except (use);
2140 if (new_last < point)
2142 gcc_assert (num_uses < ARRAY_SIZE (uses));
2143 uses[num_uses].last_use = new_last;
2144 uses[num_uses].regno = use->regno;
2145 /* This register is no longer live after POINT - 1. */
2146 mark_regno_birth_or_death (NULL, delta, use->regno, false);
2147 num_uses++;
2148 if (new_last >= 0)
2149 num_pending_births++;
2153 /* Update the MODEL_REF_PRESSURE and MODEL_MAX_PRESSURE for POINT.
2154 Also set each group pressure limit for POINT. */
2155 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2157 cl = ira_pressure_classes[pci];
2158 model_start_update_pressure (&model_before_pressure,
2159 point, pci, delta[cl]);
2162 /* Walk the model schedule backwards, starting immediately before POINT. */
2163 print_p = false;
2164 if (point != model_curr_point)
2167 point--;
2168 insn = MODEL_INSN (point);
2169 queue = QUEUE_INDEX (insn);
2171 if (queue != QUEUE_SCHEDULED)
2173 /* DELTA describes the effect of the move on the register pressure
2174 after POINT. Make it describe the effect on the pressure
2175 before POINT. */
2176 i = 0;
2177 while (i < num_uses)
2179 if (uses[i].last_use == point)
2181 /* This register is now live again. */
2182 mark_regno_birth_or_death (NULL, delta,
2183 uses[i].regno, true);
2185 /* Remove this use from the array. */
2186 uses[i] = uses[num_uses - 1];
2187 num_uses--;
2188 num_pending_births--;
2190 else
2191 i++;
2194 if (sched_verbose >= 5)
2196 if (!print_p)
2198 fprintf (sched_dump, MODEL_BAR);
2199 fprintf (sched_dump, ";;\t\t| New pressure for model"
2200 " schedule\n");
2201 fprintf (sched_dump, MODEL_BAR);
2202 print_p = true;
2205 fprintf (sched_dump, ";;\t\t| %3d %4d %-30s ",
2206 point, INSN_UID (insn),
2207 str_pattern_slim (PATTERN (insn)));
2208 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2210 cl = ira_pressure_classes[pci];
2211 ref_pressure = MODEL_REF_PRESSURE (&model_before_pressure,
2212 point, pci);
2213 fprintf (sched_dump, " %s:[%d->%d]",
2214 reg_class_names[ira_pressure_classes[pci]],
2215 ref_pressure, ref_pressure + delta[cl]);
2217 fprintf (sched_dump, "\n");
2221 /* Adjust the pressure at POINT. Set MIX to nonzero if POINT - 1
2222 might have changed as well. */
2223 mix = num_pending_births;
2224 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2226 cl = ira_pressure_classes[pci];
2227 mix |= delta[cl];
2228 mix |= model_update_pressure (&model_before_pressure,
2229 point, pci, delta[cl]);
2232 while (mix && point > model_curr_point);
2234 if (print_p)
2235 fprintf (sched_dump, MODEL_BAR);
2238 /* After DEP, which was cancelled, has been resolved for insn NEXT,
2239 check whether the insn's pattern needs restoring. */
2240 static bool
2241 must_restore_pattern_p (rtx next, dep_t dep)
2243 if (QUEUE_INDEX (next) == QUEUE_SCHEDULED)
2244 return false;
2246 if (DEP_TYPE (dep) == REG_DEP_CONTROL)
2248 gcc_assert (ORIG_PAT (next) != NULL_RTX);
2249 gcc_assert (next == DEP_CON (dep));
2251 else
2253 struct dep_replacement *desc = DEP_REPLACE (dep);
2254 if (desc->insn != next)
2256 gcc_assert (*desc->loc == desc->orig);
2257 return false;
2260 return true;
2263 /* model_spill_cost (CL, P, P') returns the cost of increasing the
2264 pressure on CL from P to P'. We use this to calculate a "base ECC",
2265 baseECC (CL, X), for each pressure class CL and each instruction X.
2266 Supposing X changes the pressure on CL from P to P', and that the
2267 maximum pressure on CL in the current model schedule is MP', then:
2269 * if X occurs before or at the next point of maximum pressure in
2270 the model schedule and P' > MP', then:
2272 baseECC (CL, X) = model_spill_cost (CL, MP, P')
2274 The idea is that the pressure after scheduling a fixed set of
2275 instructions -- in this case, the set up to and including the
2276 next maximum pressure point -- is going to be the same regardless
2277 of the order; we simply want to keep the intermediate pressure
2278 under control. Thus X has a cost of zero unless scheduling it
2279 now would exceed MP'.
2281 If all increases in the set are by the same amount, no zero-cost
2282 instruction will ever cause the pressure to exceed MP'. However,
2283 if X is instead moved past an instruction X' with pressure in the
2284 range (MP' - (P' - P), MP'), the pressure at X' will increase
2285 beyond MP'. Since baseECC is very much a heuristic anyway,
2286 it doesn't seem worth the overhead of tracking cases like these.
2288 The cost of exceeding MP' is always based on the original maximum
2289 pressure MP. This is so that going 2 registers over the original
2290 limit has the same cost regardless of whether it comes from two
2291 separate +1 deltas or from a single +2 delta.
2293 * if X occurs after the next point of maximum pressure in the model
2294 schedule and P' > P, then:
2296 baseECC (CL, X) = model_spill_cost (CL, MP, MP' + (P' - P))
2298 That is, if we move X forward across a point of maximum pressure,
2299 and if X increases the pressure by P' - P, then we conservatively
2300 assume that scheduling X next would increase the maximum pressure
2301 by P' - P. Again, the cost of doing this is based on the original
2302 maximum pressure MP, for the same reason as above.
2304 * if P' < P, P > MP, and X occurs at or after the next point of
2305 maximum pressure, then:
2307 baseECC (CL, X) = -model_spill_cost (CL, MAX (MP, P'), P)
2309 That is, if we have already exceeded the original maximum pressure MP,
2310 and if X might reduce the maximum pressure again -- or at least push
2311 it further back, and thus allow more scheduling freedom -- it is given
2312 a negative cost to reflect the improvement.
2314 * otherwise,
2316 baseECC (CL, X) = 0
2318 In this case, X is not expected to affect the maximum pressure MP',
2319 so it has zero cost.
2321 We then create a combined value baseECC (X) that is the sum of
2322 baseECC (CL, X) for each pressure class CL.
2324 baseECC (X) could itself be used as the ECC value described above.
2325 However, this is often too conservative, in the sense that it
2326 tends to make high-priority instructions that increase pressure
2327 wait too long in cases where introducing a spill would be better.
2328 For this reason the final ECC is a priority-adjusted form of
2329 baseECC (X). Specifically, we calculate:
2331 P (X) = INSN_PRIORITY (X) - insn_delay (X) - baseECC (X)
2332 baseP = MAX { P (X) | baseECC (X) <= 0 }
2334 Then:
2336 ECC (X) = MAX (MIN (baseP - P (X), baseECC (X)), 0)
2338 Thus an instruction's effect on pressure is ignored if it has a high
2339 enough priority relative to the ones that don't increase pressure.
2340 Negative values of baseECC (X) do not increase the priority of X
2341 itself, but they do make it harder for other instructions to
2342 increase the pressure further.
2344 This pressure cost is deliberately timid. The intention has been
2345 to choose a heuristic that rarely interferes with the normal list
2346 scheduler in cases where that scheduler would produce good code.
2347 We simply want to curb some of its worst excesses. */
2349 /* Return the cost of increasing the pressure in class CL from FROM to TO.
2351 Here we use the very simplistic cost model that every register above
2352 ira_class_hard_regs_num[CL] has a spill cost of 1. We could use other
2353 measures instead, such as one based on MEMORY_MOVE_COST. However:
2355 (1) In order for an instruction to be scheduled, the higher cost
2356 would need to be justified in a single saving of that many stalls.
2357 This is overly pessimistic, because the benefit of spilling is
2358 often to avoid a sequence of several short stalls rather than
2359 a single long one.
2361 (2) The cost is still arbitrary. Because we are not allocating
2362 registers during scheduling, we have no way of knowing for
2363 sure how many memory accesses will be required by each spill,
2364 where the spills will be placed within the block, or even
2365 which block(s) will contain the spills.
2367 So a higher cost than 1 is often too conservative in practice,
2368 forcing blocks to contain unnecessary stalls instead of spill code.
2369 The simple cost below seems to be the best compromise. It reduces
2370 the interference with the normal list scheduler, which helps make
2371 it more suitable for a default-on option. */
2373 static int
2374 model_spill_cost (int cl, int from, int to)
2376 from = MAX (from, ira_class_hard_regs_num[cl]);
2377 return MAX (to, from) - from;
2380 /* Return baseECC (ira_pressure_classes[PCI], POINT), given that
2381 P = curr_reg_pressure[ira_pressure_classes[PCI]] and that
2382 P' = P + DELTA. */
2384 static int
2385 model_excess_group_cost (struct model_pressure_group *group,
2386 int point, int pci, int delta)
2388 int pressure, cl;
2390 cl = ira_pressure_classes[pci];
2391 if (delta < 0 && point >= group->limits[pci].point)
2393 pressure = MAX (group->limits[pci].orig_pressure,
2394 curr_reg_pressure[cl] + delta);
2395 return -model_spill_cost (cl, pressure, curr_reg_pressure[cl]);
2398 if (delta > 0)
2400 if (point > group->limits[pci].point)
2401 pressure = group->limits[pci].pressure + delta;
2402 else
2403 pressure = curr_reg_pressure[cl] + delta;
2405 if (pressure > group->limits[pci].pressure)
2406 return model_spill_cost (cl, group->limits[pci].orig_pressure,
2407 pressure);
2410 return 0;
2413 /* Return baseECC (MODEL_INSN (INSN)). Dump the costs to sched_dump
2414 if PRINT_P. */
2416 static int
2417 model_excess_cost (rtx insn, bool print_p)
2419 int point, pci, cl, cost, this_cost, delta;
2420 struct reg_pressure_data *insn_reg_pressure;
2421 int insn_death[N_REG_CLASSES];
2423 calculate_reg_deaths (insn, insn_death);
2424 point = model_index (insn);
2425 insn_reg_pressure = INSN_REG_PRESSURE (insn);
2426 cost = 0;
2428 if (print_p)
2429 fprintf (sched_dump, ";;\t\t| %3d %4d | %4d %+3d |", point,
2430 INSN_UID (insn), INSN_PRIORITY (insn), insn_delay (insn));
2432 /* Sum up the individual costs for each register class. */
2433 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2435 cl = ira_pressure_classes[pci];
2436 delta = insn_reg_pressure[pci].set_increase - insn_death[cl];
2437 this_cost = model_excess_group_cost (&model_before_pressure,
2438 point, pci, delta);
2439 cost += this_cost;
2440 if (print_p)
2441 fprintf (sched_dump, " %s:[%d base cost %d]",
2442 reg_class_names[cl], delta, this_cost);
2445 if (print_p)
2446 fprintf (sched_dump, "\n");
2448 return cost;
2451 /* Dump the next points of maximum pressure for GROUP. */
2453 static void
2454 model_dump_pressure_points (struct model_pressure_group *group)
2456 int pci, cl;
2458 fprintf (sched_dump, ";;\t\t| pressure points");
2459 for (pci = 0; pci < ira_pressure_classes_num; pci++)
2461 cl = ira_pressure_classes[pci];
2462 fprintf (sched_dump, " %s:[%d->%d at ", reg_class_names[cl],
2463 curr_reg_pressure[cl], group->limits[pci].pressure);
2464 if (group->limits[pci].point < model_num_insns)
2465 fprintf (sched_dump, "%d:%d]", group->limits[pci].point,
2466 INSN_UID (MODEL_INSN (group->limits[pci].point)));
2467 else
2468 fprintf (sched_dump, "end]");
2470 fprintf (sched_dump, "\n");
2473 /* Set INSN_REG_PRESSURE_EXCESS_COST_CHANGE for INSNS[0...COUNT-1]. */
2475 static void
2476 model_set_excess_costs (rtx *insns, int count)
2478 int i, cost, priority_base, priority;
2479 bool print_p;
2481 /* Record the baseECC value for each instruction in the model schedule,
2482 except that negative costs are converted to zero ones now rather thatn
2483 later. Do not assign a cost to debug instructions, since they must
2484 not change code-generation decisions. Experiments suggest we also
2485 get better results by not assigning a cost to instructions from
2486 a different block.
2488 Set PRIORITY_BASE to baseP in the block comment above. This is the
2489 maximum priority of the "cheap" instructions, which should always
2490 include the next model instruction. */
2491 priority_base = 0;
2492 print_p = false;
2493 for (i = 0; i < count; i++)
2494 if (INSN_MODEL_INDEX (insns[i]))
2496 if (sched_verbose >= 6 && !print_p)
2498 fprintf (sched_dump, MODEL_BAR);
2499 fprintf (sched_dump, ";;\t\t| Pressure costs for ready queue\n");
2500 model_dump_pressure_points (&model_before_pressure);
2501 fprintf (sched_dump, MODEL_BAR);
2502 print_p = true;
2504 cost = model_excess_cost (insns[i], print_p);
2505 if (cost <= 0)
2507 priority = INSN_PRIORITY (insns[i]) - insn_delay (insns[i]) - cost;
2508 priority_base = MAX (priority_base, priority);
2509 cost = 0;
2511 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i]) = cost;
2513 if (print_p)
2514 fprintf (sched_dump, MODEL_BAR);
2516 /* Use MAX (baseECC, 0) and baseP to calculcate ECC for each
2517 instruction. */
2518 for (i = 0; i < count; i++)
2520 cost = INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i]);
2521 priority = INSN_PRIORITY (insns[i]) - insn_delay (insns[i]);
2522 if (cost > 0 && priority > priority_base)
2524 cost += priority_base - priority;
2525 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (insns[i]) = MAX (cost, 0);
2530 /* Returns a positive value if x is preferred; returns a negative value if
2531 y is preferred. Should never return 0, since that will make the sort
2532 unstable. */
2534 static int
2535 rank_for_schedule (const void *x, const void *y)
2537 rtx tmp = *(const rtx *) y;
2538 rtx tmp2 = *(const rtx *) x;
2539 int tmp_class, tmp2_class;
2540 int val, priority_val, info_val, diff;
2542 if (MAY_HAVE_DEBUG_INSNS)
2544 /* Schedule debug insns as early as possible. */
2545 if (DEBUG_INSN_P (tmp) && !DEBUG_INSN_P (tmp2))
2546 return -1;
2547 else if (!DEBUG_INSN_P (tmp) && DEBUG_INSN_P (tmp2))
2548 return 1;
2549 else if (DEBUG_INSN_P (tmp) && DEBUG_INSN_P (tmp2))
2550 return INSN_LUID (tmp) - INSN_LUID (tmp2);
2553 if (live_range_shrinkage_p)
2555 /* Don't use SCHED_PRESSURE_MODEL -- it results in much worse
2556 code. */
2557 gcc_assert (sched_pressure == SCHED_PRESSURE_WEIGHTED);
2558 if ((INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp) < 0
2559 || INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2) < 0)
2560 && (diff = (INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp)
2561 - INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2))) != 0)
2562 return diff;
2563 /* Sort by INSN_LUID (original insn order), so that we make the
2564 sort stable. This minimizes instruction movement, thus
2565 minimizing sched's effect on debugging and cross-jumping. */
2566 return INSN_LUID (tmp) - INSN_LUID (tmp2);
2569 /* The insn in a schedule group should be issued the first. */
2570 if (flag_sched_group_heuristic &&
2571 SCHED_GROUP_P (tmp) != SCHED_GROUP_P (tmp2))
2572 return SCHED_GROUP_P (tmp2) ? 1 : -1;
2574 /* Make sure that priority of TMP and TMP2 are initialized. */
2575 gcc_assert (INSN_PRIORITY_KNOWN (tmp) && INSN_PRIORITY_KNOWN (tmp2));
2577 if (sched_pressure != SCHED_PRESSURE_NONE)
2579 /* Prefer insn whose scheduling results in the smallest register
2580 pressure excess. */
2581 if ((diff = (INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp)
2582 + insn_delay (tmp)
2583 - INSN_REG_PRESSURE_EXCESS_COST_CHANGE (tmp2)
2584 - insn_delay (tmp2))))
2585 return diff;
2588 if (sched_pressure != SCHED_PRESSURE_NONE
2589 && (INSN_TICK (tmp2) > clock_var || INSN_TICK (tmp) > clock_var))
2591 if (INSN_TICK (tmp) <= clock_var)
2592 return -1;
2593 else if (INSN_TICK (tmp2) <= clock_var)
2594 return 1;
2595 else
2596 return INSN_TICK (tmp) - INSN_TICK (tmp2);
2599 /* If we are doing backtracking in this schedule, prefer insns that
2600 have forward dependencies with negative cost against an insn that
2601 was already scheduled. */
2602 if (current_sched_info->flags & DO_BACKTRACKING)
2604 priority_val = FEEDS_BACKTRACK_INSN (tmp2) - FEEDS_BACKTRACK_INSN (tmp);
2605 if (priority_val)
2606 return priority_val;
2609 /* Prefer insn with higher priority. */
2610 priority_val = INSN_PRIORITY (tmp2) - INSN_PRIORITY (tmp);
2612 if (flag_sched_critical_path_heuristic && priority_val)
2613 return priority_val;
2615 /* Prefer speculative insn with greater dependencies weakness. */
2616 if (flag_sched_spec_insn_heuristic && spec_info)
2618 ds_t ds1, ds2;
2619 dw_t dw1, dw2;
2620 int dw;
2622 ds1 = TODO_SPEC (tmp) & SPECULATIVE;
2623 if (ds1)
2624 dw1 = ds_weak (ds1);
2625 else
2626 dw1 = NO_DEP_WEAK;
2628 ds2 = TODO_SPEC (tmp2) & SPECULATIVE;
2629 if (ds2)
2630 dw2 = ds_weak (ds2);
2631 else
2632 dw2 = NO_DEP_WEAK;
2634 dw = dw2 - dw1;
2635 if (dw > (NO_DEP_WEAK / 8) || dw < -(NO_DEP_WEAK / 8))
2636 return dw;
2639 info_val = (*current_sched_info->rank) (tmp, tmp2);
2640 if (flag_sched_rank_heuristic && info_val)
2641 return info_val;
2643 /* Compare insns based on their relation to the last scheduled
2644 non-debug insn. */
2645 if (flag_sched_last_insn_heuristic && last_nondebug_scheduled_insn)
2647 dep_t dep1;
2648 dep_t dep2;
2649 rtx last = last_nondebug_scheduled_insn;
2651 /* Classify the instructions into three classes:
2652 1) Data dependent on last schedule insn.
2653 2) Anti/Output dependent on last scheduled insn.
2654 3) Independent of last scheduled insn, or has latency of one.
2655 Choose the insn from the highest numbered class if different. */
2656 dep1 = sd_find_dep_between (last, tmp, true);
2658 if (dep1 == NULL || dep_cost (dep1) == 1)
2659 tmp_class = 3;
2660 else if (/* Data dependence. */
2661 DEP_TYPE (dep1) == REG_DEP_TRUE)
2662 tmp_class = 1;
2663 else
2664 tmp_class = 2;
2666 dep2 = sd_find_dep_between (last, tmp2, true);
2668 if (dep2 == NULL || dep_cost (dep2) == 1)
2669 tmp2_class = 3;
2670 else if (/* Data dependence. */
2671 DEP_TYPE (dep2) == REG_DEP_TRUE)
2672 tmp2_class = 1;
2673 else
2674 tmp2_class = 2;
2676 if ((val = tmp2_class - tmp_class))
2677 return val;
2680 /* Prefer instructions that occur earlier in the model schedule. */
2681 if (sched_pressure == SCHED_PRESSURE_MODEL)
2683 int diff;
2685 diff = model_index (tmp) - model_index (tmp2);
2686 if (diff != 0)
2687 return diff;
2690 /* Prefer the insn which has more later insns that depend on it.
2691 This gives the scheduler more freedom when scheduling later
2692 instructions at the expense of added register pressure. */
2694 val = (dep_list_size (tmp2, SD_LIST_FORW)
2695 - dep_list_size (tmp, SD_LIST_FORW));
2697 if (flag_sched_dep_count_heuristic && val != 0)
2698 return val;
2700 /* If insns are equally good, sort by INSN_LUID (original insn order),
2701 so that we make the sort stable. This minimizes instruction movement,
2702 thus minimizing sched's effect on debugging and cross-jumping. */
2703 return INSN_LUID (tmp) - INSN_LUID (tmp2);
2706 /* Resort the array A in which only element at index N may be out of order. */
2708 HAIFA_INLINE static void
2709 swap_sort (rtx *a, int n)
2711 rtx insn = a[n - 1];
2712 int i = n - 2;
2714 while (i >= 0 && rank_for_schedule (a + i, &insn) >= 0)
2716 a[i + 1] = a[i];
2717 i -= 1;
2719 a[i + 1] = insn;
2722 /* Add INSN to the insn queue so that it can be executed at least
2723 N_CYCLES after the currently executing insn. Preserve insns
2724 chain for debugging purposes. REASON will be printed in debugging
2725 output. */
2727 HAIFA_INLINE static void
2728 queue_insn (rtx insn, int n_cycles, const char *reason)
2730 int next_q = NEXT_Q_AFTER (q_ptr, n_cycles);
2731 rtx link = alloc_INSN_LIST (insn, insn_queue[next_q]);
2732 int new_tick;
2734 gcc_assert (n_cycles <= max_insn_queue_index);
2735 gcc_assert (!DEBUG_INSN_P (insn));
2737 insn_queue[next_q] = link;
2738 q_size += 1;
2740 if (sched_verbose >= 2)
2742 fprintf (sched_dump, ";;\t\tReady-->Q: insn %s: ",
2743 (*current_sched_info->print_insn) (insn, 0));
2745 fprintf (sched_dump, "queued for %d cycles (%s).\n", n_cycles, reason);
2748 QUEUE_INDEX (insn) = next_q;
2750 if (current_sched_info->flags & DO_BACKTRACKING)
2752 new_tick = clock_var + n_cycles;
2753 if (INSN_TICK (insn) == INVALID_TICK || INSN_TICK (insn) < new_tick)
2754 INSN_TICK (insn) = new_tick;
2756 if (INSN_EXACT_TICK (insn) != INVALID_TICK
2757 && INSN_EXACT_TICK (insn) < clock_var + n_cycles)
2759 must_backtrack = true;
2760 if (sched_verbose >= 2)
2761 fprintf (sched_dump, ";;\t\tcausing a backtrack.\n");
2766 /* Remove INSN from queue. */
2767 static void
2768 queue_remove (rtx insn)
2770 gcc_assert (QUEUE_INDEX (insn) >= 0);
2771 remove_free_INSN_LIST_elem (insn, &insn_queue[QUEUE_INDEX (insn)]);
2772 q_size--;
2773 QUEUE_INDEX (insn) = QUEUE_NOWHERE;
2776 /* Return a pointer to the bottom of the ready list, i.e. the insn
2777 with the lowest priority. */
2779 rtx *
2780 ready_lastpos (struct ready_list *ready)
2782 gcc_assert (ready->n_ready >= 1);
2783 return ready->vec + ready->first - ready->n_ready + 1;
2786 /* Add an element INSN to the ready list so that it ends up with the
2787 lowest/highest priority depending on FIRST_P. */
2789 HAIFA_INLINE static void
2790 ready_add (struct ready_list *ready, rtx insn, bool first_p)
2792 if (!first_p)
2794 if (ready->first == ready->n_ready)
2796 memmove (ready->vec + ready->veclen - ready->n_ready,
2797 ready_lastpos (ready),
2798 ready->n_ready * sizeof (rtx));
2799 ready->first = ready->veclen - 1;
2801 ready->vec[ready->first - ready->n_ready] = insn;
2803 else
2805 if (ready->first == ready->veclen - 1)
2807 if (ready->n_ready)
2808 /* ready_lastpos() fails when called with (ready->n_ready == 0). */
2809 memmove (ready->vec + ready->veclen - ready->n_ready - 1,
2810 ready_lastpos (ready),
2811 ready->n_ready * sizeof (rtx));
2812 ready->first = ready->veclen - 2;
2814 ready->vec[++(ready->first)] = insn;
2817 ready->n_ready++;
2818 if (DEBUG_INSN_P (insn))
2819 ready->n_debug++;
2821 gcc_assert (QUEUE_INDEX (insn) != QUEUE_READY);
2822 QUEUE_INDEX (insn) = QUEUE_READY;
2824 if (INSN_EXACT_TICK (insn) != INVALID_TICK
2825 && INSN_EXACT_TICK (insn) < clock_var)
2827 must_backtrack = true;
2831 /* Remove the element with the highest priority from the ready list and
2832 return it. */
2834 HAIFA_INLINE static rtx
2835 ready_remove_first (struct ready_list *ready)
2837 rtx t;
2839 gcc_assert (ready->n_ready);
2840 t = ready->vec[ready->first--];
2841 ready->n_ready--;
2842 if (DEBUG_INSN_P (t))
2843 ready->n_debug--;
2844 /* If the queue becomes empty, reset it. */
2845 if (ready->n_ready == 0)
2846 ready->first = ready->veclen - 1;
2848 gcc_assert (QUEUE_INDEX (t) == QUEUE_READY);
2849 QUEUE_INDEX (t) = QUEUE_NOWHERE;
2851 return t;
2854 /* The following code implements multi-pass scheduling for the first
2855 cycle. In other words, we will try to choose ready insn which
2856 permits to start maximum number of insns on the same cycle. */
2858 /* Return a pointer to the element INDEX from the ready. INDEX for
2859 insn with the highest priority is 0, and the lowest priority has
2860 N_READY - 1. */
2863 ready_element (struct ready_list *ready, int index)
2865 gcc_assert (ready->n_ready && index < ready->n_ready);
2867 return ready->vec[ready->first - index];
2870 /* Remove the element INDEX from the ready list and return it. INDEX
2871 for insn with the highest priority is 0, and the lowest priority
2872 has N_READY - 1. */
2874 HAIFA_INLINE static rtx
2875 ready_remove (struct ready_list *ready, int index)
2877 rtx t;
2878 int i;
2880 if (index == 0)
2881 return ready_remove_first (ready);
2882 gcc_assert (ready->n_ready && index < ready->n_ready);
2883 t = ready->vec[ready->first - index];
2884 ready->n_ready--;
2885 if (DEBUG_INSN_P (t))
2886 ready->n_debug--;
2887 for (i = index; i < ready->n_ready; i++)
2888 ready->vec[ready->first - i] = ready->vec[ready->first - i - 1];
2889 QUEUE_INDEX (t) = QUEUE_NOWHERE;
2890 return t;
2893 /* Remove INSN from the ready list. */
2894 static void
2895 ready_remove_insn (rtx insn)
2897 int i;
2899 for (i = 0; i < readyp->n_ready; i++)
2900 if (ready_element (readyp, i) == insn)
2902 ready_remove (readyp, i);
2903 return;
2905 gcc_unreachable ();
2908 /* Sort the ready list READY by ascending priority, using the SCHED_SORT
2909 macro. */
2911 void
2912 ready_sort (struct ready_list *ready)
2914 int i;
2915 rtx *first = ready_lastpos (ready);
2917 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
2919 for (i = 0; i < ready->n_ready; i++)
2920 if (!DEBUG_INSN_P (first[i]))
2921 setup_insn_reg_pressure_info (first[i]);
2923 if (sched_pressure == SCHED_PRESSURE_MODEL
2924 && model_curr_point < model_num_insns)
2925 model_set_excess_costs (first, ready->n_ready);
2926 SCHED_SORT (first, ready->n_ready);
2929 /* PREV is an insn that is ready to execute. Adjust its priority if that
2930 will help shorten or lengthen register lifetimes as appropriate. Also
2931 provide a hook for the target to tweak itself. */
2933 HAIFA_INLINE static void
2934 adjust_priority (rtx prev)
2936 /* ??? There used to be code here to try and estimate how an insn
2937 affected register lifetimes, but it did it by looking at REG_DEAD
2938 notes, which we removed in schedule_region. Nor did it try to
2939 take into account register pressure or anything useful like that.
2941 Revisit when we have a machine model to work with and not before. */
2943 if (targetm.sched.adjust_priority)
2944 INSN_PRIORITY (prev) =
2945 targetm.sched.adjust_priority (prev, INSN_PRIORITY (prev));
2948 /* Advance DFA state STATE on one cycle. */
2949 void
2950 advance_state (state_t state)
2952 if (targetm.sched.dfa_pre_advance_cycle)
2953 targetm.sched.dfa_pre_advance_cycle ();
2955 if (targetm.sched.dfa_pre_cycle_insn)
2956 state_transition (state,
2957 targetm.sched.dfa_pre_cycle_insn ());
2959 state_transition (state, NULL);
2961 if (targetm.sched.dfa_post_cycle_insn)
2962 state_transition (state,
2963 targetm.sched.dfa_post_cycle_insn ());
2965 if (targetm.sched.dfa_post_advance_cycle)
2966 targetm.sched.dfa_post_advance_cycle ();
2969 /* Advance time on one cycle. */
2970 HAIFA_INLINE static void
2971 advance_one_cycle (void)
2973 advance_state (curr_state);
2974 if (sched_verbose >= 6)
2975 fprintf (sched_dump, ";;\tAdvanced a state.\n");
2978 /* Update register pressure after scheduling INSN. */
2979 static void
2980 update_register_pressure (rtx insn)
2982 struct reg_use_data *use;
2983 struct reg_set_data *set;
2985 gcc_checking_assert (!DEBUG_INSN_P (insn));
2987 for (use = INSN_REG_USE_LIST (insn); use != NULL; use = use->next_insn_use)
2988 if (dying_use_p (use))
2989 mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure,
2990 use->regno, false);
2991 for (set = INSN_REG_SET_LIST (insn); set != NULL; set = set->next_insn_set)
2992 mark_regno_birth_or_death (curr_reg_live, curr_reg_pressure,
2993 set->regno, true);
2996 /* Set up or update (if UPDATE_P) max register pressure (see its
2997 meaning in sched-int.h::_haifa_insn_data) for all current BB insns
2998 after insn AFTER. */
2999 static void
3000 setup_insn_max_reg_pressure (rtx after, bool update_p)
3002 int i, p;
3003 bool eq_p;
3004 rtx insn;
3005 static int max_reg_pressure[N_REG_CLASSES];
3007 save_reg_pressure ();
3008 for (i = 0; i < ira_pressure_classes_num; i++)
3009 max_reg_pressure[ira_pressure_classes[i]]
3010 = curr_reg_pressure[ira_pressure_classes[i]];
3011 for (insn = NEXT_INSN (after);
3012 insn != NULL_RTX && ! BARRIER_P (insn)
3013 && BLOCK_FOR_INSN (insn) == BLOCK_FOR_INSN (after);
3014 insn = NEXT_INSN (insn))
3015 if (NONDEBUG_INSN_P (insn))
3017 eq_p = true;
3018 for (i = 0; i < ira_pressure_classes_num; i++)
3020 p = max_reg_pressure[ira_pressure_classes[i]];
3021 if (INSN_MAX_REG_PRESSURE (insn)[i] != p)
3023 eq_p = false;
3024 INSN_MAX_REG_PRESSURE (insn)[i]
3025 = max_reg_pressure[ira_pressure_classes[i]];
3028 if (update_p && eq_p)
3029 break;
3030 update_register_pressure (insn);
3031 for (i = 0; i < ira_pressure_classes_num; i++)
3032 if (max_reg_pressure[ira_pressure_classes[i]]
3033 < curr_reg_pressure[ira_pressure_classes[i]])
3034 max_reg_pressure[ira_pressure_classes[i]]
3035 = curr_reg_pressure[ira_pressure_classes[i]];
3037 restore_reg_pressure ();
3040 /* Update the current register pressure after scheduling INSN. Update
3041 also max register pressure for unscheduled insns of the current
3042 BB. */
3043 static void
3044 update_reg_and_insn_max_reg_pressure (rtx insn)
3046 int i;
3047 int before[N_REG_CLASSES];
3049 for (i = 0; i < ira_pressure_classes_num; i++)
3050 before[i] = curr_reg_pressure[ira_pressure_classes[i]];
3051 update_register_pressure (insn);
3052 for (i = 0; i < ira_pressure_classes_num; i++)
3053 if (curr_reg_pressure[ira_pressure_classes[i]] != before[i])
3054 break;
3055 if (i < ira_pressure_classes_num)
3056 setup_insn_max_reg_pressure (insn, true);
3059 /* Set up register pressure at the beginning of basic block BB whose
3060 insns starting after insn AFTER. Set up also max register pressure
3061 for all insns of the basic block. */
3062 void
3063 sched_setup_bb_reg_pressure_info (basic_block bb, rtx after)
3065 gcc_assert (sched_pressure == SCHED_PRESSURE_WEIGHTED);
3066 initiate_bb_reg_pressure_info (bb);
3067 setup_insn_max_reg_pressure (after, false);
3070 /* If doing predication while scheduling, verify whether INSN, which
3071 has just been scheduled, clobbers the conditions of any
3072 instructions that must be predicated in order to break their
3073 dependencies. If so, remove them from the queues so that they will
3074 only be scheduled once their control dependency is resolved. */
3076 static void
3077 check_clobbered_conditions (rtx insn)
3079 HARD_REG_SET t;
3080 int i;
3082 if ((current_sched_info->flags & DO_PREDICATION) == 0)
3083 return;
3085 find_all_hard_reg_sets (insn, &t);
3087 restart:
3088 for (i = 0; i < ready.n_ready; i++)
3090 rtx x = ready_element (&ready, i);
3091 if (TODO_SPEC (x) == DEP_CONTROL && cond_clobbered_p (x, t))
3093 ready_remove_insn (x);
3094 goto restart;
3097 for (i = 0; i <= max_insn_queue_index; i++)
3099 rtx link;
3100 int q = NEXT_Q_AFTER (q_ptr, i);
3102 restart_queue:
3103 for (link = insn_queue[q]; link; link = XEXP (link, 1))
3105 rtx x = XEXP (link, 0);
3106 if (TODO_SPEC (x) == DEP_CONTROL && cond_clobbered_p (x, t))
3108 queue_remove (x);
3109 goto restart_queue;
3115 /* Return (in order):
3117 - positive if INSN adversely affects the pressure on one
3118 register class
3120 - negative if INSN reduces the pressure on one register class
3122 - 0 if INSN doesn't affect the pressure on any register class. */
3124 static int
3125 model_classify_pressure (struct model_insn_info *insn)
3127 struct reg_pressure_data *reg_pressure;
3128 int death[N_REG_CLASSES];
3129 int pci, cl, sum;
3131 calculate_reg_deaths (insn->insn, death);
3132 reg_pressure = INSN_REG_PRESSURE (insn->insn);
3133 sum = 0;
3134 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3136 cl = ira_pressure_classes[pci];
3137 if (death[cl] < reg_pressure[pci].set_increase)
3138 return 1;
3139 sum += reg_pressure[pci].set_increase - death[cl];
3141 return sum;
3144 /* Return true if INSN1 should come before INSN2 in the model schedule. */
3146 static int
3147 model_order_p (struct model_insn_info *insn1, struct model_insn_info *insn2)
3149 unsigned int height1, height2;
3150 unsigned int priority1, priority2;
3152 /* Prefer instructions with a higher model priority. */
3153 if (insn1->model_priority != insn2->model_priority)
3154 return insn1->model_priority > insn2->model_priority;
3156 /* Combine the length of the longest path of satisfied true dependencies
3157 that leads to each instruction (depth) with the length of the longest
3158 path of any dependencies that leads from the instruction (alap).
3159 Prefer instructions with the greatest combined length. If the combined
3160 lengths are equal, prefer instructions with the greatest depth.
3162 The idea is that, if we have a set S of "equal" instructions that each
3163 have ALAP value X, and we pick one such instruction I, any true-dependent
3164 successors of I that have ALAP value X - 1 should be preferred over S.
3165 This encourages the schedule to be "narrow" rather than "wide".
3166 However, if I is a low-priority instruction that we decided to
3167 schedule because of its model_classify_pressure, and if there
3168 is a set of higher-priority instructions T, the aforementioned
3169 successors of I should not have the edge over T. */
3170 height1 = insn1->depth + insn1->alap;
3171 height2 = insn2->depth + insn2->alap;
3172 if (height1 != height2)
3173 return height1 > height2;
3174 if (insn1->depth != insn2->depth)
3175 return insn1->depth > insn2->depth;
3177 /* We have no real preference between INSN1 an INSN2 as far as attempts
3178 to reduce pressure go. Prefer instructions with higher priorities. */
3179 priority1 = INSN_PRIORITY (insn1->insn);
3180 priority2 = INSN_PRIORITY (insn2->insn);
3181 if (priority1 != priority2)
3182 return priority1 > priority2;
3184 /* Use the original rtl sequence as a tie-breaker. */
3185 return insn1 < insn2;
3188 /* Add INSN to the model worklist immediately after PREV. Add it to the
3189 beginning of the list if PREV is null. */
3191 static void
3192 model_add_to_worklist_at (struct model_insn_info *insn,
3193 struct model_insn_info *prev)
3195 gcc_assert (QUEUE_INDEX (insn->insn) == QUEUE_NOWHERE);
3196 QUEUE_INDEX (insn->insn) = QUEUE_READY;
3198 insn->prev = prev;
3199 if (prev)
3201 insn->next = prev->next;
3202 prev->next = insn;
3204 else
3206 insn->next = model_worklist;
3207 model_worklist = insn;
3209 if (insn->next)
3210 insn->next->prev = insn;
3213 /* Remove INSN from the model worklist. */
3215 static void
3216 model_remove_from_worklist (struct model_insn_info *insn)
3218 gcc_assert (QUEUE_INDEX (insn->insn) == QUEUE_READY);
3219 QUEUE_INDEX (insn->insn) = QUEUE_NOWHERE;
3221 if (insn->prev)
3222 insn->prev->next = insn->next;
3223 else
3224 model_worklist = insn->next;
3225 if (insn->next)
3226 insn->next->prev = insn->prev;
3229 /* Add INSN to the model worklist. Start looking for a suitable position
3230 between neighbors PREV and NEXT, testing at most MAX_SCHED_READY_INSNS
3231 insns either side. A null PREV indicates the beginning of the list and
3232 a null NEXT indicates the end. */
3234 static void
3235 model_add_to_worklist (struct model_insn_info *insn,
3236 struct model_insn_info *prev,
3237 struct model_insn_info *next)
3239 int count;
3241 count = MAX_SCHED_READY_INSNS;
3242 if (count > 0 && prev && model_order_p (insn, prev))
3245 count--;
3246 prev = prev->prev;
3248 while (count > 0 && prev && model_order_p (insn, prev));
3249 else
3250 while (count > 0 && next && model_order_p (next, insn))
3252 count--;
3253 prev = next;
3254 next = next->next;
3256 model_add_to_worklist_at (insn, prev);
3259 /* INSN may now have a higher priority (in the model_order_p sense)
3260 than before. Move it up the worklist if necessary. */
3262 static void
3263 model_promote_insn (struct model_insn_info *insn)
3265 struct model_insn_info *prev;
3266 int count;
3268 prev = insn->prev;
3269 count = MAX_SCHED_READY_INSNS;
3270 while (count > 0 && prev && model_order_p (insn, prev))
3272 count--;
3273 prev = prev->prev;
3275 if (prev != insn->prev)
3277 model_remove_from_worklist (insn);
3278 model_add_to_worklist_at (insn, prev);
3282 /* Add INSN to the end of the model schedule. */
3284 static void
3285 model_add_to_schedule (rtx insn)
3287 unsigned int point;
3289 gcc_assert (QUEUE_INDEX (insn) == QUEUE_NOWHERE);
3290 QUEUE_INDEX (insn) = QUEUE_SCHEDULED;
3292 point = model_schedule.length ();
3293 model_schedule.quick_push (insn);
3294 INSN_MODEL_INDEX (insn) = point + 1;
3297 /* Analyze the instructions that are to be scheduled, setting up
3298 MODEL_INSN_INFO (...) and model_num_insns accordingly. Add ready
3299 instructions to model_worklist. */
3301 static void
3302 model_analyze_insns (void)
3304 rtx start, end, iter;
3305 sd_iterator_def sd_it;
3306 dep_t dep;
3307 struct model_insn_info *insn, *con;
3309 model_num_insns = 0;
3310 start = PREV_INSN (current_sched_info->next_tail);
3311 end = current_sched_info->prev_head;
3312 for (iter = start; iter != end; iter = PREV_INSN (iter))
3313 if (NONDEBUG_INSN_P (iter))
3315 insn = MODEL_INSN_INFO (iter);
3316 insn->insn = iter;
3317 FOR_EACH_DEP (iter, SD_LIST_FORW, sd_it, dep)
3319 con = MODEL_INSN_INFO (DEP_CON (dep));
3320 if (con->insn && insn->alap < con->alap + 1)
3321 insn->alap = con->alap + 1;
3324 insn->old_queue = QUEUE_INDEX (iter);
3325 QUEUE_INDEX (iter) = QUEUE_NOWHERE;
3327 insn->unscheduled_preds = dep_list_size (iter, SD_LIST_HARD_BACK);
3328 if (insn->unscheduled_preds == 0)
3329 model_add_to_worklist (insn, NULL, model_worklist);
3331 model_num_insns++;
3335 /* The global state describes the register pressure at the start of the
3336 model schedule. Initialize GROUP accordingly. */
3338 static void
3339 model_init_pressure_group (struct model_pressure_group *group)
3341 int pci, cl;
3343 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3345 cl = ira_pressure_classes[pci];
3346 group->limits[pci].pressure = curr_reg_pressure[cl];
3347 group->limits[pci].point = 0;
3349 /* Use index model_num_insns to record the state after the last
3350 instruction in the model schedule. */
3351 group->model = XNEWVEC (struct model_pressure_data,
3352 (model_num_insns + 1) * ira_pressure_classes_num);
3355 /* Record that MODEL_REF_PRESSURE (GROUP, POINT, PCI) is PRESSURE.
3356 Update the maximum pressure for the whole schedule. */
3358 static void
3359 model_record_pressure (struct model_pressure_group *group,
3360 int point, int pci, int pressure)
3362 MODEL_REF_PRESSURE (group, point, pci) = pressure;
3363 if (group->limits[pci].pressure < pressure)
3365 group->limits[pci].pressure = pressure;
3366 group->limits[pci].point = point;
3370 /* INSN has just been added to the end of the model schedule. Record its
3371 register-pressure information. */
3373 static void
3374 model_record_pressures (struct model_insn_info *insn)
3376 struct reg_pressure_data *reg_pressure;
3377 int point, pci, cl, delta;
3378 int death[N_REG_CLASSES];
3380 point = model_index (insn->insn);
3381 if (sched_verbose >= 2)
3383 if (point == 0)
3385 fprintf (sched_dump, "\n;;\tModel schedule:\n;;\n");
3386 fprintf (sched_dump, ";;\t| idx insn | mpri hght dpth prio |\n");
3388 fprintf (sched_dump, ";;\t| %3d %4d | %4d %4d %4d %4d | %-30s ",
3389 point, INSN_UID (insn->insn), insn->model_priority,
3390 insn->depth + insn->alap, insn->depth,
3391 INSN_PRIORITY (insn->insn),
3392 str_pattern_slim (PATTERN (insn->insn)));
3394 calculate_reg_deaths (insn->insn, death);
3395 reg_pressure = INSN_REG_PRESSURE (insn->insn);
3396 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3398 cl = ira_pressure_classes[pci];
3399 delta = reg_pressure[pci].set_increase - death[cl];
3400 if (sched_verbose >= 2)
3401 fprintf (sched_dump, " %s:[%d,%+d]", reg_class_names[cl],
3402 curr_reg_pressure[cl], delta);
3403 model_record_pressure (&model_before_pressure, point, pci,
3404 curr_reg_pressure[cl]);
3406 if (sched_verbose >= 2)
3407 fprintf (sched_dump, "\n");
3410 /* All instructions have been added to the model schedule. Record the
3411 final register pressure in GROUP and set up all MODEL_MAX_PRESSUREs. */
3413 static void
3414 model_record_final_pressures (struct model_pressure_group *group)
3416 int point, pci, max_pressure, ref_pressure, cl;
3418 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3420 /* Record the final pressure for this class. */
3421 cl = ira_pressure_classes[pci];
3422 point = model_num_insns;
3423 ref_pressure = curr_reg_pressure[cl];
3424 model_record_pressure (group, point, pci, ref_pressure);
3426 /* Record the original maximum pressure. */
3427 group->limits[pci].orig_pressure = group->limits[pci].pressure;
3429 /* Update the MODEL_MAX_PRESSURE for every point of the schedule. */
3430 max_pressure = ref_pressure;
3431 MODEL_MAX_PRESSURE (group, point, pci) = max_pressure;
3432 while (point > 0)
3434 point--;
3435 ref_pressure = MODEL_REF_PRESSURE (group, point, pci);
3436 max_pressure = MAX (max_pressure, ref_pressure);
3437 MODEL_MAX_PRESSURE (group, point, pci) = max_pressure;
3442 /* Update all successors of INSN, given that INSN has just been scheduled. */
3444 static void
3445 model_add_successors_to_worklist (struct model_insn_info *insn)
3447 sd_iterator_def sd_it;
3448 struct model_insn_info *con;
3449 dep_t dep;
3451 FOR_EACH_DEP (insn->insn, SD_LIST_FORW, sd_it, dep)
3453 con = MODEL_INSN_INFO (DEP_CON (dep));
3454 /* Ignore debug instructions, and instructions from other blocks. */
3455 if (con->insn)
3457 con->unscheduled_preds--;
3459 /* Update the depth field of each true-dependent successor.
3460 Increasing the depth gives them a higher priority than
3461 before. */
3462 if (DEP_TYPE (dep) == REG_DEP_TRUE && con->depth < insn->depth + 1)
3464 con->depth = insn->depth + 1;
3465 if (QUEUE_INDEX (con->insn) == QUEUE_READY)
3466 model_promote_insn (con);
3469 /* If this is a true dependency, or if there are no remaining
3470 dependencies for CON (meaning that CON only had non-true
3471 dependencies), make sure that CON is on the worklist.
3472 We don't bother otherwise because it would tend to fill the
3473 worklist with a lot of low-priority instructions that are not
3474 yet ready to issue. */
3475 if ((con->depth > 0 || con->unscheduled_preds == 0)
3476 && QUEUE_INDEX (con->insn) == QUEUE_NOWHERE)
3477 model_add_to_worklist (con, insn, insn->next);
3482 /* Give INSN a higher priority than any current instruction, then give
3483 unscheduled predecessors of INSN a higher priority still. If any of
3484 those predecessors are not on the model worklist, do the same for its
3485 predecessors, and so on. */
3487 static void
3488 model_promote_predecessors (struct model_insn_info *insn)
3490 struct model_insn_info *pro, *first;
3491 sd_iterator_def sd_it;
3492 dep_t dep;
3494 if (sched_verbose >= 7)
3495 fprintf (sched_dump, ";;\t+--- priority of %d = %d, priority of",
3496 INSN_UID (insn->insn), model_next_priority);
3497 insn->model_priority = model_next_priority++;
3498 model_remove_from_worklist (insn);
3499 model_add_to_worklist_at (insn, NULL);
3501 first = NULL;
3502 for (;;)
3504 FOR_EACH_DEP (insn->insn, SD_LIST_HARD_BACK, sd_it, dep)
3506 pro = MODEL_INSN_INFO (DEP_PRO (dep));
3507 /* The first test is to ignore debug instructions, and instructions
3508 from other blocks. */
3509 if (pro->insn
3510 && pro->model_priority != model_next_priority
3511 && QUEUE_INDEX (pro->insn) != QUEUE_SCHEDULED)
3513 pro->model_priority = model_next_priority;
3514 if (sched_verbose >= 7)
3515 fprintf (sched_dump, " %d", INSN_UID (pro->insn));
3516 if (QUEUE_INDEX (pro->insn) == QUEUE_READY)
3518 /* PRO is already in the worklist, but it now has
3519 a higher priority than before. Move it at the
3520 appropriate place. */
3521 model_remove_from_worklist (pro);
3522 model_add_to_worklist (pro, NULL, model_worklist);
3524 else
3526 /* PRO isn't in the worklist. Recursively process
3527 its predecessors until we find one that is. */
3528 pro->next = first;
3529 first = pro;
3533 if (!first)
3534 break;
3535 insn = first;
3536 first = insn->next;
3538 if (sched_verbose >= 7)
3539 fprintf (sched_dump, " = %d\n", model_next_priority);
3540 model_next_priority++;
3543 /* Pick one instruction from model_worklist and process it. */
3545 static void
3546 model_choose_insn (void)
3548 struct model_insn_info *insn, *fallback;
3549 int count;
3551 if (sched_verbose >= 7)
3553 fprintf (sched_dump, ";;\t+--- worklist:\n");
3554 insn = model_worklist;
3555 count = MAX_SCHED_READY_INSNS;
3556 while (count > 0 && insn)
3558 fprintf (sched_dump, ";;\t+--- %d [%d, %d, %d, %d]\n",
3559 INSN_UID (insn->insn), insn->model_priority,
3560 insn->depth + insn->alap, insn->depth,
3561 INSN_PRIORITY (insn->insn));
3562 count--;
3563 insn = insn->next;
3567 /* Look for a ready instruction whose model_classify_priority is zero
3568 or negative, picking the highest-priority one. Adding such an
3569 instruction to the schedule now should do no harm, and may actually
3570 do some good.
3572 Failing that, see whether there is an instruction with the highest
3573 extant model_priority that is not yet ready, but which would reduce
3574 pressure if it became ready. This is designed to catch cases like:
3576 (set (mem (reg R1)) (reg R2))
3578 where the instruction is the last remaining use of R1 and where the
3579 value of R2 is not yet available (or vice versa). The death of R1
3580 means that this instruction already reduces pressure. It is of
3581 course possible that the computation of R2 involves other registers
3582 that are hard to kill, but such cases are rare enough for this
3583 heuristic to be a win in general.
3585 Failing that, just pick the highest-priority instruction in the
3586 worklist. */
3587 count = MAX_SCHED_READY_INSNS;
3588 insn = model_worklist;
3589 fallback = 0;
3590 for (;;)
3592 if (count == 0 || !insn)
3594 insn = fallback ? fallback : model_worklist;
3595 break;
3597 if (insn->unscheduled_preds)
3599 if (model_worklist->model_priority == insn->model_priority
3600 && !fallback
3601 && model_classify_pressure (insn) < 0)
3602 fallback = insn;
3604 else
3606 if (model_classify_pressure (insn) <= 0)
3607 break;
3609 count--;
3610 insn = insn->next;
3613 if (sched_verbose >= 7 && insn != model_worklist)
3615 if (insn->unscheduled_preds)
3616 fprintf (sched_dump, ";;\t+--- promoting insn %d, with dependencies\n",
3617 INSN_UID (insn->insn));
3618 else
3619 fprintf (sched_dump, ";;\t+--- promoting insn %d, which is ready\n",
3620 INSN_UID (insn->insn));
3622 if (insn->unscheduled_preds)
3623 /* INSN isn't yet ready to issue. Give all its predecessors the
3624 highest priority. */
3625 model_promote_predecessors (insn);
3626 else
3628 /* INSN is ready. Add it to the end of model_schedule and
3629 process its successors. */
3630 model_add_successors_to_worklist (insn);
3631 model_remove_from_worklist (insn);
3632 model_add_to_schedule (insn->insn);
3633 model_record_pressures (insn);
3634 update_register_pressure (insn->insn);
3638 /* Restore all QUEUE_INDEXs to the values that they had before
3639 model_start_schedule was called. */
3641 static void
3642 model_reset_queue_indices (void)
3644 unsigned int i;
3645 rtx insn;
3647 FOR_EACH_VEC_ELT (model_schedule, i, insn)
3648 QUEUE_INDEX (insn) = MODEL_INSN_INFO (insn)->old_queue;
3651 /* We have calculated the model schedule and spill costs. Print a summary
3652 to sched_dump. */
3654 static void
3655 model_dump_pressure_summary (void)
3657 int pci, cl;
3659 fprintf (sched_dump, ";; Pressure summary:");
3660 for (pci = 0; pci < ira_pressure_classes_num; pci++)
3662 cl = ira_pressure_classes[pci];
3663 fprintf (sched_dump, " %s:%d", reg_class_names[cl],
3664 model_before_pressure.limits[pci].pressure);
3666 fprintf (sched_dump, "\n\n");
3669 /* Initialize the SCHED_PRESSURE_MODEL information for the current
3670 scheduling region. */
3672 static void
3673 model_start_schedule (void)
3675 basic_block bb;
3677 model_next_priority = 1;
3678 model_schedule.create (sched_max_luid);
3679 model_insns = XCNEWVEC (struct model_insn_info, sched_max_luid);
3681 bb = BLOCK_FOR_INSN (NEXT_INSN (current_sched_info->prev_head));
3682 initiate_reg_pressure_info (df_get_live_in (bb));
3684 model_analyze_insns ();
3685 model_init_pressure_group (&model_before_pressure);
3686 while (model_worklist)
3687 model_choose_insn ();
3688 gcc_assert (model_num_insns == (int) model_schedule.length ());
3689 if (sched_verbose >= 2)
3690 fprintf (sched_dump, "\n");
3692 model_record_final_pressures (&model_before_pressure);
3693 model_reset_queue_indices ();
3695 XDELETEVEC (model_insns);
3697 model_curr_point = 0;
3698 initiate_reg_pressure_info (df_get_live_in (bb));
3699 if (sched_verbose >= 1)
3700 model_dump_pressure_summary ();
3703 /* Free the information associated with GROUP. */
3705 static void
3706 model_finalize_pressure_group (struct model_pressure_group *group)
3708 XDELETEVEC (group->model);
3711 /* Free the information created by model_start_schedule. */
3713 static void
3714 model_end_schedule (void)
3716 model_finalize_pressure_group (&model_before_pressure);
3717 model_schedule.release ();
3720 /* A structure that holds local state for the loop in schedule_block. */
3721 struct sched_block_state
3723 /* True if no real insns have been scheduled in the current cycle. */
3724 bool first_cycle_insn_p;
3725 /* True if a shadow insn has been scheduled in the current cycle, which
3726 means that no more normal insns can be issued. */
3727 bool shadows_only_p;
3728 /* True if we're winding down a modulo schedule, which means that we only
3729 issue insns with INSN_EXACT_TICK set. */
3730 bool modulo_epilogue;
3731 /* Initialized with the machine's issue rate every cycle, and updated
3732 by calls to the variable_issue hook. */
3733 int can_issue_more;
3736 /* INSN is the "currently executing insn". Launch each insn which was
3737 waiting on INSN. READY is the ready list which contains the insns
3738 that are ready to fire. CLOCK is the current cycle. The function
3739 returns necessary cycle advance after issuing the insn (it is not
3740 zero for insns in a schedule group). */
3742 static int
3743 schedule_insn (rtx insn)
3745 sd_iterator_def sd_it;
3746 dep_t dep;
3747 int i;
3748 int advance = 0;
3750 if (sched_verbose >= 1)
3752 struct reg_pressure_data *pressure_info;
3753 fprintf (sched_dump, ";;\t%3i--> %s%-40s:",
3754 clock_var, (*current_sched_info->print_insn) (insn, 1),
3755 str_pattern_slim (PATTERN (insn)));
3757 if (recog_memoized (insn) < 0)
3758 fprintf (sched_dump, "nothing");
3759 else
3760 print_reservation (sched_dump, insn);
3761 pressure_info = INSN_REG_PRESSURE (insn);
3762 if (pressure_info != NULL)
3764 fputc (':', sched_dump);
3765 for (i = 0; i < ira_pressure_classes_num; i++)
3766 fprintf (sched_dump, "%s%s%+d(%d)",
3767 scheduled_insns.length () > 1
3768 && INSN_LUID (insn)
3769 < INSN_LUID (scheduled_insns[scheduled_insns.length () - 2]) ? "@" : "",
3770 reg_class_names[ira_pressure_classes[i]],
3771 pressure_info[i].set_increase, pressure_info[i].change);
3773 if (sched_pressure == SCHED_PRESSURE_MODEL
3774 && model_curr_point < model_num_insns
3775 && model_index (insn) == model_curr_point)
3776 fprintf (sched_dump, ":model %d", model_curr_point);
3777 fputc ('\n', sched_dump);
3780 if (sched_pressure == SCHED_PRESSURE_WEIGHTED && !DEBUG_INSN_P (insn))
3781 update_reg_and_insn_max_reg_pressure (insn);
3783 /* Scheduling instruction should have all its dependencies resolved and
3784 should have been removed from the ready list. */
3785 gcc_assert (sd_lists_empty_p (insn, SD_LIST_HARD_BACK));
3787 /* Reset debug insns invalidated by moving this insn. */
3788 if (MAY_HAVE_DEBUG_INSNS && !DEBUG_INSN_P (insn))
3789 for (sd_it = sd_iterator_start (insn, SD_LIST_BACK);
3790 sd_iterator_cond (&sd_it, &dep);)
3792 rtx dbg = DEP_PRO (dep);
3793 struct reg_use_data *use, *next;
3795 if (DEP_STATUS (dep) & DEP_CANCELLED)
3797 sd_iterator_next (&sd_it);
3798 continue;
3801 gcc_assert (DEBUG_INSN_P (dbg));
3803 if (sched_verbose >= 6)
3804 fprintf (sched_dump, ";;\t\tresetting: debug insn %d\n",
3805 INSN_UID (dbg));
3807 /* ??? Rather than resetting the debug insn, we might be able
3808 to emit a debug temp before the just-scheduled insn, but
3809 this would involve checking that the expression at the
3810 point of the debug insn is equivalent to the expression
3811 before the just-scheduled insn. They might not be: the
3812 expression in the debug insn may depend on other insns not
3813 yet scheduled that set MEMs, REGs or even other debug
3814 insns. It's not clear that attempting to preserve debug
3815 information in these cases is worth the effort, given how
3816 uncommon these resets are and the likelihood that the debug
3817 temps introduced won't survive the schedule change. */
3818 INSN_VAR_LOCATION_LOC (dbg) = gen_rtx_UNKNOWN_VAR_LOC ();
3819 df_insn_rescan (dbg);
3821 /* Unknown location doesn't use any registers. */
3822 for (use = INSN_REG_USE_LIST (dbg); use != NULL; use = next)
3824 struct reg_use_data *prev = use;
3826 /* Remove use from the cyclic next_regno_use chain first. */
3827 while (prev->next_regno_use != use)
3828 prev = prev->next_regno_use;
3829 prev->next_regno_use = use->next_regno_use;
3830 next = use->next_insn_use;
3831 free (use);
3833 INSN_REG_USE_LIST (dbg) = NULL;
3835 /* We delete rather than resolve these deps, otherwise we
3836 crash in sched_free_deps(), because forward deps are
3837 expected to be released before backward deps. */
3838 sd_delete_dep (sd_it);
3841 gcc_assert (QUEUE_INDEX (insn) == QUEUE_NOWHERE);
3842 QUEUE_INDEX (insn) = QUEUE_SCHEDULED;
3844 if (sched_pressure == SCHED_PRESSURE_MODEL
3845 && model_curr_point < model_num_insns
3846 && NONDEBUG_INSN_P (insn))
3848 if (model_index (insn) == model_curr_point)
3850 model_curr_point++;
3851 while (model_curr_point < model_num_insns
3852 && (QUEUE_INDEX (MODEL_INSN (model_curr_point))
3853 == QUEUE_SCHEDULED));
3854 else
3855 model_recompute (insn);
3856 model_update_limit_points ();
3857 update_register_pressure (insn);
3858 if (sched_verbose >= 2)
3859 print_curr_reg_pressure ();
3862 gcc_assert (INSN_TICK (insn) >= MIN_TICK);
3863 if (INSN_TICK (insn) > clock_var)
3864 /* INSN has been prematurely moved from the queue to the ready list.
3865 This is possible only if following flag is set. */
3866 gcc_assert (flag_sched_stalled_insns);
3868 /* ??? Probably, if INSN is scheduled prematurely, we should leave
3869 INSN_TICK untouched. This is a machine-dependent issue, actually. */
3870 INSN_TICK (insn) = clock_var;
3872 check_clobbered_conditions (insn);
3874 /* Update dependent instructions. First, see if by scheduling this insn
3875 now we broke a dependence in a way that requires us to change another
3876 insn. */
3877 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
3878 sd_iterator_cond (&sd_it, &dep); sd_iterator_next (&sd_it))
3880 struct dep_replacement *desc = DEP_REPLACE (dep);
3881 rtx pro = DEP_PRO (dep);
3882 if (QUEUE_INDEX (pro) != QUEUE_SCHEDULED
3883 && desc != NULL && desc->insn == pro)
3884 apply_replacement (dep, false);
3887 /* Go through and resolve forward dependencies. */
3888 for (sd_it = sd_iterator_start (insn, SD_LIST_FORW);
3889 sd_iterator_cond (&sd_it, &dep);)
3891 rtx next = DEP_CON (dep);
3892 bool cancelled = (DEP_STATUS (dep) & DEP_CANCELLED) != 0;
3894 /* Resolve the dependence between INSN and NEXT.
3895 sd_resolve_dep () moves current dep to another list thus
3896 advancing the iterator. */
3897 sd_resolve_dep (sd_it);
3899 if (cancelled)
3901 if (must_restore_pattern_p (next, dep))
3902 restore_pattern (dep, false);
3903 continue;
3906 /* Don't bother trying to mark next as ready if insn is a debug
3907 insn. If insn is the last hard dependency, it will have
3908 already been discounted. */
3909 if (DEBUG_INSN_P (insn) && !DEBUG_INSN_P (next))
3910 continue;
3912 if (!IS_SPECULATION_BRANCHY_CHECK_P (insn))
3914 int effective_cost;
3916 effective_cost = try_ready (next);
3918 if (effective_cost >= 0
3919 && SCHED_GROUP_P (next)
3920 && advance < effective_cost)
3921 advance = effective_cost;
3923 else
3924 /* Check always has only one forward dependence (to the first insn in
3925 the recovery block), therefore, this will be executed only once. */
3927 gcc_assert (sd_lists_empty_p (insn, SD_LIST_FORW));
3928 fix_recovery_deps (RECOVERY_BLOCK (insn));
3932 /* Annotate the instruction with issue information -- TImode
3933 indicates that the instruction is expected not to be able
3934 to issue on the same cycle as the previous insn. A machine
3935 may use this information to decide how the instruction should
3936 be aligned. */
3937 if (issue_rate > 1
3938 && GET_CODE (PATTERN (insn)) != USE
3939 && GET_CODE (PATTERN (insn)) != CLOBBER
3940 && !DEBUG_INSN_P (insn))
3942 if (reload_completed)
3943 PUT_MODE (insn, clock_var > last_clock_var ? TImode : VOIDmode);
3944 last_clock_var = clock_var;
3947 return advance;
3950 /* Functions for handling of notes. */
3952 /* Add note list that ends on FROM_END to the end of TO_ENDP. */
3953 void
3954 concat_note_lists (rtx from_end, rtx *to_endp)
3956 rtx from_start;
3958 /* It's easy when have nothing to concat. */
3959 if (from_end == NULL)
3960 return;
3962 /* It's also easy when destination is empty. */
3963 if (*to_endp == NULL)
3965 *to_endp = from_end;
3966 return;
3969 from_start = from_end;
3970 while (PREV_INSN (from_start) != NULL)
3971 from_start = PREV_INSN (from_start);
3973 PREV_INSN (from_start) = *to_endp;
3974 NEXT_INSN (*to_endp) = from_start;
3975 *to_endp = from_end;
3978 /* Delete notes between HEAD and TAIL and put them in the chain
3979 of notes ended by NOTE_LIST. */
3980 void
3981 remove_notes (rtx head, rtx tail)
3983 rtx next_tail, insn, next;
3985 note_list = 0;
3986 if (head == tail && !INSN_P (head))
3987 return;
3989 next_tail = NEXT_INSN (tail);
3990 for (insn = head; insn != next_tail; insn = next)
3992 next = NEXT_INSN (insn);
3993 if (!NOTE_P (insn))
3994 continue;
3996 switch (NOTE_KIND (insn))
3998 case NOTE_INSN_BASIC_BLOCK:
3999 continue;
4001 case NOTE_INSN_EPILOGUE_BEG:
4002 if (insn != tail)
4004 remove_insn (insn);
4005 add_reg_note (next, REG_SAVE_NOTE,
4006 GEN_INT (NOTE_INSN_EPILOGUE_BEG));
4007 break;
4009 /* FALLTHRU */
4011 default:
4012 remove_insn (insn);
4014 /* Add the note to list that ends at NOTE_LIST. */
4015 PREV_INSN (insn) = note_list;
4016 NEXT_INSN (insn) = NULL_RTX;
4017 if (note_list)
4018 NEXT_INSN (note_list) = insn;
4019 note_list = insn;
4020 break;
4023 gcc_assert ((sel_sched_p () || insn != tail) && insn != head);
4027 /* A structure to record enough data to allow us to backtrack the scheduler to
4028 a previous state. */
4029 struct haifa_saved_data
4031 /* Next entry on the list. */
4032 struct haifa_saved_data *next;
4034 /* Backtracking is associated with scheduling insns that have delay slots.
4035 DELAY_PAIR points to the structure that contains the insns involved, and
4036 the number of cycles between them. */
4037 struct delay_pair *delay_pair;
4039 /* Data used by the frontend (e.g. sched-ebb or sched-rgn). */
4040 void *fe_saved_data;
4041 /* Data used by the backend. */
4042 void *be_saved_data;
4044 /* Copies of global state. */
4045 int clock_var, last_clock_var;
4046 struct ready_list ready;
4047 state_t curr_state;
4049 rtx last_scheduled_insn;
4050 rtx last_nondebug_scheduled_insn;
4051 int cycle_issued_insns;
4053 /* Copies of state used in the inner loop of schedule_block. */
4054 struct sched_block_state sched_block;
4056 /* We don't need to save q_ptr, as its value is arbitrary and we can set it
4057 to 0 when restoring. */
4058 int q_size;
4059 rtx *insn_queue;
4061 /* Describe pattern replacements that occurred since this backtrack point
4062 was queued. */
4063 vec<dep_t> replacement_deps;
4064 vec<int> replace_apply;
4066 /* A copy of the next-cycle replacement vectors at the time of the backtrack
4067 point. */
4068 vec<dep_t> next_cycle_deps;
4069 vec<int> next_cycle_apply;
4072 /* A record, in reverse order, of all scheduled insns which have delay slots
4073 and may require backtracking. */
4074 static struct haifa_saved_data *backtrack_queue;
4076 /* For every dependency of INSN, set the FEEDS_BACKTRACK_INSN bit according
4077 to SET_P. */
4078 static void
4079 mark_backtrack_feeds (rtx insn, int set_p)
4081 sd_iterator_def sd_it;
4082 dep_t dep;
4083 FOR_EACH_DEP (insn, SD_LIST_HARD_BACK, sd_it, dep)
4085 FEEDS_BACKTRACK_INSN (DEP_PRO (dep)) = set_p;
4089 /* Save the current scheduler state so that we can backtrack to it
4090 later if necessary. PAIR gives the insns that make it necessary to
4091 save this point. SCHED_BLOCK is the local state of schedule_block
4092 that need to be saved. */
4093 static void
4094 save_backtrack_point (struct delay_pair *pair,
4095 struct sched_block_state sched_block)
4097 int i;
4098 struct haifa_saved_data *save = XNEW (struct haifa_saved_data);
4100 save->curr_state = xmalloc (dfa_state_size);
4101 memcpy (save->curr_state, curr_state, dfa_state_size);
4103 save->ready.first = ready.first;
4104 save->ready.n_ready = ready.n_ready;
4105 save->ready.n_debug = ready.n_debug;
4106 save->ready.veclen = ready.veclen;
4107 save->ready.vec = XNEWVEC (rtx, ready.veclen);
4108 memcpy (save->ready.vec, ready.vec, ready.veclen * sizeof (rtx));
4110 save->insn_queue = XNEWVEC (rtx, max_insn_queue_index + 1);
4111 save->q_size = q_size;
4112 for (i = 0; i <= max_insn_queue_index; i++)
4114 int q = NEXT_Q_AFTER (q_ptr, i);
4115 save->insn_queue[i] = copy_INSN_LIST (insn_queue[q]);
4118 save->clock_var = clock_var;
4119 save->last_clock_var = last_clock_var;
4120 save->cycle_issued_insns = cycle_issued_insns;
4121 save->last_scheduled_insn = last_scheduled_insn;
4122 save->last_nondebug_scheduled_insn = last_nondebug_scheduled_insn;
4124 save->sched_block = sched_block;
4126 save->replacement_deps.create (0);
4127 save->replace_apply.create (0);
4128 save->next_cycle_deps = next_cycle_replace_deps.copy ();
4129 save->next_cycle_apply = next_cycle_apply.copy ();
4131 if (current_sched_info->save_state)
4132 save->fe_saved_data = (*current_sched_info->save_state) ();
4134 if (targetm.sched.alloc_sched_context)
4136 save->be_saved_data = targetm.sched.alloc_sched_context ();
4137 targetm.sched.init_sched_context (save->be_saved_data, false);
4139 else
4140 save->be_saved_data = NULL;
4142 save->delay_pair = pair;
4144 save->next = backtrack_queue;
4145 backtrack_queue = save;
4147 while (pair)
4149 mark_backtrack_feeds (pair->i2, 1);
4150 INSN_TICK (pair->i2) = INVALID_TICK;
4151 INSN_EXACT_TICK (pair->i2) = clock_var + pair_delay (pair);
4152 SHADOW_P (pair->i2) = pair->stages == 0;
4153 pair = pair->next_same_i1;
4157 /* Walk the ready list and all queues. If any insns have unresolved backwards
4158 dependencies, these must be cancelled deps, broken by predication. Set or
4159 clear (depending on SET) the DEP_CANCELLED bit in DEP_STATUS. */
4161 static void
4162 toggle_cancelled_flags (bool set)
4164 int i;
4165 sd_iterator_def sd_it;
4166 dep_t dep;
4168 if (ready.n_ready > 0)
4170 rtx *first = ready_lastpos (&ready);
4171 for (i = 0; i < ready.n_ready; i++)
4172 FOR_EACH_DEP (first[i], SD_LIST_BACK, sd_it, dep)
4173 if (!DEBUG_INSN_P (DEP_PRO (dep)))
4175 if (set)
4176 DEP_STATUS (dep) |= DEP_CANCELLED;
4177 else
4178 DEP_STATUS (dep) &= ~DEP_CANCELLED;
4181 for (i = 0; i <= max_insn_queue_index; i++)
4183 int q = NEXT_Q_AFTER (q_ptr, i);
4184 rtx link;
4185 for (link = insn_queue[q]; link; link = XEXP (link, 1))
4187 rtx insn = XEXP (link, 0);
4188 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
4189 if (!DEBUG_INSN_P (DEP_PRO (dep)))
4191 if (set)
4192 DEP_STATUS (dep) |= DEP_CANCELLED;
4193 else
4194 DEP_STATUS (dep) &= ~DEP_CANCELLED;
4200 /* Undo the replacements that have occurred after backtrack point SAVE
4201 was placed. */
4202 static void
4203 undo_replacements_for_backtrack (struct haifa_saved_data *save)
4205 while (!save->replacement_deps.is_empty ())
4207 dep_t dep = save->replacement_deps.pop ();
4208 int apply_p = save->replace_apply.pop ();
4210 if (apply_p)
4211 restore_pattern (dep, true);
4212 else
4213 apply_replacement (dep, true);
4215 save->replacement_deps.release ();
4216 save->replace_apply.release ();
4219 /* Pop entries from the SCHEDULED_INSNS vector up to and including INSN.
4220 Restore their dependencies to an unresolved state, and mark them as
4221 queued nowhere. */
4223 static void
4224 unschedule_insns_until (rtx insn)
4226 vec<rtx> recompute_vec = vNULL;
4228 /* Make two passes over the insns to be unscheduled. First, we clear out
4229 dependencies and other trivial bookkeeping. */
4230 for (;;)
4232 rtx last;
4233 sd_iterator_def sd_it;
4234 dep_t dep;
4236 last = scheduled_insns.pop ();
4238 /* This will be changed by restore_backtrack_point if the insn is in
4239 any queue. */
4240 QUEUE_INDEX (last) = QUEUE_NOWHERE;
4241 if (last != insn)
4242 INSN_TICK (last) = INVALID_TICK;
4244 if (modulo_ii > 0 && INSN_UID (last) < modulo_iter0_max_uid)
4245 modulo_insns_scheduled--;
4247 for (sd_it = sd_iterator_start (last, SD_LIST_RES_FORW);
4248 sd_iterator_cond (&sd_it, &dep);)
4250 rtx con = DEP_CON (dep);
4251 sd_unresolve_dep (sd_it);
4252 if (!MUST_RECOMPUTE_SPEC_P (con))
4254 MUST_RECOMPUTE_SPEC_P (con) = 1;
4255 recompute_vec.safe_push (con);
4259 if (last == insn)
4260 break;
4263 /* A second pass, to update ready and speculation status for insns
4264 depending on the unscheduled ones. The first pass must have
4265 popped the scheduled_insns vector up to the point where we
4266 restart scheduling, as recompute_todo_spec requires it to be
4267 up-to-date. */
4268 while (!recompute_vec.is_empty ())
4270 rtx con;
4272 con = recompute_vec.pop ();
4273 MUST_RECOMPUTE_SPEC_P (con) = 0;
4274 if (!sd_lists_empty_p (con, SD_LIST_HARD_BACK))
4276 TODO_SPEC (con) = HARD_DEP;
4277 INSN_TICK (con) = INVALID_TICK;
4278 if (PREDICATED_PAT (con) != NULL_RTX)
4279 haifa_change_pattern (con, ORIG_PAT (con));
4281 else if (QUEUE_INDEX (con) != QUEUE_SCHEDULED)
4282 TODO_SPEC (con) = recompute_todo_spec (con, true);
4284 recompute_vec.release ();
4287 /* Restore scheduler state from the topmost entry on the backtracking queue.
4288 PSCHED_BLOCK_P points to the local data of schedule_block that we must
4289 overwrite with the saved data.
4290 The caller must already have called unschedule_insns_until. */
4292 static void
4293 restore_last_backtrack_point (struct sched_block_state *psched_block)
4295 rtx link;
4296 int i;
4297 struct haifa_saved_data *save = backtrack_queue;
4299 backtrack_queue = save->next;
4301 if (current_sched_info->restore_state)
4302 (*current_sched_info->restore_state) (save->fe_saved_data);
4304 if (targetm.sched.alloc_sched_context)
4306 targetm.sched.set_sched_context (save->be_saved_data);
4307 targetm.sched.free_sched_context (save->be_saved_data);
4310 /* Do this first since it clobbers INSN_TICK of the involved
4311 instructions. */
4312 undo_replacements_for_backtrack (save);
4314 /* Clear the QUEUE_INDEX of everything in the ready list or one
4315 of the queues. */
4316 if (ready.n_ready > 0)
4318 rtx *first = ready_lastpos (&ready);
4319 for (i = 0; i < ready.n_ready; i++)
4321 rtx insn = first[i];
4322 QUEUE_INDEX (insn) = QUEUE_NOWHERE;
4323 INSN_TICK (insn) = INVALID_TICK;
4326 for (i = 0; i <= max_insn_queue_index; i++)
4328 int q = NEXT_Q_AFTER (q_ptr, i);
4330 for (link = insn_queue[q]; link; link = XEXP (link, 1))
4332 rtx x = XEXP (link, 0);
4333 QUEUE_INDEX (x) = QUEUE_NOWHERE;
4334 INSN_TICK (x) = INVALID_TICK;
4336 free_INSN_LIST_list (&insn_queue[q]);
4339 free (ready.vec);
4340 ready = save->ready;
4342 if (ready.n_ready > 0)
4344 rtx *first = ready_lastpos (&ready);
4345 for (i = 0; i < ready.n_ready; i++)
4347 rtx insn = first[i];
4348 QUEUE_INDEX (insn) = QUEUE_READY;
4349 TODO_SPEC (insn) = recompute_todo_spec (insn, true);
4350 INSN_TICK (insn) = save->clock_var;
4354 q_ptr = 0;
4355 q_size = save->q_size;
4356 for (i = 0; i <= max_insn_queue_index; i++)
4358 int q = NEXT_Q_AFTER (q_ptr, i);
4360 insn_queue[q] = save->insn_queue[q];
4362 for (link = insn_queue[q]; link; link = XEXP (link, 1))
4364 rtx x = XEXP (link, 0);
4365 QUEUE_INDEX (x) = i;
4366 TODO_SPEC (x) = recompute_todo_spec (x, true);
4367 INSN_TICK (x) = save->clock_var + i;
4370 free (save->insn_queue);
4372 toggle_cancelled_flags (true);
4374 clock_var = save->clock_var;
4375 last_clock_var = save->last_clock_var;
4376 cycle_issued_insns = save->cycle_issued_insns;
4377 last_scheduled_insn = save->last_scheduled_insn;
4378 last_nondebug_scheduled_insn = save->last_nondebug_scheduled_insn;
4380 *psched_block = save->sched_block;
4382 memcpy (curr_state, save->curr_state, dfa_state_size);
4383 free (save->curr_state);
4385 mark_backtrack_feeds (save->delay_pair->i2, 0);
4387 gcc_assert (next_cycle_replace_deps.is_empty ());
4388 next_cycle_replace_deps = save->next_cycle_deps.copy ();
4389 next_cycle_apply = save->next_cycle_apply.copy ();
4391 free (save);
4393 for (save = backtrack_queue; save; save = save->next)
4395 mark_backtrack_feeds (save->delay_pair->i2, 1);
4399 /* Discard all data associated with the topmost entry in the backtrack
4400 queue. If RESET_TICK is false, we just want to free the data. If true,
4401 we are doing this because we discovered a reason to backtrack. In the
4402 latter case, also reset the INSN_TICK for the shadow insn. */
4403 static void
4404 free_topmost_backtrack_point (bool reset_tick)
4406 struct haifa_saved_data *save = backtrack_queue;
4407 int i;
4409 backtrack_queue = save->next;
4411 if (reset_tick)
4413 struct delay_pair *pair = save->delay_pair;
4414 while (pair)
4416 INSN_TICK (pair->i2) = INVALID_TICK;
4417 INSN_EXACT_TICK (pair->i2) = INVALID_TICK;
4418 pair = pair->next_same_i1;
4420 undo_replacements_for_backtrack (save);
4422 else
4424 save->replacement_deps.release ();
4425 save->replace_apply.release ();
4428 if (targetm.sched.free_sched_context)
4429 targetm.sched.free_sched_context (save->be_saved_data);
4430 if (current_sched_info->restore_state)
4431 free (save->fe_saved_data);
4432 for (i = 0; i <= max_insn_queue_index; i++)
4433 free_INSN_LIST_list (&save->insn_queue[i]);
4434 free (save->insn_queue);
4435 free (save->curr_state);
4436 free (save->ready.vec);
4437 free (save);
4440 /* Free the entire backtrack queue. */
4441 static void
4442 free_backtrack_queue (void)
4444 while (backtrack_queue)
4445 free_topmost_backtrack_point (false);
4448 /* Apply a replacement described by DESC. If IMMEDIATELY is false, we
4449 may have to postpone the replacement until the start of the next cycle,
4450 at which point we will be called again with IMMEDIATELY true. This is
4451 only done for machines which have instruction packets with explicit
4452 parallelism however. */
4453 static void
4454 apply_replacement (dep_t dep, bool immediately)
4456 struct dep_replacement *desc = DEP_REPLACE (dep);
4457 if (!immediately && targetm.sched.exposed_pipeline && reload_completed)
4459 next_cycle_replace_deps.safe_push (dep);
4460 next_cycle_apply.safe_push (1);
4462 else
4464 bool success;
4466 if (QUEUE_INDEX (desc->insn) == QUEUE_SCHEDULED)
4467 return;
4469 if (sched_verbose >= 5)
4470 fprintf (sched_dump, "applying replacement for insn %d\n",
4471 INSN_UID (desc->insn));
4473 success = validate_change (desc->insn, desc->loc, desc->newval, 0);
4474 gcc_assert (success);
4476 update_insn_after_change (desc->insn);
4477 if ((TODO_SPEC (desc->insn) & (HARD_DEP | DEP_POSTPONED)) == 0)
4478 fix_tick_ready (desc->insn);
4480 if (backtrack_queue != NULL)
4482 backtrack_queue->replacement_deps.safe_push (dep);
4483 backtrack_queue->replace_apply.safe_push (1);
4488 /* We have determined that a pattern involved in DEP must be restored.
4489 If IMMEDIATELY is false, we may have to postpone the replacement
4490 until the start of the next cycle, at which point we will be called
4491 again with IMMEDIATELY true. */
4492 static void
4493 restore_pattern (dep_t dep, bool immediately)
4495 rtx next = DEP_CON (dep);
4496 int tick = INSN_TICK (next);
4498 /* If we already scheduled the insn, the modified version is
4499 correct. */
4500 if (QUEUE_INDEX (next) == QUEUE_SCHEDULED)
4501 return;
4503 if (!immediately && targetm.sched.exposed_pipeline && reload_completed)
4505 next_cycle_replace_deps.safe_push (dep);
4506 next_cycle_apply.safe_push (0);
4507 return;
4511 if (DEP_TYPE (dep) == REG_DEP_CONTROL)
4513 if (sched_verbose >= 5)
4514 fprintf (sched_dump, "restoring pattern for insn %d\n",
4515 INSN_UID (next));
4516 haifa_change_pattern (next, ORIG_PAT (next));
4518 else
4520 struct dep_replacement *desc = DEP_REPLACE (dep);
4521 bool success;
4523 if (sched_verbose >= 5)
4524 fprintf (sched_dump, "restoring pattern for insn %d\n",
4525 INSN_UID (desc->insn));
4526 tick = INSN_TICK (desc->insn);
4528 success = validate_change (desc->insn, desc->loc, desc->orig, 0);
4529 gcc_assert (success);
4530 update_insn_after_change (desc->insn);
4531 if (backtrack_queue != NULL)
4533 backtrack_queue->replacement_deps.safe_push (dep);
4534 backtrack_queue->replace_apply.safe_push (0);
4537 INSN_TICK (next) = tick;
4538 if (TODO_SPEC (next) == DEP_POSTPONED)
4539 return;
4541 if (sd_lists_empty_p (next, SD_LIST_BACK))
4542 TODO_SPEC (next) = 0;
4543 else if (!sd_lists_empty_p (next, SD_LIST_HARD_BACK))
4544 TODO_SPEC (next) = HARD_DEP;
4547 /* Perform pattern replacements that were queued up until the next
4548 cycle. */
4549 static void
4550 perform_replacements_new_cycle (void)
4552 int i;
4553 dep_t dep;
4554 FOR_EACH_VEC_ELT (next_cycle_replace_deps, i, dep)
4556 int apply_p = next_cycle_apply[i];
4557 if (apply_p)
4558 apply_replacement (dep, true);
4559 else
4560 restore_pattern (dep, true);
4562 next_cycle_replace_deps.truncate (0);
4563 next_cycle_apply.truncate (0);
4566 /* Compute INSN_TICK_ESTIMATE for INSN. PROCESSED is a bitmap of
4567 instructions we've previously encountered, a set bit prevents
4568 recursion. BUDGET is a limit on how far ahead we look, it is
4569 reduced on recursive calls. Return true if we produced a good
4570 estimate, or false if we exceeded the budget. */
4571 static bool
4572 estimate_insn_tick (bitmap processed, rtx insn, int budget)
4574 sd_iterator_def sd_it;
4575 dep_t dep;
4576 int earliest = INSN_TICK (insn);
4578 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
4580 rtx pro = DEP_PRO (dep);
4581 int t;
4583 if (DEP_STATUS (dep) & DEP_CANCELLED)
4584 continue;
4586 if (QUEUE_INDEX (pro) == QUEUE_SCHEDULED)
4587 gcc_assert (INSN_TICK (pro) + dep_cost (dep) <= INSN_TICK (insn));
4588 else
4590 int cost = dep_cost (dep);
4591 if (cost >= budget)
4592 return false;
4593 if (!bitmap_bit_p (processed, INSN_LUID (pro)))
4595 if (!estimate_insn_tick (processed, pro, budget - cost))
4596 return false;
4598 gcc_assert (INSN_TICK_ESTIMATE (pro) != INVALID_TICK);
4599 t = INSN_TICK_ESTIMATE (pro) + cost;
4600 if (earliest == INVALID_TICK || t > earliest)
4601 earliest = t;
4604 bitmap_set_bit (processed, INSN_LUID (insn));
4605 INSN_TICK_ESTIMATE (insn) = earliest;
4606 return true;
4609 /* Examine the pair of insns in P, and estimate (optimistically, assuming
4610 infinite resources) the cycle in which the delayed shadow can be issued.
4611 Return the number of cycles that must pass before the real insn can be
4612 issued in order to meet this constraint. */
4613 static int
4614 estimate_shadow_tick (struct delay_pair *p)
4616 bitmap_head processed;
4617 int t;
4618 bool cutoff;
4619 bitmap_initialize (&processed, 0);
4621 cutoff = !estimate_insn_tick (&processed, p->i2,
4622 max_insn_queue_index + pair_delay (p));
4623 bitmap_clear (&processed);
4624 if (cutoff)
4625 return max_insn_queue_index;
4626 t = INSN_TICK_ESTIMATE (p->i2) - (clock_var + pair_delay (p) + 1);
4627 if (t > 0)
4628 return t;
4629 return 0;
4632 /* If INSN has no unresolved backwards dependencies, add it to the schedule and
4633 recursively resolve all its forward dependencies. */
4634 static void
4635 resolve_dependencies (rtx insn)
4637 sd_iterator_def sd_it;
4638 dep_t dep;
4640 /* Don't use sd_lists_empty_p; it ignores debug insns. */
4641 if (DEPS_LIST_FIRST (INSN_HARD_BACK_DEPS (insn)) != NULL
4642 || DEPS_LIST_FIRST (INSN_SPEC_BACK_DEPS (insn)) != NULL)
4643 return;
4645 if (sched_verbose >= 4)
4646 fprintf (sched_dump, ";;\tquickly resolving %d\n", INSN_UID (insn));
4648 if (QUEUE_INDEX (insn) >= 0)
4649 queue_remove (insn);
4651 scheduled_insns.safe_push (insn);
4653 /* Update dependent instructions. */
4654 for (sd_it = sd_iterator_start (insn, SD_LIST_FORW);
4655 sd_iterator_cond (&sd_it, &dep);)
4657 rtx next = DEP_CON (dep);
4659 if (sched_verbose >= 4)
4660 fprintf (sched_dump, ";;\t\tdep %d against %d\n", INSN_UID (insn),
4661 INSN_UID (next));
4663 /* Resolve the dependence between INSN and NEXT.
4664 sd_resolve_dep () moves current dep to another list thus
4665 advancing the iterator. */
4666 sd_resolve_dep (sd_it);
4668 if (!IS_SPECULATION_BRANCHY_CHECK_P (insn))
4670 resolve_dependencies (next);
4672 else
4673 /* Check always has only one forward dependence (to the first insn in
4674 the recovery block), therefore, this will be executed only once. */
4676 gcc_assert (sd_lists_empty_p (insn, SD_LIST_FORW));
4682 /* Return the head and tail pointers of ebb starting at BEG and ending
4683 at END. */
4684 void
4685 get_ebb_head_tail (basic_block beg, basic_block end, rtx *headp, rtx *tailp)
4687 rtx beg_head = BB_HEAD (beg);
4688 rtx beg_tail = BB_END (beg);
4689 rtx end_head = BB_HEAD (end);
4690 rtx end_tail = BB_END (end);
4692 /* Don't include any notes or labels at the beginning of the BEG
4693 basic block, or notes at the end of the END basic blocks. */
4695 if (LABEL_P (beg_head))
4696 beg_head = NEXT_INSN (beg_head);
4698 while (beg_head != beg_tail)
4699 if (NOTE_P (beg_head))
4700 beg_head = NEXT_INSN (beg_head);
4701 else if (DEBUG_INSN_P (beg_head))
4703 rtx note, next;
4705 for (note = NEXT_INSN (beg_head);
4706 note != beg_tail;
4707 note = next)
4709 next = NEXT_INSN (note);
4710 if (NOTE_P (note))
4712 if (sched_verbose >= 9)
4713 fprintf (sched_dump, "reorder %i\n", INSN_UID (note));
4715 reorder_insns_nobb (note, note, PREV_INSN (beg_head));
4717 if (BLOCK_FOR_INSN (note) != beg)
4718 df_insn_change_bb (note, beg);
4720 else if (!DEBUG_INSN_P (note))
4721 break;
4724 break;
4726 else
4727 break;
4729 *headp = beg_head;
4731 if (beg == end)
4732 end_head = beg_head;
4733 else if (LABEL_P (end_head))
4734 end_head = NEXT_INSN (end_head);
4736 while (end_head != end_tail)
4737 if (NOTE_P (end_tail))
4738 end_tail = PREV_INSN (end_tail);
4739 else if (DEBUG_INSN_P (end_tail))
4741 rtx note, prev;
4743 for (note = PREV_INSN (end_tail);
4744 note != end_head;
4745 note = prev)
4747 prev = PREV_INSN (note);
4748 if (NOTE_P (note))
4750 if (sched_verbose >= 9)
4751 fprintf (sched_dump, "reorder %i\n", INSN_UID (note));
4753 reorder_insns_nobb (note, note, end_tail);
4755 if (end_tail == BB_END (end))
4756 BB_END (end) = note;
4758 if (BLOCK_FOR_INSN (note) != end)
4759 df_insn_change_bb (note, end);
4761 else if (!DEBUG_INSN_P (note))
4762 break;
4765 break;
4767 else
4768 break;
4770 *tailp = end_tail;
4773 /* Return nonzero if there are no real insns in the range [ HEAD, TAIL ]. */
4776 no_real_insns_p (const_rtx head, const_rtx tail)
4778 while (head != NEXT_INSN (tail))
4780 if (!NOTE_P (head) && !LABEL_P (head))
4781 return 0;
4782 head = NEXT_INSN (head);
4784 return 1;
4787 /* Restore-other-notes: NOTE_LIST is the end of a chain of notes
4788 previously found among the insns. Insert them just before HEAD. */
4790 restore_other_notes (rtx head, basic_block head_bb)
4792 if (note_list != 0)
4794 rtx note_head = note_list;
4796 if (head)
4797 head_bb = BLOCK_FOR_INSN (head);
4798 else
4799 head = NEXT_INSN (bb_note (head_bb));
4801 while (PREV_INSN (note_head))
4803 set_block_for_insn (note_head, head_bb);
4804 note_head = PREV_INSN (note_head);
4806 /* In the above cycle we've missed this note. */
4807 set_block_for_insn (note_head, head_bb);
4809 PREV_INSN (note_head) = PREV_INSN (head);
4810 NEXT_INSN (PREV_INSN (head)) = note_head;
4811 PREV_INSN (head) = note_list;
4812 NEXT_INSN (note_list) = head;
4814 if (BLOCK_FOR_INSN (head) != head_bb)
4815 BB_END (head_bb) = note_list;
4817 head = note_head;
4820 return head;
4823 /* When we know we are going to discard the schedule due to a failed attempt
4824 at modulo scheduling, undo all replacements. */
4825 static void
4826 undo_all_replacements (void)
4828 rtx insn;
4829 int i;
4831 FOR_EACH_VEC_ELT (scheduled_insns, i, insn)
4833 sd_iterator_def sd_it;
4834 dep_t dep;
4836 /* See if we must undo a replacement. */
4837 for (sd_it = sd_iterator_start (insn, SD_LIST_RES_FORW);
4838 sd_iterator_cond (&sd_it, &dep); sd_iterator_next (&sd_it))
4840 struct dep_replacement *desc = DEP_REPLACE (dep);
4841 if (desc != NULL)
4842 validate_change (desc->insn, desc->loc, desc->orig, 0);
4847 /* Move insns that became ready to fire from queue to ready list. */
4849 static void
4850 queue_to_ready (struct ready_list *ready)
4852 rtx insn;
4853 rtx link;
4854 rtx skip_insn;
4856 q_ptr = NEXT_Q (q_ptr);
4858 if (dbg_cnt (sched_insn) == false)
4860 /* If debug counter is activated do not requeue the first
4861 nonscheduled insn. */
4862 skip_insn = nonscheduled_insns_begin;
4865 skip_insn = next_nonnote_nondebug_insn (skip_insn);
4867 while (QUEUE_INDEX (skip_insn) == QUEUE_SCHEDULED);
4869 else
4870 skip_insn = NULL_RTX;
4872 /* Add all pending insns that can be scheduled without stalls to the
4873 ready list. */
4874 for (link = insn_queue[q_ptr]; link; link = XEXP (link, 1))
4876 insn = XEXP (link, 0);
4877 q_size -= 1;
4879 if (sched_verbose >= 2)
4880 fprintf (sched_dump, ";;\t\tQ-->Ready: insn %s: ",
4881 (*current_sched_info->print_insn) (insn, 0));
4883 /* If the ready list is full, delay the insn for 1 cycle.
4884 See the comment in schedule_block for the rationale. */
4885 if (!reload_completed
4886 && (ready->n_ready - ready->n_debug > MAX_SCHED_READY_INSNS
4887 || (sched_pressure == SCHED_PRESSURE_MODEL
4888 /* Limit pressure recalculations to MAX_SCHED_READY_INSNS
4889 instructions too. */
4890 && model_index (insn) > (model_curr_point
4891 + MAX_SCHED_READY_INSNS)))
4892 && !(sched_pressure == SCHED_PRESSURE_MODEL
4893 && model_curr_point < model_num_insns
4894 /* Always allow the next model instruction to issue. */
4895 && model_index (insn) == model_curr_point)
4896 && !SCHED_GROUP_P (insn)
4897 && insn != skip_insn)
4898 queue_insn (insn, 1, "ready full");
4899 else
4901 ready_add (ready, insn, false);
4902 if (sched_verbose >= 2)
4903 fprintf (sched_dump, "moving to ready without stalls\n");
4906 free_INSN_LIST_list (&insn_queue[q_ptr]);
4908 /* If there are no ready insns, stall until one is ready and add all
4909 of the pending insns at that point to the ready list. */
4910 if (ready->n_ready == 0)
4912 int stalls;
4914 for (stalls = 1; stalls <= max_insn_queue_index; stalls++)
4916 if ((link = insn_queue[NEXT_Q_AFTER (q_ptr, stalls)]))
4918 for (; link; link = XEXP (link, 1))
4920 insn = XEXP (link, 0);
4921 q_size -= 1;
4923 if (sched_verbose >= 2)
4924 fprintf (sched_dump, ";;\t\tQ-->Ready: insn %s: ",
4925 (*current_sched_info->print_insn) (insn, 0));
4927 ready_add (ready, insn, false);
4928 if (sched_verbose >= 2)
4929 fprintf (sched_dump, "moving to ready with %d stalls\n", stalls);
4931 free_INSN_LIST_list (&insn_queue[NEXT_Q_AFTER (q_ptr, stalls)]);
4933 advance_one_cycle ();
4935 break;
4938 advance_one_cycle ();
4941 q_ptr = NEXT_Q_AFTER (q_ptr, stalls);
4942 clock_var += stalls;
4946 /* Used by early_queue_to_ready. Determines whether it is "ok" to
4947 prematurely move INSN from the queue to the ready list. Currently,
4948 if a target defines the hook 'is_costly_dependence', this function
4949 uses the hook to check whether there exist any dependences which are
4950 considered costly by the target, between INSN and other insns that
4951 have already been scheduled. Dependences are checked up to Y cycles
4952 back, with default Y=1; The flag -fsched-stalled-insns-dep=Y allows
4953 controlling this value.
4954 (Other considerations could be taken into account instead (or in
4955 addition) depending on user flags and target hooks. */
4957 static bool
4958 ok_for_early_queue_removal (rtx insn)
4960 if (targetm.sched.is_costly_dependence)
4962 rtx prev_insn;
4963 int n_cycles;
4964 int i = scheduled_insns.length ();
4965 for (n_cycles = flag_sched_stalled_insns_dep; n_cycles; n_cycles--)
4967 while (i-- > 0)
4969 int cost;
4971 prev_insn = scheduled_insns[i];
4973 if (!NOTE_P (prev_insn))
4975 dep_t dep;
4977 dep = sd_find_dep_between (prev_insn, insn, true);
4979 if (dep != NULL)
4981 cost = dep_cost (dep);
4983 if (targetm.sched.is_costly_dependence (dep, cost,
4984 flag_sched_stalled_insns_dep - n_cycles))
4985 return false;
4989 if (GET_MODE (prev_insn) == TImode) /* end of dispatch group */
4990 break;
4993 if (i == 0)
4994 break;
4998 return true;
5002 /* Remove insns from the queue, before they become "ready" with respect
5003 to FU latency considerations. */
5005 static int
5006 early_queue_to_ready (state_t state, struct ready_list *ready)
5008 rtx insn;
5009 rtx link;
5010 rtx next_link;
5011 rtx prev_link;
5012 bool move_to_ready;
5013 int cost;
5014 state_t temp_state = alloca (dfa_state_size);
5015 int stalls;
5016 int insns_removed = 0;
5019 Flag '-fsched-stalled-insns=X' determines the aggressiveness of this
5020 function:
5022 X == 0: There is no limit on how many queued insns can be removed
5023 prematurely. (flag_sched_stalled_insns = -1).
5025 X >= 1: Only X queued insns can be removed prematurely in each
5026 invocation. (flag_sched_stalled_insns = X).
5028 Otherwise: Early queue removal is disabled.
5029 (flag_sched_stalled_insns = 0)
5032 if (! flag_sched_stalled_insns)
5033 return 0;
5035 for (stalls = 0; stalls <= max_insn_queue_index; stalls++)
5037 if ((link = insn_queue[NEXT_Q_AFTER (q_ptr, stalls)]))
5039 if (sched_verbose > 6)
5040 fprintf (sched_dump, ";; look at index %d + %d\n", q_ptr, stalls);
5042 prev_link = 0;
5043 while (link)
5045 next_link = XEXP (link, 1);
5046 insn = XEXP (link, 0);
5047 if (insn && sched_verbose > 6)
5048 print_rtl_single (sched_dump, insn);
5050 memcpy (temp_state, state, dfa_state_size);
5051 if (recog_memoized (insn) < 0)
5052 /* non-negative to indicate that it's not ready
5053 to avoid infinite Q->R->Q->R... */
5054 cost = 0;
5055 else
5056 cost = state_transition (temp_state, insn);
5058 if (sched_verbose >= 6)
5059 fprintf (sched_dump, "transition cost = %d\n", cost);
5061 move_to_ready = false;
5062 if (cost < 0)
5064 move_to_ready = ok_for_early_queue_removal (insn);
5065 if (move_to_ready == true)
5067 /* move from Q to R */
5068 q_size -= 1;
5069 ready_add (ready, insn, false);
5071 if (prev_link)
5072 XEXP (prev_link, 1) = next_link;
5073 else
5074 insn_queue[NEXT_Q_AFTER (q_ptr, stalls)] = next_link;
5076 free_INSN_LIST_node (link);
5078 if (sched_verbose >= 2)
5079 fprintf (sched_dump, ";;\t\tEarly Q-->Ready: insn %s\n",
5080 (*current_sched_info->print_insn) (insn, 0));
5082 insns_removed++;
5083 if (insns_removed == flag_sched_stalled_insns)
5084 /* Remove no more than flag_sched_stalled_insns insns
5085 from Q at a time. */
5086 return insns_removed;
5090 if (move_to_ready == false)
5091 prev_link = link;
5093 link = next_link;
5094 } /* while link */
5095 } /* if link */
5097 } /* for stalls.. */
5099 return insns_removed;
5103 /* Print the ready list for debugging purposes. Callable from debugger. */
5105 static void
5106 debug_ready_list (struct ready_list *ready)
5108 rtx *p;
5109 int i;
5111 if (ready->n_ready == 0)
5113 fprintf (sched_dump, "\n");
5114 return;
5117 p = ready_lastpos (ready);
5118 for (i = 0; i < ready->n_ready; i++)
5120 fprintf (sched_dump, " %s:%d",
5121 (*current_sched_info->print_insn) (p[i], 0),
5122 INSN_LUID (p[i]));
5123 if (sched_pressure != SCHED_PRESSURE_NONE)
5124 fprintf (sched_dump, "(cost=%d",
5125 INSN_REG_PRESSURE_EXCESS_COST_CHANGE (p[i]));
5126 if (INSN_TICK (p[i]) > clock_var)
5127 fprintf (sched_dump, ":delay=%d", INSN_TICK (p[i]) - clock_var);
5128 if (sched_pressure != SCHED_PRESSURE_NONE)
5129 fprintf (sched_dump, ")");
5131 fprintf (sched_dump, "\n");
5134 /* Search INSN for REG_SAVE_NOTE notes and convert them back into insn
5135 NOTEs. This is used for NOTE_INSN_EPILOGUE_BEG, so that sched-ebb
5136 replaces the epilogue note in the correct basic block. */
5137 void
5138 reemit_notes (rtx insn)
5140 rtx note, last = insn;
5142 for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
5144 if (REG_NOTE_KIND (note) == REG_SAVE_NOTE)
5146 enum insn_note note_type = (enum insn_note) INTVAL (XEXP (note, 0));
5148 last = emit_note_before (note_type, last);
5149 remove_note (insn, note);
5154 /* Move INSN. Reemit notes if needed. Update CFG, if needed. */
5155 static void
5156 move_insn (rtx insn, rtx last, rtx nt)
5158 if (PREV_INSN (insn) != last)
5160 basic_block bb;
5161 rtx note;
5162 int jump_p = 0;
5164 bb = BLOCK_FOR_INSN (insn);
5166 /* BB_HEAD is either LABEL or NOTE. */
5167 gcc_assert (BB_HEAD (bb) != insn);
5169 if (BB_END (bb) == insn)
5170 /* If this is last instruction in BB, move end marker one
5171 instruction up. */
5173 /* Jumps are always placed at the end of basic block. */
5174 jump_p = control_flow_insn_p (insn);
5176 gcc_assert (!jump_p
5177 || ((common_sched_info->sched_pass_id == SCHED_RGN_PASS)
5178 && IS_SPECULATION_BRANCHY_CHECK_P (insn))
5179 || (common_sched_info->sched_pass_id
5180 == SCHED_EBB_PASS));
5182 gcc_assert (BLOCK_FOR_INSN (PREV_INSN (insn)) == bb);
5184 BB_END (bb) = PREV_INSN (insn);
5187 gcc_assert (BB_END (bb) != last);
5189 if (jump_p)
5190 /* We move the block note along with jump. */
5192 gcc_assert (nt);
5194 note = NEXT_INSN (insn);
5195 while (NOTE_NOT_BB_P (note) && note != nt)
5196 note = NEXT_INSN (note);
5198 if (note != nt
5199 && (LABEL_P (note)
5200 || BARRIER_P (note)))
5201 note = NEXT_INSN (note);
5203 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
5205 else
5206 note = insn;
5208 NEXT_INSN (PREV_INSN (insn)) = NEXT_INSN (note);
5209 PREV_INSN (NEXT_INSN (note)) = PREV_INSN (insn);
5211 NEXT_INSN (note) = NEXT_INSN (last);
5212 PREV_INSN (NEXT_INSN (last)) = note;
5214 NEXT_INSN (last) = insn;
5215 PREV_INSN (insn) = last;
5217 bb = BLOCK_FOR_INSN (last);
5219 if (jump_p)
5221 fix_jump_move (insn);
5223 if (BLOCK_FOR_INSN (insn) != bb)
5224 move_block_after_check (insn);
5226 gcc_assert (BB_END (bb) == last);
5229 df_insn_change_bb (insn, bb);
5231 /* Update BB_END, if needed. */
5232 if (BB_END (bb) == last)
5233 BB_END (bb) = insn;
5236 SCHED_GROUP_P (insn) = 0;
5239 /* Return true if scheduling INSN will finish current clock cycle. */
5240 static bool
5241 insn_finishes_cycle_p (rtx insn)
5243 if (SCHED_GROUP_P (insn))
5244 /* After issuing INSN, rest of the sched_group will be forced to issue
5245 in order. Don't make any plans for the rest of cycle. */
5246 return true;
5248 /* Finishing the block will, apparently, finish the cycle. */
5249 if (current_sched_info->insn_finishes_block_p
5250 && current_sched_info->insn_finishes_block_p (insn))
5251 return true;
5253 return false;
5256 /* Define type for target data used in multipass scheduling. */
5257 #ifndef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DATA_T
5258 # define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DATA_T int
5259 #endif
5260 typedef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DATA_T first_cycle_multipass_data_t;
5262 /* The following structure describe an entry of the stack of choices. */
5263 struct choice_entry
5265 /* Ordinal number of the issued insn in the ready queue. */
5266 int index;
5267 /* The number of the rest insns whose issues we should try. */
5268 int rest;
5269 /* The number of issued essential insns. */
5270 int n;
5271 /* State after issuing the insn. */
5272 state_t state;
5273 /* Target-specific data. */
5274 first_cycle_multipass_data_t target_data;
5277 /* The following array is used to implement a stack of choices used in
5278 function max_issue. */
5279 static struct choice_entry *choice_stack;
5281 /* This holds the value of the target dfa_lookahead hook. */
5282 int dfa_lookahead;
5284 /* The following variable value is maximal number of tries of issuing
5285 insns for the first cycle multipass insn scheduling. We define
5286 this value as constant*(DFA_LOOKAHEAD**ISSUE_RATE). We would not
5287 need this constraint if all real insns (with non-negative codes)
5288 had reservations because in this case the algorithm complexity is
5289 O(DFA_LOOKAHEAD**ISSUE_RATE). Unfortunately, the dfa descriptions
5290 might be incomplete and such insn might occur. For such
5291 descriptions, the complexity of algorithm (without the constraint)
5292 could achieve DFA_LOOKAHEAD ** N , where N is the queue length. */
5293 static int max_lookahead_tries;
5295 /* The following value is value of hook
5296 `first_cycle_multipass_dfa_lookahead' at the last call of
5297 `max_issue'. */
5298 static int cached_first_cycle_multipass_dfa_lookahead = 0;
5300 /* The following value is value of `issue_rate' at the last call of
5301 `sched_init'. */
5302 static int cached_issue_rate = 0;
5304 /* The following function returns maximal (or close to maximal) number
5305 of insns which can be issued on the same cycle and one of which
5306 insns is insns with the best rank (the first insn in READY). To
5307 make this function tries different samples of ready insns. READY
5308 is current queue `ready'. Global array READY_TRY reflects what
5309 insns are already issued in this try. The function stops immediately,
5310 if it reached the such a solution, that all instruction can be issued.
5311 INDEX will contain index of the best insn in READY. The following
5312 function is used only for first cycle multipass scheduling.
5314 PRIVILEGED_N >= 0
5316 This function expects recognized insns only. All USEs,
5317 CLOBBERs, etc must be filtered elsewhere. */
5319 max_issue (struct ready_list *ready, int privileged_n, state_t state,
5320 bool first_cycle_insn_p, int *index)
5322 int n, i, all, n_ready, best, delay, tries_num;
5323 int more_issue;
5324 struct choice_entry *top;
5325 rtx insn;
5327 n_ready = ready->n_ready;
5328 gcc_assert (dfa_lookahead >= 1 && privileged_n >= 0
5329 && privileged_n <= n_ready);
5331 /* Init MAX_LOOKAHEAD_TRIES. */
5332 if (cached_first_cycle_multipass_dfa_lookahead != dfa_lookahead)
5334 cached_first_cycle_multipass_dfa_lookahead = dfa_lookahead;
5335 max_lookahead_tries = 100;
5336 for (i = 0; i < issue_rate; i++)
5337 max_lookahead_tries *= dfa_lookahead;
5340 /* Init max_points. */
5341 more_issue = issue_rate - cycle_issued_insns;
5342 gcc_assert (more_issue >= 0);
5344 /* The number of the issued insns in the best solution. */
5345 best = 0;
5347 top = choice_stack;
5349 /* Set initial state of the search. */
5350 memcpy (top->state, state, dfa_state_size);
5351 top->rest = dfa_lookahead;
5352 top->n = 0;
5353 if (targetm.sched.first_cycle_multipass_begin)
5354 targetm.sched.first_cycle_multipass_begin (&top->target_data,
5355 ready_try, n_ready,
5356 first_cycle_insn_p);
5358 /* Count the number of the insns to search among. */
5359 for (all = i = 0; i < n_ready; i++)
5360 if (!ready_try [i])
5361 all++;
5363 /* I is the index of the insn to try next. */
5364 i = 0;
5365 tries_num = 0;
5366 for (;;)
5368 if (/* If we've reached a dead end or searched enough of what we have
5369 been asked... */
5370 top->rest == 0
5371 /* or have nothing else to try... */
5372 || i >= n_ready
5373 /* or should not issue more. */
5374 || top->n >= more_issue)
5376 /* ??? (... || i == n_ready). */
5377 gcc_assert (i <= n_ready);
5379 /* We should not issue more than issue_rate instructions. */
5380 gcc_assert (top->n <= more_issue);
5382 if (top == choice_stack)
5383 break;
5385 if (best < top - choice_stack)
5387 if (privileged_n)
5389 n = privileged_n;
5390 /* Try to find issued privileged insn. */
5391 while (n && !ready_try[--n])
5395 if (/* If all insns are equally good... */
5396 privileged_n == 0
5397 /* Or a privileged insn will be issued. */
5398 || ready_try[n])
5399 /* Then we have a solution. */
5401 best = top - choice_stack;
5402 /* This is the index of the insn issued first in this
5403 solution. */
5404 *index = choice_stack [1].index;
5405 if (top->n == more_issue || best == all)
5406 break;
5410 /* Set ready-list index to point to the last insn
5411 ('i++' below will advance it to the next insn). */
5412 i = top->index;
5414 /* Backtrack. */
5415 ready_try [i] = 0;
5417 if (targetm.sched.first_cycle_multipass_backtrack)
5418 targetm.sched.first_cycle_multipass_backtrack (&top->target_data,
5419 ready_try, n_ready);
5421 top--;
5422 memcpy (state, top->state, dfa_state_size);
5424 else if (!ready_try [i])
5426 tries_num++;
5427 if (tries_num > max_lookahead_tries)
5428 break;
5429 insn = ready_element (ready, i);
5430 delay = state_transition (state, insn);
5431 if (delay < 0)
5433 if (state_dead_lock_p (state)
5434 || insn_finishes_cycle_p (insn))
5435 /* We won't issue any more instructions in the next
5436 choice_state. */
5437 top->rest = 0;
5438 else
5439 top->rest--;
5441 n = top->n;
5442 if (memcmp (top->state, state, dfa_state_size) != 0)
5443 n++;
5445 /* Advance to the next choice_entry. */
5446 top++;
5447 /* Initialize it. */
5448 top->rest = dfa_lookahead;
5449 top->index = i;
5450 top->n = n;
5451 memcpy (top->state, state, dfa_state_size);
5452 ready_try [i] = 1;
5454 if (targetm.sched.first_cycle_multipass_issue)
5455 targetm.sched.first_cycle_multipass_issue (&top->target_data,
5456 ready_try, n_ready,
5457 insn,
5458 &((top - 1)
5459 ->target_data));
5461 i = -1;
5465 /* Increase ready-list index. */
5466 i++;
5469 if (targetm.sched.first_cycle_multipass_end)
5470 targetm.sched.first_cycle_multipass_end (best != 0
5471 ? &choice_stack[1].target_data
5472 : NULL);
5474 /* Restore the original state of the DFA. */
5475 memcpy (state, choice_stack->state, dfa_state_size);
5477 return best;
5480 /* The following function chooses insn from READY and modifies
5481 READY. The following function is used only for first
5482 cycle multipass scheduling.
5483 Return:
5484 -1 if cycle should be advanced,
5485 0 if INSN_PTR is set to point to the desirable insn,
5486 1 if choose_ready () should be restarted without advancing the cycle. */
5487 static int
5488 choose_ready (struct ready_list *ready, bool first_cycle_insn_p,
5489 rtx *insn_ptr)
5491 int lookahead;
5493 if (dbg_cnt (sched_insn) == false)
5495 rtx insn = nonscheduled_insns_begin;
5498 insn = next_nonnote_insn (insn);
5500 while (QUEUE_INDEX (insn) == QUEUE_SCHEDULED);
5502 if (QUEUE_INDEX (insn) == QUEUE_READY)
5503 /* INSN is in the ready_list. */
5505 nonscheduled_insns_begin = insn;
5506 ready_remove_insn (insn);
5507 *insn_ptr = insn;
5508 return 0;
5511 /* INSN is in the queue. Advance cycle to move it to the ready list. */
5512 return -1;
5515 lookahead = 0;
5517 if (targetm.sched.first_cycle_multipass_dfa_lookahead)
5518 lookahead = targetm.sched.first_cycle_multipass_dfa_lookahead ();
5519 if (lookahead <= 0 || SCHED_GROUP_P (ready_element (ready, 0))
5520 || DEBUG_INSN_P (ready_element (ready, 0)))
5522 if (targetm.sched.dispatch (NULL_RTX, IS_DISPATCH_ON))
5523 *insn_ptr = ready_remove_first_dispatch (ready);
5524 else
5525 *insn_ptr = ready_remove_first (ready);
5527 return 0;
5529 else
5531 /* Try to choose the better insn. */
5532 int index = 0, i, n;
5533 rtx insn;
5534 int try_data = 1, try_control = 1;
5535 ds_t ts;
5537 insn = ready_element (ready, 0);
5538 if (INSN_CODE (insn) < 0)
5540 *insn_ptr = ready_remove_first (ready);
5541 return 0;
5544 if (spec_info
5545 && spec_info->flags & (PREFER_NON_DATA_SPEC
5546 | PREFER_NON_CONTROL_SPEC))
5548 for (i = 0, n = ready->n_ready; i < n; i++)
5550 rtx x;
5551 ds_t s;
5553 x = ready_element (ready, i);
5554 s = TODO_SPEC (x);
5556 if (spec_info->flags & PREFER_NON_DATA_SPEC
5557 && !(s & DATA_SPEC))
5559 try_data = 0;
5560 if (!(spec_info->flags & PREFER_NON_CONTROL_SPEC)
5561 || !try_control)
5562 break;
5565 if (spec_info->flags & PREFER_NON_CONTROL_SPEC
5566 && !(s & CONTROL_SPEC))
5568 try_control = 0;
5569 if (!(spec_info->flags & PREFER_NON_DATA_SPEC) || !try_data)
5570 break;
5575 ts = TODO_SPEC (insn);
5576 if ((ts & SPECULATIVE)
5577 && (((!try_data && (ts & DATA_SPEC))
5578 || (!try_control && (ts & CONTROL_SPEC)))
5579 || (targetm.sched.first_cycle_multipass_dfa_lookahead_guard_spec
5580 && !targetm.sched
5581 .first_cycle_multipass_dfa_lookahead_guard_spec (insn))))
5582 /* Discard speculative instruction that stands first in the ready
5583 list. */
5585 change_queue_index (insn, 1);
5586 return 1;
5589 ready_try[0] = 0;
5591 for (i = 1; i < ready->n_ready; i++)
5593 insn = ready_element (ready, i);
5595 ready_try [i]
5596 = ((!try_data && (TODO_SPEC (insn) & DATA_SPEC))
5597 || (!try_control && (TODO_SPEC (insn) & CONTROL_SPEC)));
5600 /* Let the target filter the search space. */
5601 for (i = 1; i < ready->n_ready; i++)
5602 if (!ready_try[i])
5604 insn = ready_element (ready, i);
5606 /* If this insn is recognizable we should have already
5607 recognized it earlier.
5608 ??? Not very clear where this is supposed to be done.
5609 See dep_cost_1. */
5610 gcc_checking_assert (INSN_CODE (insn) >= 0
5611 || recog_memoized (insn) < 0);
5613 ready_try [i]
5614 = (/* INSN_CODE check can be omitted here as it is also done later
5615 in max_issue (). */
5616 INSN_CODE (insn) < 0
5617 || (targetm.sched.first_cycle_multipass_dfa_lookahead_guard
5618 && !targetm.sched.first_cycle_multipass_dfa_lookahead_guard
5619 (insn)));
5622 if (max_issue (ready, 1, curr_state, first_cycle_insn_p, &index) == 0)
5624 *insn_ptr = ready_remove_first (ready);
5625 if (sched_verbose >= 4)
5626 fprintf (sched_dump, ";;\t\tChosen insn (but can't issue) : %s \n",
5627 (*current_sched_info->print_insn) (*insn_ptr, 0));
5628 return 0;
5630 else
5632 if (sched_verbose >= 4)
5633 fprintf (sched_dump, ";;\t\tChosen insn : %s\n",
5634 (*current_sched_info->print_insn)
5635 (ready_element (ready, index), 0));
5637 *insn_ptr = ready_remove (ready, index);
5638 return 0;
5643 /* This function is called when we have successfully scheduled a
5644 block. It uses the schedule stored in the scheduled_insns vector
5645 to rearrange the RTL. PREV_HEAD is used as the anchor to which we
5646 append the scheduled insns; TAIL is the insn after the scheduled
5647 block. TARGET_BB is the argument passed to schedule_block. */
5649 static void
5650 commit_schedule (rtx prev_head, rtx tail, basic_block *target_bb)
5652 unsigned int i;
5653 rtx insn;
5655 last_scheduled_insn = prev_head;
5656 for (i = 0;
5657 scheduled_insns.iterate (i, &insn);
5658 i++)
5660 if (control_flow_insn_p (last_scheduled_insn)
5661 || current_sched_info->advance_target_bb (*target_bb, insn))
5663 *target_bb = current_sched_info->advance_target_bb (*target_bb, 0);
5665 if (sched_verbose)
5667 rtx x;
5669 x = next_real_insn (last_scheduled_insn);
5670 gcc_assert (x);
5671 dump_new_block_header (1, *target_bb, x, tail);
5674 last_scheduled_insn = bb_note (*target_bb);
5677 if (current_sched_info->begin_move_insn)
5678 (*current_sched_info->begin_move_insn) (insn, last_scheduled_insn);
5679 move_insn (insn, last_scheduled_insn,
5680 current_sched_info->next_tail);
5681 if (!DEBUG_INSN_P (insn))
5682 reemit_notes (insn);
5683 last_scheduled_insn = insn;
5686 scheduled_insns.truncate (0);
5689 /* Examine all insns on the ready list and queue those which can't be
5690 issued in this cycle. TEMP_STATE is temporary scheduler state we
5691 can use as scratch space. If FIRST_CYCLE_INSN_P is true, no insns
5692 have been issued for the current cycle, which means it is valid to
5693 issue an asm statement.
5695 If SHADOWS_ONLY_P is true, we eliminate all real insns and only
5696 leave those for which SHADOW_P is true. If MODULO_EPILOGUE is true,
5697 we only leave insns which have an INSN_EXACT_TICK. */
5699 static void
5700 prune_ready_list (state_t temp_state, bool first_cycle_insn_p,
5701 bool shadows_only_p, bool modulo_epilogue_p)
5703 int i, pass;
5704 bool sched_group_found = false;
5705 int min_cost_group = 1;
5707 for (i = 0; i < ready.n_ready; i++)
5709 rtx insn = ready_element (&ready, i);
5710 if (SCHED_GROUP_P (insn))
5712 sched_group_found = true;
5713 break;
5717 /* Make two passes if there's a SCHED_GROUP_P insn; make sure to handle
5718 such an insn first and note its cost, then schedule all other insns
5719 for one cycle later. */
5720 for (pass = sched_group_found ? 0 : 1; pass < 2; )
5722 int n = ready.n_ready;
5723 for (i = 0; i < n; i++)
5725 rtx insn = ready_element (&ready, i);
5726 int cost = 0;
5727 const char *reason = "resource conflict";
5729 if (DEBUG_INSN_P (insn))
5730 continue;
5732 if (sched_group_found && !SCHED_GROUP_P (insn))
5734 if (pass == 0)
5735 continue;
5736 cost = min_cost_group;
5737 reason = "not in sched group";
5739 else if (modulo_epilogue_p
5740 && INSN_EXACT_TICK (insn) == INVALID_TICK)
5742 cost = max_insn_queue_index;
5743 reason = "not an epilogue insn";
5745 else if (shadows_only_p && !SHADOW_P (insn))
5747 cost = 1;
5748 reason = "not a shadow";
5750 else if (recog_memoized (insn) < 0)
5752 if (!first_cycle_insn_p
5753 && (GET_CODE (PATTERN (insn)) == ASM_INPUT
5754 || asm_noperands (PATTERN (insn)) >= 0))
5755 cost = 1;
5756 reason = "asm";
5758 else if (sched_pressure != SCHED_PRESSURE_NONE)
5760 if (sched_pressure == SCHED_PRESSURE_MODEL
5761 && INSN_TICK (insn) <= clock_var)
5763 memcpy (temp_state, curr_state, dfa_state_size);
5764 if (state_transition (temp_state, insn) >= 0)
5765 INSN_TICK (insn) = clock_var + 1;
5767 cost = 0;
5769 else
5771 int delay_cost = 0;
5773 if (delay_htab.is_created ())
5775 struct delay_pair *delay_entry;
5776 delay_entry
5777 = delay_htab.find_with_hash (insn,
5778 htab_hash_pointer (insn));
5779 while (delay_entry && delay_cost == 0)
5781 delay_cost = estimate_shadow_tick (delay_entry);
5782 if (delay_cost > max_insn_queue_index)
5783 delay_cost = max_insn_queue_index;
5784 delay_entry = delay_entry->next_same_i1;
5788 memcpy (temp_state, curr_state, dfa_state_size);
5789 cost = state_transition (temp_state, insn);
5790 if (cost < 0)
5791 cost = 0;
5792 else if (cost == 0)
5793 cost = 1;
5794 if (cost < delay_cost)
5796 cost = delay_cost;
5797 reason = "shadow tick";
5800 if (cost >= 1)
5802 if (SCHED_GROUP_P (insn) && cost > min_cost_group)
5803 min_cost_group = cost;
5804 ready_remove (&ready, i);
5805 queue_insn (insn, cost, reason);
5806 if (i + 1 < n)
5807 break;
5810 if (i == n)
5811 pass++;
5815 /* Called when we detect that the schedule is impossible. We examine the
5816 backtrack queue to find the earliest insn that caused this condition. */
5818 static struct haifa_saved_data *
5819 verify_shadows (void)
5821 struct haifa_saved_data *save, *earliest_fail = NULL;
5822 for (save = backtrack_queue; save; save = save->next)
5824 int t;
5825 struct delay_pair *pair = save->delay_pair;
5826 rtx i1 = pair->i1;
5828 for (; pair; pair = pair->next_same_i1)
5830 rtx i2 = pair->i2;
5832 if (QUEUE_INDEX (i2) == QUEUE_SCHEDULED)
5833 continue;
5835 t = INSN_TICK (i1) + pair_delay (pair);
5836 if (t < clock_var)
5838 if (sched_verbose >= 2)
5839 fprintf (sched_dump,
5840 ";;\t\tfailed delay requirements for %d/%d (%d->%d)"
5841 ", not ready\n",
5842 INSN_UID (pair->i1), INSN_UID (pair->i2),
5843 INSN_TICK (pair->i1), INSN_EXACT_TICK (pair->i2));
5844 earliest_fail = save;
5845 break;
5847 if (QUEUE_INDEX (i2) >= 0)
5849 int queued_for = INSN_TICK (i2);
5851 if (t < queued_for)
5853 if (sched_verbose >= 2)
5854 fprintf (sched_dump,
5855 ";;\t\tfailed delay requirements for %d/%d"
5856 " (%d->%d), queued too late\n",
5857 INSN_UID (pair->i1), INSN_UID (pair->i2),
5858 INSN_TICK (pair->i1), INSN_EXACT_TICK (pair->i2));
5859 earliest_fail = save;
5860 break;
5866 return earliest_fail;
5869 /* Use forward list scheduling to rearrange insns of block pointed to by
5870 TARGET_BB, possibly bringing insns from subsequent blocks in the same
5871 region. */
5873 bool
5874 schedule_block (basic_block *target_bb, state_t init_state)
5876 int i;
5877 bool success = modulo_ii == 0;
5878 struct sched_block_state ls;
5879 state_t temp_state = NULL; /* It is used for multipass scheduling. */
5880 int sort_p, advance, start_clock_var;
5882 /* Head/tail info for this block. */
5883 rtx prev_head = current_sched_info->prev_head;
5884 rtx next_tail = current_sched_info->next_tail;
5885 rtx head = NEXT_INSN (prev_head);
5886 rtx tail = PREV_INSN (next_tail);
5888 if ((current_sched_info->flags & DONT_BREAK_DEPENDENCIES) == 0
5889 && sched_pressure != SCHED_PRESSURE_MODEL)
5890 find_modifiable_mems (head, tail);
5892 /* We used to have code to avoid getting parameters moved from hard
5893 argument registers into pseudos.
5895 However, it was removed when it proved to be of marginal benefit
5896 and caused problems because schedule_block and compute_forward_dependences
5897 had different notions of what the "head" insn was. */
5899 gcc_assert (head != tail || INSN_P (head));
5901 haifa_recovery_bb_recently_added_p = false;
5903 backtrack_queue = NULL;
5905 /* Debug info. */
5906 if (sched_verbose)
5907 dump_new_block_header (0, *target_bb, head, tail);
5909 if (init_state == NULL)
5910 state_reset (curr_state);
5911 else
5912 memcpy (curr_state, init_state, dfa_state_size);
5914 /* Clear the ready list. */
5915 ready.first = ready.veclen - 1;
5916 ready.n_ready = 0;
5917 ready.n_debug = 0;
5919 /* It is used for first cycle multipass scheduling. */
5920 temp_state = alloca (dfa_state_size);
5922 if (targetm.sched.init)
5923 targetm.sched.init (sched_dump, sched_verbose, ready.veclen);
5925 /* We start inserting insns after PREV_HEAD. */
5926 last_scheduled_insn = nonscheduled_insns_begin = prev_head;
5927 last_nondebug_scheduled_insn = NULL_RTX;
5929 gcc_assert ((NOTE_P (last_scheduled_insn)
5930 || DEBUG_INSN_P (last_scheduled_insn))
5931 && BLOCK_FOR_INSN (last_scheduled_insn) == *target_bb);
5933 /* Initialize INSN_QUEUE. Q_SIZE is the total number of insns in the
5934 queue. */
5935 q_ptr = 0;
5936 q_size = 0;
5938 insn_queue = XALLOCAVEC (rtx, max_insn_queue_index + 1);
5939 memset (insn_queue, 0, (max_insn_queue_index + 1) * sizeof (rtx));
5941 /* Start just before the beginning of time. */
5942 clock_var = -1;
5944 /* We need queue and ready lists and clock_var be initialized
5945 in try_ready () (which is called through init_ready_list ()). */
5946 (*current_sched_info->init_ready_list) ();
5948 if (sched_pressure == SCHED_PRESSURE_MODEL)
5949 model_start_schedule ();
5951 /* The algorithm is O(n^2) in the number of ready insns at any given
5952 time in the worst case. Before reload we are more likely to have
5953 big lists so truncate them to a reasonable size. */
5954 if (!reload_completed
5955 && ready.n_ready - ready.n_debug > MAX_SCHED_READY_INSNS)
5957 ready_sort (&ready);
5959 /* Find first free-standing insn past MAX_SCHED_READY_INSNS.
5960 If there are debug insns, we know they're first. */
5961 for (i = MAX_SCHED_READY_INSNS + ready.n_debug; i < ready.n_ready; i++)
5962 if (!SCHED_GROUP_P (ready_element (&ready, i)))
5963 break;
5965 if (sched_verbose >= 2)
5967 fprintf (sched_dump,
5968 ";;\t\tReady list on entry: %d insns\n", ready.n_ready);
5969 fprintf (sched_dump,
5970 ";;\t\t before reload => truncated to %d insns\n", i);
5973 /* Delay all insns past it for 1 cycle. If debug counter is
5974 activated make an exception for the insn right after
5975 nonscheduled_insns_begin. */
5977 rtx skip_insn;
5979 if (dbg_cnt (sched_insn) == false)
5980 skip_insn = next_nonnote_insn (nonscheduled_insns_begin);
5981 else
5982 skip_insn = NULL_RTX;
5984 while (i < ready.n_ready)
5986 rtx insn;
5988 insn = ready_remove (&ready, i);
5990 if (insn != skip_insn)
5991 queue_insn (insn, 1, "list truncated");
5993 if (skip_insn)
5994 ready_add (&ready, skip_insn, true);
5998 /* Now we can restore basic block notes and maintain precise cfg. */
5999 restore_bb_notes (*target_bb);
6001 last_clock_var = -1;
6003 advance = 0;
6005 gcc_assert (scheduled_insns.length () == 0);
6006 sort_p = TRUE;
6007 must_backtrack = false;
6008 modulo_insns_scheduled = 0;
6010 ls.modulo_epilogue = false;
6012 /* Loop until all the insns in BB are scheduled. */
6013 while ((*current_sched_info->schedule_more_p) ())
6015 perform_replacements_new_cycle ();
6018 start_clock_var = clock_var;
6020 clock_var++;
6022 advance_one_cycle ();
6024 /* Add to the ready list all pending insns that can be issued now.
6025 If there are no ready insns, increment clock until one
6026 is ready and add all pending insns at that point to the ready
6027 list. */
6028 queue_to_ready (&ready);
6030 gcc_assert (ready.n_ready);
6032 if (sched_verbose >= 2)
6034 fprintf (sched_dump, ";;\t\tReady list after queue_to_ready: ");
6035 debug_ready_list (&ready);
6037 advance -= clock_var - start_clock_var;
6039 while (advance > 0);
6041 if (ls.modulo_epilogue)
6043 int stage = clock_var / modulo_ii;
6044 if (stage > modulo_last_stage * 2 + 2)
6046 if (sched_verbose >= 2)
6047 fprintf (sched_dump,
6048 ";;\t\tmodulo scheduled succeeded at II %d\n",
6049 modulo_ii);
6050 success = true;
6051 goto end_schedule;
6054 else if (modulo_ii > 0)
6056 int stage = clock_var / modulo_ii;
6057 if (stage > modulo_max_stages)
6059 if (sched_verbose >= 2)
6060 fprintf (sched_dump,
6061 ";;\t\tfailing schedule due to excessive stages\n");
6062 goto end_schedule;
6064 if (modulo_n_insns == modulo_insns_scheduled
6065 && stage > modulo_last_stage)
6067 if (sched_verbose >= 2)
6068 fprintf (sched_dump,
6069 ";;\t\tfound kernel after %d stages, II %d\n",
6070 stage, modulo_ii);
6071 ls.modulo_epilogue = true;
6075 prune_ready_list (temp_state, true, false, ls.modulo_epilogue);
6076 if (ready.n_ready == 0)
6077 continue;
6078 if (must_backtrack)
6079 goto do_backtrack;
6081 ls.first_cycle_insn_p = true;
6082 ls.shadows_only_p = false;
6083 cycle_issued_insns = 0;
6084 ls.can_issue_more = issue_rate;
6085 for (;;)
6087 rtx insn;
6088 int cost;
6089 bool asm_p;
6091 if (sort_p && ready.n_ready > 0)
6093 /* Sort the ready list based on priority. This must be
6094 done every iteration through the loop, as schedule_insn
6095 may have readied additional insns that will not be
6096 sorted correctly. */
6097 ready_sort (&ready);
6099 if (sched_verbose >= 2)
6101 fprintf (sched_dump, ";;\t\tReady list after ready_sort: ");
6102 debug_ready_list (&ready);
6106 /* We don't want md sched reorder to even see debug isns, so put
6107 them out right away. */
6108 if (ready.n_ready && DEBUG_INSN_P (ready_element (&ready, 0))
6109 && (*current_sched_info->schedule_more_p) ())
6111 while (ready.n_ready && DEBUG_INSN_P (ready_element (&ready, 0)))
6113 rtx insn = ready_remove_first (&ready);
6114 gcc_assert (DEBUG_INSN_P (insn));
6115 (*current_sched_info->begin_schedule_ready) (insn);
6116 scheduled_insns.safe_push (insn);
6117 last_scheduled_insn = insn;
6118 advance = schedule_insn (insn);
6119 gcc_assert (advance == 0);
6120 if (ready.n_ready > 0)
6121 ready_sort (&ready);
6125 if (ls.first_cycle_insn_p && !ready.n_ready)
6126 break;
6128 resume_after_backtrack:
6129 /* Allow the target to reorder the list, typically for
6130 better instruction bundling. */
6131 if (sort_p
6132 && (ready.n_ready == 0
6133 || !SCHED_GROUP_P (ready_element (&ready, 0))))
6135 if (ls.first_cycle_insn_p && targetm.sched.reorder)
6136 ls.can_issue_more
6137 = targetm.sched.reorder (sched_dump, sched_verbose,
6138 ready_lastpos (&ready),
6139 &ready.n_ready, clock_var);
6140 else if (!ls.first_cycle_insn_p && targetm.sched.reorder2)
6141 ls.can_issue_more
6142 = targetm.sched.reorder2 (sched_dump, sched_verbose,
6143 ready.n_ready
6144 ? ready_lastpos (&ready) : NULL,
6145 &ready.n_ready, clock_var);
6148 restart_choose_ready:
6149 if (sched_verbose >= 2)
6151 fprintf (sched_dump, ";;\tReady list (t = %3d): ",
6152 clock_var);
6153 debug_ready_list (&ready);
6154 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
6155 print_curr_reg_pressure ();
6158 if (ready.n_ready == 0
6159 && ls.can_issue_more
6160 && reload_completed)
6162 /* Allow scheduling insns directly from the queue in case
6163 there's nothing better to do (ready list is empty) but
6164 there are still vacant dispatch slots in the current cycle. */
6165 if (sched_verbose >= 6)
6166 fprintf (sched_dump,";;\t\tSecond chance\n");
6167 memcpy (temp_state, curr_state, dfa_state_size);
6168 if (early_queue_to_ready (temp_state, &ready))
6169 ready_sort (&ready);
6172 if (ready.n_ready == 0
6173 || !ls.can_issue_more
6174 || state_dead_lock_p (curr_state)
6175 || !(*current_sched_info->schedule_more_p) ())
6176 break;
6178 /* Select and remove the insn from the ready list. */
6179 if (sort_p)
6181 int res;
6183 insn = NULL_RTX;
6184 res = choose_ready (&ready, ls.first_cycle_insn_p, &insn);
6186 if (res < 0)
6187 /* Finish cycle. */
6188 break;
6189 if (res > 0)
6190 goto restart_choose_ready;
6192 gcc_assert (insn != NULL_RTX);
6194 else
6195 insn = ready_remove_first (&ready);
6197 if (sched_pressure != SCHED_PRESSURE_NONE
6198 && INSN_TICK (insn) > clock_var)
6200 ready_add (&ready, insn, true);
6201 advance = 1;
6202 break;
6205 if (targetm.sched.dfa_new_cycle
6206 && targetm.sched.dfa_new_cycle (sched_dump, sched_verbose,
6207 insn, last_clock_var,
6208 clock_var, &sort_p))
6209 /* SORT_P is used by the target to override sorting
6210 of the ready list. This is needed when the target
6211 has modified its internal structures expecting that
6212 the insn will be issued next. As we need the insn
6213 to have the highest priority (so it will be returned by
6214 the ready_remove_first call above), we invoke
6215 ready_add (&ready, insn, true).
6216 But, still, there is one issue: INSN can be later
6217 discarded by scheduler's front end through
6218 current_sched_info->can_schedule_ready_p, hence, won't
6219 be issued next. */
6221 ready_add (&ready, insn, true);
6222 break;
6225 sort_p = TRUE;
6227 if (current_sched_info->can_schedule_ready_p
6228 && ! (*current_sched_info->can_schedule_ready_p) (insn))
6229 /* We normally get here only if we don't want to move
6230 insn from the split block. */
6232 TODO_SPEC (insn) = DEP_POSTPONED;
6233 goto restart_choose_ready;
6236 if (delay_htab.is_created ())
6238 /* If this insn is the first part of a delay-slot pair, record a
6239 backtrack point. */
6240 struct delay_pair *delay_entry;
6241 delay_entry
6242 = delay_htab.find_with_hash (insn, htab_hash_pointer (insn));
6243 if (delay_entry)
6245 save_backtrack_point (delay_entry, ls);
6246 if (sched_verbose >= 2)
6247 fprintf (sched_dump, ";;\t\tsaving backtrack point\n");
6251 /* DECISION is made. */
6253 if (modulo_ii > 0 && INSN_UID (insn) < modulo_iter0_max_uid)
6255 modulo_insns_scheduled++;
6256 modulo_last_stage = clock_var / modulo_ii;
6258 if (TODO_SPEC (insn) & SPECULATIVE)
6259 generate_recovery_code (insn);
6261 if (targetm.sched.dispatch (NULL_RTX, IS_DISPATCH_ON))
6262 targetm.sched.dispatch_do (insn, ADD_TO_DISPATCH_WINDOW);
6264 /* Update counters, etc in the scheduler's front end. */
6265 (*current_sched_info->begin_schedule_ready) (insn);
6266 scheduled_insns.safe_push (insn);
6267 gcc_assert (NONDEBUG_INSN_P (insn));
6268 last_nondebug_scheduled_insn = last_scheduled_insn = insn;
6270 if (recog_memoized (insn) >= 0)
6272 memcpy (temp_state, curr_state, dfa_state_size);
6273 cost = state_transition (curr_state, insn);
6274 if (sched_pressure != SCHED_PRESSURE_WEIGHTED)
6275 gcc_assert (cost < 0);
6276 if (memcmp (temp_state, curr_state, dfa_state_size) != 0)
6277 cycle_issued_insns++;
6278 asm_p = false;
6280 else
6281 asm_p = (GET_CODE (PATTERN (insn)) == ASM_INPUT
6282 || asm_noperands (PATTERN (insn)) >= 0);
6284 if (targetm.sched.variable_issue)
6285 ls.can_issue_more =
6286 targetm.sched.variable_issue (sched_dump, sched_verbose,
6287 insn, ls.can_issue_more);
6288 /* A naked CLOBBER or USE generates no instruction, so do
6289 not count them against the issue rate. */
6290 else if (GET_CODE (PATTERN (insn)) != USE
6291 && GET_CODE (PATTERN (insn)) != CLOBBER)
6292 ls.can_issue_more--;
6293 advance = schedule_insn (insn);
6295 if (SHADOW_P (insn))
6296 ls.shadows_only_p = true;
6298 /* After issuing an asm insn we should start a new cycle. */
6299 if (advance == 0 && asm_p)
6300 advance = 1;
6302 if (must_backtrack)
6303 break;
6305 if (advance != 0)
6306 break;
6308 ls.first_cycle_insn_p = false;
6309 if (ready.n_ready > 0)
6310 prune_ready_list (temp_state, false, ls.shadows_only_p,
6311 ls.modulo_epilogue);
6314 do_backtrack:
6315 if (!must_backtrack)
6316 for (i = 0; i < ready.n_ready; i++)
6318 rtx insn = ready_element (&ready, i);
6319 if (INSN_EXACT_TICK (insn) == clock_var)
6321 must_backtrack = true;
6322 clock_var++;
6323 break;
6326 if (must_backtrack && modulo_ii > 0)
6328 if (modulo_backtracks_left == 0)
6329 goto end_schedule;
6330 modulo_backtracks_left--;
6332 while (must_backtrack)
6334 struct haifa_saved_data *failed;
6335 rtx failed_insn;
6337 must_backtrack = false;
6338 failed = verify_shadows ();
6339 gcc_assert (failed);
6341 failed_insn = failed->delay_pair->i1;
6342 /* Clear these queues. */
6343 perform_replacements_new_cycle ();
6344 toggle_cancelled_flags (false);
6345 unschedule_insns_until (failed_insn);
6346 while (failed != backtrack_queue)
6347 free_topmost_backtrack_point (true);
6348 restore_last_backtrack_point (&ls);
6349 if (sched_verbose >= 2)
6350 fprintf (sched_dump, ";;\t\trewind to cycle %d\n", clock_var);
6351 /* Delay by at least a cycle. This could cause additional
6352 backtracking. */
6353 queue_insn (failed_insn, 1, "backtracked");
6354 advance = 0;
6355 if (must_backtrack)
6356 continue;
6357 if (ready.n_ready > 0)
6358 goto resume_after_backtrack;
6359 else
6361 if (clock_var == 0 && ls.first_cycle_insn_p)
6362 goto end_schedule;
6363 advance = 1;
6364 break;
6368 if (ls.modulo_epilogue)
6369 success = true;
6370 end_schedule:
6371 advance_one_cycle ();
6372 perform_replacements_new_cycle ();
6373 if (modulo_ii > 0)
6375 /* Once again, debug insn suckiness: they can be on the ready list
6376 even if they have unresolved dependencies. To make our view
6377 of the world consistent, remove such "ready" insns. */
6378 restart_debug_insn_loop:
6379 for (i = ready.n_ready - 1; i >= 0; i--)
6381 rtx x;
6383 x = ready_element (&ready, i);
6384 if (DEPS_LIST_FIRST (INSN_HARD_BACK_DEPS (x)) != NULL
6385 || DEPS_LIST_FIRST (INSN_SPEC_BACK_DEPS (x)) != NULL)
6387 ready_remove (&ready, i);
6388 goto restart_debug_insn_loop;
6391 for (i = ready.n_ready - 1; i >= 0; i--)
6393 rtx x;
6395 x = ready_element (&ready, i);
6396 resolve_dependencies (x);
6398 for (i = 0; i <= max_insn_queue_index; i++)
6400 rtx link;
6401 while ((link = insn_queue[i]) != NULL)
6403 rtx x = XEXP (link, 0);
6404 insn_queue[i] = XEXP (link, 1);
6405 QUEUE_INDEX (x) = QUEUE_NOWHERE;
6406 free_INSN_LIST_node (link);
6407 resolve_dependencies (x);
6412 if (!success)
6413 undo_all_replacements ();
6415 /* Debug info. */
6416 if (sched_verbose)
6418 fprintf (sched_dump, ";;\tReady list (final): ");
6419 debug_ready_list (&ready);
6422 if (modulo_ii == 0 && current_sched_info->queue_must_finish_empty)
6423 /* Sanity check -- queue must be empty now. Meaningless if region has
6424 multiple bbs. */
6425 gcc_assert (!q_size && !ready.n_ready && !ready.n_debug);
6426 else if (modulo_ii == 0)
6428 /* We must maintain QUEUE_INDEX between blocks in region. */
6429 for (i = ready.n_ready - 1; i >= 0; i--)
6431 rtx x;
6433 x = ready_element (&ready, i);
6434 QUEUE_INDEX (x) = QUEUE_NOWHERE;
6435 TODO_SPEC (x) = HARD_DEP;
6438 if (q_size)
6439 for (i = 0; i <= max_insn_queue_index; i++)
6441 rtx link;
6442 for (link = insn_queue[i]; link; link = XEXP (link, 1))
6444 rtx x;
6446 x = XEXP (link, 0);
6447 QUEUE_INDEX (x) = QUEUE_NOWHERE;
6448 TODO_SPEC (x) = HARD_DEP;
6450 free_INSN_LIST_list (&insn_queue[i]);
6454 if (sched_pressure == SCHED_PRESSURE_MODEL)
6455 model_end_schedule ();
6457 if (success)
6459 commit_schedule (prev_head, tail, target_bb);
6460 if (sched_verbose)
6461 fprintf (sched_dump, ";; total time = %d\n", clock_var);
6463 else
6464 last_scheduled_insn = tail;
6466 scheduled_insns.truncate (0);
6468 if (!current_sched_info->queue_must_finish_empty
6469 || haifa_recovery_bb_recently_added_p)
6471 /* INSN_TICK (minimum clock tick at which the insn becomes
6472 ready) may be not correct for the insn in the subsequent
6473 blocks of the region. We should use a correct value of
6474 `clock_var' or modify INSN_TICK. It is better to keep
6475 clock_var value equal to 0 at the start of a basic block.
6476 Therefore we modify INSN_TICK here. */
6477 fix_inter_tick (NEXT_INSN (prev_head), last_scheduled_insn);
6480 if (targetm.sched.finish)
6482 targetm.sched.finish (sched_dump, sched_verbose);
6483 /* Target might have added some instructions to the scheduled block
6484 in its md_finish () hook. These new insns don't have any data
6485 initialized and to identify them we extend h_i_d so that they'll
6486 get zero luids. */
6487 sched_extend_luids ();
6490 if (sched_verbose)
6491 fprintf (sched_dump, ";; new head = %d\n;; new tail = %d\n\n",
6492 INSN_UID (head), INSN_UID (tail));
6494 /* Update head/tail boundaries. */
6495 head = NEXT_INSN (prev_head);
6496 tail = last_scheduled_insn;
6498 head = restore_other_notes (head, NULL);
6500 current_sched_info->head = head;
6501 current_sched_info->tail = tail;
6503 free_backtrack_queue ();
6505 return success;
6508 /* Set_priorities: compute priority of each insn in the block. */
6511 set_priorities (rtx head, rtx tail)
6513 rtx insn;
6514 int n_insn;
6515 int sched_max_insns_priority =
6516 current_sched_info->sched_max_insns_priority;
6517 rtx prev_head;
6519 if (head == tail && ! INSN_P (head))
6520 gcc_unreachable ();
6522 n_insn = 0;
6524 prev_head = PREV_INSN (head);
6525 for (insn = tail; insn != prev_head; insn = PREV_INSN (insn))
6527 if (!INSN_P (insn))
6528 continue;
6530 n_insn++;
6531 (void) priority (insn);
6533 gcc_assert (INSN_PRIORITY_KNOWN (insn));
6535 sched_max_insns_priority = MAX (sched_max_insns_priority,
6536 INSN_PRIORITY (insn));
6539 current_sched_info->sched_max_insns_priority = sched_max_insns_priority;
6541 return n_insn;
6544 /* Set dump and sched_verbose for the desired debugging output. If no
6545 dump-file was specified, but -fsched-verbose=N (any N), print to stderr.
6546 For -fsched-verbose=N, N>=10, print everything to stderr. */
6547 void
6548 setup_sched_dump (void)
6550 sched_verbose = sched_verbose_param;
6551 if (sched_verbose_param == 0 && dump_file)
6552 sched_verbose = 1;
6553 sched_dump = ((sched_verbose_param >= 10 || !dump_file)
6554 ? stderr : dump_file);
6557 /* Try to group comparison and the following conditional jump INSN if
6558 they're already adjacent. This is to prevent scheduler from scheduling
6559 them apart. */
6561 static void
6562 try_group_insn (rtx insn)
6564 unsigned int condreg1, condreg2;
6565 rtx cc_reg_1;
6566 rtx prev;
6568 if (!any_condjump_p (insn))
6569 return;
6571 targetm.fixed_condition_code_regs (&condreg1, &condreg2);
6572 cc_reg_1 = gen_rtx_REG (CCmode, condreg1);
6573 prev = prev_nonnote_nondebug_insn (insn);
6574 if (!reg_referenced_p (cc_reg_1, PATTERN (insn))
6575 || !prev
6576 || !modified_in_p (cc_reg_1, prev))
6577 return;
6579 /* Different microarchitectures support macro fusions for different
6580 combinations of insn pairs. */
6581 if (!targetm.sched.macro_fusion_pair_p
6582 || !targetm.sched.macro_fusion_pair_p (prev, insn))
6583 return;
6585 SCHED_GROUP_P (insn) = 1;
6588 /* If the last cond jump and the cond register defining insn are consecutive
6589 before scheduling, we want them to be in a schedule group. This is good
6590 for performance on microarchitectures supporting macro-fusion. */
6592 static void
6593 group_insns_for_macro_fusion ()
6595 basic_block bb;
6597 FOR_EACH_BB (bb)
6598 try_group_insn (BB_END (bb));
6601 /* Initialize some global state for the scheduler. This function works
6602 with the common data shared between all the schedulers. It is called
6603 from the scheduler specific initialization routine. */
6605 void
6606 sched_init (void)
6608 /* Disable speculative loads in their presence if cc0 defined. */
6609 #ifdef HAVE_cc0
6610 flag_schedule_speculative_load = 0;
6611 #endif
6613 if (targetm.sched.dispatch (NULL_RTX, IS_DISPATCH_ON))
6614 targetm.sched.dispatch_do (NULL_RTX, DISPATCH_INIT);
6616 if (live_range_shrinkage_p)
6617 sched_pressure = SCHED_PRESSURE_WEIGHTED;
6618 else if (flag_sched_pressure
6619 && !reload_completed
6620 && common_sched_info->sched_pass_id == SCHED_RGN_PASS)
6621 sched_pressure = ((enum sched_pressure_algorithm)
6622 PARAM_VALUE (PARAM_SCHED_PRESSURE_ALGORITHM));
6623 else
6624 sched_pressure = SCHED_PRESSURE_NONE;
6626 if (sched_pressure != SCHED_PRESSURE_NONE)
6627 ira_setup_eliminable_regset (false);
6629 /* Initialize SPEC_INFO. */
6630 if (targetm.sched.set_sched_flags)
6632 spec_info = &spec_info_var;
6633 targetm.sched.set_sched_flags (spec_info);
6635 if (spec_info->mask != 0)
6637 spec_info->data_weakness_cutoff =
6638 (PARAM_VALUE (PARAM_SCHED_SPEC_PROB_CUTOFF) * MAX_DEP_WEAK) / 100;
6639 spec_info->control_weakness_cutoff =
6640 (PARAM_VALUE (PARAM_SCHED_SPEC_PROB_CUTOFF)
6641 * REG_BR_PROB_BASE) / 100;
6643 else
6644 /* So we won't read anything accidentally. */
6645 spec_info = NULL;
6648 else
6649 /* So we won't read anything accidentally. */
6650 spec_info = 0;
6652 /* Initialize issue_rate. */
6653 if (targetm.sched.issue_rate)
6654 issue_rate = targetm.sched.issue_rate ();
6655 else
6656 issue_rate = 1;
6658 if (cached_issue_rate != issue_rate)
6660 cached_issue_rate = issue_rate;
6661 /* To invalidate max_lookahead_tries: */
6662 cached_first_cycle_multipass_dfa_lookahead = 0;
6665 if (targetm.sched.first_cycle_multipass_dfa_lookahead)
6666 dfa_lookahead = targetm.sched.first_cycle_multipass_dfa_lookahead ();
6667 else
6668 dfa_lookahead = 0;
6670 if (targetm.sched.init_dfa_pre_cycle_insn)
6671 targetm.sched.init_dfa_pre_cycle_insn ();
6673 if (targetm.sched.init_dfa_post_cycle_insn)
6674 targetm.sched.init_dfa_post_cycle_insn ();
6676 dfa_start ();
6677 dfa_state_size = state_size ();
6679 init_alias_analysis ();
6681 if (!sched_no_dce)
6682 df_set_flags (DF_LR_RUN_DCE);
6683 df_note_add_problem ();
6685 /* More problems needed for interloop dep calculation in SMS. */
6686 if (common_sched_info->sched_pass_id == SCHED_SMS_PASS)
6688 df_rd_add_problem ();
6689 df_chain_add_problem (DF_DU_CHAIN + DF_UD_CHAIN);
6692 df_analyze ();
6694 /* Do not run DCE after reload, as this can kill nops inserted
6695 by bundling. */
6696 if (reload_completed)
6697 df_clear_flags (DF_LR_RUN_DCE);
6699 regstat_compute_calls_crossed ();
6701 if (targetm.sched.init_global)
6702 targetm.sched.init_global (sched_dump, sched_verbose, get_max_uid () + 1);
6704 if (sched_pressure != SCHED_PRESSURE_NONE)
6706 int i, max_regno = max_reg_num ();
6708 if (sched_dump != NULL)
6709 /* We need info about pseudos for rtl dumps about pseudo
6710 classes and costs. */
6711 regstat_init_n_sets_and_refs ();
6712 ira_set_pseudo_classes (true, sched_verbose ? sched_dump : NULL);
6713 sched_regno_pressure_class
6714 = (enum reg_class *) xmalloc (max_regno * sizeof (enum reg_class));
6715 for (i = 0; i < max_regno; i++)
6716 sched_regno_pressure_class[i]
6717 = (i < FIRST_PSEUDO_REGISTER
6718 ? ira_pressure_class_translate[REGNO_REG_CLASS (i)]
6719 : ira_pressure_class_translate[reg_allocno_class (i)]);
6720 curr_reg_live = BITMAP_ALLOC (NULL);
6721 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
6723 saved_reg_live = BITMAP_ALLOC (NULL);
6724 region_ref_regs = BITMAP_ALLOC (NULL);
6728 curr_state = xmalloc (dfa_state_size);
6730 /* Group compare and branch insns for macro-fusion. */
6731 if (targetm.sched.macro_fusion_p
6732 && targetm.sched.macro_fusion_p ())
6733 group_insns_for_macro_fusion ();
6736 static void haifa_init_only_bb (basic_block, basic_block);
6738 /* Initialize data structures specific to the Haifa scheduler. */
6739 void
6740 haifa_sched_init (void)
6742 setup_sched_dump ();
6743 sched_init ();
6745 scheduled_insns.create (0);
6747 if (spec_info != NULL)
6749 sched_deps_info->use_deps_list = 1;
6750 sched_deps_info->generate_spec_deps = 1;
6753 /* Initialize luids, dependency caches, target and h_i_d for the
6754 whole function. */
6756 bb_vec_t bbs;
6757 bbs.create (n_basic_blocks_for_fn (cfun));
6758 basic_block bb;
6760 sched_init_bbs ();
6762 FOR_EACH_BB (bb)
6763 bbs.quick_push (bb);
6764 sched_init_luids (bbs);
6765 sched_deps_init (true);
6766 sched_extend_target ();
6767 haifa_init_h_i_d (bbs);
6769 bbs.release ();
6772 sched_init_only_bb = haifa_init_only_bb;
6773 sched_split_block = sched_split_block_1;
6774 sched_create_empty_bb = sched_create_empty_bb_1;
6775 haifa_recovery_bb_ever_added_p = false;
6777 nr_begin_data = nr_begin_control = nr_be_in_data = nr_be_in_control = 0;
6778 before_recovery = 0;
6779 after_recovery = 0;
6781 modulo_ii = 0;
6784 /* Finish work with the data specific to the Haifa scheduler. */
6785 void
6786 haifa_sched_finish (void)
6788 sched_create_empty_bb = NULL;
6789 sched_split_block = NULL;
6790 sched_init_only_bb = NULL;
6792 if (spec_info && spec_info->dump)
6794 char c = reload_completed ? 'a' : 'b';
6796 fprintf (spec_info->dump,
6797 ";; %s:\n", current_function_name ());
6799 fprintf (spec_info->dump,
6800 ";; Procedure %cr-begin-data-spec motions == %d\n",
6801 c, nr_begin_data);
6802 fprintf (spec_info->dump,
6803 ";; Procedure %cr-be-in-data-spec motions == %d\n",
6804 c, nr_be_in_data);
6805 fprintf (spec_info->dump,
6806 ";; Procedure %cr-begin-control-spec motions == %d\n",
6807 c, nr_begin_control);
6808 fprintf (spec_info->dump,
6809 ";; Procedure %cr-be-in-control-spec motions == %d\n",
6810 c, nr_be_in_control);
6813 scheduled_insns.release ();
6815 /* Finalize h_i_d, dependency caches, and luids for the whole
6816 function. Target will be finalized in md_global_finish (). */
6817 sched_deps_finish ();
6818 sched_finish_luids ();
6819 current_sched_info = NULL;
6820 sched_finish ();
6823 /* Free global data used during insn scheduling. This function works with
6824 the common data shared between the schedulers. */
6826 void
6827 sched_finish (void)
6829 haifa_finish_h_i_d ();
6830 if (sched_pressure != SCHED_PRESSURE_NONE)
6832 if (regstat_n_sets_and_refs != NULL)
6833 regstat_free_n_sets_and_refs ();
6834 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
6836 BITMAP_FREE (region_ref_regs);
6837 BITMAP_FREE (saved_reg_live);
6839 BITMAP_FREE (curr_reg_live);
6840 free (sched_regno_pressure_class);
6842 free (curr_state);
6844 if (targetm.sched.finish_global)
6845 targetm.sched.finish_global (sched_dump, sched_verbose);
6847 end_alias_analysis ();
6849 regstat_free_calls_crossed ();
6851 dfa_finish ();
6854 /* Free all delay_pair structures that were recorded. */
6855 void
6856 free_delay_pairs (void)
6858 if (delay_htab.is_created ())
6860 delay_htab.empty ();
6861 delay_htab_i2.empty ();
6865 /* Fix INSN_TICKs of the instructions in the current block as well as
6866 INSN_TICKs of their dependents.
6867 HEAD and TAIL are the begin and the end of the current scheduled block. */
6868 static void
6869 fix_inter_tick (rtx head, rtx tail)
6871 /* Set of instructions with corrected INSN_TICK. */
6872 bitmap_head processed;
6873 /* ??? It is doubtful if we should assume that cycle advance happens on
6874 basic block boundaries. Basically insns that are unconditionally ready
6875 on the start of the block are more preferable then those which have
6876 a one cycle dependency over insn from the previous block. */
6877 int next_clock = clock_var + 1;
6879 bitmap_initialize (&processed, 0);
6881 /* Iterates over scheduled instructions and fix their INSN_TICKs and
6882 INSN_TICKs of dependent instructions, so that INSN_TICKs are consistent
6883 across different blocks. */
6884 for (tail = NEXT_INSN (tail); head != tail; head = NEXT_INSN (head))
6886 if (INSN_P (head))
6888 int tick;
6889 sd_iterator_def sd_it;
6890 dep_t dep;
6892 tick = INSN_TICK (head);
6893 gcc_assert (tick >= MIN_TICK);
6895 /* Fix INSN_TICK of instruction from just scheduled block. */
6896 if (bitmap_set_bit (&processed, INSN_LUID (head)))
6898 tick -= next_clock;
6900 if (tick < MIN_TICK)
6901 tick = MIN_TICK;
6903 INSN_TICK (head) = tick;
6906 if (DEBUG_INSN_P (head))
6907 continue;
6909 FOR_EACH_DEP (head, SD_LIST_RES_FORW, sd_it, dep)
6911 rtx next;
6913 next = DEP_CON (dep);
6914 tick = INSN_TICK (next);
6916 if (tick != INVALID_TICK
6917 /* If NEXT has its INSN_TICK calculated, fix it.
6918 If not - it will be properly calculated from
6919 scratch later in fix_tick_ready. */
6920 && bitmap_set_bit (&processed, INSN_LUID (next)))
6922 tick -= next_clock;
6924 if (tick < MIN_TICK)
6925 tick = MIN_TICK;
6927 if (tick > INTER_TICK (next))
6928 INTER_TICK (next) = tick;
6929 else
6930 tick = INTER_TICK (next);
6932 INSN_TICK (next) = tick;
6937 bitmap_clear (&processed);
6940 /* Check if NEXT is ready to be added to the ready or queue list.
6941 If "yes", add it to the proper list.
6942 Returns:
6943 -1 - is not ready yet,
6944 0 - added to the ready list,
6945 0 < N - queued for N cycles. */
6947 try_ready (rtx next)
6949 ds_t old_ts, new_ts;
6951 old_ts = TODO_SPEC (next);
6953 gcc_assert (!(old_ts & ~(SPECULATIVE | HARD_DEP | DEP_CONTROL | DEP_POSTPONED))
6954 && (old_ts == HARD_DEP
6955 || old_ts == DEP_POSTPONED
6956 || (old_ts & SPECULATIVE)
6957 || old_ts == DEP_CONTROL));
6959 new_ts = recompute_todo_spec (next, false);
6961 if (new_ts & (HARD_DEP | DEP_POSTPONED))
6962 gcc_assert (new_ts == old_ts
6963 && QUEUE_INDEX (next) == QUEUE_NOWHERE);
6964 else if (current_sched_info->new_ready)
6965 new_ts = current_sched_info->new_ready (next, new_ts);
6967 /* * if !(old_ts & SPECULATIVE) (e.g. HARD_DEP or 0), then insn might
6968 have its original pattern or changed (speculative) one. This is due
6969 to changing ebb in region scheduling.
6970 * But if (old_ts & SPECULATIVE), then we are pretty sure that insn
6971 has speculative pattern.
6973 We can't assert (!(new_ts & HARD_DEP) || new_ts == old_ts) here because
6974 control-speculative NEXT could have been discarded by sched-rgn.c
6975 (the same case as when discarded by can_schedule_ready_p ()). */
6977 if ((new_ts & SPECULATIVE)
6978 /* If (old_ts == new_ts), then (old_ts & SPECULATIVE) and we don't
6979 need to change anything. */
6980 && new_ts != old_ts)
6982 int res;
6983 rtx new_pat;
6985 gcc_assert ((new_ts & SPECULATIVE) && !(new_ts & ~SPECULATIVE));
6987 res = haifa_speculate_insn (next, new_ts, &new_pat);
6989 switch (res)
6991 case -1:
6992 /* It would be nice to change DEP_STATUS of all dependences,
6993 which have ((DEP_STATUS & SPECULATIVE) == new_ts) to HARD_DEP,
6994 so we won't reanalyze anything. */
6995 new_ts = HARD_DEP;
6996 break;
6998 case 0:
6999 /* We follow the rule, that every speculative insn
7000 has non-null ORIG_PAT. */
7001 if (!ORIG_PAT (next))
7002 ORIG_PAT (next) = PATTERN (next);
7003 break;
7005 case 1:
7006 if (!ORIG_PAT (next))
7007 /* If we gonna to overwrite the original pattern of insn,
7008 save it. */
7009 ORIG_PAT (next) = PATTERN (next);
7011 res = haifa_change_pattern (next, new_pat);
7012 gcc_assert (res);
7013 break;
7015 default:
7016 gcc_unreachable ();
7020 /* We need to restore pattern only if (new_ts == 0), because otherwise it is
7021 either correct (new_ts & SPECULATIVE),
7022 or we simply don't care (new_ts & HARD_DEP). */
7024 gcc_assert (!ORIG_PAT (next)
7025 || !IS_SPECULATION_BRANCHY_CHECK_P (next));
7027 TODO_SPEC (next) = new_ts;
7029 if (new_ts & (HARD_DEP | DEP_POSTPONED))
7031 /* We can't assert (QUEUE_INDEX (next) == QUEUE_NOWHERE) here because
7032 control-speculative NEXT could have been discarded by sched-rgn.c
7033 (the same case as when discarded by can_schedule_ready_p ()). */
7034 /*gcc_assert (QUEUE_INDEX (next) == QUEUE_NOWHERE);*/
7036 change_queue_index (next, QUEUE_NOWHERE);
7038 return -1;
7040 else if (!(new_ts & BEGIN_SPEC)
7041 && ORIG_PAT (next) && PREDICATED_PAT (next) == NULL_RTX
7042 && !IS_SPECULATION_CHECK_P (next))
7043 /* We should change pattern of every previously speculative
7044 instruction - and we determine if NEXT was speculative by using
7045 ORIG_PAT field. Except one case - speculation checks have ORIG_PAT
7046 pat too, so skip them. */
7048 bool success = haifa_change_pattern (next, ORIG_PAT (next));
7049 gcc_assert (success);
7050 ORIG_PAT (next) = 0;
7053 if (sched_verbose >= 2)
7055 fprintf (sched_dump, ";;\t\tdependencies resolved: insn %s",
7056 (*current_sched_info->print_insn) (next, 0));
7058 if (spec_info && spec_info->dump)
7060 if (new_ts & BEGIN_DATA)
7061 fprintf (spec_info->dump, "; data-spec;");
7062 if (new_ts & BEGIN_CONTROL)
7063 fprintf (spec_info->dump, "; control-spec;");
7064 if (new_ts & BE_IN_CONTROL)
7065 fprintf (spec_info->dump, "; in-control-spec;");
7067 if (TODO_SPEC (next) & DEP_CONTROL)
7068 fprintf (sched_dump, " predicated");
7069 fprintf (sched_dump, "\n");
7072 adjust_priority (next);
7074 return fix_tick_ready (next);
7077 /* Calculate INSN_TICK of NEXT and add it to either ready or queue list. */
7078 static int
7079 fix_tick_ready (rtx next)
7081 int tick, delay;
7083 if (!DEBUG_INSN_P (next) && !sd_lists_empty_p (next, SD_LIST_RES_BACK))
7085 int full_p;
7086 sd_iterator_def sd_it;
7087 dep_t dep;
7089 tick = INSN_TICK (next);
7090 /* if tick is not equal to INVALID_TICK, then update
7091 INSN_TICK of NEXT with the most recent resolved dependence
7092 cost. Otherwise, recalculate from scratch. */
7093 full_p = (tick == INVALID_TICK);
7095 FOR_EACH_DEP (next, SD_LIST_RES_BACK, sd_it, dep)
7097 rtx pro = DEP_PRO (dep);
7098 int tick1;
7100 gcc_assert (INSN_TICK (pro) >= MIN_TICK);
7102 tick1 = INSN_TICK (pro) + dep_cost (dep);
7103 if (tick1 > tick)
7104 tick = tick1;
7106 if (!full_p)
7107 break;
7110 else
7111 tick = -1;
7113 INSN_TICK (next) = tick;
7115 delay = tick - clock_var;
7116 if (delay <= 0 || sched_pressure != SCHED_PRESSURE_NONE)
7117 delay = QUEUE_READY;
7119 change_queue_index (next, delay);
7121 return delay;
7124 /* Move NEXT to the proper queue list with (DELAY >= 1),
7125 or add it to the ready list (DELAY == QUEUE_READY),
7126 or remove it from ready and queue lists at all (DELAY == QUEUE_NOWHERE). */
7127 static void
7128 change_queue_index (rtx next, int delay)
7130 int i = QUEUE_INDEX (next);
7132 gcc_assert (QUEUE_NOWHERE <= delay && delay <= max_insn_queue_index
7133 && delay != 0);
7134 gcc_assert (i != QUEUE_SCHEDULED);
7136 if ((delay > 0 && NEXT_Q_AFTER (q_ptr, delay) == i)
7137 || (delay < 0 && delay == i))
7138 /* We have nothing to do. */
7139 return;
7141 /* Remove NEXT from wherever it is now. */
7142 if (i == QUEUE_READY)
7143 ready_remove_insn (next);
7144 else if (i >= 0)
7145 queue_remove (next);
7147 /* Add it to the proper place. */
7148 if (delay == QUEUE_READY)
7149 ready_add (readyp, next, false);
7150 else if (delay >= 1)
7151 queue_insn (next, delay, "change queue index");
7153 if (sched_verbose >= 2)
7155 fprintf (sched_dump, ";;\t\ttick updated: insn %s",
7156 (*current_sched_info->print_insn) (next, 0));
7158 if (delay == QUEUE_READY)
7159 fprintf (sched_dump, " into ready\n");
7160 else if (delay >= 1)
7161 fprintf (sched_dump, " into queue with cost=%d\n", delay);
7162 else
7163 fprintf (sched_dump, " removed from ready or queue lists\n");
7167 static int sched_ready_n_insns = -1;
7169 /* Initialize per region data structures. */
7170 void
7171 sched_extend_ready_list (int new_sched_ready_n_insns)
7173 int i;
7175 if (sched_ready_n_insns == -1)
7176 /* At the first call we need to initialize one more choice_stack
7177 entry. */
7179 i = 0;
7180 sched_ready_n_insns = 0;
7181 scheduled_insns.reserve (new_sched_ready_n_insns);
7183 else
7184 i = sched_ready_n_insns + 1;
7186 ready.veclen = new_sched_ready_n_insns + issue_rate;
7187 ready.vec = XRESIZEVEC (rtx, ready.vec, ready.veclen);
7189 gcc_assert (new_sched_ready_n_insns >= sched_ready_n_insns);
7191 ready_try = (char *) xrecalloc (ready_try, new_sched_ready_n_insns,
7192 sched_ready_n_insns, sizeof (*ready_try));
7194 /* We allocate +1 element to save initial state in the choice_stack[0]
7195 entry. */
7196 choice_stack = XRESIZEVEC (struct choice_entry, choice_stack,
7197 new_sched_ready_n_insns + 1);
7199 for (; i <= new_sched_ready_n_insns; i++)
7201 choice_stack[i].state = xmalloc (dfa_state_size);
7203 if (targetm.sched.first_cycle_multipass_init)
7204 targetm.sched.first_cycle_multipass_init (&(choice_stack[i]
7205 .target_data));
7208 sched_ready_n_insns = new_sched_ready_n_insns;
7211 /* Free per region data structures. */
7212 void
7213 sched_finish_ready_list (void)
7215 int i;
7217 free (ready.vec);
7218 ready.vec = NULL;
7219 ready.veclen = 0;
7221 free (ready_try);
7222 ready_try = NULL;
7224 for (i = 0; i <= sched_ready_n_insns; i++)
7226 if (targetm.sched.first_cycle_multipass_fini)
7227 targetm.sched.first_cycle_multipass_fini (&(choice_stack[i]
7228 .target_data));
7230 free (choice_stack [i].state);
7232 free (choice_stack);
7233 choice_stack = NULL;
7235 sched_ready_n_insns = -1;
7238 static int
7239 haifa_luid_for_non_insn (rtx x)
7241 gcc_assert (NOTE_P (x) || LABEL_P (x));
7243 return 0;
7246 /* Generates recovery code for INSN. */
7247 static void
7248 generate_recovery_code (rtx insn)
7250 if (TODO_SPEC (insn) & BEGIN_SPEC)
7251 begin_speculative_block (insn);
7253 /* Here we have insn with no dependencies to
7254 instructions other then CHECK_SPEC ones. */
7256 if (TODO_SPEC (insn) & BE_IN_SPEC)
7257 add_to_speculative_block (insn);
7260 /* Helper function.
7261 Tries to add speculative dependencies of type FS between instructions
7262 in deps_list L and TWIN. */
7263 static void
7264 process_insn_forw_deps_be_in_spec (rtx insn, rtx twin, ds_t fs)
7266 sd_iterator_def sd_it;
7267 dep_t dep;
7269 FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep)
7271 ds_t ds;
7272 rtx consumer;
7274 consumer = DEP_CON (dep);
7276 ds = DEP_STATUS (dep);
7278 if (/* If we want to create speculative dep. */
7280 /* And we can do that because this is a true dep. */
7281 && (ds & DEP_TYPES) == DEP_TRUE)
7283 gcc_assert (!(ds & BE_IN_SPEC));
7285 if (/* If this dep can be overcome with 'begin speculation'. */
7286 ds & BEGIN_SPEC)
7287 /* Then we have a choice: keep the dep 'begin speculative'
7288 or transform it into 'be in speculative'. */
7290 if (/* In try_ready we assert that if insn once became ready
7291 it can be removed from the ready (or queue) list only
7292 due to backend decision. Hence we can't let the
7293 probability of the speculative dep to decrease. */
7294 ds_weak (ds) <= ds_weak (fs))
7296 ds_t new_ds;
7298 new_ds = (ds & ~BEGIN_SPEC) | fs;
7300 if (/* consumer can 'be in speculative'. */
7301 sched_insn_is_legitimate_for_speculation_p (consumer,
7302 new_ds))
7303 /* Transform it to be in speculative. */
7304 ds = new_ds;
7307 else
7308 /* Mark the dep as 'be in speculative'. */
7309 ds |= fs;
7313 dep_def _new_dep, *new_dep = &_new_dep;
7315 init_dep_1 (new_dep, twin, consumer, DEP_TYPE (dep), ds);
7316 sd_add_dep (new_dep, false);
7321 /* Generates recovery code for BEGIN speculative INSN. */
7322 static void
7323 begin_speculative_block (rtx insn)
7325 if (TODO_SPEC (insn) & BEGIN_DATA)
7326 nr_begin_data++;
7327 if (TODO_SPEC (insn) & BEGIN_CONTROL)
7328 nr_begin_control++;
7330 create_check_block_twin (insn, false);
7332 TODO_SPEC (insn) &= ~BEGIN_SPEC;
7335 static void haifa_init_insn (rtx);
7337 /* Generates recovery code for BE_IN speculative INSN. */
7338 static void
7339 add_to_speculative_block (rtx insn)
7341 ds_t ts;
7342 sd_iterator_def sd_it;
7343 dep_t dep;
7344 rtx twins = NULL;
7345 rtx_vec_t priorities_roots;
7347 ts = TODO_SPEC (insn);
7348 gcc_assert (!(ts & ~BE_IN_SPEC));
7350 if (ts & BE_IN_DATA)
7351 nr_be_in_data++;
7352 if (ts & BE_IN_CONTROL)
7353 nr_be_in_control++;
7355 TODO_SPEC (insn) &= ~BE_IN_SPEC;
7356 gcc_assert (!TODO_SPEC (insn));
7358 DONE_SPEC (insn) |= ts;
7360 /* First we convert all simple checks to branchy. */
7361 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
7362 sd_iterator_cond (&sd_it, &dep);)
7364 rtx check = DEP_PRO (dep);
7366 if (IS_SPECULATION_SIMPLE_CHECK_P (check))
7368 create_check_block_twin (check, true);
7370 /* Restart search. */
7371 sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
7373 else
7374 /* Continue search. */
7375 sd_iterator_next (&sd_it);
7378 priorities_roots.create (0);
7379 clear_priorities (insn, &priorities_roots);
7381 while (1)
7383 rtx check, twin;
7384 basic_block rec;
7386 /* Get the first backward dependency of INSN. */
7387 sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
7388 if (!sd_iterator_cond (&sd_it, &dep))
7389 /* INSN has no backward dependencies left. */
7390 break;
7392 gcc_assert ((DEP_STATUS (dep) & BEGIN_SPEC) == 0
7393 && (DEP_STATUS (dep) & BE_IN_SPEC) != 0
7394 && (DEP_STATUS (dep) & DEP_TYPES) == DEP_TRUE);
7396 check = DEP_PRO (dep);
7398 gcc_assert (!IS_SPECULATION_CHECK_P (check) && !ORIG_PAT (check)
7399 && QUEUE_INDEX (check) == QUEUE_NOWHERE);
7401 rec = BLOCK_FOR_INSN (check);
7403 twin = emit_insn_before (copy_insn (PATTERN (insn)), BB_END (rec));
7404 haifa_init_insn (twin);
7406 sd_copy_back_deps (twin, insn, true);
7408 if (sched_verbose && spec_info->dump)
7409 /* INSN_BB (insn) isn't determined for twin insns yet.
7410 So we can't use current_sched_info->print_insn. */
7411 fprintf (spec_info->dump, ";;\t\tGenerated twin insn : %d/rec%d\n",
7412 INSN_UID (twin), rec->index);
7414 twins = alloc_INSN_LIST (twin, twins);
7416 /* Add dependences between TWIN and all appropriate
7417 instructions from REC. */
7418 FOR_EACH_DEP (insn, SD_LIST_SPEC_BACK, sd_it, dep)
7420 rtx pro = DEP_PRO (dep);
7422 gcc_assert (DEP_TYPE (dep) == REG_DEP_TRUE);
7424 /* INSN might have dependencies from the instructions from
7425 several recovery blocks. At this iteration we process those
7426 producers that reside in REC. */
7427 if (BLOCK_FOR_INSN (pro) == rec)
7429 dep_def _new_dep, *new_dep = &_new_dep;
7431 init_dep (new_dep, pro, twin, REG_DEP_TRUE);
7432 sd_add_dep (new_dep, false);
7436 process_insn_forw_deps_be_in_spec (insn, twin, ts);
7438 /* Remove all dependencies between INSN and insns in REC. */
7439 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
7440 sd_iterator_cond (&sd_it, &dep);)
7442 rtx pro = DEP_PRO (dep);
7444 if (BLOCK_FOR_INSN (pro) == rec)
7445 sd_delete_dep (sd_it);
7446 else
7447 sd_iterator_next (&sd_it);
7451 /* We couldn't have added the dependencies between INSN and TWINS earlier
7452 because that would make TWINS appear in the INSN_BACK_DEPS (INSN). */
7453 while (twins)
7455 rtx twin;
7457 twin = XEXP (twins, 0);
7460 dep_def _new_dep, *new_dep = &_new_dep;
7462 init_dep (new_dep, insn, twin, REG_DEP_OUTPUT);
7463 sd_add_dep (new_dep, false);
7466 twin = XEXP (twins, 1);
7467 free_INSN_LIST_node (twins);
7468 twins = twin;
7471 calc_priorities (priorities_roots);
7472 priorities_roots.release ();
7475 /* Extends and fills with zeros (only the new part) array pointed to by P. */
7476 void *
7477 xrecalloc (void *p, size_t new_nmemb, size_t old_nmemb, size_t size)
7479 gcc_assert (new_nmemb >= old_nmemb);
7480 p = XRESIZEVAR (void, p, new_nmemb * size);
7481 memset (((char *) p) + old_nmemb * size, 0, (new_nmemb - old_nmemb) * size);
7482 return p;
7485 /* Helper function.
7486 Find fallthru edge from PRED. */
7487 edge
7488 find_fallthru_edge_from (basic_block pred)
7490 edge e;
7491 basic_block succ;
7493 succ = pred->next_bb;
7494 gcc_assert (succ->prev_bb == pred);
7496 if (EDGE_COUNT (pred->succs) <= EDGE_COUNT (succ->preds))
7498 e = find_fallthru_edge (pred->succs);
7500 if (e)
7502 gcc_assert (e->dest == succ);
7503 return e;
7506 else
7508 e = find_fallthru_edge (succ->preds);
7510 if (e)
7512 gcc_assert (e->src == pred);
7513 return e;
7517 return NULL;
7520 /* Extend per basic block data structures. */
7521 static void
7522 sched_extend_bb (void)
7524 /* The following is done to keep current_sched_info->next_tail non null. */
7525 rtx end = BB_END (EXIT_BLOCK_PTR->prev_bb);
7526 rtx insn = DEBUG_INSN_P (end) ? prev_nondebug_insn (end) : end;
7527 if (NEXT_INSN (end) == 0
7528 || (!NOTE_P (insn)
7529 && !LABEL_P (insn)
7530 /* Don't emit a NOTE if it would end up before a BARRIER. */
7531 && !BARRIER_P (NEXT_INSN (end))))
7533 rtx note = emit_note_after (NOTE_INSN_DELETED, end);
7534 /* Make note appear outside BB. */
7535 set_block_for_insn (note, NULL);
7536 BB_END (EXIT_BLOCK_PTR->prev_bb) = end;
7540 /* Init per basic block data structures. */
7541 void
7542 sched_init_bbs (void)
7544 sched_extend_bb ();
7547 /* Initialize BEFORE_RECOVERY variable. */
7548 static void
7549 init_before_recovery (basic_block *before_recovery_ptr)
7551 basic_block last;
7552 edge e;
7554 last = EXIT_BLOCK_PTR->prev_bb;
7555 e = find_fallthru_edge_from (last);
7557 if (e)
7559 /* We create two basic blocks:
7560 1. Single instruction block is inserted right after E->SRC
7561 and has jump to
7562 2. Empty block right before EXIT_BLOCK.
7563 Between these two blocks recovery blocks will be emitted. */
7565 basic_block single, empty;
7566 rtx x, label;
7568 /* If the fallthrough edge to exit we've found is from the block we've
7569 created before, don't do anything more. */
7570 if (last == after_recovery)
7571 return;
7573 adding_bb_to_current_region_p = false;
7575 single = sched_create_empty_bb (last);
7576 empty = sched_create_empty_bb (single);
7578 /* Add new blocks to the root loop. */
7579 if (current_loops != NULL)
7581 add_bb_to_loop (single, (*current_loops->larray)[0]);
7582 add_bb_to_loop (empty, (*current_loops->larray)[0]);
7585 single->count = last->count;
7586 empty->count = last->count;
7587 single->frequency = last->frequency;
7588 empty->frequency = last->frequency;
7589 BB_COPY_PARTITION (single, last);
7590 BB_COPY_PARTITION (empty, last);
7592 redirect_edge_succ (e, single);
7593 make_single_succ_edge (single, empty, 0);
7594 make_single_succ_edge (empty, EXIT_BLOCK_PTR, EDGE_FALLTHRU);
7596 label = block_label (empty);
7597 x = emit_jump_insn_after (gen_jump (label), BB_END (single));
7598 JUMP_LABEL (x) = label;
7599 LABEL_NUSES (label)++;
7600 haifa_init_insn (x);
7602 emit_barrier_after (x);
7604 sched_init_only_bb (empty, NULL);
7605 sched_init_only_bb (single, NULL);
7606 sched_extend_bb ();
7608 adding_bb_to_current_region_p = true;
7609 before_recovery = single;
7610 after_recovery = empty;
7612 if (before_recovery_ptr)
7613 *before_recovery_ptr = before_recovery;
7615 if (sched_verbose >= 2 && spec_info->dump)
7616 fprintf (spec_info->dump,
7617 ";;\t\tFixed fallthru to EXIT : %d->>%d->%d->>EXIT\n",
7618 last->index, single->index, empty->index);
7620 else
7621 before_recovery = last;
7624 /* Returns new recovery block. */
7625 basic_block
7626 sched_create_recovery_block (basic_block *before_recovery_ptr)
7628 rtx label;
7629 rtx barrier;
7630 basic_block rec;
7632 haifa_recovery_bb_recently_added_p = true;
7633 haifa_recovery_bb_ever_added_p = true;
7635 init_before_recovery (before_recovery_ptr);
7637 barrier = get_last_bb_insn (before_recovery);
7638 gcc_assert (BARRIER_P (barrier));
7640 label = emit_label_after (gen_label_rtx (), barrier);
7642 rec = create_basic_block (label, label, before_recovery);
7644 /* A recovery block always ends with an unconditional jump. */
7645 emit_barrier_after (BB_END (rec));
7647 if (BB_PARTITION (before_recovery) != BB_UNPARTITIONED)
7648 BB_SET_PARTITION (rec, BB_COLD_PARTITION);
7650 if (sched_verbose && spec_info->dump)
7651 fprintf (spec_info->dump, ";;\t\tGenerated recovery block rec%d\n",
7652 rec->index);
7654 return rec;
7657 /* Create edges: FIRST_BB -> REC; FIRST_BB -> SECOND_BB; REC -> SECOND_BB
7658 and emit necessary jumps. */
7659 void
7660 sched_create_recovery_edges (basic_block first_bb, basic_block rec,
7661 basic_block second_bb)
7663 rtx label;
7664 rtx jump;
7665 int edge_flags;
7667 /* This is fixing of incoming edge. */
7668 /* ??? Which other flags should be specified? */
7669 if (BB_PARTITION (first_bb) != BB_PARTITION (rec))
7670 /* Partition type is the same, if it is "unpartitioned". */
7671 edge_flags = EDGE_CROSSING;
7672 else
7673 edge_flags = 0;
7675 make_edge (first_bb, rec, edge_flags);
7676 label = block_label (second_bb);
7677 jump = emit_jump_insn_after (gen_jump (label), BB_END (rec));
7678 JUMP_LABEL (jump) = label;
7679 LABEL_NUSES (label)++;
7681 if (BB_PARTITION (second_bb) != BB_PARTITION (rec))
7682 /* Partition type is the same, if it is "unpartitioned". */
7684 /* Rewritten from cfgrtl.c. */
7685 if (flag_reorder_blocks_and_partition
7686 && targetm_common.have_named_sections)
7688 /* We don't need the same note for the check because
7689 any_condjump_p (check) == true. */
7690 add_reg_note (jump, REG_CROSSING_JUMP, NULL_RTX);
7692 edge_flags = EDGE_CROSSING;
7694 else
7695 edge_flags = 0;
7697 make_single_succ_edge (rec, second_bb, edge_flags);
7698 if (dom_info_available_p (CDI_DOMINATORS))
7699 set_immediate_dominator (CDI_DOMINATORS, rec, first_bb);
7702 /* This function creates recovery code for INSN. If MUTATE_P is nonzero,
7703 INSN is a simple check, that should be converted to branchy one. */
7704 static void
7705 create_check_block_twin (rtx insn, bool mutate_p)
7707 basic_block rec;
7708 rtx label, check, twin;
7709 ds_t fs;
7710 sd_iterator_def sd_it;
7711 dep_t dep;
7712 dep_def _new_dep, *new_dep = &_new_dep;
7713 ds_t todo_spec;
7715 gcc_assert (ORIG_PAT (insn) != NULL_RTX);
7717 if (!mutate_p)
7718 todo_spec = TODO_SPEC (insn);
7719 else
7721 gcc_assert (IS_SPECULATION_SIMPLE_CHECK_P (insn)
7722 && (TODO_SPEC (insn) & SPECULATIVE) == 0);
7724 todo_spec = CHECK_SPEC (insn);
7727 todo_spec &= SPECULATIVE;
7729 /* Create recovery block. */
7730 if (mutate_p || targetm.sched.needs_block_p (todo_spec))
7732 rec = sched_create_recovery_block (NULL);
7733 label = BB_HEAD (rec);
7735 else
7737 rec = EXIT_BLOCK_PTR;
7738 label = NULL_RTX;
7741 /* Emit CHECK. */
7742 check = targetm.sched.gen_spec_check (insn, label, todo_spec);
7744 if (rec != EXIT_BLOCK_PTR)
7746 /* To have mem_reg alive at the beginning of second_bb,
7747 we emit check BEFORE insn, so insn after splitting
7748 insn will be at the beginning of second_bb, which will
7749 provide us with the correct life information. */
7750 check = emit_jump_insn_before (check, insn);
7751 JUMP_LABEL (check) = label;
7752 LABEL_NUSES (label)++;
7754 else
7755 check = emit_insn_before (check, insn);
7757 /* Extend data structures. */
7758 haifa_init_insn (check);
7760 /* CHECK is being added to current region. Extend ready list. */
7761 gcc_assert (sched_ready_n_insns != -1);
7762 sched_extend_ready_list (sched_ready_n_insns + 1);
7764 if (current_sched_info->add_remove_insn)
7765 current_sched_info->add_remove_insn (insn, 0);
7767 RECOVERY_BLOCK (check) = rec;
7769 if (sched_verbose && spec_info->dump)
7770 fprintf (spec_info->dump, ";;\t\tGenerated check insn : %s\n",
7771 (*current_sched_info->print_insn) (check, 0));
7773 gcc_assert (ORIG_PAT (insn));
7775 /* Initialize TWIN (twin is a duplicate of original instruction
7776 in the recovery block). */
7777 if (rec != EXIT_BLOCK_PTR)
7779 sd_iterator_def sd_it;
7780 dep_t dep;
7782 FOR_EACH_DEP (insn, SD_LIST_RES_BACK, sd_it, dep)
7783 if ((DEP_STATUS (dep) & DEP_OUTPUT) != 0)
7785 struct _dep _dep2, *dep2 = &_dep2;
7787 init_dep (dep2, DEP_PRO (dep), check, REG_DEP_TRUE);
7789 sd_add_dep (dep2, true);
7792 twin = emit_insn_after (ORIG_PAT (insn), BB_END (rec));
7793 haifa_init_insn (twin);
7795 if (sched_verbose && spec_info->dump)
7796 /* INSN_BB (insn) isn't determined for twin insns yet.
7797 So we can't use current_sched_info->print_insn. */
7798 fprintf (spec_info->dump, ";;\t\tGenerated twin insn : %d/rec%d\n",
7799 INSN_UID (twin), rec->index);
7801 else
7803 ORIG_PAT (check) = ORIG_PAT (insn);
7804 HAS_INTERNAL_DEP (check) = 1;
7805 twin = check;
7806 /* ??? We probably should change all OUTPUT dependencies to
7807 (TRUE | OUTPUT). */
7810 /* Copy all resolved back dependencies of INSN to TWIN. This will
7811 provide correct value for INSN_TICK (TWIN). */
7812 sd_copy_back_deps (twin, insn, true);
7814 if (rec != EXIT_BLOCK_PTR)
7815 /* In case of branchy check, fix CFG. */
7817 basic_block first_bb, second_bb;
7818 rtx jump;
7820 first_bb = BLOCK_FOR_INSN (check);
7821 second_bb = sched_split_block (first_bb, check);
7823 sched_create_recovery_edges (first_bb, rec, second_bb);
7825 sched_init_only_bb (second_bb, first_bb);
7826 sched_init_only_bb (rec, EXIT_BLOCK_PTR);
7828 jump = BB_END (rec);
7829 haifa_init_insn (jump);
7832 /* Move backward dependences from INSN to CHECK and
7833 move forward dependences from INSN to TWIN. */
7835 /* First, create dependencies between INSN's producers and CHECK & TWIN. */
7836 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
7838 rtx pro = DEP_PRO (dep);
7839 ds_t ds;
7841 /* If BEGIN_DATA: [insn ~~TRUE~~> producer]:
7842 check --TRUE--> producer ??? or ANTI ???
7843 twin --TRUE--> producer
7844 twin --ANTI--> check
7846 If BEGIN_CONTROL: [insn ~~ANTI~~> producer]:
7847 check --ANTI--> producer
7848 twin --ANTI--> producer
7849 twin --ANTI--> check
7851 If BE_IN_SPEC: [insn ~~TRUE~~> producer]:
7852 check ~~TRUE~~> producer
7853 twin ~~TRUE~~> producer
7854 twin --ANTI--> check */
7856 ds = DEP_STATUS (dep);
7858 if (ds & BEGIN_SPEC)
7860 gcc_assert (!mutate_p);
7861 ds &= ~BEGIN_SPEC;
7864 init_dep_1 (new_dep, pro, check, DEP_TYPE (dep), ds);
7865 sd_add_dep (new_dep, false);
7867 if (rec != EXIT_BLOCK_PTR)
7869 DEP_CON (new_dep) = twin;
7870 sd_add_dep (new_dep, false);
7874 /* Second, remove backward dependencies of INSN. */
7875 for (sd_it = sd_iterator_start (insn, SD_LIST_SPEC_BACK);
7876 sd_iterator_cond (&sd_it, &dep);)
7878 if ((DEP_STATUS (dep) & BEGIN_SPEC)
7879 || mutate_p)
7880 /* We can delete this dep because we overcome it with
7881 BEGIN_SPECULATION. */
7882 sd_delete_dep (sd_it);
7883 else
7884 sd_iterator_next (&sd_it);
7887 /* Future Speculations. Determine what BE_IN speculations will be like. */
7888 fs = 0;
7890 /* Fields (DONE_SPEC (x) & BEGIN_SPEC) and CHECK_SPEC (x) are set only
7891 here. */
7893 gcc_assert (!DONE_SPEC (insn));
7895 if (!mutate_p)
7897 ds_t ts = TODO_SPEC (insn);
7899 DONE_SPEC (insn) = ts & BEGIN_SPEC;
7900 CHECK_SPEC (check) = ts & BEGIN_SPEC;
7902 /* Luckiness of future speculations solely depends upon initial
7903 BEGIN speculation. */
7904 if (ts & BEGIN_DATA)
7905 fs = set_dep_weak (fs, BE_IN_DATA, get_dep_weak (ts, BEGIN_DATA));
7906 if (ts & BEGIN_CONTROL)
7907 fs = set_dep_weak (fs, BE_IN_CONTROL,
7908 get_dep_weak (ts, BEGIN_CONTROL));
7910 else
7911 CHECK_SPEC (check) = CHECK_SPEC (insn);
7913 /* Future speculations: call the helper. */
7914 process_insn_forw_deps_be_in_spec (insn, twin, fs);
7916 if (rec != EXIT_BLOCK_PTR)
7918 /* Which types of dependencies should we use here is,
7919 generally, machine-dependent question... But, for now,
7920 it is not. */
7922 if (!mutate_p)
7924 init_dep (new_dep, insn, check, REG_DEP_TRUE);
7925 sd_add_dep (new_dep, false);
7927 init_dep (new_dep, insn, twin, REG_DEP_OUTPUT);
7928 sd_add_dep (new_dep, false);
7930 else
7932 if (spec_info->dump)
7933 fprintf (spec_info->dump, ";;\t\tRemoved simple check : %s\n",
7934 (*current_sched_info->print_insn) (insn, 0));
7936 /* Remove all dependencies of the INSN. */
7938 sd_it = sd_iterator_start (insn, (SD_LIST_FORW
7939 | SD_LIST_BACK
7940 | SD_LIST_RES_BACK));
7941 while (sd_iterator_cond (&sd_it, &dep))
7942 sd_delete_dep (sd_it);
7945 /* If former check (INSN) already was moved to the ready (or queue)
7946 list, add new check (CHECK) there too. */
7947 if (QUEUE_INDEX (insn) != QUEUE_NOWHERE)
7948 try_ready (check);
7950 /* Remove old check from instruction stream and free its
7951 data. */
7952 sched_remove_insn (insn);
7955 init_dep (new_dep, check, twin, REG_DEP_ANTI);
7956 sd_add_dep (new_dep, false);
7958 else
7960 init_dep_1 (new_dep, insn, check, REG_DEP_TRUE, DEP_TRUE | DEP_OUTPUT);
7961 sd_add_dep (new_dep, false);
7964 if (!mutate_p)
7965 /* Fix priorities. If MUTATE_P is nonzero, this is not necessary,
7966 because it'll be done later in add_to_speculative_block. */
7968 rtx_vec_t priorities_roots = rtx_vec_t ();
7970 clear_priorities (twin, &priorities_roots);
7971 calc_priorities (priorities_roots);
7972 priorities_roots.release ();
7976 /* Removes dependency between instructions in the recovery block REC
7977 and usual region instructions. It keeps inner dependences so it
7978 won't be necessary to recompute them. */
7979 static void
7980 fix_recovery_deps (basic_block rec)
7982 rtx note, insn, jump, ready_list = 0;
7983 bitmap_head in_ready;
7984 rtx link;
7986 bitmap_initialize (&in_ready, 0);
7988 /* NOTE - a basic block note. */
7989 note = NEXT_INSN (BB_HEAD (rec));
7990 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
7991 insn = BB_END (rec);
7992 gcc_assert (JUMP_P (insn));
7993 insn = PREV_INSN (insn);
7997 sd_iterator_def sd_it;
7998 dep_t dep;
8000 for (sd_it = sd_iterator_start (insn, SD_LIST_FORW);
8001 sd_iterator_cond (&sd_it, &dep);)
8003 rtx consumer = DEP_CON (dep);
8005 if (BLOCK_FOR_INSN (consumer) != rec)
8007 sd_delete_dep (sd_it);
8009 if (bitmap_set_bit (&in_ready, INSN_LUID (consumer)))
8010 ready_list = alloc_INSN_LIST (consumer, ready_list);
8012 else
8014 gcc_assert ((DEP_STATUS (dep) & DEP_TYPES) == DEP_TRUE);
8016 sd_iterator_next (&sd_it);
8020 insn = PREV_INSN (insn);
8022 while (insn != note);
8024 bitmap_clear (&in_ready);
8026 /* Try to add instructions to the ready or queue list. */
8027 for (link = ready_list; link; link = XEXP (link, 1))
8028 try_ready (XEXP (link, 0));
8029 free_INSN_LIST_list (&ready_list);
8031 /* Fixing jump's dependences. */
8032 insn = BB_HEAD (rec);
8033 jump = BB_END (rec);
8035 gcc_assert (LABEL_P (insn));
8036 insn = NEXT_INSN (insn);
8038 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (insn));
8039 add_jump_dependencies (insn, jump);
8042 /* Change pattern of INSN to NEW_PAT. Invalidate cached haifa
8043 instruction data. */
8044 static bool
8045 haifa_change_pattern (rtx insn, rtx new_pat)
8047 int t;
8049 t = validate_change (insn, &PATTERN (insn), new_pat, 0);
8050 if (!t)
8051 return false;
8053 update_insn_after_change (insn);
8054 return true;
8057 /* -1 - can't speculate,
8058 0 - for speculation with REQUEST mode it is OK to use
8059 current instruction pattern,
8060 1 - need to change pattern for *NEW_PAT to be speculative. */
8062 sched_speculate_insn (rtx insn, ds_t request, rtx *new_pat)
8064 gcc_assert (current_sched_info->flags & DO_SPECULATION
8065 && (request & SPECULATIVE)
8066 && sched_insn_is_legitimate_for_speculation_p (insn, request));
8068 if ((request & spec_info->mask) != request)
8069 return -1;
8071 if (request & BE_IN_SPEC
8072 && !(request & BEGIN_SPEC))
8073 return 0;
8075 return targetm.sched.speculate_insn (insn, request, new_pat);
8078 static int
8079 haifa_speculate_insn (rtx insn, ds_t request, rtx *new_pat)
8081 gcc_assert (sched_deps_info->generate_spec_deps
8082 && !IS_SPECULATION_CHECK_P (insn));
8084 if (HAS_INTERNAL_DEP (insn)
8085 || SCHED_GROUP_P (insn))
8086 return -1;
8088 return sched_speculate_insn (insn, request, new_pat);
8091 /* Print some information about block BB, which starts with HEAD and
8092 ends with TAIL, before scheduling it.
8093 I is zero, if scheduler is about to start with the fresh ebb. */
8094 static void
8095 dump_new_block_header (int i, basic_block bb, rtx head, rtx tail)
8097 if (!i)
8098 fprintf (sched_dump,
8099 ";; ======================================================\n");
8100 else
8101 fprintf (sched_dump,
8102 ";; =====================ADVANCING TO=====================\n");
8103 fprintf (sched_dump,
8104 ";; -- basic block %d from %d to %d -- %s reload\n",
8105 bb->index, INSN_UID (head), INSN_UID (tail),
8106 (reload_completed ? "after" : "before"));
8107 fprintf (sched_dump,
8108 ";; ======================================================\n");
8109 fprintf (sched_dump, "\n");
8112 /* Unlink basic block notes and labels and saves them, so they
8113 can be easily restored. We unlink basic block notes in EBB to
8114 provide back-compatibility with the previous code, as target backends
8115 assume, that there'll be only instructions between
8116 current_sched_info->{head and tail}. We restore these notes as soon
8117 as we can.
8118 FIRST (LAST) is the first (last) basic block in the ebb.
8119 NB: In usual case (FIRST == LAST) nothing is really done. */
8120 void
8121 unlink_bb_notes (basic_block first, basic_block last)
8123 /* We DON'T unlink basic block notes of the first block in the ebb. */
8124 if (first == last)
8125 return;
8127 bb_header = XNEWVEC (rtx, last_basic_block);
8129 /* Make a sentinel. */
8130 if (last->next_bb != EXIT_BLOCK_PTR)
8131 bb_header[last->next_bb->index] = 0;
8133 first = first->next_bb;
8136 rtx prev, label, note, next;
8138 label = BB_HEAD (last);
8139 if (LABEL_P (label))
8140 note = NEXT_INSN (label);
8141 else
8142 note = label;
8143 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
8145 prev = PREV_INSN (label);
8146 next = NEXT_INSN (note);
8147 gcc_assert (prev && next);
8149 NEXT_INSN (prev) = next;
8150 PREV_INSN (next) = prev;
8152 bb_header[last->index] = label;
8154 if (last == first)
8155 break;
8157 last = last->prev_bb;
8159 while (1);
8162 /* Restore basic block notes.
8163 FIRST is the first basic block in the ebb. */
8164 static void
8165 restore_bb_notes (basic_block first)
8167 if (!bb_header)
8168 return;
8170 /* We DON'T unlink basic block notes of the first block in the ebb. */
8171 first = first->next_bb;
8172 /* Remember: FIRST is actually a second basic block in the ebb. */
8174 while (first != EXIT_BLOCK_PTR
8175 && bb_header[first->index])
8177 rtx prev, label, note, next;
8179 label = bb_header[first->index];
8180 prev = PREV_INSN (label);
8181 next = NEXT_INSN (prev);
8183 if (LABEL_P (label))
8184 note = NEXT_INSN (label);
8185 else
8186 note = label;
8187 gcc_assert (NOTE_INSN_BASIC_BLOCK_P (note));
8189 bb_header[first->index] = 0;
8191 NEXT_INSN (prev) = label;
8192 NEXT_INSN (note) = next;
8193 PREV_INSN (next) = note;
8195 first = first->next_bb;
8198 free (bb_header);
8199 bb_header = 0;
8202 /* Helper function.
8203 Fix CFG after both in- and inter-block movement of
8204 control_flow_insn_p JUMP. */
8205 static void
8206 fix_jump_move (rtx jump)
8208 basic_block bb, jump_bb, jump_bb_next;
8210 bb = BLOCK_FOR_INSN (PREV_INSN (jump));
8211 jump_bb = BLOCK_FOR_INSN (jump);
8212 jump_bb_next = jump_bb->next_bb;
8214 gcc_assert (common_sched_info->sched_pass_id == SCHED_EBB_PASS
8215 || IS_SPECULATION_BRANCHY_CHECK_P (jump));
8217 if (!NOTE_INSN_BASIC_BLOCK_P (BB_END (jump_bb_next)))
8218 /* if jump_bb_next is not empty. */
8219 BB_END (jump_bb) = BB_END (jump_bb_next);
8221 if (BB_END (bb) != PREV_INSN (jump))
8222 /* Then there are instruction after jump that should be placed
8223 to jump_bb_next. */
8224 BB_END (jump_bb_next) = BB_END (bb);
8225 else
8226 /* Otherwise jump_bb_next is empty. */
8227 BB_END (jump_bb_next) = NEXT_INSN (BB_HEAD (jump_bb_next));
8229 /* To make assertion in move_insn happy. */
8230 BB_END (bb) = PREV_INSN (jump);
8232 update_bb_for_insn (jump_bb_next);
8235 /* Fix CFG after interblock movement of control_flow_insn_p JUMP. */
8236 static void
8237 move_block_after_check (rtx jump)
8239 basic_block bb, jump_bb, jump_bb_next;
8240 vec<edge, va_gc> *t;
8242 bb = BLOCK_FOR_INSN (PREV_INSN (jump));
8243 jump_bb = BLOCK_FOR_INSN (jump);
8244 jump_bb_next = jump_bb->next_bb;
8246 update_bb_for_insn (jump_bb);
8248 gcc_assert (IS_SPECULATION_CHECK_P (jump)
8249 || IS_SPECULATION_CHECK_P (BB_END (jump_bb_next)));
8251 unlink_block (jump_bb_next);
8252 link_block (jump_bb_next, bb);
8254 t = bb->succs;
8255 bb->succs = 0;
8256 move_succs (&(jump_bb->succs), bb);
8257 move_succs (&(jump_bb_next->succs), jump_bb);
8258 move_succs (&t, jump_bb_next);
8260 df_mark_solutions_dirty ();
8262 common_sched_info->fix_recovery_cfg
8263 (bb->index, jump_bb->index, jump_bb_next->index);
8266 /* Helper function for move_block_after_check.
8267 This functions attaches edge vector pointed to by SUCCSP to
8268 block TO. */
8269 static void
8270 move_succs (vec<edge, va_gc> **succsp, basic_block to)
8272 edge e;
8273 edge_iterator ei;
8275 gcc_assert (to->succs == 0);
8277 to->succs = *succsp;
8279 FOR_EACH_EDGE (e, ei, to->succs)
8280 e->src = to;
8282 *succsp = 0;
8285 /* Remove INSN from the instruction stream.
8286 INSN should have any dependencies. */
8287 static void
8288 sched_remove_insn (rtx insn)
8290 sd_finish_insn (insn);
8292 change_queue_index (insn, QUEUE_NOWHERE);
8293 current_sched_info->add_remove_insn (insn, 1);
8294 delete_insn (insn);
8297 /* Clear priorities of all instructions, that are forward dependent on INSN.
8298 Store in vector pointed to by ROOTS_PTR insns on which priority () should
8299 be invoked to initialize all cleared priorities. */
8300 static void
8301 clear_priorities (rtx insn, rtx_vec_t *roots_ptr)
8303 sd_iterator_def sd_it;
8304 dep_t dep;
8305 bool insn_is_root_p = true;
8307 gcc_assert (QUEUE_INDEX (insn) != QUEUE_SCHEDULED);
8309 FOR_EACH_DEP (insn, SD_LIST_BACK, sd_it, dep)
8311 rtx pro = DEP_PRO (dep);
8313 if (INSN_PRIORITY_STATUS (pro) >= 0
8314 && QUEUE_INDEX (insn) != QUEUE_SCHEDULED)
8316 /* If DEP doesn't contribute to priority then INSN itself should
8317 be added to priority roots. */
8318 if (contributes_to_priority_p (dep))
8319 insn_is_root_p = false;
8321 INSN_PRIORITY_STATUS (pro) = -1;
8322 clear_priorities (pro, roots_ptr);
8326 if (insn_is_root_p)
8327 roots_ptr->safe_push (insn);
8330 /* Recompute priorities of instructions, whose priorities might have been
8331 changed. ROOTS is a vector of instructions whose priority computation will
8332 trigger initialization of all cleared priorities. */
8333 static void
8334 calc_priorities (rtx_vec_t roots)
8336 int i;
8337 rtx insn;
8339 FOR_EACH_VEC_ELT (roots, i, insn)
8340 priority (insn);
8344 /* Add dependences between JUMP and other instructions in the recovery
8345 block. INSN is the first insn the recovery block. */
8346 static void
8347 add_jump_dependencies (rtx insn, rtx jump)
8351 insn = NEXT_INSN (insn);
8352 if (insn == jump)
8353 break;
8355 if (dep_list_size (insn, SD_LIST_FORW) == 0)
8357 dep_def _new_dep, *new_dep = &_new_dep;
8359 init_dep (new_dep, insn, jump, REG_DEP_ANTI);
8360 sd_add_dep (new_dep, false);
8363 while (1);
8365 gcc_assert (!sd_lists_empty_p (jump, SD_LIST_BACK));
8368 /* Extend data structures for logical insn UID. */
8369 void
8370 sched_extend_luids (void)
8372 int new_luids_max_uid = get_max_uid () + 1;
8374 sched_luids.safe_grow_cleared (new_luids_max_uid);
8377 /* Initialize LUID for INSN. */
8378 void
8379 sched_init_insn_luid (rtx insn)
8381 int i = INSN_P (insn) ? 1 : common_sched_info->luid_for_non_insn (insn);
8382 int luid;
8384 if (i >= 0)
8386 luid = sched_max_luid;
8387 sched_max_luid += i;
8389 else
8390 luid = -1;
8392 SET_INSN_LUID (insn, luid);
8395 /* Initialize luids for BBS.
8396 The hook common_sched_info->luid_for_non_insn () is used to determine
8397 if notes, labels, etc. need luids. */
8398 void
8399 sched_init_luids (bb_vec_t bbs)
8401 int i;
8402 basic_block bb;
8404 sched_extend_luids ();
8405 FOR_EACH_VEC_ELT (bbs, i, bb)
8407 rtx insn;
8409 FOR_BB_INSNS (bb, insn)
8410 sched_init_insn_luid (insn);
8414 /* Free LUIDs. */
8415 void
8416 sched_finish_luids (void)
8418 sched_luids.release ();
8419 sched_max_luid = 1;
8422 /* Return logical uid of INSN. Helpful while debugging. */
8424 insn_luid (rtx insn)
8426 return INSN_LUID (insn);
8429 /* Extend per insn data in the target. */
8430 void
8431 sched_extend_target (void)
8433 if (targetm.sched.h_i_d_extended)
8434 targetm.sched.h_i_d_extended ();
8437 /* Extend global scheduler structures (those, that live across calls to
8438 schedule_block) to include information about just emitted INSN. */
8439 static void
8440 extend_h_i_d (void)
8442 int reserve = (get_max_uid () + 1 - h_i_d.length ());
8443 if (reserve > 0
8444 && ! h_i_d.space (reserve))
8446 h_i_d.safe_grow_cleared (3 * get_max_uid () / 2);
8447 sched_extend_target ();
8451 /* Initialize h_i_d entry of the INSN with default values.
8452 Values, that are not explicitly initialized here, hold zero. */
8453 static void
8454 init_h_i_d (rtx insn)
8456 if (INSN_LUID (insn) > 0)
8458 INSN_COST (insn) = -1;
8459 QUEUE_INDEX (insn) = QUEUE_NOWHERE;
8460 INSN_TICK (insn) = INVALID_TICK;
8461 INSN_EXACT_TICK (insn) = INVALID_TICK;
8462 INTER_TICK (insn) = INVALID_TICK;
8463 TODO_SPEC (insn) = HARD_DEP;
8467 /* Initialize haifa_insn_data for BBS. */
8468 void
8469 haifa_init_h_i_d (bb_vec_t bbs)
8471 int i;
8472 basic_block bb;
8474 extend_h_i_d ();
8475 FOR_EACH_VEC_ELT (bbs, i, bb)
8477 rtx insn;
8479 FOR_BB_INSNS (bb, insn)
8480 init_h_i_d (insn);
8484 /* Finalize haifa_insn_data. */
8485 void
8486 haifa_finish_h_i_d (void)
8488 int i;
8489 haifa_insn_data_t data;
8490 struct reg_use_data *use, *next;
8492 FOR_EACH_VEC_ELT (h_i_d, i, data)
8494 free (data->max_reg_pressure);
8495 free (data->reg_pressure);
8496 for (use = data->reg_use_list; use != NULL; use = next)
8498 next = use->next_insn_use;
8499 free (use);
8502 h_i_d.release ();
8505 /* Init data for the new insn INSN. */
8506 static void
8507 haifa_init_insn (rtx insn)
8509 gcc_assert (insn != NULL);
8511 sched_extend_luids ();
8512 sched_init_insn_luid (insn);
8513 sched_extend_target ();
8514 sched_deps_init (false);
8515 extend_h_i_d ();
8516 init_h_i_d (insn);
8518 if (adding_bb_to_current_region_p)
8520 sd_init_insn (insn);
8522 /* Extend dependency caches by one element. */
8523 extend_dependency_caches (1, false);
8525 if (sched_pressure != SCHED_PRESSURE_NONE)
8526 init_insn_reg_pressure_info (insn);
8529 /* Init data for the new basic block BB which comes after AFTER. */
8530 static void
8531 haifa_init_only_bb (basic_block bb, basic_block after)
8533 gcc_assert (bb != NULL);
8535 sched_init_bbs ();
8537 if (common_sched_info->add_block)
8538 /* This changes only data structures of the front-end. */
8539 common_sched_info->add_block (bb, after);
8542 /* A generic version of sched_split_block (). */
8543 basic_block
8544 sched_split_block_1 (basic_block first_bb, rtx after)
8546 edge e;
8548 e = split_block (first_bb, after);
8549 gcc_assert (e->src == first_bb);
8551 /* sched_split_block emits note if *check == BB_END. Probably it
8552 is better to rip that note off. */
8554 return e->dest;
8557 /* A generic version of sched_create_empty_bb (). */
8558 basic_block
8559 sched_create_empty_bb_1 (basic_block after)
8561 return create_empty_bb (after);
8564 /* Insert PAT as an INSN into the schedule and update the necessary data
8565 structures to account for it. */
8567 sched_emit_insn (rtx pat)
8569 rtx insn = emit_insn_before (pat, nonscheduled_insns_begin);
8570 haifa_init_insn (insn);
8572 if (current_sched_info->add_remove_insn)
8573 current_sched_info->add_remove_insn (insn, 0);
8575 (*current_sched_info->begin_schedule_ready) (insn);
8576 scheduled_insns.safe_push (insn);
8578 last_scheduled_insn = insn;
8579 return insn;
8582 /* This function returns a candidate satisfying dispatch constraints from
8583 the ready list. */
8585 static rtx
8586 ready_remove_first_dispatch (struct ready_list *ready)
8588 int i;
8589 rtx insn = ready_element (ready, 0);
8591 if (ready->n_ready == 1
8592 || !INSN_P (insn)
8593 || INSN_CODE (insn) < 0
8594 || !active_insn_p (insn)
8595 || targetm.sched.dispatch (insn, FITS_DISPATCH_WINDOW))
8596 return ready_remove_first (ready);
8598 for (i = 1; i < ready->n_ready; i++)
8600 insn = ready_element (ready, i);
8602 if (!INSN_P (insn)
8603 || INSN_CODE (insn) < 0
8604 || !active_insn_p (insn))
8605 continue;
8607 if (targetm.sched.dispatch (insn, FITS_DISPATCH_WINDOW))
8609 /* Return ith element of ready. */
8610 insn = ready_remove (ready, i);
8611 return insn;
8615 if (targetm.sched.dispatch (NULL_RTX, DISPATCH_VIOLATION))
8616 return ready_remove_first (ready);
8618 for (i = 1; i < ready->n_ready; i++)
8620 insn = ready_element (ready, i);
8622 if (!INSN_P (insn)
8623 || INSN_CODE (insn) < 0
8624 || !active_insn_p (insn))
8625 continue;
8627 /* Return i-th element of ready. */
8628 if (targetm.sched.dispatch (insn, IS_CMP))
8629 return ready_remove (ready, i);
8632 return ready_remove_first (ready);
8635 /* Get number of ready insn in the ready list. */
8638 number_in_ready (void)
8640 return ready.n_ready;
8643 /* Get number of ready's in the ready list. */
8646 get_ready_element (int i)
8648 return ready_element (&ready, i);
8651 #endif /* INSN_SCHEDULING */