re PR target/55981 (std::atomic store is split in two smaller stores)
[gcc.git] / gcc / sel-sched-ir.c
1 /* Instruction scheduling pass. Selective scheduler and pipeliner.
2 Copyright (C) 2006-2013 Free Software Foundation, Inc.
3
4 This file is part of GCC.
5
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
10
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
15
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
19
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "tm.h"
24 #include "diagnostic-core.h"
25 #include "rtl.h"
26 #include "tm_p.h"
27 #include "hard-reg-set.h"
28 #include "regs.h"
29 #include "function.h"
30 #include "flags.h"
31 #include "insn-config.h"
32 #include "insn-attr.h"
33 #include "except.h"
34 #include "recog.h"
35 #include "params.h"
36 #include "target.h"
37 #include "sched-int.h"
38 #include "ggc.h"
39 #include "tree.h"
40 #include "vec.h"
41 #include "langhooks.h"
42 #include "rtlhooks-def.h"
43 #include "emit-rtl.h" /* FIXME: Can go away once crtl is moved to rtl.h. */
44
45 #ifdef INSN_SCHEDULING
46 #include "sel-sched-ir.h"
47 /* We don't have to use it except for sel_print_insn. */
48 #include "sel-sched-dump.h"
49
50 /* A vector holding bb info for whole scheduling pass. */
51 vec<sel_global_bb_info_def>
52 sel_global_bb_info = vNULL;
53
54 /* A vector holding bb info. */
55 vec<sel_region_bb_info_def>
56 sel_region_bb_info = vNULL;
57
58 /* A pool for allocating all lists. */
59 alloc_pool sched_lists_pool;
60
61 /* This contains information about successors for compute_av_set. */
62 struct succs_info current_succs;
63
64 /* Data structure to describe interaction with the generic scheduler utils. */
65 static struct common_sched_info_def sel_common_sched_info;
66
67 /* The loop nest being pipelined. */
68 struct loop *current_loop_nest;
69
70 /* LOOP_NESTS is a vector containing the corresponding loop nest for
71 each region. */
72 static vec<loop_p> loop_nests = vNULL;
73
74 /* Saves blocks already in loop regions, indexed by bb->index. */
75 static sbitmap bbs_in_loop_rgns = NULL;
76
77 /* CFG hooks that are saved before changing create_basic_block hook. */
78 static struct cfg_hooks orig_cfg_hooks;
79 \f
80
81 /* Array containing reverse topological index of function basic blocks,
82 indexed by BB->INDEX. */
83 static int *rev_top_order_index = NULL;
84
85 /* Length of the above array. */
86 static int rev_top_order_index_len = -1;
87
88 /* A regset pool structure. */
89 static struct
90 {
91 /* The stack to which regsets are returned. */
92 regset *v;
93
94 /* Its pointer. */
95 int n;
96
97 /* Its size. */
98 int s;
99
100 /* In VV we save all generated regsets so that, when destructing the
101 pool, we can compare it with V and check that every regset was returned
102 back to pool. */
103 regset *vv;
104
105 /* The pointer of VV stack. */
106 int nn;
107
108 /* Its size. */
109 int ss;
110
111 /* The difference between allocated and returned regsets. */
112 int diff;
113 } regset_pool = { NULL, 0, 0, NULL, 0, 0, 0 };
114
115 /* This represents the nop pool. */
116 static struct
117 {
118 /* The vector which holds previously emitted nops. */
119 insn_t *v;
120
121 /* Its pointer. */
122 int n;
123
124 /* Its size. */
125 int s;
126 } nop_pool = { NULL, 0, 0 };
127
128 /* The pool for basic block notes. */
129 static rtx_vec_t bb_note_pool;
130
131 /* A NOP pattern used to emit placeholder insns. */
132 rtx nop_pattern = NULL_RTX;
133 /* A special instruction that resides in EXIT_BLOCK.
134 EXIT_INSN is successor of the insns that lead to EXIT_BLOCK. */
135 rtx exit_insn = NULL_RTX;
136
137 /* TRUE if while scheduling current region, which is loop, its preheader
138 was removed. */
139 bool preheader_removed = false;
140 \f
141
142 /* Forward static declarations. */
143 static void fence_clear (fence_t);
144
145 static void deps_init_id (idata_t, insn_t, bool);
146 static void init_id_from_df (idata_t, insn_t, bool);
147 static expr_t set_insn_init (expr_t, vinsn_t, int);
148
149 static void cfg_preds (basic_block, insn_t **, int *);
150 static void prepare_insn_expr (insn_t, int);
151 static void free_history_vect (vec<expr_history_def> &);
152
153 static void move_bb_info (basic_block, basic_block);
154 static void remove_empty_bb (basic_block, bool);
155 static void sel_merge_blocks (basic_block, basic_block);
156 static void sel_remove_loop_preheader (void);
157 static bool bb_has_removable_jump_to_p (basic_block, basic_block);
158
159 static bool insn_is_the_only_one_in_bb_p (insn_t);
160 static void create_initial_data_sets (basic_block);
161
162 static void free_av_set (basic_block);
163 static void invalidate_av_set (basic_block);
164 static void extend_insn_data (void);
165 static void sel_init_new_insn (insn_t, int);
166 static void finish_insns (void);
167 \f
168 /* Various list functions. */
169
170 /* Copy an instruction list L. */
171 ilist_t
172 ilist_copy (ilist_t l)
173 {
174 ilist_t head = NULL, *tailp = &head;
175
176 while (l)
177 {
178 ilist_add (tailp, ILIST_INSN (l));
179 tailp = &ILIST_NEXT (*tailp);
180 l = ILIST_NEXT (l);
181 }
182
183 return head;
184 }
185
186 /* Invert an instruction list L. */
187 ilist_t
188 ilist_invert (ilist_t l)
189 {
190 ilist_t res = NULL;
191
192 while (l)
193 {
194 ilist_add (&res, ILIST_INSN (l));
195 l = ILIST_NEXT (l);
196 }
197
198 return res;
199 }
200
201 /* Add a new boundary to the LP list with parameters TO, PTR, and DC. */
202 void
203 blist_add (blist_t *lp, insn_t to, ilist_t ptr, deps_t dc)
204 {
205 bnd_t bnd;
206
207 _list_add (lp);
208 bnd = BLIST_BND (*lp);
209
210 BND_TO (bnd) = to;
211 BND_PTR (bnd) = ptr;
212 BND_AV (bnd) = NULL;
213 BND_AV1 (bnd) = NULL;
214 BND_DC (bnd) = dc;
215 }
216
217 /* Remove the list note pointed to by LP. */
218 void
219 blist_remove (blist_t *lp)
220 {
221 bnd_t b = BLIST_BND (*lp);
222
223 av_set_clear (&BND_AV (b));
224 av_set_clear (&BND_AV1 (b));
225 ilist_clear (&BND_PTR (b));
226
227 _list_remove (lp);
228 }
229
230 /* Init a fence tail L. */
231 void
232 flist_tail_init (flist_tail_t l)
233 {
234 FLIST_TAIL_HEAD (l) = NULL;
235 FLIST_TAIL_TAILP (l) = &FLIST_TAIL_HEAD (l);
236 }
237
238 /* Try to find fence corresponding to INSN in L. */
239 fence_t
240 flist_lookup (flist_t l, insn_t insn)
241 {
242 while (l)
243 {
244 if (FENCE_INSN (FLIST_FENCE (l)) == insn)
245 return FLIST_FENCE (l);
246
247 l = FLIST_NEXT (l);
248 }
249
250 return NULL;
251 }
252
253 /* Init the fields of F before running fill_insns. */
254 static void
255 init_fence_for_scheduling (fence_t f)
256 {
257 FENCE_BNDS (f) = NULL;
258 FENCE_PROCESSED_P (f) = false;
259 FENCE_SCHEDULED_P (f) = false;
260 }
261
262 /* Add new fence consisting of INSN and STATE to the list pointed to by LP. */
263 static void
264 flist_add (flist_t *lp, insn_t insn, state_t state, deps_t dc, void *tc,
265 insn_t last_scheduled_insn, vec<rtx, va_gc> *executing_insns,
266 int *ready_ticks, int ready_ticks_size, insn_t sched_next,
267 int cycle, int cycle_issued_insns, int issue_more,
268 bool starts_cycle_p, bool after_stall_p)
269 {
270 fence_t f;
271
272 _list_add (lp);
273 f = FLIST_FENCE (*lp);
274
275 FENCE_INSN (f) = insn;
276
277 gcc_assert (state != NULL);
278 FENCE_STATE (f) = state;
279
280 FENCE_CYCLE (f) = cycle;
281 FENCE_ISSUED_INSNS (f) = cycle_issued_insns;
282 FENCE_STARTS_CYCLE_P (f) = starts_cycle_p;
283 FENCE_AFTER_STALL_P (f) = after_stall_p;
284
285 gcc_assert (dc != NULL);
286 FENCE_DC (f) = dc;
287
288 gcc_assert (tc != NULL || targetm.sched.alloc_sched_context == NULL);
289 FENCE_TC (f) = tc;
290
291 FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn;
292 FENCE_ISSUE_MORE (f) = issue_more;
293 FENCE_EXECUTING_INSNS (f) = executing_insns;
294 FENCE_READY_TICKS (f) = ready_ticks;
295 FENCE_READY_TICKS_SIZE (f) = ready_ticks_size;
296 FENCE_SCHED_NEXT (f) = sched_next;
297
298 init_fence_for_scheduling (f);
299 }
300
301 /* Remove the head node of the list pointed to by LP. */
302 static void
303 flist_remove (flist_t *lp)
304 {
305 if (FENCE_INSN (FLIST_FENCE (*lp)))
306 fence_clear (FLIST_FENCE (*lp));
307 _list_remove (lp);
308 }
309
310 /* Clear the fence list pointed to by LP. */
311 void
312 flist_clear (flist_t *lp)
313 {
314 while (*lp)
315 flist_remove (lp);
316 }
317
318 /* Add ORIGINAL_INSN the def list DL honoring CROSSES_CALL. */
319 void
320 def_list_add (def_list_t *dl, insn_t original_insn, bool crosses_call)
321 {
322 def_t d;
323
324 _list_add (dl);
325 d = DEF_LIST_DEF (*dl);
326
327 d->orig_insn = original_insn;
328 d->crosses_call = crosses_call;
329 }
330 \f
331
332 /* Functions to work with target contexts. */
333
334 /* Bulk target context. It is convenient for debugging purposes to ensure
335 that there are no uninitialized (null) target contexts. */
336 static tc_t bulk_tc = (tc_t) 1;
337
338 /* Target hooks wrappers. In the future we can provide some default
339 implementations for them. */
340
341 /* Allocate a store for the target context. */
342 static tc_t
343 alloc_target_context (void)
344 {
345 return (targetm.sched.alloc_sched_context
346 ? targetm.sched.alloc_sched_context () : bulk_tc);
347 }
348
349 /* Init target context TC.
350 If CLEAN_P is true, then make TC as it is beginning of the scheduler.
351 Overwise, copy current backend context to TC. */
352 static void
353 init_target_context (tc_t tc, bool clean_p)
354 {
355 if (targetm.sched.init_sched_context)
356 targetm.sched.init_sched_context (tc, clean_p);
357 }
358
359 /* Allocate and initialize a target context. Meaning of CLEAN_P is the same as
360 int init_target_context (). */
361 tc_t
362 create_target_context (bool clean_p)
363 {
364 tc_t tc = alloc_target_context ();
365
366 init_target_context (tc, clean_p);
367 return tc;
368 }
369
370 /* Copy TC to the current backend context. */
371 void
372 set_target_context (tc_t tc)
373 {
374 if (targetm.sched.set_sched_context)
375 targetm.sched.set_sched_context (tc);
376 }
377
378 /* TC is about to be destroyed. Free any internal data. */
379 static void
380 clear_target_context (tc_t tc)
381 {
382 if (targetm.sched.clear_sched_context)
383 targetm.sched.clear_sched_context (tc);
384 }
385
386 /* Clear and free it. */
387 static void
388 delete_target_context (tc_t tc)
389 {
390 clear_target_context (tc);
391
392 if (targetm.sched.free_sched_context)
393 targetm.sched.free_sched_context (tc);
394 }
395
396 /* Make a copy of FROM in TO.
397 NB: May be this should be a hook. */
398 static void
399 copy_target_context (tc_t to, tc_t from)
400 {
401 tc_t tmp = create_target_context (false);
402
403 set_target_context (from);
404 init_target_context (to, false);
405
406 set_target_context (tmp);
407 delete_target_context (tmp);
408 }
409
410 /* Create a copy of TC. */
411 static tc_t
412 create_copy_of_target_context (tc_t tc)
413 {
414 tc_t copy = alloc_target_context ();
415
416 copy_target_context (copy, tc);
417
418 return copy;
419 }
420
421 /* Clear TC and initialize it according to CLEAN_P. The meaning of CLEAN_P
422 is the same as in init_target_context (). */
423 void
424 reset_target_context (tc_t tc, bool clean_p)
425 {
426 clear_target_context (tc);
427 init_target_context (tc, clean_p);
428 }
429 \f
430 /* Functions to work with dependence contexts.
431 Dc (aka deps context, aka deps_t, aka struct deps_desc *) is short for dependence
432 context. It accumulates information about processed insns to decide if
433 current insn is dependent on the processed ones. */
434
435 /* Make a copy of FROM in TO. */
436 static void
437 copy_deps_context (deps_t to, deps_t from)
438 {
439 init_deps (to, false);
440 deps_join (to, from);
441 }
442
443 /* Allocate store for dep context. */
444 static deps_t
445 alloc_deps_context (void)
446 {
447 return XNEW (struct deps_desc);
448 }
449
450 /* Allocate and initialize dep context. */
451 static deps_t
452 create_deps_context (void)
453 {
454 deps_t dc = alloc_deps_context ();
455
456 init_deps (dc, false);
457 return dc;
458 }
459
460 /* Create a copy of FROM. */
461 static deps_t
462 create_copy_of_deps_context (deps_t from)
463 {
464 deps_t to = alloc_deps_context ();
465
466 copy_deps_context (to, from);
467 return to;
468 }
469
470 /* Clean up internal data of DC. */
471 static void
472 clear_deps_context (deps_t dc)
473 {
474 free_deps (dc);
475 }
476
477 /* Clear and free DC. */
478 static void
479 delete_deps_context (deps_t dc)
480 {
481 clear_deps_context (dc);
482 free (dc);
483 }
484
485 /* Clear and init DC. */
486 static void
487 reset_deps_context (deps_t dc)
488 {
489 clear_deps_context (dc);
490 init_deps (dc, false);
491 }
492
493 /* This structure describes the dependence analysis hooks for advancing
494 dependence context. */
495 static struct sched_deps_info_def advance_deps_context_sched_deps_info =
496 {
497 NULL,
498
499 NULL, /* start_insn */
500 NULL, /* finish_insn */
501 NULL, /* start_lhs */
502 NULL, /* finish_lhs */
503 NULL, /* start_rhs */
504 NULL, /* finish_rhs */
505 haifa_note_reg_set,
506 haifa_note_reg_clobber,
507 haifa_note_reg_use,
508 NULL, /* note_mem_dep */
509 NULL, /* note_dep */
510
511 0, 0, 0
512 };
513
514 /* Process INSN and add its impact on DC. */
515 void
516 advance_deps_context (deps_t dc, insn_t insn)
517 {
518 sched_deps_info = &advance_deps_context_sched_deps_info;
519 deps_analyze_insn (dc, insn);
520 }
521 \f
522
523 /* Functions to work with DFA states. */
524
525 /* Allocate store for a DFA state. */
526 static state_t
527 state_alloc (void)
528 {
529 return xmalloc (dfa_state_size);
530 }
531
532 /* Allocate and initialize DFA state. */
533 static state_t
534 state_create (void)
535 {
536 state_t state = state_alloc ();
537
538 state_reset (state);
539 advance_state (state);
540 return state;
541 }
542
543 /* Free DFA state. */
544 static void
545 state_free (state_t state)
546 {
547 free (state);
548 }
549
550 /* Make a copy of FROM in TO. */
551 static void
552 state_copy (state_t to, state_t from)
553 {
554 memcpy (to, from, dfa_state_size);
555 }
556
557 /* Create a copy of FROM. */
558 static state_t
559 state_create_copy (state_t from)
560 {
561 state_t to = state_alloc ();
562
563 state_copy (to, from);
564 return to;
565 }
566 \f
567
568 /* Functions to work with fences. */
569
570 /* Clear the fence. */
571 static void
572 fence_clear (fence_t f)
573 {
574 state_t s = FENCE_STATE (f);
575 deps_t dc = FENCE_DC (f);
576 void *tc = FENCE_TC (f);
577
578 ilist_clear (&FENCE_BNDS (f));
579
580 gcc_assert ((s != NULL && dc != NULL && tc != NULL)
581 || (s == NULL && dc == NULL && tc == NULL));
582
583 free (s);
584
585 if (dc != NULL)
586 delete_deps_context (dc);
587
588 if (tc != NULL)
589 delete_target_context (tc);
590 vec_free (FENCE_EXECUTING_INSNS (f));
591 free (FENCE_READY_TICKS (f));
592 FENCE_READY_TICKS (f) = NULL;
593 }
594
595 /* Init a list of fences with successors of OLD_FENCE. */
596 void
597 init_fences (insn_t old_fence)
598 {
599 insn_t succ;
600 succ_iterator si;
601 bool first = true;
602 int ready_ticks_size = get_max_uid () + 1;
603
604 FOR_EACH_SUCC_1 (succ, si, old_fence,
605 SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS)
606 {
607
608 if (first)
609 first = false;
610 else
611 gcc_assert (flag_sel_sched_pipelining_outer_loops);
612
613 flist_add (&fences, succ,
614 state_create (),
615 create_deps_context () /* dc */,
616 create_target_context (true) /* tc */,
617 NULL_RTX /* last_scheduled_insn */,
618 NULL, /* executing_insns */
619 XCNEWVEC (int, ready_ticks_size), /* ready_ticks */
620 ready_ticks_size,
621 NULL_RTX /* sched_next */,
622 1 /* cycle */, 0 /* cycle_issued_insns */,
623 issue_rate, /* issue_more */
624 1 /* starts_cycle_p */, 0 /* after_stall_p */);
625 }
626 }
627
628 /* Merges two fences (filling fields of fence F with resulting values) by
629 following rules: 1) state, target context and last scheduled insn are
630 propagated from fallthrough edge if it is available;
631 2) deps context and cycle is propagated from more probable edge;
632 3) all other fields are set to corresponding constant values.
633
634 INSN, STATE, DC, TC, LAST_SCHEDULED_INSN, EXECUTING_INSNS,
635 READY_TICKS, READY_TICKS_SIZE, SCHED_NEXT, CYCLE, ISSUE_MORE
636 and AFTER_STALL_P are the corresponding fields of the second fence. */
637 static void
638 merge_fences (fence_t f, insn_t insn,
639 state_t state, deps_t dc, void *tc,
640 rtx last_scheduled_insn, vec<rtx, va_gc> *executing_insns,
641 int *ready_ticks, int ready_ticks_size,
642 rtx sched_next, int cycle, int issue_more, bool after_stall_p)
643 {
644 insn_t last_scheduled_insn_old = FENCE_LAST_SCHEDULED_INSN (f);
645
646 gcc_assert (sel_bb_head_p (FENCE_INSN (f))
647 && !sched_next && !FENCE_SCHED_NEXT (f));
648
649 /* Check if we can decide which path fences came.
650 If we can't (or don't want to) - reset all. */
651 if (last_scheduled_insn == NULL
652 || last_scheduled_insn_old == NULL
653 /* This is a case when INSN is reachable on several paths from
654 one insn (this can happen when pipelining of outer loops is on and
655 there are two edges: one going around of inner loop and the other -
656 right through it; in such case just reset everything). */
657 || last_scheduled_insn == last_scheduled_insn_old)
658 {
659 state_reset (FENCE_STATE (f));
660 state_free (state);
661
662 reset_deps_context (FENCE_DC (f));
663 delete_deps_context (dc);
664
665 reset_target_context (FENCE_TC (f), true);
666 delete_target_context (tc);
667
668 if (cycle > FENCE_CYCLE (f))
669 FENCE_CYCLE (f) = cycle;
670
671 FENCE_LAST_SCHEDULED_INSN (f) = NULL;
672 FENCE_ISSUE_MORE (f) = issue_rate;
673 vec_free (executing_insns);
674 free (ready_ticks);
675 if (FENCE_EXECUTING_INSNS (f))
676 FENCE_EXECUTING_INSNS (f)->block_remove (0,
677 FENCE_EXECUTING_INSNS (f)->length ());
678 if (FENCE_READY_TICKS (f))
679 memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f));
680 }
681 else
682 {
683 edge edge_old = NULL, edge_new = NULL;
684 edge candidate;
685 succ_iterator si;
686 insn_t succ;
687
688 /* Find fallthrough edge. */
689 gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb);
690 candidate = find_fallthru_edge_from (BLOCK_FOR_INSN (insn)->prev_bb);
691
692 if (!candidate
693 || (candidate->src != BLOCK_FOR_INSN (last_scheduled_insn)
694 && candidate->src != BLOCK_FOR_INSN (last_scheduled_insn_old)))
695 {
696 /* No fallthrough edge leading to basic block of INSN. */
697 state_reset (FENCE_STATE (f));
698 state_free (state);
699
700 reset_target_context (FENCE_TC (f), true);
701 delete_target_context (tc);
702
703 FENCE_LAST_SCHEDULED_INSN (f) = NULL;
704 FENCE_ISSUE_MORE (f) = issue_rate;
705 }
706 else
707 if (candidate->src == BLOCK_FOR_INSN (last_scheduled_insn))
708 {
709 /* Would be weird if same insn is successor of several fallthrough
710 edges. */
711 gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb
712 != BLOCK_FOR_INSN (last_scheduled_insn_old));
713
714 state_free (FENCE_STATE (f));
715 FENCE_STATE (f) = state;
716
717 delete_target_context (FENCE_TC (f));
718 FENCE_TC (f) = tc;
719
720 FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn;
721 FENCE_ISSUE_MORE (f) = issue_more;
722 }
723 else
724 {
725 /* Leave STATE, TC and LAST_SCHEDULED_INSN fields untouched. */
726 state_free (state);
727 delete_target_context (tc);
728
729 gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb
730 != BLOCK_FOR_INSN (last_scheduled_insn));
731 }
732
733 /* Find edge of first predecessor (last_scheduled_insn_old->insn). */
734 FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn_old,
735 SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS)
736 {
737 if (succ == insn)
738 {
739 /* No same successor allowed from several edges. */
740 gcc_assert (!edge_old);
741 edge_old = si.e1;
742 }
743 }
744 /* Find edge of second predecessor (last_scheduled_insn->insn). */
745 FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn,
746 SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS)
747 {
748 if (succ == insn)
749 {
750 /* No same successor allowed from several edges. */
751 gcc_assert (!edge_new);
752 edge_new = si.e1;
753 }
754 }
755
756 /* Check if we can choose most probable predecessor. */
757 if (edge_old == NULL || edge_new == NULL)
758 {
759 reset_deps_context (FENCE_DC (f));
760 delete_deps_context (dc);
761 vec_free (executing_insns);
762 free (ready_ticks);
763
764 FENCE_CYCLE (f) = MAX (FENCE_CYCLE (f), cycle);
765 if (FENCE_EXECUTING_INSNS (f))
766 FENCE_EXECUTING_INSNS (f)->block_remove (0,
767 FENCE_EXECUTING_INSNS (f)->length ());
768 if (FENCE_READY_TICKS (f))
769 memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f));
770 }
771 else
772 if (edge_new->probability > edge_old->probability)
773 {
774 delete_deps_context (FENCE_DC (f));
775 FENCE_DC (f) = dc;
776 vec_free (FENCE_EXECUTING_INSNS (f));
777 FENCE_EXECUTING_INSNS (f) = executing_insns;
778 free (FENCE_READY_TICKS (f));
779 FENCE_READY_TICKS (f) = ready_ticks;
780 FENCE_READY_TICKS_SIZE (f) = ready_ticks_size;
781 FENCE_CYCLE (f) = cycle;
782 }
783 else
784 {
785 /* Leave DC and CYCLE untouched. */
786 delete_deps_context (dc);
787 vec_free (executing_insns);
788 free (ready_ticks);
789 }
790 }
791
792 /* Fill remaining invariant fields. */
793 if (after_stall_p)
794 FENCE_AFTER_STALL_P (f) = 1;
795
796 FENCE_ISSUED_INSNS (f) = 0;
797 FENCE_STARTS_CYCLE_P (f) = 1;
798 FENCE_SCHED_NEXT (f) = NULL;
799 }
800
801 /* Add a new fence to NEW_FENCES list, initializing it from all
802 other parameters. */
803 static void
804 add_to_fences (flist_tail_t new_fences, insn_t insn,
805 state_t state, deps_t dc, void *tc, rtx last_scheduled_insn,
806 vec<rtx, va_gc> *executing_insns, int *ready_ticks,
807 int ready_ticks_size, rtx sched_next, int cycle,
808 int cycle_issued_insns, int issue_rate,
809 bool starts_cycle_p, bool after_stall_p)
810 {
811 fence_t f = flist_lookup (FLIST_TAIL_HEAD (new_fences), insn);
812
813 if (! f)
814 {
815 flist_add (FLIST_TAIL_TAILP (new_fences), insn, state, dc, tc,
816 last_scheduled_insn, executing_insns, ready_ticks,
817 ready_ticks_size, sched_next, cycle, cycle_issued_insns,
818 issue_rate, starts_cycle_p, after_stall_p);
819
820 FLIST_TAIL_TAILP (new_fences)
821 = &FLIST_NEXT (*FLIST_TAIL_TAILP (new_fences));
822 }
823 else
824 {
825 merge_fences (f, insn, state, dc, tc, last_scheduled_insn,
826 executing_insns, ready_ticks, ready_ticks_size,
827 sched_next, cycle, issue_rate, after_stall_p);
828 }
829 }
830
831 /* Move the first fence in the OLD_FENCES list to NEW_FENCES. */
832 void
833 move_fence_to_fences (flist_t old_fences, flist_tail_t new_fences)
834 {
835 fence_t f, old;
836 flist_t *tailp = FLIST_TAIL_TAILP (new_fences);
837
838 old = FLIST_FENCE (old_fences);
839 f = flist_lookup (FLIST_TAIL_HEAD (new_fences),
840 FENCE_INSN (FLIST_FENCE (old_fences)));
841 if (f)
842 {
843 merge_fences (f, old->insn, old->state, old->dc, old->tc,
844 old->last_scheduled_insn, old->executing_insns,
845 old->ready_ticks, old->ready_ticks_size,
846 old->sched_next, old->cycle, old->issue_more,
847 old->after_stall_p);
848 }
849 else
850 {
851 _list_add (tailp);
852 FLIST_TAIL_TAILP (new_fences) = &FLIST_NEXT (*tailp);
853 *FLIST_FENCE (*tailp) = *old;
854 init_fence_for_scheduling (FLIST_FENCE (*tailp));
855 }
856 FENCE_INSN (old) = NULL;
857 }
858
859 /* Add a new fence to NEW_FENCES list and initialize most of its data
860 as a clean one. */
861 void
862 add_clean_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence)
863 {
864 int ready_ticks_size = get_max_uid () + 1;
865
866 add_to_fences (new_fences,
867 succ, state_create (), create_deps_context (),
868 create_target_context (true),
869 NULL_RTX, NULL,
870 XCNEWVEC (int, ready_ticks_size), ready_ticks_size,
871 NULL_RTX, FENCE_CYCLE (fence) + 1,
872 0, issue_rate, 1, FENCE_AFTER_STALL_P (fence));
873 }
874
875 /* Add a new fence to NEW_FENCES list and initialize all of its data
876 from FENCE and SUCC. */
877 void
878 add_dirty_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence)
879 {
880 int * new_ready_ticks
881 = XNEWVEC (int, FENCE_READY_TICKS_SIZE (fence));
882
883 memcpy (new_ready_ticks, FENCE_READY_TICKS (fence),
884 FENCE_READY_TICKS_SIZE (fence) * sizeof (int));
885 add_to_fences (new_fences,
886 succ, state_create_copy (FENCE_STATE (fence)),
887 create_copy_of_deps_context (FENCE_DC (fence)),
888 create_copy_of_target_context (FENCE_TC (fence)),
889 FENCE_LAST_SCHEDULED_INSN (fence),
890 vec_safe_copy (FENCE_EXECUTING_INSNS (fence)),
891 new_ready_ticks,
892 FENCE_READY_TICKS_SIZE (fence),
893 FENCE_SCHED_NEXT (fence),
894 FENCE_CYCLE (fence),
895 FENCE_ISSUED_INSNS (fence),
896 FENCE_ISSUE_MORE (fence),
897 FENCE_STARTS_CYCLE_P (fence),
898 FENCE_AFTER_STALL_P (fence));
899 }
900 \f
901
902 /* Functions to work with regset and nop pools. */
903
904 /* Returns the new regset from pool. It might have some of the bits set
905 from the previous usage. */
906 regset
907 get_regset_from_pool (void)
908 {
909 regset rs;
910
911 if (regset_pool.n != 0)
912 rs = regset_pool.v[--regset_pool.n];
913 else
914 /* We need to create the regset. */
915 {
916 rs = ALLOC_REG_SET (&reg_obstack);
917
918 if (regset_pool.nn == regset_pool.ss)
919 regset_pool.vv = XRESIZEVEC (regset, regset_pool.vv,
920 (regset_pool.ss = 2 * regset_pool.ss + 1));
921 regset_pool.vv[regset_pool.nn++] = rs;
922 }
923
924 regset_pool.diff++;
925
926 return rs;
927 }
928
929 /* Same as above, but returns the empty regset. */
930 regset
931 get_clear_regset_from_pool (void)
932 {
933 regset rs = get_regset_from_pool ();
934
935 CLEAR_REG_SET (rs);
936 return rs;
937 }
938
939 /* Return regset RS to the pool for future use. */
940 void
941 return_regset_to_pool (regset rs)
942 {
943 gcc_assert (rs);
944 regset_pool.diff--;
945
946 if (regset_pool.n == regset_pool.s)
947 regset_pool.v = XRESIZEVEC (regset, regset_pool.v,
948 (regset_pool.s = 2 * regset_pool.s + 1));
949 regset_pool.v[regset_pool.n++] = rs;
950 }
951
952 #ifdef ENABLE_CHECKING
953 /* This is used as a qsort callback for sorting regset pool stacks.
954 X and XX are addresses of two regsets. They are never equal. */
955 static int
956 cmp_v_in_regset_pool (const void *x, const void *xx)
957 {
958 uintptr_t r1 = (uintptr_t) *((const regset *) x);
959 uintptr_t r2 = (uintptr_t) *((const regset *) xx);
960 if (r1 > r2)
961 return 1;
962 else if (r1 < r2)
963 return -1;
964 gcc_unreachable ();
965 }
966 #endif
967
968 /* Free the regset pool possibly checking for memory leaks. */
969 void
970 free_regset_pool (void)
971 {
972 #ifdef ENABLE_CHECKING
973 {
974 regset *v = regset_pool.v;
975 int i = 0;
976 int n = regset_pool.n;
977
978 regset *vv = regset_pool.vv;
979 int ii = 0;
980 int nn = regset_pool.nn;
981
982 int diff = 0;
983
984 gcc_assert (n <= nn);
985
986 /* Sort both vectors so it will be possible to compare them. */
987 qsort (v, n, sizeof (*v), cmp_v_in_regset_pool);
988 qsort (vv, nn, sizeof (*vv), cmp_v_in_regset_pool);
989
990 while (ii < nn)
991 {
992 if (v[i] == vv[ii])
993 i++;
994 else
995 /* VV[II] was lost. */
996 diff++;
997
998 ii++;
999 }
1000
1001 gcc_assert (diff == regset_pool.diff);
1002 }
1003 #endif
1004
1005 /* If not true - we have a memory leak. */
1006 gcc_assert (regset_pool.diff == 0);
1007
1008 while (regset_pool.n)
1009 {
1010 --regset_pool.n;
1011 FREE_REG_SET (regset_pool.v[regset_pool.n]);
1012 }
1013
1014 free (regset_pool.v);
1015 regset_pool.v = NULL;
1016 regset_pool.s = 0;
1017
1018 free (regset_pool.vv);
1019 regset_pool.vv = NULL;
1020 regset_pool.nn = 0;
1021 regset_pool.ss = 0;
1022
1023 regset_pool.diff = 0;
1024 }
1025 \f
1026
1027 /* Functions to work with nop pools. NOP insns are used as temporary
1028 placeholders of the insns being scheduled to allow correct update of
1029 the data sets. When update is finished, NOPs are deleted. */
1030
1031 /* A vinsn that is used to represent a nop. This vinsn is shared among all
1032 nops sel-sched generates. */
1033 static vinsn_t nop_vinsn = NULL;
1034
1035 /* Emit a nop before INSN, taking it from pool. */
1036 insn_t
1037 get_nop_from_pool (insn_t insn)
1038 {
1039 insn_t nop;
1040 bool old_p = nop_pool.n != 0;
1041 int flags;
1042
1043 if (old_p)
1044 nop = nop_pool.v[--nop_pool.n];
1045 else
1046 nop = nop_pattern;
1047
1048 nop = emit_insn_before (nop, insn);
1049
1050 if (old_p)
1051 flags = INSN_INIT_TODO_SSID;
1052 else
1053 flags = INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID;
1054
1055 set_insn_init (INSN_EXPR (insn), nop_vinsn, INSN_SEQNO (insn));
1056 sel_init_new_insn (nop, flags);
1057
1058 return nop;
1059 }
1060
1061 /* Remove NOP from the instruction stream and return it to the pool. */
1062 void
1063 return_nop_to_pool (insn_t nop, bool full_tidying)
1064 {
1065 gcc_assert (INSN_IN_STREAM_P (nop));
1066 sel_remove_insn (nop, false, full_tidying);
1067
1068 if (nop_pool.n == nop_pool.s)
1069 nop_pool.v = XRESIZEVEC (rtx, nop_pool.v,
1070 (nop_pool.s = 2 * nop_pool.s + 1));
1071 nop_pool.v[nop_pool.n++] = nop;
1072 }
1073
1074 /* Free the nop pool. */
1075 void
1076 free_nop_pool (void)
1077 {
1078 nop_pool.n = 0;
1079 nop_pool.s = 0;
1080 free (nop_pool.v);
1081 nop_pool.v = NULL;
1082 }
1083 \f
1084
1085 /* Skip unspec to support ia64 speculation. Called from rtx_equal_p_cb.
1086 The callback is given two rtxes XX and YY and writes the new rtxes
1087 to NX and NY in case some needs to be skipped. */
1088 static int
1089 skip_unspecs_callback (const_rtx *xx, const_rtx *yy, rtx *nx, rtx* ny)
1090 {
1091 const_rtx x = *xx;
1092 const_rtx y = *yy;
1093
1094 if (GET_CODE (x) == UNSPEC
1095 && (targetm.sched.skip_rtx_p == NULL
1096 || targetm.sched.skip_rtx_p (x)))
1097 {
1098 *nx = XVECEXP (x, 0, 0);
1099 *ny = CONST_CAST_RTX (y);
1100 return 1;
1101 }
1102
1103 if (GET_CODE (y) == UNSPEC
1104 && (targetm.sched.skip_rtx_p == NULL
1105 || targetm.sched.skip_rtx_p (y)))
1106 {
1107 *nx = CONST_CAST_RTX (x);
1108 *ny = XVECEXP (y, 0, 0);
1109 return 1;
1110 }
1111
1112 return 0;
1113 }
1114
1115 /* Callback, called from hash_rtx_cb. Helps to hash UNSPEC rtx X in a correct way
1116 to support ia64 speculation. When changes are needed, new rtx X and new mode
1117 NMODE are written, and the callback returns true. */
1118 static int
1119 hash_with_unspec_callback (const_rtx x, enum machine_mode mode ATTRIBUTE_UNUSED,
1120 rtx *nx, enum machine_mode* nmode)
1121 {
1122 if (GET_CODE (x) == UNSPEC
1123 && targetm.sched.skip_rtx_p
1124 && targetm.sched.skip_rtx_p (x))
1125 {
1126 *nx = XVECEXP (x, 0 ,0);
1127 *nmode = VOIDmode;
1128 return 1;
1129 }
1130
1131 return 0;
1132 }
1133
1134 /* Returns LHS and RHS are ok to be scheduled separately. */
1135 static bool
1136 lhs_and_rhs_separable_p (rtx lhs, rtx rhs)
1137 {
1138 if (lhs == NULL || rhs == NULL)
1139 return false;
1140
1141 /* Do not schedule CONST, CONST_INT and CONST_DOUBLE etc as rhs: no point
1142 to use reg, if const can be used. Moreover, scheduling const as rhs may
1143 lead to mode mismatch cause consts don't have modes but they could be
1144 merged from branches where the same const used in different modes. */
1145 if (CONSTANT_P (rhs))
1146 return false;
1147
1148 /* ??? Do not rename predicate registers to avoid ICEs in bundling. */
1149 if (COMPARISON_P (rhs))
1150 return false;
1151
1152 /* Do not allow single REG to be an rhs. */
1153 if (REG_P (rhs))
1154 return false;
1155
1156 /* See comment at find_used_regs_1 (*1) for explanation of this
1157 restriction. */
1158 /* FIXME: remove this later. */
1159 if (MEM_P (lhs))
1160 return false;
1161
1162 /* This will filter all tricky things like ZERO_EXTRACT etc.
1163 For now we don't handle it. */
1164 if (!REG_P (lhs) && !MEM_P (lhs))
1165 return false;
1166
1167 return true;
1168 }
1169
1170 /* Initialize vinsn VI for INSN. Only for use from vinsn_create (). When
1171 FORCE_UNIQUE_P is true, the resulting vinsn will not be clonable. This is
1172 used e.g. for insns from recovery blocks. */
1173 static void
1174 vinsn_init (vinsn_t vi, insn_t insn, bool force_unique_p)
1175 {
1176 hash_rtx_callback_function hrcf;
1177 int insn_class;
1178
1179 VINSN_INSN_RTX (vi) = insn;
1180 VINSN_COUNT (vi) = 0;
1181 vi->cost = -1;
1182
1183 if (INSN_NOP_P (insn))
1184 return;
1185
1186 if (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL)
1187 init_id_from_df (VINSN_ID (vi), insn, force_unique_p);
1188 else
1189 deps_init_id (VINSN_ID (vi), insn, force_unique_p);
1190
1191 /* Hash vinsn depending on whether it is separable or not. */
1192 hrcf = targetm.sched.skip_rtx_p ? hash_with_unspec_callback : NULL;
1193 if (VINSN_SEPARABLE_P (vi))
1194 {
1195 rtx rhs = VINSN_RHS (vi);
1196
1197 VINSN_HASH (vi) = hash_rtx_cb (rhs, GET_MODE (rhs),
1198 NULL, NULL, false, hrcf);
1199 VINSN_HASH_RTX (vi) = hash_rtx_cb (VINSN_PATTERN (vi),
1200 VOIDmode, NULL, NULL,
1201 false, hrcf);
1202 }
1203 else
1204 {
1205 VINSN_HASH (vi) = hash_rtx_cb (VINSN_PATTERN (vi), VOIDmode,
1206 NULL, NULL, false, hrcf);
1207 VINSN_HASH_RTX (vi) = VINSN_HASH (vi);
1208 }
1209
1210 insn_class = haifa_classify_insn (insn);
1211 if (insn_class >= 2
1212 && (!targetm.sched.get_insn_spec_ds
1213 || ((targetm.sched.get_insn_spec_ds (insn) & BEGIN_CONTROL)
1214 == 0)))
1215 VINSN_MAY_TRAP_P (vi) = true;
1216 else
1217 VINSN_MAY_TRAP_P (vi) = false;
1218 }
1219
1220 /* Indicate that VI has become the part of an rtx object. */
1221 void
1222 vinsn_attach (vinsn_t vi)
1223 {
1224 /* Assert that VI is not pending for deletion. */
1225 gcc_assert (VINSN_INSN_RTX (vi));
1226
1227 VINSN_COUNT (vi)++;
1228 }
1229
1230 /* Create and init VI from the INSN. Use UNIQUE_P for determining the correct
1231 VINSN_TYPE (VI). */
1232 static vinsn_t
1233 vinsn_create (insn_t insn, bool force_unique_p)
1234 {
1235 vinsn_t vi = XCNEW (struct vinsn_def);
1236
1237 vinsn_init (vi, insn, force_unique_p);
1238 return vi;
1239 }
1240
1241 /* Return a copy of VI. When REATTACH_P is true, detach VI and attach
1242 the copy. */
1243 vinsn_t
1244 vinsn_copy (vinsn_t vi, bool reattach_p)
1245 {
1246 rtx copy;
1247 bool unique = VINSN_UNIQUE_P (vi);
1248 vinsn_t new_vi;
1249
1250 copy = create_copy_of_insn_rtx (VINSN_INSN_RTX (vi));
1251 new_vi = create_vinsn_from_insn_rtx (copy, unique);
1252 if (reattach_p)
1253 {
1254 vinsn_detach (vi);
1255 vinsn_attach (new_vi);
1256 }
1257
1258 return new_vi;
1259 }
1260
1261 /* Delete the VI vinsn and free its data. */
1262 static void
1263 vinsn_delete (vinsn_t vi)
1264 {
1265 gcc_assert (VINSN_COUNT (vi) == 0);
1266
1267 if (!INSN_NOP_P (VINSN_INSN_RTX (vi)))
1268 {
1269 return_regset_to_pool (VINSN_REG_SETS (vi));
1270 return_regset_to_pool (VINSN_REG_USES (vi));
1271 return_regset_to_pool (VINSN_REG_CLOBBERS (vi));
1272 }
1273
1274 free (vi);
1275 }
1276
1277 /* Indicate that VI is no longer a part of some rtx object.
1278 Remove VI if it is no longer needed. */
1279 void
1280 vinsn_detach (vinsn_t vi)
1281 {
1282 gcc_assert (VINSN_COUNT (vi) > 0);
1283
1284 if (--VINSN_COUNT (vi) == 0)
1285 vinsn_delete (vi);
1286 }
1287
1288 /* Returns TRUE if VI is a branch. */
1289 bool
1290 vinsn_cond_branch_p (vinsn_t vi)
1291 {
1292 insn_t insn;
1293
1294 if (!VINSN_UNIQUE_P (vi))
1295 return false;
1296
1297 insn = VINSN_INSN_RTX (vi);
1298 if (BB_END (BLOCK_FOR_INSN (insn)) != insn)
1299 return false;
1300
1301 return control_flow_insn_p (insn);
1302 }
1303
1304 /* Return latency of INSN. */
1305 static int
1306 sel_insn_rtx_cost (rtx insn)
1307 {
1308 int cost;
1309
1310 /* A USE insn, or something else we don't need to
1311 understand. We can't pass these directly to
1312 result_ready_cost or insn_default_latency because it will
1313 trigger a fatal error for unrecognizable insns. */
1314 if (recog_memoized (insn) < 0)
1315 cost = 0;
1316 else
1317 {
1318 cost = insn_default_latency (insn);
1319
1320 if (cost < 0)
1321 cost = 0;
1322 }
1323
1324 return cost;
1325 }
1326
1327 /* Return the cost of the VI.
1328 !!! FIXME: Unify with haifa-sched.c: insn_cost (). */
1329 int
1330 sel_vinsn_cost (vinsn_t vi)
1331 {
1332 int cost = vi->cost;
1333
1334 if (cost < 0)
1335 {
1336 cost = sel_insn_rtx_cost (VINSN_INSN_RTX (vi));
1337 vi->cost = cost;
1338 }
1339
1340 return cost;
1341 }
1342 \f
1343
1344 /* Functions for insn emitting. */
1345
1346 /* Emit new insn after AFTER based on PATTERN and initialize its data from
1347 EXPR and SEQNO. */
1348 insn_t
1349 sel_gen_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, insn_t after)
1350 {
1351 insn_t new_insn;
1352
1353 gcc_assert (EXPR_TARGET_AVAILABLE (expr) == true);
1354
1355 new_insn = emit_insn_after (pattern, after);
1356 set_insn_init (expr, NULL, seqno);
1357 sel_init_new_insn (new_insn, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID);
1358
1359 return new_insn;
1360 }
1361
1362 /* Force newly generated vinsns to be unique. */
1363 static bool init_insn_force_unique_p = false;
1364
1365 /* Emit new speculation recovery insn after AFTER based on PATTERN and
1366 initialize its data from EXPR and SEQNO. */
1367 insn_t
1368 sel_gen_recovery_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno,
1369 insn_t after)
1370 {
1371 insn_t insn;
1372
1373 gcc_assert (!init_insn_force_unique_p);
1374
1375 init_insn_force_unique_p = true;
1376 insn = sel_gen_insn_from_rtx_after (pattern, expr, seqno, after);
1377 CANT_MOVE (insn) = 1;
1378 init_insn_force_unique_p = false;
1379
1380 return insn;
1381 }
1382
1383 /* Emit new insn after AFTER based on EXPR and SEQNO. If VINSN is not NULL,
1384 take it as a new vinsn instead of EXPR's vinsn.
1385 We simplify insns later, after scheduling region in
1386 simplify_changed_insns. */
1387 insn_t
1388 sel_gen_insn_from_expr_after (expr_t expr, vinsn_t vinsn, int seqno,
1389 insn_t after)
1390 {
1391 expr_t emit_expr;
1392 insn_t insn;
1393 int flags;
1394
1395 emit_expr = set_insn_init (expr, vinsn ? vinsn : EXPR_VINSN (expr),
1396 seqno);
1397 insn = EXPR_INSN_RTX (emit_expr);
1398 add_insn_after (insn, after, BLOCK_FOR_INSN (insn));
1399
1400 flags = INSN_INIT_TODO_SSID;
1401 if (INSN_LUID (insn) == 0)
1402 flags |= INSN_INIT_TODO_LUID;
1403 sel_init_new_insn (insn, flags);
1404
1405 return insn;
1406 }
1407
1408 /* Move insn from EXPR after AFTER. */
1409 insn_t
1410 sel_move_insn (expr_t expr, int seqno, insn_t after)
1411 {
1412 insn_t insn = EXPR_INSN_RTX (expr);
1413 basic_block bb = BLOCK_FOR_INSN (after);
1414 insn_t next = NEXT_INSN (after);
1415
1416 /* Assert that in move_op we disconnected this insn properly. */
1417 gcc_assert (EXPR_VINSN (INSN_EXPR (insn)) != NULL);
1418 PREV_INSN (insn) = after;
1419 NEXT_INSN (insn) = next;
1420
1421 NEXT_INSN (after) = insn;
1422 PREV_INSN (next) = insn;
1423
1424 /* Update links from insn to bb and vice versa. */
1425 df_insn_change_bb (insn, bb);
1426 if (BB_END (bb) == after)
1427 BB_END (bb) = insn;
1428
1429 prepare_insn_expr (insn, seqno);
1430 return insn;
1431 }
1432
1433 \f
1434 /* Functions to work with right-hand sides. */
1435
1436 /* Search for a hash value determined by UID/NEW_VINSN in a sorted vector
1437 VECT and return true when found. Use NEW_VINSN for comparison only when
1438 COMPARE_VINSNS is true. Write to INDP the index on which
1439 the search has stopped, such that inserting the new element at INDP will
1440 retain VECT's sort order. */
1441 static bool
1442 find_in_history_vect_1 (vec<expr_history_def> vect,
1443 unsigned uid, vinsn_t new_vinsn,
1444 bool compare_vinsns, int *indp)
1445 {
1446 expr_history_def *arr;
1447 int i, j, len = vect.length ();
1448
1449 if (len == 0)
1450 {
1451 *indp = 0;
1452 return false;
1453 }
1454
1455 arr = vect.address ();
1456 i = 0, j = len - 1;
1457
1458 while (i <= j)
1459 {
1460 unsigned auid = arr[i].uid;
1461 vinsn_t avinsn = arr[i].new_expr_vinsn;
1462
1463 if (auid == uid
1464 /* When undoing transformation on a bookkeeping copy, the new vinsn
1465 may not be exactly equal to the one that is saved in the vector.
1466 This is because the insn whose copy we're checking was possibly
1467 substituted itself. */
1468 && (! compare_vinsns
1469 || vinsn_equal_p (avinsn, new_vinsn)))
1470 {
1471 *indp = i;
1472 return true;
1473 }
1474 else if (auid > uid)
1475 break;
1476 i++;
1477 }
1478
1479 *indp = i;
1480 return false;
1481 }
1482
1483 /* Search for a uid of INSN and NEW_VINSN in a sorted vector VECT. Return
1484 the position found or -1, if no such value is in vector.
1485 Search also for UIDs of insn's originators, if ORIGINATORS_P is true. */
1486 int
1487 find_in_history_vect (vec<expr_history_def> vect, rtx insn,
1488 vinsn_t new_vinsn, bool originators_p)
1489 {
1490 int ind;
1491
1492 if (find_in_history_vect_1 (vect, INSN_UID (insn), new_vinsn,
1493 false, &ind))
1494 return ind;
1495
1496 if (INSN_ORIGINATORS (insn) && originators_p)
1497 {
1498 unsigned uid;
1499 bitmap_iterator bi;
1500
1501 EXECUTE_IF_SET_IN_BITMAP (INSN_ORIGINATORS (insn), 0, uid, bi)
1502 if (find_in_history_vect_1 (vect, uid, new_vinsn, false, &ind))
1503 return ind;
1504 }
1505
1506 return -1;
1507 }
1508
1509 /* Insert new element in a sorted history vector pointed to by PVECT,
1510 if it is not there already. The element is searched using
1511 UID/NEW_EXPR_VINSN pair. TYPE, OLD_EXPR_VINSN and SPEC_DS save
1512 the history of a transformation. */
1513 void
1514 insert_in_history_vect (vec<expr_history_def> *pvect,
1515 unsigned uid, enum local_trans_type type,
1516 vinsn_t old_expr_vinsn, vinsn_t new_expr_vinsn,
1517 ds_t spec_ds)
1518 {
1519 vec<expr_history_def> vect = *pvect;
1520 expr_history_def temp;
1521 bool res;
1522 int ind;
1523
1524 res = find_in_history_vect_1 (vect, uid, new_expr_vinsn, true, &ind);
1525
1526 if (res)
1527 {
1528 expr_history_def *phist = &vect[ind];
1529
1530 /* It is possible that speculation types of expressions that were
1531 propagated through different paths will be different here. In this
1532 case, merge the status to get the correct check later. */
1533 if (phist->spec_ds != spec_ds)
1534 phist->spec_ds = ds_max_merge (phist->spec_ds, spec_ds);
1535 return;
1536 }
1537
1538 temp.uid = uid;
1539 temp.old_expr_vinsn = old_expr_vinsn;
1540 temp.new_expr_vinsn = new_expr_vinsn;
1541 temp.spec_ds = spec_ds;
1542 temp.type = type;
1543
1544 vinsn_attach (old_expr_vinsn);
1545 vinsn_attach (new_expr_vinsn);
1546 vect.safe_insert (ind, temp);
1547 *pvect = vect;
1548 }
1549
1550 /* Free history vector PVECT. */
1551 static void
1552 free_history_vect (vec<expr_history_def> &pvect)
1553 {
1554 unsigned i;
1555 expr_history_def *phist;
1556
1557 if (! pvect.exists ())
1558 return;
1559
1560 for (i = 0; pvect.iterate (i, &phist); i++)
1561 {
1562 vinsn_detach (phist->old_expr_vinsn);
1563 vinsn_detach (phist->new_expr_vinsn);
1564 }
1565
1566 pvect.release ();
1567 }
1568
1569 /* Merge vector FROM to PVECT. */
1570 static void
1571 merge_history_vect (vec<expr_history_def> *pvect,
1572 vec<expr_history_def> from)
1573 {
1574 expr_history_def *phist;
1575 int i;
1576
1577 /* We keep this vector sorted. */
1578 for (i = 0; from.iterate (i, &phist); i++)
1579 insert_in_history_vect (pvect, phist->uid, phist->type,
1580 phist->old_expr_vinsn, phist->new_expr_vinsn,
1581 phist->spec_ds);
1582 }
1583
1584 /* Compare two vinsns as rhses if possible and as vinsns otherwise. */
1585 bool
1586 vinsn_equal_p (vinsn_t x, vinsn_t y)
1587 {
1588 rtx_equal_p_callback_function repcf;
1589
1590 if (x == y)
1591 return true;
1592
1593 if (VINSN_TYPE (x) != VINSN_TYPE (y))
1594 return false;
1595
1596 if (VINSN_HASH (x) != VINSN_HASH (y))
1597 return false;
1598
1599 repcf = targetm.sched.skip_rtx_p ? skip_unspecs_callback : NULL;
1600 if (VINSN_SEPARABLE_P (x))
1601 {
1602 /* Compare RHSes of VINSNs. */
1603 gcc_assert (VINSN_RHS (x));
1604 gcc_assert (VINSN_RHS (y));
1605
1606 return rtx_equal_p_cb (VINSN_RHS (x), VINSN_RHS (y), repcf);
1607 }
1608
1609 return rtx_equal_p_cb (VINSN_PATTERN (x), VINSN_PATTERN (y), repcf);
1610 }
1611 \f
1612
1613 /* Functions for working with expressions. */
1614
1615 /* Initialize EXPR. */
1616 static void
1617 init_expr (expr_t expr, vinsn_t vi, int spec, int use, int priority,
1618 int sched_times, int orig_bb_index, ds_t spec_done_ds,
1619 ds_t spec_to_check_ds, int orig_sched_cycle,
1620 vec<expr_history_def> history,
1621 signed char target_available,
1622 bool was_substituted, bool was_renamed, bool needs_spec_check_p,
1623 bool cant_move)
1624 {
1625 vinsn_attach (vi);
1626
1627 EXPR_VINSN (expr) = vi;
1628 EXPR_SPEC (expr) = spec;
1629 EXPR_USEFULNESS (expr) = use;
1630 EXPR_PRIORITY (expr) = priority;
1631 EXPR_PRIORITY_ADJ (expr) = 0;
1632 EXPR_SCHED_TIMES (expr) = sched_times;
1633 EXPR_ORIG_BB_INDEX (expr) = orig_bb_index;
1634 EXPR_ORIG_SCHED_CYCLE (expr) = orig_sched_cycle;
1635 EXPR_SPEC_DONE_DS (expr) = spec_done_ds;
1636 EXPR_SPEC_TO_CHECK_DS (expr) = spec_to_check_ds;
1637
1638 if (history.exists ())
1639 EXPR_HISTORY_OF_CHANGES (expr) = history;
1640 else
1641 EXPR_HISTORY_OF_CHANGES (expr).create (0);
1642
1643 EXPR_TARGET_AVAILABLE (expr) = target_available;
1644 EXPR_WAS_SUBSTITUTED (expr) = was_substituted;
1645 EXPR_WAS_RENAMED (expr) = was_renamed;
1646 EXPR_NEEDS_SPEC_CHECK_P (expr) = needs_spec_check_p;
1647 EXPR_CANT_MOVE (expr) = cant_move;
1648 }
1649
1650 /* Make a copy of the expr FROM into the expr TO. */
1651 void
1652 copy_expr (expr_t to, expr_t from)
1653 {
1654 vec<expr_history_def> temp = vNULL;
1655
1656 if (EXPR_HISTORY_OF_CHANGES (from).exists ())
1657 {
1658 unsigned i;
1659 expr_history_def *phist;
1660
1661 temp = EXPR_HISTORY_OF_CHANGES (from).copy ();
1662 for (i = 0;
1663 temp.iterate (i, &phist);
1664 i++)
1665 {
1666 vinsn_attach (phist->old_expr_vinsn);
1667 vinsn_attach (phist->new_expr_vinsn);
1668 }
1669 }
1670
1671 init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from),
1672 EXPR_USEFULNESS (from), EXPR_PRIORITY (from),
1673 EXPR_SCHED_TIMES (from), EXPR_ORIG_BB_INDEX (from),
1674 EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from),
1675 EXPR_ORIG_SCHED_CYCLE (from), temp,
1676 EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from),
1677 EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from),
1678 EXPR_CANT_MOVE (from));
1679 }
1680
1681 /* Same, but the final expr will not ever be in av sets, so don't copy
1682 "uninteresting" data such as bitmap cache. */
1683 void
1684 copy_expr_onside (expr_t to, expr_t from)
1685 {
1686 init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from), EXPR_USEFULNESS (from),
1687 EXPR_PRIORITY (from), EXPR_SCHED_TIMES (from), 0,
1688 EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), 0,
1689 vNULL,
1690 EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from),
1691 EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from),
1692 EXPR_CANT_MOVE (from));
1693 }
1694
1695 /* Prepare the expr of INSN for scheduling. Used when moving insn and when
1696 initializing new insns. */
1697 static void
1698 prepare_insn_expr (insn_t insn, int seqno)
1699 {
1700 expr_t expr = INSN_EXPR (insn);
1701 ds_t ds;
1702
1703 INSN_SEQNO (insn) = seqno;
1704 EXPR_ORIG_BB_INDEX (expr) = BLOCK_NUM (insn);
1705 EXPR_SPEC (expr) = 0;
1706 EXPR_ORIG_SCHED_CYCLE (expr) = 0;
1707 EXPR_WAS_SUBSTITUTED (expr) = 0;
1708 EXPR_WAS_RENAMED (expr) = 0;
1709 EXPR_TARGET_AVAILABLE (expr) = 1;
1710 INSN_LIVE_VALID_P (insn) = false;
1711
1712 /* ??? If this expression is speculative, make its dependence
1713 as weak as possible. We can filter this expression later
1714 in process_spec_exprs, because we do not distinguish
1715 between the status we got during compute_av_set and the
1716 existing status. To be fixed. */
1717 ds = EXPR_SPEC_DONE_DS (expr);
1718 if (ds)
1719 EXPR_SPEC_DONE_DS (expr) = ds_get_max_dep_weak (ds);
1720
1721 free_history_vect (EXPR_HISTORY_OF_CHANGES (expr));
1722 }
1723
1724 /* Update target_available bits when merging exprs TO and FROM. SPLIT_POINT
1725 is non-null when expressions are merged from different successors at
1726 a split point. */
1727 static void
1728 update_target_availability (expr_t to, expr_t from, insn_t split_point)
1729 {
1730 if (EXPR_TARGET_AVAILABLE (to) < 0
1731 || EXPR_TARGET_AVAILABLE (from) < 0)
1732 EXPR_TARGET_AVAILABLE (to) = -1;
1733 else
1734 {
1735 /* We try to detect the case when one of the expressions
1736 can only be reached through another one. In this case,
1737 we can do better. */
1738 if (split_point == NULL)
1739 {
1740 int toind, fromind;
1741
1742 toind = EXPR_ORIG_BB_INDEX (to);
1743 fromind = EXPR_ORIG_BB_INDEX (from);
1744
1745 if (toind && toind == fromind)
1746 /* Do nothing -- everything is done in
1747 merge_with_other_exprs. */
1748 ;
1749 else
1750 EXPR_TARGET_AVAILABLE (to) = -1;
1751 }
1752 else if (EXPR_TARGET_AVAILABLE (from) == 0
1753 && EXPR_LHS (from)
1754 && REG_P (EXPR_LHS (from))
1755 && REGNO (EXPR_LHS (to)) != REGNO (EXPR_LHS (from)))
1756 EXPR_TARGET_AVAILABLE (to) = -1;
1757 else
1758 EXPR_TARGET_AVAILABLE (to) &= EXPR_TARGET_AVAILABLE (from);
1759 }
1760 }
1761
1762 /* Update speculation bits when merging exprs TO and FROM. SPLIT_POINT
1763 is non-null when expressions are merged from different successors at
1764 a split point. */
1765 static void
1766 update_speculative_bits (expr_t to, expr_t from, insn_t split_point)
1767 {
1768 ds_t old_to_ds, old_from_ds;
1769
1770 old_to_ds = EXPR_SPEC_DONE_DS (to);
1771 old_from_ds = EXPR_SPEC_DONE_DS (from);
1772
1773 EXPR_SPEC_DONE_DS (to) = ds_max_merge (old_to_ds, old_from_ds);
1774 EXPR_SPEC_TO_CHECK_DS (to) |= EXPR_SPEC_TO_CHECK_DS (from);
1775 EXPR_NEEDS_SPEC_CHECK_P (to) |= EXPR_NEEDS_SPEC_CHECK_P (from);
1776
1777 /* When merging e.g. control & data speculative exprs, or a control
1778 speculative with a control&data speculative one, we really have
1779 to change vinsn too. Also, when speculative status is changed,
1780 we also need to record this as a transformation in expr's history. */
1781 if ((old_to_ds & SPECULATIVE) || (old_from_ds & SPECULATIVE))
1782 {
1783 old_to_ds = ds_get_speculation_types (old_to_ds);
1784 old_from_ds = ds_get_speculation_types (old_from_ds);
1785
1786 if (old_to_ds != old_from_ds)
1787 {
1788 ds_t record_ds;
1789
1790 /* When both expressions are speculative, we need to change
1791 the vinsn first. */
1792 if ((old_to_ds & SPECULATIVE) && (old_from_ds & SPECULATIVE))
1793 {
1794 int res;
1795
1796 res = speculate_expr (to, EXPR_SPEC_DONE_DS (to));
1797 gcc_assert (res >= 0);
1798 }
1799
1800 if (split_point != NULL)
1801 {
1802 /* Record the change with proper status. */
1803 record_ds = EXPR_SPEC_DONE_DS (to) & SPECULATIVE;
1804 record_ds &= ~(old_to_ds & SPECULATIVE);
1805 record_ds &= ~(old_from_ds & SPECULATIVE);
1806
1807 insert_in_history_vect (&EXPR_HISTORY_OF_CHANGES (to),
1808 INSN_UID (split_point), TRANS_SPECULATION,
1809 EXPR_VINSN (from), EXPR_VINSN (to),
1810 record_ds);
1811 }
1812 }
1813 }
1814 }
1815
1816
1817 /* Merge bits of FROM expr to TO expr. When SPLIT_POINT is not NULL,
1818 this is done along different paths. */
1819 void
1820 merge_expr_data (expr_t to, expr_t from, insn_t split_point)
1821 {
1822 /* Choose the maximum of the specs of merged exprs. This is required
1823 for correctness of bookkeeping. */
1824 if (EXPR_SPEC (to) < EXPR_SPEC (from))
1825 EXPR_SPEC (to) = EXPR_SPEC (from);
1826
1827 if (split_point)
1828 EXPR_USEFULNESS (to) += EXPR_USEFULNESS (from);
1829 else
1830 EXPR_USEFULNESS (to) = MAX (EXPR_USEFULNESS (to),
1831 EXPR_USEFULNESS (from));
1832
1833 if (EXPR_PRIORITY (to) < EXPR_PRIORITY (from))
1834 EXPR_PRIORITY (to) = EXPR_PRIORITY (from);
1835
1836 if (EXPR_SCHED_TIMES (to) > EXPR_SCHED_TIMES (from))
1837 EXPR_SCHED_TIMES (to) = EXPR_SCHED_TIMES (from);
1838
1839 if (EXPR_ORIG_BB_INDEX (to) != EXPR_ORIG_BB_INDEX (from))
1840 EXPR_ORIG_BB_INDEX (to) = 0;
1841
1842 EXPR_ORIG_SCHED_CYCLE (to) = MIN (EXPR_ORIG_SCHED_CYCLE (to),
1843 EXPR_ORIG_SCHED_CYCLE (from));
1844
1845 EXPR_WAS_SUBSTITUTED (to) |= EXPR_WAS_SUBSTITUTED (from);
1846 EXPR_WAS_RENAMED (to) |= EXPR_WAS_RENAMED (from);
1847 EXPR_CANT_MOVE (to) |= EXPR_CANT_MOVE (from);
1848
1849 merge_history_vect (&EXPR_HISTORY_OF_CHANGES (to),
1850 EXPR_HISTORY_OF_CHANGES (from));
1851 update_target_availability (to, from, split_point);
1852 update_speculative_bits (to, from, split_point);
1853 }
1854
1855 /* Merge bits of FROM expr to TO expr. Vinsns in the exprs should be equal
1856 in terms of vinsn_equal_p. SPLIT_POINT is non-null when expressions
1857 are merged from different successors at a split point. */
1858 void
1859 merge_expr (expr_t to, expr_t from, insn_t split_point)
1860 {
1861 vinsn_t to_vi = EXPR_VINSN (to);
1862 vinsn_t from_vi = EXPR_VINSN (from);
1863
1864 gcc_assert (vinsn_equal_p (to_vi, from_vi));
1865
1866 /* Make sure that speculative pattern is propagated into exprs that
1867 have non-speculative one. This will provide us with consistent
1868 speculative bits and speculative patterns inside expr. */
1869 if (EXPR_SPEC_DONE_DS (to) == 0
1870 && EXPR_SPEC_DONE_DS (from) != 0)
1871 change_vinsn_in_expr (to, EXPR_VINSN (from));
1872
1873 merge_expr_data (to, from, split_point);
1874 gcc_assert (EXPR_USEFULNESS (to) <= REG_BR_PROB_BASE);
1875 }
1876
1877 /* Clear the information of this EXPR. */
1878 void
1879 clear_expr (expr_t expr)
1880 {
1881
1882 vinsn_detach (EXPR_VINSN (expr));
1883 EXPR_VINSN (expr) = NULL;
1884
1885 free_history_vect (EXPR_HISTORY_OF_CHANGES (expr));
1886 }
1887
1888 /* For a given LV_SET, mark EXPR having unavailable target register. */
1889 static void
1890 set_unavailable_target_for_expr (expr_t expr, regset lv_set)
1891 {
1892 if (EXPR_SEPARABLE_P (expr))
1893 {
1894 if (REG_P (EXPR_LHS (expr))
1895 && register_unavailable_p (lv_set, EXPR_LHS (expr)))
1896 {
1897 /* If it's an insn like r1 = use (r1, ...), and it exists in
1898 different forms in each of the av_sets being merged, we can't say
1899 whether original destination register is available or not.
1900 However, this still works if destination register is not used
1901 in the original expression: if the branch at which LV_SET we're
1902 looking here is not actually 'other branch' in sense that same
1903 expression is available through it (but it can't be determined
1904 at computation stage because of transformations on one of the
1905 branches), it still won't affect the availability.
1906 Liveness of a register somewhere on a code motion path means
1907 it's either read somewhere on a codemotion path, live on
1908 'other' branch, live at the point immediately following
1909 the original operation, or is read by the original operation.
1910 The latter case is filtered out in the condition below.
1911 It still doesn't cover the case when register is defined and used
1912 somewhere within the code motion path, and in this case we could
1913 miss a unifying code motion along both branches using a renamed
1914 register, but it won't affect a code correctness since upon
1915 an actual code motion a bookkeeping code would be generated. */
1916 if (register_unavailable_p (VINSN_REG_USES (EXPR_VINSN (expr)),
1917 EXPR_LHS (expr)))
1918 EXPR_TARGET_AVAILABLE (expr) = -1;
1919 else
1920 EXPR_TARGET_AVAILABLE (expr) = false;
1921 }
1922 }
1923 else
1924 {
1925 unsigned regno;
1926 reg_set_iterator rsi;
1927
1928 EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_SETS (EXPR_VINSN (expr)),
1929 0, regno, rsi)
1930 if (bitmap_bit_p (lv_set, regno))
1931 {
1932 EXPR_TARGET_AVAILABLE (expr) = false;
1933 break;
1934 }
1935
1936 EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_CLOBBERS (EXPR_VINSN (expr)),
1937 0, regno, rsi)
1938 if (bitmap_bit_p (lv_set, regno))
1939 {
1940 EXPR_TARGET_AVAILABLE (expr) = false;
1941 break;
1942 }
1943 }
1944 }
1945
1946 /* Try to make EXPR speculative. Return 1 when EXPR's pattern
1947 or dependence status have changed, 2 when also the target register
1948 became unavailable, 0 if nothing had to be changed. */
1949 int
1950 speculate_expr (expr_t expr, ds_t ds)
1951 {
1952 int res;
1953 rtx orig_insn_rtx;
1954 rtx spec_pat;
1955 ds_t target_ds, current_ds;
1956
1957 /* Obtain the status we need to put on EXPR. */
1958 target_ds = (ds & SPECULATIVE);
1959 current_ds = EXPR_SPEC_DONE_DS (expr);
1960 ds = ds_full_merge (current_ds, target_ds, NULL_RTX, NULL_RTX);
1961
1962 orig_insn_rtx = EXPR_INSN_RTX (expr);
1963
1964 res = sched_speculate_insn (orig_insn_rtx, ds, &spec_pat);
1965
1966 switch (res)
1967 {
1968 case 0:
1969 EXPR_SPEC_DONE_DS (expr) = ds;
1970 return current_ds != ds ? 1 : 0;
1971
1972 case 1:
1973 {
1974 rtx spec_insn_rtx = create_insn_rtx_from_pattern (spec_pat, NULL_RTX);
1975 vinsn_t spec_vinsn = create_vinsn_from_insn_rtx (spec_insn_rtx, false);
1976
1977 change_vinsn_in_expr (expr, spec_vinsn);
1978 EXPR_SPEC_DONE_DS (expr) = ds;
1979 EXPR_NEEDS_SPEC_CHECK_P (expr) = true;
1980
1981 /* Do not allow clobbering the address register of speculative
1982 insns. */
1983 if (register_unavailable_p (VINSN_REG_USES (EXPR_VINSN (expr)),
1984 expr_dest_reg (expr)))
1985 {
1986 EXPR_TARGET_AVAILABLE (expr) = false;
1987 return 2;
1988 }
1989
1990 return 1;
1991 }
1992
1993 case -1:
1994 return -1;
1995
1996 default:
1997 gcc_unreachable ();
1998 return -1;
1999 }
2000 }
2001
2002 /* Return a destination register, if any, of EXPR. */
2003 rtx
2004 expr_dest_reg (expr_t expr)
2005 {
2006 rtx dest = VINSN_LHS (EXPR_VINSN (expr));
2007
2008 if (dest != NULL_RTX && REG_P (dest))
2009 return dest;
2010
2011 return NULL_RTX;
2012 }
2013
2014 /* Returns the REGNO of the R's destination. */
2015 unsigned
2016 expr_dest_regno (expr_t expr)
2017 {
2018 rtx dest = expr_dest_reg (expr);
2019
2020 gcc_assert (dest != NULL_RTX);
2021 return REGNO (dest);
2022 }
2023
2024 /* For a given LV_SET, mark all expressions in JOIN_SET, but not present in
2025 AV_SET having unavailable target register. */
2026 void
2027 mark_unavailable_targets (av_set_t join_set, av_set_t av_set, regset lv_set)
2028 {
2029 expr_t expr;
2030 av_set_iterator avi;
2031
2032 FOR_EACH_EXPR (expr, avi, join_set)
2033 if (av_set_lookup (av_set, EXPR_VINSN (expr)) == NULL)
2034 set_unavailable_target_for_expr (expr, lv_set);
2035 }
2036 \f
2037
2038 /* Returns true if REG (at least partially) is present in REGS. */
2039 bool
2040 register_unavailable_p (regset regs, rtx reg)
2041 {
2042 unsigned regno, end_regno;
2043
2044 regno = REGNO (reg);
2045 if (bitmap_bit_p (regs, regno))
2046 return true;
2047
2048 end_regno = END_REGNO (reg);
2049
2050 while (++regno < end_regno)
2051 if (bitmap_bit_p (regs, regno))
2052 return true;
2053
2054 return false;
2055 }
2056
2057 /* Av set functions. */
2058
2059 /* Add a new element to av set SETP.
2060 Return the element added. */
2061 static av_set_t
2062 av_set_add_element (av_set_t *setp)
2063 {
2064 /* Insert at the beginning of the list. */
2065 _list_add (setp);
2066 return *setp;
2067 }
2068
2069 /* Add EXPR to SETP. */
2070 void
2071 av_set_add (av_set_t *setp, expr_t expr)
2072 {
2073 av_set_t elem;
2074
2075 gcc_assert (!INSN_NOP_P (EXPR_INSN_RTX (expr)));
2076 elem = av_set_add_element (setp);
2077 copy_expr (_AV_SET_EXPR (elem), expr);
2078 }
2079
2080 /* Same, but do not copy EXPR. */
2081 static void
2082 av_set_add_nocopy (av_set_t *setp, expr_t expr)
2083 {
2084 av_set_t elem;
2085
2086 elem = av_set_add_element (setp);
2087 *_AV_SET_EXPR (elem) = *expr;
2088 }
2089
2090 /* Remove expr pointed to by IP from the av_set. */
2091 void
2092 av_set_iter_remove (av_set_iterator *ip)
2093 {
2094 clear_expr (_AV_SET_EXPR (*ip->lp));
2095 _list_iter_remove (ip);
2096 }
2097
2098 /* Search for an expr in SET, such that it's equivalent to SOUGHT_VINSN in the
2099 sense of vinsn_equal_p function. Return NULL if no such expr is
2100 in SET was found. */
2101 expr_t
2102 av_set_lookup (av_set_t set, vinsn_t sought_vinsn)
2103 {
2104 expr_t expr;
2105 av_set_iterator i;
2106
2107 FOR_EACH_EXPR (expr, i, set)
2108 if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn))
2109 return expr;
2110 return NULL;
2111 }
2112
2113 /* Same, but also remove the EXPR found. */
2114 static expr_t
2115 av_set_lookup_and_remove (av_set_t *setp, vinsn_t sought_vinsn)
2116 {
2117 expr_t expr;
2118 av_set_iterator i;
2119
2120 FOR_EACH_EXPR_1 (expr, i, setp)
2121 if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn))
2122 {
2123 _list_iter_remove_nofree (&i);
2124 return expr;
2125 }
2126 return NULL;
2127 }
2128
2129 /* Search for an expr in SET, such that it's equivalent to EXPR in the
2130 sense of vinsn_equal_p function of their vinsns, but not EXPR itself.
2131 Returns NULL if no such expr is in SET was found. */
2132 static expr_t
2133 av_set_lookup_other_equiv_expr (av_set_t set, expr_t expr)
2134 {
2135 expr_t cur_expr;
2136 av_set_iterator i;
2137
2138 FOR_EACH_EXPR (cur_expr, i, set)
2139 {
2140 if (cur_expr == expr)
2141 continue;
2142 if (vinsn_equal_p (EXPR_VINSN (cur_expr), EXPR_VINSN (expr)))
2143 return cur_expr;
2144 }
2145
2146 return NULL;
2147 }
2148
2149 /* If other expression is already in AVP, remove one of them. */
2150 expr_t
2151 merge_with_other_exprs (av_set_t *avp, av_set_iterator *ip, expr_t expr)
2152 {
2153 expr_t expr2;
2154
2155 expr2 = av_set_lookup_other_equiv_expr (*avp, expr);
2156 if (expr2 != NULL)
2157 {
2158 /* Reset target availability on merge, since taking it only from one
2159 of the exprs would be controversial for different code. */
2160 EXPR_TARGET_AVAILABLE (expr2) = -1;
2161 EXPR_USEFULNESS (expr2) = 0;
2162
2163 merge_expr (expr2, expr, NULL);
2164
2165 /* Fix usefulness as it should be now REG_BR_PROB_BASE. */
2166 EXPR_USEFULNESS (expr2) = REG_BR_PROB_BASE;
2167
2168 av_set_iter_remove (ip);
2169 return expr2;
2170 }
2171
2172 return expr;
2173 }
2174
2175 /* Return true if there is an expr that correlates to VI in SET. */
2176 bool
2177 av_set_is_in_p (av_set_t set, vinsn_t vi)
2178 {
2179 return av_set_lookup (set, vi) != NULL;
2180 }
2181
2182 /* Return a copy of SET. */
2183 av_set_t
2184 av_set_copy (av_set_t set)
2185 {
2186 expr_t expr;
2187 av_set_iterator i;
2188 av_set_t res = NULL;
2189
2190 FOR_EACH_EXPR (expr, i, set)
2191 av_set_add (&res, expr);
2192
2193 return res;
2194 }
2195
2196 /* Join two av sets that do not have common elements by attaching second set
2197 (pointed to by FROMP) to the end of first set (TO_TAILP must point to
2198 _AV_SET_NEXT of first set's last element). */
2199 static void
2200 join_distinct_sets (av_set_t *to_tailp, av_set_t *fromp)
2201 {
2202 gcc_assert (*to_tailp == NULL);
2203 *to_tailp = *fromp;
2204 *fromp = NULL;
2205 }
2206
2207 /* Makes set pointed to by TO to be the union of TO and FROM. Clear av_set
2208 pointed to by FROMP afterwards. */
2209 void
2210 av_set_union_and_clear (av_set_t *top, av_set_t *fromp, insn_t insn)
2211 {
2212 expr_t expr1;
2213 av_set_iterator i;
2214
2215 /* Delete from TOP all exprs, that present in FROMP. */
2216 FOR_EACH_EXPR_1 (expr1, i, top)
2217 {
2218 expr_t expr2 = av_set_lookup (*fromp, EXPR_VINSN (expr1));
2219
2220 if (expr2)
2221 {
2222 merge_expr (expr2, expr1, insn);
2223 av_set_iter_remove (&i);
2224 }
2225 }
2226
2227 join_distinct_sets (i.lp, fromp);
2228 }
2229
2230 /* Same as above, but also update availability of target register in
2231 TOP judging by TO_LV_SET and FROM_LV_SET. */
2232 void
2233 av_set_union_and_live (av_set_t *top, av_set_t *fromp, regset to_lv_set,
2234 regset from_lv_set, insn_t insn)
2235 {
2236 expr_t expr1;
2237 av_set_iterator i;
2238 av_set_t *to_tailp, in_both_set = NULL;
2239
2240 /* Delete from TOP all expres, that present in FROMP. */
2241 FOR_EACH_EXPR_1 (expr1, i, top)
2242 {
2243 expr_t expr2 = av_set_lookup_and_remove (fromp, EXPR_VINSN (expr1));
2244
2245 if (expr2)
2246 {
2247 /* It may be that the expressions have different destination
2248 registers, in which case we need to check liveness here. */
2249 if (EXPR_SEPARABLE_P (expr1))
2250 {
2251 int regno1 = (REG_P (EXPR_LHS (expr1))
2252 ? (int) expr_dest_regno (expr1) : -1);
2253 int regno2 = (REG_P (EXPR_LHS (expr2))
2254 ? (int) expr_dest_regno (expr2) : -1);
2255
2256 /* ??? We don't have a way to check restrictions for
2257 *other* register on the current path, we did it only
2258 for the current target register. Give up. */
2259 if (regno1 != regno2)
2260 EXPR_TARGET_AVAILABLE (expr2) = -1;
2261 }
2262 else if (EXPR_INSN_RTX (expr1) != EXPR_INSN_RTX (expr2))
2263 EXPR_TARGET_AVAILABLE (expr2) = -1;
2264
2265 merge_expr (expr2, expr1, insn);
2266 av_set_add_nocopy (&in_both_set, expr2);
2267 av_set_iter_remove (&i);
2268 }
2269 else
2270 /* EXPR1 is present in TOP, but not in FROMP. Check it on
2271 FROM_LV_SET. */
2272 set_unavailable_target_for_expr (expr1, from_lv_set);
2273 }
2274 to_tailp = i.lp;
2275
2276 /* These expressions are not present in TOP. Check liveness
2277 restrictions on TO_LV_SET. */
2278 FOR_EACH_EXPR (expr1, i, *fromp)
2279 set_unavailable_target_for_expr (expr1, to_lv_set);
2280
2281 join_distinct_sets (i.lp, &in_both_set);
2282 join_distinct_sets (to_tailp, fromp);
2283 }
2284
2285 /* Clear av_set pointed to by SETP. */
2286 void
2287 av_set_clear (av_set_t *setp)
2288 {
2289 expr_t expr;
2290 av_set_iterator i;
2291
2292 FOR_EACH_EXPR_1 (expr, i, setp)
2293 av_set_iter_remove (&i);
2294
2295 gcc_assert (*setp == NULL);
2296 }
2297
2298 /* Leave only one non-speculative element in the SETP. */
2299 void
2300 av_set_leave_one_nonspec (av_set_t *setp)
2301 {
2302 expr_t expr;
2303 av_set_iterator i;
2304 bool has_one_nonspec = false;
2305
2306 /* Keep all speculative exprs, and leave one non-speculative
2307 (the first one). */
2308 FOR_EACH_EXPR_1 (expr, i, setp)
2309 {
2310 if (!EXPR_SPEC_DONE_DS (expr))
2311 {
2312 if (has_one_nonspec)
2313 av_set_iter_remove (&i);
2314 else
2315 has_one_nonspec = true;
2316 }
2317 }
2318 }
2319
2320 /* Return the N'th element of the SET. */
2321 expr_t
2322 av_set_element (av_set_t set, int n)
2323 {
2324 expr_t expr;
2325 av_set_iterator i;
2326
2327 FOR_EACH_EXPR (expr, i, set)
2328 if (n-- == 0)
2329 return expr;
2330
2331 gcc_unreachable ();
2332 return NULL;
2333 }
2334
2335 /* Deletes all expressions from AVP that are conditional branches (IFs). */
2336 void
2337 av_set_substract_cond_branches (av_set_t *avp)
2338 {
2339 av_set_iterator i;
2340 expr_t expr;
2341
2342 FOR_EACH_EXPR_1 (expr, i, avp)
2343 if (vinsn_cond_branch_p (EXPR_VINSN (expr)))
2344 av_set_iter_remove (&i);
2345 }
2346
2347 /* Multiplies usefulness attribute of each member of av-set *AVP by
2348 value PROB / ALL_PROB. */
2349 void
2350 av_set_split_usefulness (av_set_t av, int prob, int all_prob)
2351 {
2352 av_set_iterator i;
2353 expr_t expr;
2354
2355 FOR_EACH_EXPR (expr, i, av)
2356 EXPR_USEFULNESS (expr) = (all_prob
2357 ? (EXPR_USEFULNESS (expr) * prob) / all_prob
2358 : 0);
2359 }
2360
2361 /* Leave in AVP only those expressions, which are present in AV,
2362 and return it, merging history expressions. */
2363 void
2364 av_set_code_motion_filter (av_set_t *avp, av_set_t av)
2365 {
2366 av_set_iterator i;
2367 expr_t expr, expr2;
2368
2369 FOR_EACH_EXPR_1 (expr, i, avp)
2370 if ((expr2 = av_set_lookup (av, EXPR_VINSN (expr))) == NULL)
2371 av_set_iter_remove (&i);
2372 else
2373 /* When updating av sets in bookkeeping blocks, we can add more insns
2374 there which will be transformed but the upper av sets will not
2375 reflect those transformations. We then fail to undo those
2376 when searching for such insns. So merge the history saved
2377 in the av set of the block we are processing. */
2378 merge_history_vect (&EXPR_HISTORY_OF_CHANGES (expr),
2379 EXPR_HISTORY_OF_CHANGES (expr2));
2380 }
2381
2382 \f
2383
2384 /* Dependence hooks to initialize insn data. */
2385
2386 /* This is used in hooks callable from dependence analysis when initializing
2387 instruction's data. */
2388 static struct
2389 {
2390 /* Where the dependence was found (lhs/rhs). */
2391 deps_where_t where;
2392
2393 /* The actual data object to initialize. */
2394 idata_t id;
2395
2396 /* True when the insn should not be made clonable. */
2397 bool force_unique_p;
2398
2399 /* True when insn should be treated as of type USE, i.e. never renamed. */
2400 bool force_use_p;
2401 } deps_init_id_data;
2402
2403
2404 /* Setup ID for INSN. FORCE_UNIQUE_P is true when INSN should not be
2405 clonable. */
2406 static void
2407 setup_id_for_insn (idata_t id, insn_t insn, bool force_unique_p)
2408 {
2409 int type;
2410
2411 /* Determine whether INSN could be cloned and return appropriate vinsn type.
2412 That clonable insns which can be separated into lhs and rhs have type SET.
2413 Other clonable insns have type USE. */
2414 type = GET_CODE (insn);
2415
2416 /* Only regular insns could be cloned. */
2417 if (type == INSN && !force_unique_p)
2418 type = SET;
2419 else if (type == JUMP_INSN && simplejump_p (insn))
2420 type = PC;
2421 else if (type == DEBUG_INSN)
2422 type = !force_unique_p ? USE : INSN;
2423
2424 IDATA_TYPE (id) = type;
2425 IDATA_REG_SETS (id) = get_clear_regset_from_pool ();
2426 IDATA_REG_USES (id) = get_clear_regset_from_pool ();
2427 IDATA_REG_CLOBBERS (id) = get_clear_regset_from_pool ();
2428 }
2429
2430 /* Start initializing insn data. */
2431 static void
2432 deps_init_id_start_insn (insn_t insn)
2433 {
2434 gcc_assert (deps_init_id_data.where == DEPS_IN_NOWHERE);
2435
2436 setup_id_for_insn (deps_init_id_data.id, insn,
2437 deps_init_id_data.force_unique_p);
2438 deps_init_id_data.where = DEPS_IN_INSN;
2439 }
2440
2441 /* Start initializing lhs data. */
2442 static void
2443 deps_init_id_start_lhs (rtx lhs)
2444 {
2445 gcc_assert (deps_init_id_data.where == DEPS_IN_INSN);
2446 gcc_assert (IDATA_LHS (deps_init_id_data.id) == NULL);
2447
2448 if (IDATA_TYPE (deps_init_id_data.id) == SET)
2449 {
2450 IDATA_LHS (deps_init_id_data.id) = lhs;
2451 deps_init_id_data.where = DEPS_IN_LHS;
2452 }
2453 }
2454
2455 /* Finish initializing lhs data. */
2456 static void
2457 deps_init_id_finish_lhs (void)
2458 {
2459 deps_init_id_data.where = DEPS_IN_INSN;
2460 }
2461
2462 /* Note a set of REGNO. */
2463 static void
2464 deps_init_id_note_reg_set (int regno)
2465 {
2466 haifa_note_reg_set (regno);
2467
2468 if (deps_init_id_data.where == DEPS_IN_RHS)
2469 deps_init_id_data.force_use_p = true;
2470
2471 if (IDATA_TYPE (deps_init_id_data.id) != PC)
2472 SET_REGNO_REG_SET (IDATA_REG_SETS (deps_init_id_data.id), regno);
2473
2474 #ifdef STACK_REGS
2475 /* Make instructions that set stack registers to be ineligible for
2476 renaming to avoid issues with find_used_regs. */
2477 if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG))
2478 deps_init_id_data.force_use_p = true;
2479 #endif
2480 }
2481
2482 /* Note a clobber of REGNO. */
2483 static void
2484 deps_init_id_note_reg_clobber (int regno)
2485 {
2486 haifa_note_reg_clobber (regno);
2487
2488 if (deps_init_id_data.where == DEPS_IN_RHS)
2489 deps_init_id_data.force_use_p = true;
2490
2491 if (IDATA_TYPE (deps_init_id_data.id) != PC)
2492 SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (deps_init_id_data.id), regno);
2493 }
2494
2495 /* Note a use of REGNO. */
2496 static void
2497 deps_init_id_note_reg_use (int regno)
2498 {
2499 haifa_note_reg_use (regno);
2500
2501 if (IDATA_TYPE (deps_init_id_data.id) != PC)
2502 SET_REGNO_REG_SET (IDATA_REG_USES (deps_init_id_data.id), regno);
2503 }
2504
2505 /* Start initializing rhs data. */
2506 static void
2507 deps_init_id_start_rhs (rtx rhs)
2508 {
2509 gcc_assert (deps_init_id_data.where == DEPS_IN_INSN);
2510
2511 /* And there was no sel_deps_reset_to_insn (). */
2512 if (IDATA_LHS (deps_init_id_data.id) != NULL)
2513 {
2514 IDATA_RHS (deps_init_id_data.id) = rhs;
2515 deps_init_id_data.where = DEPS_IN_RHS;
2516 }
2517 }
2518
2519 /* Finish initializing rhs data. */
2520 static void
2521 deps_init_id_finish_rhs (void)
2522 {
2523 gcc_assert (deps_init_id_data.where == DEPS_IN_RHS
2524 || deps_init_id_data.where == DEPS_IN_INSN);
2525 deps_init_id_data.where = DEPS_IN_INSN;
2526 }
2527
2528 /* Finish initializing insn data. */
2529 static void
2530 deps_init_id_finish_insn (void)
2531 {
2532 gcc_assert (deps_init_id_data.where == DEPS_IN_INSN);
2533
2534 if (IDATA_TYPE (deps_init_id_data.id) == SET)
2535 {
2536 rtx lhs = IDATA_LHS (deps_init_id_data.id);
2537 rtx rhs = IDATA_RHS (deps_init_id_data.id);
2538
2539 if (lhs == NULL || rhs == NULL || !lhs_and_rhs_separable_p (lhs, rhs)
2540 || deps_init_id_data.force_use_p)
2541 {
2542 /* This should be a USE, as we don't want to schedule its RHS
2543 separately. However, we still want to have them recorded
2544 for the purposes of substitution. That's why we don't
2545 simply call downgrade_to_use () here. */
2546 gcc_assert (IDATA_TYPE (deps_init_id_data.id) == SET);
2547 gcc_assert (!lhs == !rhs);
2548
2549 IDATA_TYPE (deps_init_id_data.id) = USE;
2550 }
2551 }
2552
2553 deps_init_id_data.where = DEPS_IN_NOWHERE;
2554 }
2555
2556 /* This is dependence info used for initializing insn's data. */
2557 static struct sched_deps_info_def deps_init_id_sched_deps_info;
2558
2559 /* This initializes most of the static part of the above structure. */
2560 static const struct sched_deps_info_def const_deps_init_id_sched_deps_info =
2561 {
2562 NULL,
2563
2564 deps_init_id_start_insn,
2565 deps_init_id_finish_insn,
2566 deps_init_id_start_lhs,
2567 deps_init_id_finish_lhs,
2568 deps_init_id_start_rhs,
2569 deps_init_id_finish_rhs,
2570 deps_init_id_note_reg_set,
2571 deps_init_id_note_reg_clobber,
2572 deps_init_id_note_reg_use,
2573 NULL, /* note_mem_dep */
2574 NULL, /* note_dep */
2575
2576 0, /* use_cselib */
2577 0, /* use_deps_list */
2578 0 /* generate_spec_deps */
2579 };
2580
2581 /* Initialize INSN's lhs and rhs in ID. When FORCE_UNIQUE_P is true,
2582 we don't actually need information about lhs and rhs. */
2583 static void
2584 setup_id_lhs_rhs (idata_t id, insn_t insn, bool force_unique_p)
2585 {
2586 rtx pat = PATTERN (insn);
2587
2588 if (NONJUMP_INSN_P (insn)
2589 && GET_CODE (pat) == SET
2590 && !force_unique_p)
2591 {
2592 IDATA_RHS (id) = SET_SRC (pat);
2593 IDATA_LHS (id) = SET_DEST (pat);
2594 }
2595 else
2596 IDATA_LHS (id) = IDATA_RHS (id) = NULL;
2597 }
2598
2599 /* Possibly downgrade INSN to USE. */
2600 static void
2601 maybe_downgrade_id_to_use (idata_t id, insn_t insn)
2602 {
2603 bool must_be_use = false;
2604 unsigned uid = INSN_UID (insn);
2605 df_ref *rec;
2606 rtx lhs = IDATA_LHS (id);
2607 rtx rhs = IDATA_RHS (id);
2608
2609 /* We downgrade only SETs. */
2610 if (IDATA_TYPE (id) != SET)
2611 return;
2612
2613 if (!lhs || !lhs_and_rhs_separable_p (lhs, rhs))
2614 {
2615 IDATA_TYPE (id) = USE;
2616 return;
2617 }
2618
2619 for (rec = DF_INSN_UID_DEFS (uid); *rec; rec++)
2620 {
2621 df_ref def = *rec;
2622
2623 if (DF_REF_INSN (def)
2624 && DF_REF_FLAGS_IS_SET (def, DF_REF_PRE_POST_MODIFY)
2625 && loc_mentioned_in_p (DF_REF_LOC (def), IDATA_RHS (id)))
2626 {
2627 must_be_use = true;
2628 break;
2629 }
2630
2631 #ifdef STACK_REGS
2632 /* Make instructions that set stack registers to be ineligible for
2633 renaming to avoid issues with find_used_regs. */
2634 if (IN_RANGE (DF_REF_REGNO (def), FIRST_STACK_REG, LAST_STACK_REG))
2635 {
2636 must_be_use = true;
2637 break;
2638 }
2639 #endif
2640 }
2641
2642 if (must_be_use)
2643 IDATA_TYPE (id) = USE;
2644 }
2645
2646 /* Setup register sets describing INSN in ID. */
2647 static void
2648 setup_id_reg_sets (idata_t id, insn_t insn)
2649 {
2650 unsigned uid = INSN_UID (insn);
2651 df_ref *rec;
2652 regset tmp = get_clear_regset_from_pool ();
2653
2654 for (rec = DF_INSN_UID_DEFS (uid); *rec; rec++)
2655 {
2656 df_ref def = *rec;
2657 unsigned int regno = DF_REF_REGNO (def);
2658
2659 /* Post modifies are treated like clobbers by sched-deps.c. */
2660 if (DF_REF_FLAGS_IS_SET (def, (DF_REF_MUST_CLOBBER
2661 | DF_REF_PRE_POST_MODIFY)))
2662 SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (id), regno);
2663 else if (! DF_REF_FLAGS_IS_SET (def, DF_REF_MAY_CLOBBER))
2664 {
2665 SET_REGNO_REG_SET (IDATA_REG_SETS (id), regno);
2666
2667 #ifdef STACK_REGS
2668 /* For stack registers, treat writes to them as writes
2669 to the first one to be consistent with sched-deps.c. */
2670 if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG))
2671 SET_REGNO_REG_SET (IDATA_REG_SETS (id), FIRST_STACK_REG);
2672 #endif
2673 }
2674 /* Mark special refs that generate read/write def pair. */
2675 if (DF_REF_FLAGS_IS_SET (def, DF_REF_CONDITIONAL)
2676 || regno == STACK_POINTER_REGNUM)
2677 bitmap_set_bit (tmp, regno);
2678 }
2679
2680 for (rec = DF_INSN_UID_USES (uid); *rec; rec++)
2681 {
2682 df_ref use = *rec;
2683 unsigned int regno = DF_REF_REGNO (use);
2684
2685 /* When these refs are met for the first time, skip them, as
2686 these uses are just counterparts of some defs. */
2687 if (bitmap_bit_p (tmp, regno))
2688 bitmap_clear_bit (tmp, regno);
2689 else if (! DF_REF_FLAGS_IS_SET (use, DF_REF_CALL_STACK_USAGE))
2690 {
2691 SET_REGNO_REG_SET (IDATA_REG_USES (id), regno);
2692
2693 #ifdef STACK_REGS
2694 /* For stack registers, treat reads from them as reads from
2695 the first one to be consistent with sched-deps.c. */
2696 if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG))
2697 SET_REGNO_REG_SET (IDATA_REG_USES (id), FIRST_STACK_REG);
2698 #endif
2699 }
2700 }
2701
2702 return_regset_to_pool (tmp);
2703 }
2704
2705 /* Initialize instruction data for INSN in ID using DF's data. */
2706 static void
2707 init_id_from_df (idata_t id, insn_t insn, bool force_unique_p)
2708 {
2709 gcc_assert (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL);
2710
2711 setup_id_for_insn (id, insn, force_unique_p);
2712 setup_id_lhs_rhs (id, insn, force_unique_p);
2713
2714 if (INSN_NOP_P (insn))
2715 return;
2716
2717 maybe_downgrade_id_to_use (id, insn);
2718 setup_id_reg_sets (id, insn);
2719 }
2720
2721 /* Initialize instruction data for INSN in ID. */
2722 static void
2723 deps_init_id (idata_t id, insn_t insn, bool force_unique_p)
2724 {
2725 struct deps_desc _dc, *dc = &_dc;
2726
2727 deps_init_id_data.where = DEPS_IN_NOWHERE;
2728 deps_init_id_data.id = id;
2729 deps_init_id_data.force_unique_p = force_unique_p;
2730 deps_init_id_data.force_use_p = false;
2731
2732 init_deps (dc, false);
2733
2734 memcpy (&deps_init_id_sched_deps_info,
2735 &const_deps_init_id_sched_deps_info,
2736 sizeof (deps_init_id_sched_deps_info));
2737
2738 if (spec_info != NULL)
2739 deps_init_id_sched_deps_info.generate_spec_deps = 1;
2740
2741 sched_deps_info = &deps_init_id_sched_deps_info;
2742
2743 deps_analyze_insn (dc, insn);
2744
2745 free_deps (dc);
2746
2747 deps_init_id_data.id = NULL;
2748 }
2749
2750 \f
2751 struct sched_scan_info_def
2752 {
2753 /* This hook notifies scheduler frontend to extend its internal per basic
2754 block data structures. This hook should be called once before a series of
2755 calls to bb_init (). */
2756 void (*extend_bb) (void);
2757
2758 /* This hook makes scheduler frontend to initialize its internal data
2759 structures for the passed basic block. */
2760 void (*init_bb) (basic_block);
2761
2762 /* This hook notifies scheduler frontend to extend its internal per insn data
2763 structures. This hook should be called once before a series of calls to
2764 insn_init (). */
2765 void (*extend_insn) (void);
2766
2767 /* This hook makes scheduler frontend to initialize its internal data
2768 structures for the passed insn. */
2769 void (*init_insn) (rtx);
2770 };
2771
2772 /* A driver function to add a set of basic blocks (BBS) to the
2773 scheduling region. */
2774 static void
2775 sched_scan (const struct sched_scan_info_def *ssi, bb_vec_t bbs)
2776 {
2777 unsigned i;
2778 basic_block bb;
2779
2780 if (ssi->extend_bb)
2781 ssi->extend_bb ();
2782
2783 if (ssi->init_bb)
2784 FOR_EACH_VEC_ELT (bbs, i, bb)
2785 ssi->init_bb (bb);
2786
2787 if (ssi->extend_insn)
2788 ssi->extend_insn ();
2789
2790 if (ssi->init_insn)
2791 FOR_EACH_VEC_ELT (bbs, i, bb)
2792 {
2793 rtx insn;
2794
2795 FOR_BB_INSNS (bb, insn)
2796 ssi->init_insn (insn);
2797 }
2798 }
2799
2800 /* Implement hooks for collecting fundamental insn properties like if insn is
2801 an ASM or is within a SCHED_GROUP. */
2802
2803 /* True when a "one-time init" data for INSN was already inited. */
2804 static bool
2805 first_time_insn_init (insn_t insn)
2806 {
2807 return INSN_LIVE (insn) == NULL;
2808 }
2809
2810 /* Hash an entry in a transformed_insns hashtable. */
2811 static hashval_t
2812 hash_transformed_insns (const void *p)
2813 {
2814 return VINSN_HASH_RTX (((const struct transformed_insns *) p)->vinsn_old);
2815 }
2816
2817 /* Compare the entries in a transformed_insns hashtable. */
2818 static int
2819 eq_transformed_insns (const void *p, const void *q)
2820 {
2821 rtx i1 = VINSN_INSN_RTX (((const struct transformed_insns *) p)->vinsn_old);
2822 rtx i2 = VINSN_INSN_RTX (((const struct transformed_insns *) q)->vinsn_old);
2823
2824 if (INSN_UID (i1) == INSN_UID (i2))
2825 return 1;
2826 return rtx_equal_p (PATTERN (i1), PATTERN (i2));
2827 }
2828
2829 /* Free an entry in a transformed_insns hashtable. */
2830 static void
2831 free_transformed_insns (void *p)
2832 {
2833 struct transformed_insns *pti = (struct transformed_insns *) p;
2834
2835 vinsn_detach (pti->vinsn_old);
2836 vinsn_detach (pti->vinsn_new);
2837 free (pti);
2838 }
2839
2840 /* Init the s_i_d data for INSN which should be inited just once, when
2841 we first see the insn. */
2842 static void
2843 init_first_time_insn_data (insn_t insn)
2844 {
2845 /* This should not be set if this is the first time we init data for
2846 insn. */
2847 gcc_assert (first_time_insn_init (insn));
2848
2849 /* These are needed for nops too. */
2850 INSN_LIVE (insn) = get_regset_from_pool ();
2851 INSN_LIVE_VALID_P (insn) = false;
2852
2853 if (!INSN_NOP_P (insn))
2854 {
2855 INSN_ANALYZED_DEPS (insn) = BITMAP_ALLOC (NULL);
2856 INSN_FOUND_DEPS (insn) = BITMAP_ALLOC (NULL);
2857 INSN_TRANSFORMED_INSNS (insn)
2858 = htab_create (16, hash_transformed_insns,
2859 eq_transformed_insns, free_transformed_insns);
2860 init_deps (&INSN_DEPS_CONTEXT (insn), true);
2861 }
2862 }
2863
2864 /* Free almost all above data for INSN that is scheduled already.
2865 Used for extra-large basic blocks. */
2866 void
2867 free_data_for_scheduled_insn (insn_t insn)
2868 {
2869 gcc_assert (! first_time_insn_init (insn));
2870
2871 if (! INSN_ANALYZED_DEPS (insn))
2872 return;
2873
2874 BITMAP_FREE (INSN_ANALYZED_DEPS (insn));
2875 BITMAP_FREE (INSN_FOUND_DEPS (insn));
2876 htab_delete (INSN_TRANSFORMED_INSNS (insn));
2877
2878 /* This is allocated only for bookkeeping insns. */
2879 if (INSN_ORIGINATORS (insn))
2880 BITMAP_FREE (INSN_ORIGINATORS (insn));
2881 free_deps (&INSN_DEPS_CONTEXT (insn));
2882
2883 INSN_ANALYZED_DEPS (insn) = NULL;
2884
2885 /* Clear the readonly flag so we would ICE when trying to recalculate
2886 the deps context (as we believe that it should not happen). */
2887 (&INSN_DEPS_CONTEXT (insn))->readonly = 0;
2888 }
2889
2890 /* Free the same data as above for INSN. */
2891 static void
2892 free_first_time_insn_data (insn_t insn)
2893 {
2894 gcc_assert (! first_time_insn_init (insn));
2895
2896 free_data_for_scheduled_insn (insn);
2897 return_regset_to_pool (INSN_LIVE (insn));
2898 INSN_LIVE (insn) = NULL;
2899 INSN_LIVE_VALID_P (insn) = false;
2900 }
2901
2902 /* Initialize region-scope data structures for basic blocks. */
2903 static void
2904 init_global_and_expr_for_bb (basic_block bb)
2905 {
2906 if (sel_bb_empty_p (bb))
2907 return;
2908
2909 invalidate_av_set (bb);
2910 }
2911
2912 /* Data for global dependency analysis (to initialize CANT_MOVE and
2913 SCHED_GROUP_P). */
2914 static struct
2915 {
2916 /* Previous insn. */
2917 insn_t prev_insn;
2918 } init_global_data;
2919
2920 /* Determine if INSN is in the sched_group, is an asm or should not be
2921 cloned. After that initialize its expr. */
2922 static void
2923 init_global_and_expr_for_insn (insn_t insn)
2924 {
2925 if (LABEL_P (insn))
2926 return;
2927
2928 if (NOTE_INSN_BASIC_BLOCK_P (insn))
2929 {
2930 init_global_data.prev_insn = NULL_RTX;
2931 return;
2932 }
2933
2934 gcc_assert (INSN_P (insn));
2935
2936 if (SCHED_GROUP_P (insn))
2937 /* Setup a sched_group. */
2938 {
2939 insn_t prev_insn = init_global_data.prev_insn;
2940
2941 if (prev_insn)
2942 INSN_SCHED_NEXT (prev_insn) = insn;
2943
2944 init_global_data.prev_insn = insn;
2945 }
2946 else
2947 init_global_data.prev_insn = NULL_RTX;
2948
2949 if (GET_CODE (PATTERN (insn)) == ASM_INPUT
2950 || asm_noperands (PATTERN (insn)) >= 0)
2951 /* Mark INSN as an asm. */
2952 INSN_ASM_P (insn) = true;
2953
2954 {
2955 bool force_unique_p;
2956 ds_t spec_done_ds;
2957
2958 /* Certain instructions cannot be cloned, and frame related insns and
2959 the insn adjacent to NOTE_INSN_EPILOGUE_BEG cannot be moved out of
2960 their block. */
2961 if (prologue_epilogue_contains (insn))
2962 {
2963 if (RTX_FRAME_RELATED_P (insn))
2964 CANT_MOVE (insn) = 1;
2965 else
2966 {
2967 rtx note;
2968 for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
2969 if (REG_NOTE_KIND (note) == REG_SAVE_NOTE
2970 && ((enum insn_note) INTVAL (XEXP (note, 0))
2971 == NOTE_INSN_EPILOGUE_BEG))
2972 {
2973 CANT_MOVE (insn) = 1;
2974 break;
2975 }
2976 }
2977 force_unique_p = true;
2978 }
2979 else
2980 if (CANT_MOVE (insn)
2981 || INSN_ASM_P (insn)
2982 || SCHED_GROUP_P (insn)
2983 || CALL_P (insn)
2984 /* Exception handling insns are always unique. */
2985 || (cfun->can_throw_non_call_exceptions && can_throw_internal (insn))
2986 /* TRAP_IF though have an INSN code is control_flow_insn_p (). */
2987 || control_flow_insn_p (insn)
2988 || volatile_insn_p (PATTERN (insn))
2989 || (targetm.cannot_copy_insn_p
2990 && targetm.cannot_copy_insn_p (insn)))
2991 force_unique_p = true;
2992 else
2993 force_unique_p = false;
2994
2995 if (targetm.sched.get_insn_spec_ds)
2996 {
2997 spec_done_ds = targetm.sched.get_insn_spec_ds (insn);
2998 spec_done_ds = ds_get_max_dep_weak (spec_done_ds);
2999 }
3000 else
3001 spec_done_ds = 0;
3002
3003 /* Initialize INSN's expr. */
3004 init_expr (INSN_EXPR (insn), vinsn_create (insn, force_unique_p), 0,
3005 REG_BR_PROB_BASE, INSN_PRIORITY (insn), 0, BLOCK_NUM (insn),
3006 spec_done_ds, 0, 0, vNULL, true,
3007 false, false, false, CANT_MOVE (insn));
3008 }
3009
3010 init_first_time_insn_data (insn);
3011 }
3012
3013 /* Scan the region and initialize instruction data for basic blocks BBS. */
3014 void
3015 sel_init_global_and_expr (bb_vec_t bbs)
3016 {
3017 /* ??? It would be nice to implement push / pop scheme for sched_infos. */
3018 const struct sched_scan_info_def ssi =
3019 {
3020 NULL, /* extend_bb */
3021 init_global_and_expr_for_bb, /* init_bb */
3022 extend_insn_data, /* extend_insn */
3023 init_global_and_expr_for_insn /* init_insn */
3024 };
3025
3026 sched_scan (&ssi, bbs);
3027 }
3028
3029 /* Finalize region-scope data structures for basic blocks. */
3030 static void
3031 finish_global_and_expr_for_bb (basic_block bb)
3032 {
3033 av_set_clear (&BB_AV_SET (bb));
3034 BB_AV_LEVEL (bb) = 0;
3035 }
3036
3037 /* Finalize INSN's data. */
3038 static void
3039 finish_global_and_expr_insn (insn_t insn)
3040 {
3041 if (LABEL_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
3042 return;
3043
3044 gcc_assert (INSN_P (insn));
3045
3046 if (INSN_LUID (insn) > 0)
3047 {
3048 free_first_time_insn_data (insn);
3049 INSN_WS_LEVEL (insn) = 0;
3050 CANT_MOVE (insn) = 0;
3051
3052 /* We can no longer assert this, as vinsns of this insn could be
3053 easily live in other insn's caches. This should be changed to
3054 a counter-like approach among all vinsns. */
3055 gcc_assert (true || VINSN_COUNT (INSN_VINSN (insn)) == 1);
3056 clear_expr (INSN_EXPR (insn));
3057 }
3058 }
3059
3060 /* Finalize per instruction data for the whole region. */
3061 void
3062 sel_finish_global_and_expr (void)
3063 {
3064 {
3065 bb_vec_t bbs;
3066 int i;
3067
3068 bbs.create (current_nr_blocks);
3069
3070 for (i = 0; i < current_nr_blocks; i++)
3071 bbs.quick_push (BASIC_BLOCK (BB_TO_BLOCK (i)));
3072
3073 /* Clear AV_SETs and INSN_EXPRs. */
3074 {
3075 const struct sched_scan_info_def ssi =
3076 {
3077 NULL, /* extend_bb */
3078 finish_global_and_expr_for_bb, /* init_bb */
3079 NULL, /* extend_insn */
3080 finish_global_and_expr_insn /* init_insn */
3081 };
3082
3083 sched_scan (&ssi, bbs);
3084 }
3085
3086 bbs.release ();
3087 }
3088
3089 finish_insns ();
3090 }
3091 \f
3092
3093 /* In the below hooks, we merely calculate whether or not a dependence
3094 exists, and in what part of insn. However, we will need more data
3095 when we'll start caching dependence requests. */
3096
3097 /* Container to hold information for dependency analysis. */
3098 static struct
3099 {
3100 deps_t dc;
3101
3102 /* A variable to track which part of rtx we are scanning in
3103 sched-deps.c: sched_analyze_insn (). */
3104 deps_where_t where;
3105
3106 /* Current producer. */
3107 insn_t pro;
3108
3109 /* Current consumer. */
3110 vinsn_t con;
3111
3112 /* Is SEL_DEPS_HAS_DEP_P[DEPS_IN_X] is true, then X has a dependence.
3113 X is from { INSN, LHS, RHS }. */
3114 ds_t has_dep_p[DEPS_IN_NOWHERE];
3115 } has_dependence_data;
3116
3117 /* Start analyzing dependencies of INSN. */
3118 static void
3119 has_dependence_start_insn (insn_t insn ATTRIBUTE_UNUSED)
3120 {
3121 gcc_assert (has_dependence_data.where == DEPS_IN_NOWHERE);
3122
3123 has_dependence_data.where = DEPS_IN_INSN;
3124 }
3125
3126 /* Finish analyzing dependencies of an insn. */
3127 static void
3128 has_dependence_finish_insn (void)
3129 {
3130 gcc_assert (has_dependence_data.where == DEPS_IN_INSN);
3131
3132 has_dependence_data.where = DEPS_IN_NOWHERE;
3133 }
3134
3135 /* Start analyzing dependencies of LHS. */
3136 static void
3137 has_dependence_start_lhs (rtx lhs ATTRIBUTE_UNUSED)
3138 {
3139 gcc_assert (has_dependence_data.where == DEPS_IN_INSN);
3140
3141 if (VINSN_LHS (has_dependence_data.con) != NULL)
3142 has_dependence_data.where = DEPS_IN_LHS;
3143 }
3144
3145 /* Finish analyzing dependencies of an lhs. */
3146 static void
3147 has_dependence_finish_lhs (void)
3148 {
3149 has_dependence_data.where = DEPS_IN_INSN;
3150 }
3151
3152 /* Start analyzing dependencies of RHS. */
3153 static void
3154 has_dependence_start_rhs (rtx rhs ATTRIBUTE_UNUSED)
3155 {
3156 gcc_assert (has_dependence_data.where == DEPS_IN_INSN);
3157
3158 if (VINSN_RHS (has_dependence_data.con) != NULL)
3159 has_dependence_data.where = DEPS_IN_RHS;
3160 }
3161
3162 /* Start analyzing dependencies of an rhs. */
3163 static void
3164 has_dependence_finish_rhs (void)
3165 {
3166 gcc_assert (has_dependence_data.where == DEPS_IN_RHS
3167 || has_dependence_data.where == DEPS_IN_INSN);
3168
3169 has_dependence_data.where = DEPS_IN_INSN;
3170 }
3171
3172 /* Note a set of REGNO. */
3173 static void
3174 has_dependence_note_reg_set (int regno)
3175 {
3176 struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno];
3177
3178 if (!sched_insns_conditions_mutex_p (has_dependence_data.pro,
3179 VINSN_INSN_RTX
3180 (has_dependence_data.con)))
3181 {
3182 ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where];
3183
3184 if (reg_last->sets != NULL
3185 || reg_last->clobbers != NULL)
3186 *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT;
3187
3188 if (reg_last->uses || reg_last->implicit_sets)
3189 *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI;
3190 }
3191 }
3192
3193 /* Note a clobber of REGNO. */
3194 static void
3195 has_dependence_note_reg_clobber (int regno)
3196 {
3197 struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno];
3198
3199 if (!sched_insns_conditions_mutex_p (has_dependence_data.pro,
3200 VINSN_INSN_RTX
3201 (has_dependence_data.con)))
3202 {
3203 ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where];
3204
3205 if (reg_last->sets)
3206 *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT;
3207
3208 if (reg_last->uses || reg_last->implicit_sets)
3209 *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI;
3210 }
3211 }
3212
3213 /* Note a use of REGNO. */
3214 static void
3215 has_dependence_note_reg_use (int regno)
3216 {
3217 struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno];
3218
3219 if (!sched_insns_conditions_mutex_p (has_dependence_data.pro,
3220 VINSN_INSN_RTX
3221 (has_dependence_data.con)))
3222 {
3223 ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where];
3224
3225 if (reg_last->sets)
3226 *dsp = (*dsp & ~SPECULATIVE) | DEP_TRUE;
3227
3228 if (reg_last->clobbers || reg_last->implicit_sets)
3229 *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI;
3230
3231 /* Merge BE_IN_SPEC bits into *DSP when the dependency producer
3232 is actually a check insn. We need to do this for any register
3233 read-read dependency with the check unless we track properly
3234 all registers written by BE_IN_SPEC-speculated insns, as
3235 we don't have explicit dependence lists. See PR 53975. */
3236 if (reg_last->uses)
3237 {
3238 ds_t pro_spec_checked_ds;
3239
3240 pro_spec_checked_ds = INSN_SPEC_CHECKED_DS (has_dependence_data.pro);
3241 pro_spec_checked_ds = ds_get_max_dep_weak (pro_spec_checked_ds);
3242
3243 if (pro_spec_checked_ds != 0)
3244 *dsp = ds_full_merge (*dsp, pro_spec_checked_ds,
3245 NULL_RTX, NULL_RTX);
3246 }
3247 }
3248 }
3249
3250 /* Note a memory dependence. */
3251 static void
3252 has_dependence_note_mem_dep (rtx mem ATTRIBUTE_UNUSED,
3253 rtx pending_mem ATTRIBUTE_UNUSED,
3254 insn_t pending_insn ATTRIBUTE_UNUSED,
3255 ds_t ds ATTRIBUTE_UNUSED)
3256 {
3257 if (!sched_insns_conditions_mutex_p (has_dependence_data.pro,
3258 VINSN_INSN_RTX (has_dependence_data.con)))
3259 {
3260 ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where];
3261
3262 *dsp = ds_full_merge (ds, *dsp, pending_mem, mem);
3263 }
3264 }
3265
3266 /* Note a dependence. */
3267 static void
3268 has_dependence_note_dep (insn_t pro ATTRIBUTE_UNUSED,
3269 ds_t ds ATTRIBUTE_UNUSED)
3270 {
3271 if (!sched_insns_conditions_mutex_p (has_dependence_data.pro,
3272 VINSN_INSN_RTX (has_dependence_data.con)))
3273 {
3274 ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where];
3275
3276 *dsp = ds_full_merge (ds, *dsp, NULL_RTX, NULL_RTX);
3277 }
3278 }
3279
3280 /* Mark the insn as having a hard dependence that prevents speculation. */
3281 void
3282 sel_mark_hard_insn (rtx insn)
3283 {
3284 int i;
3285
3286 /* Only work when we're in has_dependence_p mode.
3287 ??? This is a hack, this should actually be a hook. */
3288 if (!has_dependence_data.dc || !has_dependence_data.pro)
3289 return;
3290
3291 gcc_assert (insn == VINSN_INSN_RTX (has_dependence_data.con));
3292 gcc_assert (has_dependence_data.where == DEPS_IN_INSN);
3293
3294 for (i = 0; i < DEPS_IN_NOWHERE; i++)
3295 has_dependence_data.has_dep_p[i] &= ~SPECULATIVE;
3296 }
3297
3298 /* This structure holds the hooks for the dependency analysis used when
3299 actually processing dependencies in the scheduler. */
3300 static struct sched_deps_info_def has_dependence_sched_deps_info;
3301
3302 /* This initializes most of the fields of the above structure. */
3303 static const struct sched_deps_info_def const_has_dependence_sched_deps_info =
3304 {
3305 NULL,
3306
3307 has_dependence_start_insn,
3308 has_dependence_finish_insn,
3309 has_dependence_start_lhs,
3310 has_dependence_finish_lhs,
3311 has_dependence_start_rhs,
3312 has_dependence_finish_rhs,
3313 has_dependence_note_reg_set,
3314 has_dependence_note_reg_clobber,
3315 has_dependence_note_reg_use,
3316 has_dependence_note_mem_dep,
3317 has_dependence_note_dep,
3318
3319 0, /* use_cselib */
3320 0, /* use_deps_list */
3321 0 /* generate_spec_deps */
3322 };
3323
3324 /* Initialize has_dependence_sched_deps_info with extra spec field. */
3325 static void
3326 setup_has_dependence_sched_deps_info (void)
3327 {
3328 memcpy (&has_dependence_sched_deps_info,
3329 &const_has_dependence_sched_deps_info,
3330 sizeof (has_dependence_sched_deps_info));
3331
3332 if (spec_info != NULL)
3333 has_dependence_sched_deps_info.generate_spec_deps = 1;
3334
3335 sched_deps_info = &has_dependence_sched_deps_info;
3336 }
3337
3338 /* Remove all dependences found and recorded in has_dependence_data array. */
3339 void
3340 sel_clear_has_dependence (void)
3341 {
3342 int i;
3343
3344 for (i = 0; i < DEPS_IN_NOWHERE; i++)
3345 has_dependence_data.has_dep_p[i] = 0;
3346 }
3347
3348 /* Return nonzero if EXPR has is dependent upon PRED. Return the pointer
3349 to the dependence information array in HAS_DEP_PP. */
3350 ds_t
3351 has_dependence_p (expr_t expr, insn_t pred, ds_t **has_dep_pp)
3352 {
3353 int i;
3354 ds_t ds;
3355 struct deps_desc *dc;
3356
3357 if (INSN_SIMPLEJUMP_P (pred))
3358 /* Unconditional jump is just a transfer of control flow.
3359 Ignore it. */
3360 return false;
3361
3362 dc = &INSN_DEPS_CONTEXT (pred);
3363
3364 /* We init this field lazily. */
3365 if (dc->reg_last == NULL)
3366 init_deps_reg_last (dc);
3367
3368 if (!dc->readonly)
3369 {
3370 has_dependence_data.pro = NULL;
3371 /* Initialize empty dep context with information about PRED. */
3372 advance_deps_context (dc, pred);
3373 dc->readonly = 1;
3374 }
3375
3376 has_dependence_data.where = DEPS_IN_NOWHERE;
3377 has_dependence_data.pro = pred;
3378 has_dependence_data.con = EXPR_VINSN (expr);
3379 has_dependence_data.dc = dc;
3380
3381 sel_clear_has_dependence ();
3382
3383 /* Now catch all dependencies that would be generated between PRED and
3384 INSN. */
3385 setup_has_dependence_sched_deps_info ();
3386 deps_analyze_insn (dc, EXPR_INSN_RTX (expr));
3387 has_dependence_data.dc = NULL;
3388
3389 /* When a barrier was found, set DEPS_IN_INSN bits. */
3390 if (dc->last_reg_pending_barrier == TRUE_BARRIER)
3391 has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_TRUE;
3392 else if (dc->last_reg_pending_barrier == MOVE_BARRIER)
3393 has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI;
3394
3395 /* Do not allow stores to memory to move through checks. Currently
3396 we don't move this to sched-deps.c as the check doesn't have
3397 obvious places to which this dependence can be attached.
3398 FIMXE: this should go to a hook. */
3399 if (EXPR_LHS (expr)
3400 && MEM_P (EXPR_LHS (expr))
3401 && sel_insn_is_speculation_check (pred))
3402 has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI;
3403
3404 *has_dep_pp = has_dependence_data.has_dep_p;
3405 ds = 0;
3406 for (i = 0; i < DEPS_IN_NOWHERE; i++)
3407 ds = ds_full_merge (ds, has_dependence_data.has_dep_p[i],
3408 NULL_RTX, NULL_RTX);
3409
3410 return ds;
3411 }
3412 \f
3413
3414 /* Dependence hooks implementation that checks dependence latency constraints
3415 on the insns being scheduled. The entry point for these routines is
3416 tick_check_p predicate. */
3417
3418 static struct
3419 {
3420 /* An expr we are currently checking. */
3421 expr_t expr;
3422
3423 /* A minimal cycle for its scheduling. */
3424 int cycle;
3425
3426 /* Whether we have seen a true dependence while checking. */
3427 bool seen_true_dep_p;
3428 } tick_check_data;
3429
3430 /* Update minimal scheduling cycle for tick_check_insn given that it depends
3431 on PRO with status DS and weight DW. */
3432 static void
3433 tick_check_dep_with_dw (insn_t pro_insn, ds_t ds, dw_t dw)
3434 {
3435 expr_t con_expr = tick_check_data.expr;
3436 insn_t con_insn = EXPR_INSN_RTX (con_expr);
3437
3438 if (con_insn != pro_insn)
3439 {
3440 enum reg_note dt;
3441 int tick;
3442
3443 if (/* PROducer was removed from above due to pipelining. */
3444 !INSN_IN_STREAM_P (pro_insn)
3445 /* Or PROducer was originally on the next iteration regarding the
3446 CONsumer. */
3447 || (INSN_SCHED_TIMES (pro_insn)
3448 - EXPR_SCHED_TIMES (con_expr)) > 1)
3449 /* Don't count this dependence. */
3450 return;
3451
3452 dt = ds_to_dt (ds);
3453 if (dt == REG_DEP_TRUE)
3454 tick_check_data.seen_true_dep_p = true;
3455
3456 gcc_assert (INSN_SCHED_CYCLE (pro_insn) > 0);
3457
3458 {
3459 dep_def _dep, *dep = &_dep;
3460
3461 init_dep (dep, pro_insn, con_insn, dt);
3462
3463 tick = INSN_SCHED_CYCLE (pro_insn) + dep_cost_1 (dep, dw);
3464 }
3465
3466 /* When there are several kinds of dependencies between pro and con,
3467 only REG_DEP_TRUE should be taken into account. */
3468 if (tick > tick_check_data.cycle
3469 && (dt == REG_DEP_TRUE || !tick_check_data.seen_true_dep_p))
3470 tick_check_data.cycle = tick;
3471 }
3472 }
3473
3474 /* An implementation of note_dep hook. */
3475 static void
3476 tick_check_note_dep (insn_t pro, ds_t ds)
3477 {
3478 tick_check_dep_with_dw (pro, ds, 0);
3479 }
3480
3481 /* An implementation of note_mem_dep hook. */
3482 static void
3483 tick_check_note_mem_dep (rtx mem1, rtx mem2, insn_t pro, ds_t ds)
3484 {
3485 dw_t dw;
3486
3487 dw = (ds_to_dt (ds) == REG_DEP_TRUE
3488 ? estimate_dep_weak (mem1, mem2)
3489 : 0);
3490
3491 tick_check_dep_with_dw (pro, ds, dw);
3492 }
3493
3494 /* This structure contains hooks for dependence analysis used when determining
3495 whether an insn is ready for scheduling. */
3496 static struct sched_deps_info_def tick_check_sched_deps_info =
3497 {
3498 NULL,
3499
3500 NULL,
3501 NULL,
3502 NULL,
3503 NULL,
3504 NULL,
3505 NULL,
3506 haifa_note_reg_set,
3507 haifa_note_reg_clobber,
3508 haifa_note_reg_use,
3509 tick_check_note_mem_dep,
3510 tick_check_note_dep,
3511
3512 0, 0, 0
3513 };
3514
3515 /* Estimate number of cycles from the current cycle of FENCE until EXPR can be
3516 scheduled. Return 0 if all data from producers in DC is ready. */
3517 int
3518 tick_check_p (expr_t expr, deps_t dc, fence_t fence)
3519 {
3520 int cycles_left;
3521 /* Initialize variables. */
3522 tick_check_data.expr = expr;
3523 tick_check_data.cycle = 0;
3524 tick_check_data.seen_true_dep_p = false;
3525 sched_deps_info = &tick_check_sched_deps_info;
3526
3527 gcc_assert (!dc->readonly);
3528 dc->readonly = 1;
3529 deps_analyze_insn (dc, EXPR_INSN_RTX (expr));
3530 dc->readonly = 0;
3531
3532 cycles_left = tick_check_data.cycle - FENCE_CYCLE (fence);
3533
3534 return cycles_left >= 0 ? cycles_left : 0;
3535 }
3536 \f
3537
3538 /* Functions to work with insns. */
3539
3540 /* Returns true if LHS of INSN is the same as DEST of an insn
3541 being moved. */
3542 bool
3543 lhs_of_insn_equals_to_dest_p (insn_t insn, rtx dest)
3544 {
3545 rtx lhs = INSN_LHS (insn);
3546
3547 if (lhs == NULL || dest == NULL)
3548 return false;
3549
3550 return rtx_equal_p (lhs, dest);
3551 }
3552
3553 /* Return s_i_d entry of INSN. Callable from debugger. */
3554 sel_insn_data_def
3555 insn_sid (insn_t insn)
3556 {
3557 return *SID (insn);
3558 }
3559
3560 /* True when INSN is a speculative check. We can tell this by looking
3561 at the data structures of the selective scheduler, not by examining
3562 the pattern. */
3563 bool
3564 sel_insn_is_speculation_check (rtx insn)
3565 {
3566 return s_i_d.exists () && !! INSN_SPEC_CHECKED_DS (insn);
3567 }
3568
3569 /* Extracts machine mode MODE and destination location DST_LOC
3570 for given INSN. */
3571 void
3572 get_dest_and_mode (rtx insn, rtx *dst_loc, enum machine_mode *mode)
3573 {
3574 rtx pat = PATTERN (insn);
3575
3576 gcc_assert (dst_loc);
3577 gcc_assert (GET_CODE (pat) == SET);
3578
3579 *dst_loc = SET_DEST (pat);
3580
3581 gcc_assert (*dst_loc);
3582 gcc_assert (MEM_P (*dst_loc) || REG_P (*dst_loc));
3583
3584 if (mode)
3585 *mode = GET_MODE (*dst_loc);
3586 }
3587
3588 /* Returns true when moving through JUMP will result in bookkeeping
3589 creation. */
3590 bool
3591 bookkeeping_can_be_created_if_moved_through_p (insn_t jump)
3592 {
3593 insn_t succ;
3594 succ_iterator si;
3595
3596 FOR_EACH_SUCC (succ, si, jump)
3597 if (sel_num_cfg_preds_gt_1 (succ))
3598 return true;
3599
3600 return false;
3601 }
3602
3603 /* Return 'true' if INSN is the only one in its basic block. */
3604 static bool
3605 insn_is_the_only_one_in_bb_p (insn_t insn)
3606 {
3607 return sel_bb_head_p (insn) && sel_bb_end_p (insn);
3608 }
3609
3610 #ifdef ENABLE_CHECKING
3611 /* Check that the region we're scheduling still has at most one
3612 backedge. */
3613 static void
3614 verify_backedges (void)
3615 {
3616 if (pipelining_p)
3617 {
3618 int i, n = 0;
3619 edge e;
3620 edge_iterator ei;
3621
3622 for (i = 0; i < current_nr_blocks; i++)
3623 FOR_EACH_EDGE (e, ei, BASIC_BLOCK (BB_TO_BLOCK (i))->succs)
3624 if (in_current_region_p (e->dest)
3625 && BLOCK_TO_BB (e->dest->index) < i)
3626 n++;
3627
3628 gcc_assert (n <= 1);
3629 }
3630 }
3631 #endif
3632 \f
3633
3634 /* Functions to work with control flow. */
3635
3636 /* Recompute BLOCK_TO_BB and BB_FOR_BLOCK for current region so that blocks
3637 are sorted in topological order (it might have been invalidated by
3638 redirecting an edge). */
3639 static void
3640 sel_recompute_toporder (void)
3641 {
3642 int i, n, rgn;
3643 int *postorder, n_blocks;
3644
3645 postorder = XALLOCAVEC (int, n_basic_blocks);
3646 n_blocks = post_order_compute (postorder, false, false);
3647
3648 rgn = CONTAINING_RGN (BB_TO_BLOCK (0));
3649 for (n = 0, i = n_blocks - 1; i >= 0; i--)
3650 if (CONTAINING_RGN (postorder[i]) == rgn)
3651 {
3652 BLOCK_TO_BB (postorder[i]) = n;
3653 BB_TO_BLOCK (n) = postorder[i];
3654 n++;
3655 }
3656
3657 /* Assert that we updated info for all blocks. We may miss some blocks if
3658 this function is called when redirecting an edge made a block
3659 unreachable, but that block is not deleted yet. */
3660 gcc_assert (n == RGN_NR_BLOCKS (rgn));
3661 }
3662
3663 /* Tidy the possibly empty block BB. */
3664 static bool
3665 maybe_tidy_empty_bb (basic_block bb)
3666 {
3667 basic_block succ_bb, pred_bb, note_bb;
3668 vec<basic_block> dom_bbs;
3669 edge e;
3670 edge_iterator ei;
3671 bool rescan_p;
3672
3673 /* Keep empty bb only if this block immediately precedes EXIT and
3674 has incoming non-fallthrough edge, or it has no predecessors or
3675 successors. Otherwise remove it. */
3676 if (!sel_bb_empty_p (bb)
3677 || (single_succ_p (bb)
3678 && single_succ (bb) == EXIT_BLOCK_PTR
3679 && (!single_pred_p (bb)
3680 || !(single_pred_edge (bb)->flags & EDGE_FALLTHRU)))
3681 || EDGE_COUNT (bb->preds) == 0
3682 || EDGE_COUNT (bb->succs) == 0)
3683 return false;
3684
3685 /* Do not attempt to redirect complex edges. */
3686 FOR_EACH_EDGE (e, ei, bb->preds)
3687 if (e->flags & EDGE_COMPLEX)
3688 return false;
3689 else if (e->flags & EDGE_FALLTHRU)
3690 {
3691 rtx note;
3692 /* If prev bb ends with asm goto, see if any of the
3693 ASM_OPERANDS_LABELs don't point to the fallthru
3694 label. Do not attempt to redirect it in that case. */
3695 if (JUMP_P (BB_END (e->src))
3696 && (note = extract_asm_operands (PATTERN (BB_END (e->src)))))
3697 {
3698 int i, n = ASM_OPERANDS_LABEL_LENGTH (note);
3699
3700 for (i = 0; i < n; ++i)
3701 if (XEXP (ASM_OPERANDS_LABEL (note, i), 0) == BB_HEAD (bb))
3702 return false;
3703 }
3704 }
3705
3706 free_data_sets (bb);
3707
3708 /* Do not delete BB if it has more than one successor.
3709 That can occur when we moving a jump. */
3710 if (!single_succ_p (bb))
3711 {
3712 gcc_assert (can_merge_blocks_p (bb->prev_bb, bb));
3713 sel_merge_blocks (bb->prev_bb, bb);
3714 return true;
3715 }
3716
3717 succ_bb = single_succ (bb);
3718 rescan_p = true;
3719 pred_bb = NULL;
3720 dom_bbs.create (0);
3721
3722 /* Save a pred/succ from the current region to attach the notes to. */
3723 note_bb = NULL;
3724 FOR_EACH_EDGE (e, ei, bb->preds)
3725 if (in_current_region_p (e->src))
3726 {
3727 note_bb = e->src;
3728 break;
3729 }
3730 if (note_bb == NULL)
3731 note_bb = succ_bb;
3732
3733 /* Redirect all non-fallthru edges to the next bb. */
3734 while (rescan_p)
3735 {
3736 rescan_p = false;
3737
3738 FOR_EACH_EDGE (e, ei, bb->preds)
3739 {
3740 pred_bb = e->src;
3741
3742 if (!(e->flags & EDGE_FALLTHRU))
3743 {
3744 /* We can not invalidate computed topological order by moving
3745 the edge destination block (E->SUCC) along a fallthru edge.
3746
3747 We will update dominators here only when we'll get
3748 an unreachable block when redirecting, otherwise
3749 sel_redirect_edge_and_branch will take care of it. */
3750 if (e->dest != bb
3751 && single_pred_p (e->dest))
3752 dom_bbs.safe_push (e->dest);
3753 sel_redirect_edge_and_branch (e, succ_bb);
3754 rescan_p = true;
3755 break;
3756 }
3757 /* If the edge is fallthru, but PRED_BB ends in a conditional jump
3758 to BB (so there is no non-fallthru edge from PRED_BB to BB), we
3759 still have to adjust it. */
3760 else if (single_succ_p (pred_bb) && any_condjump_p (BB_END (pred_bb)))
3761 {
3762 /* If possible, try to remove the unneeded conditional jump. */
3763 if (INSN_SCHED_TIMES (BB_END (pred_bb)) == 0
3764 && !IN_CURRENT_FENCE_P (BB_END (pred_bb)))
3765 {
3766 if (!sel_remove_insn (BB_END (pred_bb), false, false))
3767 tidy_fallthru_edge (e);
3768 }
3769 else
3770 sel_redirect_edge_and_branch (e, succ_bb);
3771 rescan_p = true;
3772 break;
3773 }
3774 }
3775 }
3776
3777 if (can_merge_blocks_p (bb->prev_bb, bb))
3778 sel_merge_blocks (bb->prev_bb, bb);
3779 else
3780 {
3781 /* This is a block without fallthru predecessor. Just delete it. */
3782 gcc_assert (note_bb);
3783 move_bb_info (note_bb, bb);
3784 remove_empty_bb (bb, true);
3785 }
3786
3787 if (!dom_bbs.is_empty ())
3788 {
3789 dom_bbs.safe_push (succ_bb);
3790 iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, false);
3791 dom_bbs.release ();
3792 }
3793
3794 return true;
3795 }
3796
3797 /* Tidy the control flow after we have removed original insn from
3798 XBB. Return true if we have removed some blocks. When FULL_TIDYING
3799 is true, also try to optimize control flow on non-empty blocks. */
3800 bool
3801 tidy_control_flow (basic_block xbb, bool full_tidying)
3802 {
3803 bool changed = true;
3804 insn_t first, last;
3805
3806 /* First check whether XBB is empty. */
3807 changed = maybe_tidy_empty_bb (xbb);
3808 if (changed || !full_tidying)
3809 return changed;
3810
3811 /* Check if there is a unnecessary jump after insn left. */
3812 if (bb_has_removable_jump_to_p (xbb, xbb->next_bb)
3813 && INSN_SCHED_TIMES (BB_END (xbb)) == 0
3814 && !IN_CURRENT_FENCE_P (BB_END (xbb)))
3815 {
3816 if (sel_remove_insn (BB_END (xbb), false, false))
3817 return true;
3818 tidy_fallthru_edge (EDGE_SUCC (xbb, 0));
3819 }
3820
3821 first = sel_bb_head (xbb);
3822 last = sel_bb_end (xbb);
3823 if (MAY_HAVE_DEBUG_INSNS)
3824 {
3825 if (first != last && DEBUG_INSN_P (first))
3826 do
3827 first = NEXT_INSN (first);
3828 while (first != last && (DEBUG_INSN_P (first) || NOTE_P (first)));
3829
3830 if (first != last && DEBUG_INSN_P (last))
3831 do
3832 last = PREV_INSN (last);
3833 while (first != last && (DEBUG_INSN_P (last) || NOTE_P (last)));
3834 }
3835 /* Check if there is an unnecessary jump in previous basic block leading
3836 to next basic block left after removing INSN from stream.
3837 If it is so, remove that jump and redirect edge to current
3838 basic block (where there was INSN before deletion). This way
3839 when NOP will be deleted several instructions later with its
3840 basic block we will not get a jump to next instruction, which
3841 can be harmful. */
3842 if (first == last
3843 && !sel_bb_empty_p (xbb)
3844 && INSN_NOP_P (last)
3845 /* Flow goes fallthru from current block to the next. */
3846 && EDGE_COUNT (xbb->succs) == 1
3847 && (EDGE_SUCC (xbb, 0)->flags & EDGE_FALLTHRU)
3848 /* When successor is an EXIT block, it may not be the next block. */
3849 && single_succ (xbb) != EXIT_BLOCK_PTR
3850 /* And unconditional jump in previous basic block leads to
3851 next basic block of XBB and this jump can be safely removed. */
3852 && in_current_region_p (xbb->prev_bb)
3853 && bb_has_removable_jump_to_p (xbb->prev_bb, xbb->next_bb)
3854 && INSN_SCHED_TIMES (BB_END (xbb->prev_bb)) == 0
3855 /* Also this jump is not at the scheduling boundary. */
3856 && !IN_CURRENT_FENCE_P (BB_END (xbb->prev_bb)))
3857 {
3858 bool recompute_toporder_p;
3859 /* Clear data structures of jump - jump itself will be removed
3860 by sel_redirect_edge_and_branch. */
3861 clear_expr (INSN_EXPR (BB_END (xbb->prev_bb)));
3862 recompute_toporder_p
3863 = sel_redirect_edge_and_branch (EDGE_SUCC (xbb->prev_bb, 0), xbb);
3864
3865 gcc_assert (EDGE_SUCC (xbb->prev_bb, 0)->flags & EDGE_FALLTHRU);
3866
3867 /* It can turn out that after removing unused jump, basic block
3868 that contained that jump, becomes empty too. In such case
3869 remove it too. */
3870 if (sel_bb_empty_p (xbb->prev_bb))
3871 changed = maybe_tidy_empty_bb (xbb->prev_bb);
3872 if (recompute_toporder_p)
3873 sel_recompute_toporder ();
3874 }
3875
3876 #ifdef ENABLE_CHECKING
3877 verify_backedges ();
3878 verify_dominators (CDI_DOMINATORS);
3879 #endif
3880
3881 return changed;
3882 }
3883
3884 /* Purge meaningless empty blocks in the middle of a region. */
3885 void
3886 purge_empty_blocks (void)
3887 {
3888 int i;
3889
3890 /* Do not attempt to delete the first basic block in the region. */
3891 for (i = 1; i < current_nr_blocks; )
3892 {
3893 basic_block b = BASIC_BLOCK (BB_TO_BLOCK (i));
3894
3895 if (maybe_tidy_empty_bb (b))
3896 continue;
3897
3898 i++;
3899 }
3900 }
3901
3902 /* Rip-off INSN from the insn stream. When ONLY_DISCONNECT is true,
3903 do not delete insn's data, because it will be later re-emitted.
3904 Return true if we have removed some blocks afterwards. */
3905 bool
3906 sel_remove_insn (insn_t insn, bool only_disconnect, bool full_tidying)
3907 {
3908 basic_block bb = BLOCK_FOR_INSN (insn);
3909
3910 gcc_assert (INSN_IN_STREAM_P (insn));
3911
3912 if (DEBUG_INSN_P (insn) && BB_AV_SET_VALID_P (bb))
3913 {
3914 expr_t expr;
3915 av_set_iterator i;
3916
3917 /* When we remove a debug insn that is head of a BB, it remains
3918 in the AV_SET of the block, but it shouldn't. */
3919 FOR_EACH_EXPR_1 (expr, i, &BB_AV_SET (bb))
3920 if (EXPR_INSN_RTX (expr) == insn)
3921 {
3922 av_set_iter_remove (&i);
3923 break;
3924 }
3925 }
3926
3927 if (only_disconnect)
3928 {
3929 insn_t prev = PREV_INSN (insn);
3930 insn_t next = NEXT_INSN (insn);
3931 basic_block bb = BLOCK_FOR_INSN (insn);
3932
3933 NEXT_INSN (prev) = next;
3934 PREV_INSN (next) = prev;
3935
3936 if (BB_HEAD (bb) == insn)
3937 {
3938 gcc_assert (BLOCK_FOR_INSN (prev) == bb);
3939 BB_HEAD (bb) = prev;
3940 }
3941 if (BB_END (bb) == insn)
3942 BB_END (bb) = prev;
3943 }
3944 else
3945 {
3946 remove_insn (insn);
3947 clear_expr (INSN_EXPR (insn));
3948 }
3949
3950 /* It is necessary to null this fields before calling add_insn (). */
3951 PREV_INSN (insn) = NULL_RTX;
3952 NEXT_INSN (insn) = NULL_RTX;
3953
3954 return tidy_control_flow (bb, full_tidying);
3955 }
3956
3957 /* Estimate number of the insns in BB. */
3958 static int
3959 sel_estimate_number_of_insns (basic_block bb)
3960 {
3961 int res = 0;
3962 insn_t insn = NEXT_INSN (BB_HEAD (bb)), next_tail = NEXT_INSN (BB_END (bb));
3963
3964 for (; insn != next_tail; insn = NEXT_INSN (insn))
3965 if (NONDEBUG_INSN_P (insn))
3966 res++;
3967
3968 return res;
3969 }
3970
3971 /* We don't need separate luids for notes or labels. */
3972 static int
3973 sel_luid_for_non_insn (rtx x)
3974 {
3975 gcc_assert (NOTE_P (x) || LABEL_P (x));
3976
3977 return -1;
3978 }
3979
3980 /* Find the proper seqno for inserting at INSN by successors.
3981 Return -1 if no successors with positive seqno exist. */
3982 static int
3983 get_seqno_by_succs (rtx insn)
3984 {
3985 basic_block bb = BLOCK_FOR_INSN (insn);
3986 rtx tmp = insn, end = BB_END (bb);
3987 int seqno;
3988 insn_t succ = NULL;
3989 succ_iterator si;
3990
3991 while (tmp != end)
3992 {
3993 tmp = NEXT_INSN (tmp);
3994 if (INSN_P (tmp))
3995 return INSN_SEQNO (tmp);
3996 }
3997
3998 seqno = INT_MAX;
3999
4000 FOR_EACH_SUCC_1 (succ, si, end, SUCCS_NORMAL)
4001 if (INSN_SEQNO (succ) > 0)
4002 seqno = MIN (seqno, INSN_SEQNO (succ));
4003
4004 if (seqno == INT_MAX)
4005 return -1;
4006
4007 return seqno;
4008 }
4009
4010 /* Compute seqno for INSN by its preds or succs. */
4011 static int
4012 get_seqno_for_a_jump (insn_t insn)
4013 {
4014 int seqno;
4015
4016 gcc_assert (INSN_SIMPLEJUMP_P (insn));
4017
4018 if (!sel_bb_head_p (insn))
4019 seqno = INSN_SEQNO (PREV_INSN (insn));
4020 else
4021 {
4022 basic_block bb = BLOCK_FOR_INSN (insn);
4023
4024 if (single_pred_p (bb)
4025 && !in_current_region_p (single_pred (bb)))
4026 {
4027 /* We can have preds outside a region when splitting edges
4028 for pipelining of an outer loop. Use succ instead.
4029 There should be only one of them. */
4030 insn_t succ = NULL;
4031 succ_iterator si;
4032 bool first = true;
4033
4034 gcc_assert (flag_sel_sched_pipelining_outer_loops
4035 && current_loop_nest);
4036 FOR_EACH_SUCC_1 (succ, si, insn,
4037 SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS)
4038 {
4039 gcc_assert (first);
4040 first = false;
4041 }
4042
4043 gcc_assert (succ != NULL);
4044 seqno = INSN_SEQNO (succ);
4045 }
4046 else
4047 {
4048 insn_t *preds;
4049 int n;
4050
4051 cfg_preds (BLOCK_FOR_INSN (insn), &preds, &n);
4052
4053 gcc_assert (n > 0);
4054 /* For one predecessor, use simple method. */
4055 if (n == 1)
4056 seqno = INSN_SEQNO (preds[0]);
4057 else
4058 seqno = get_seqno_by_preds (insn);
4059
4060 free (preds);
4061 }
4062 }
4063
4064 /* We were unable to find a good seqno among preds. */
4065 if (seqno < 0)
4066 seqno = get_seqno_by_succs (insn);
4067
4068 gcc_assert (seqno >= 0);
4069
4070 return seqno;
4071 }
4072
4073 /* Find the proper seqno for inserting at INSN. Returns -1 if no predecessors
4074 with positive seqno exist. */
4075 int
4076 get_seqno_by_preds (rtx insn)
4077 {
4078 basic_block bb = BLOCK_FOR_INSN (insn);
4079 rtx tmp = insn, head = BB_HEAD (bb);
4080 insn_t *preds;
4081 int n, i, seqno;
4082
4083 while (tmp != head)
4084 {
4085 tmp = PREV_INSN (tmp);
4086 if (INSN_P (tmp))
4087 return INSN_SEQNO (tmp);
4088 }
4089
4090 cfg_preds (bb, &preds, &n);
4091 for (i = 0, seqno = -1; i < n; i++)
4092 seqno = MAX (seqno, INSN_SEQNO (preds[i]));
4093
4094 return seqno;
4095 }
4096
4097 \f
4098
4099 /* Extend pass-scope data structures for basic blocks. */
4100 void
4101 sel_extend_global_bb_info (void)
4102 {
4103 sel_global_bb_info.safe_grow_cleared (last_basic_block);
4104 }
4105
4106 /* Extend region-scope data structures for basic blocks. */
4107 static void
4108 extend_region_bb_info (void)
4109 {
4110 sel_region_bb_info.safe_grow_cleared (last_basic_block);
4111 }
4112
4113 /* Extend all data structures to fit for all basic blocks. */
4114 static void
4115 extend_bb_info (void)
4116 {
4117 sel_extend_global_bb_info ();
4118 extend_region_bb_info ();
4119 }
4120
4121 /* Finalize pass-scope data structures for basic blocks. */
4122 void
4123 sel_finish_global_bb_info (void)
4124 {
4125 sel_global_bb_info.release ();
4126 }
4127
4128 /* Finalize region-scope data structures for basic blocks. */
4129 static void
4130 finish_region_bb_info (void)
4131 {
4132 sel_region_bb_info.release ();
4133 }
4134 \f
4135
4136 /* Data for each insn in current region. */
4137 vec<sel_insn_data_def> s_i_d = vNULL;
4138
4139 /* Extend data structures for insns from current region. */
4140 static void
4141 extend_insn_data (void)
4142 {
4143 int reserve;
4144
4145 sched_extend_target ();
4146 sched_deps_init (false);
4147
4148 /* Extend data structures for insns from current region. */
4149 reserve = (sched_max_luid + 1 - s_i_d.length ());
4150 if (reserve > 0 && ! s_i_d.space (reserve))
4151 {
4152 int size;
4153
4154 if (sched_max_luid / 2 > 1024)
4155 size = sched_max_luid + 1024;
4156 else
4157 size = 3 * sched_max_luid / 2;
4158
4159
4160 s_i_d.safe_grow_cleared (size);
4161 }
4162 }
4163
4164 /* Finalize data structures for insns from current region. */
4165 static void
4166 finish_insns (void)
4167 {
4168 unsigned i;
4169
4170 /* Clear here all dependence contexts that may have left from insns that were
4171 removed during the scheduling. */
4172 for (i = 0; i < s_i_d.length (); i++)
4173 {
4174 sel_insn_data_def *sid_entry = &s_i_d[i];
4175
4176 if (sid_entry->live)
4177 return_regset_to_pool (sid_entry->live);
4178 if (sid_entry->analyzed_deps)
4179 {
4180 BITMAP_FREE (sid_entry->analyzed_deps);
4181 BITMAP_FREE (sid_entry->found_deps);
4182 htab_delete (sid_entry->transformed_insns);
4183 free_deps (&sid_entry->deps_context);
4184 }
4185 if (EXPR_VINSN (&sid_entry->expr))
4186 {
4187 clear_expr (&sid_entry->expr);
4188
4189 /* Also, clear CANT_MOVE bit here, because we really don't want it
4190 to be passed to the next region. */
4191 CANT_MOVE_BY_LUID (i) = 0;
4192 }
4193 }
4194
4195 s_i_d.release ();
4196 }
4197
4198 /* A proxy to pass initialization data to init_insn (). */
4199 static sel_insn_data_def _insn_init_ssid;
4200 static sel_insn_data_t insn_init_ssid = &_insn_init_ssid;
4201
4202 /* If true create a new vinsn. Otherwise use the one from EXPR. */
4203 static bool insn_init_create_new_vinsn_p;
4204
4205 /* Set all necessary data for initialization of the new insn[s]. */
4206 static expr_t
4207 set_insn_init (expr_t expr, vinsn_t vi, int seqno)
4208 {
4209 expr_t x = &insn_init_ssid->expr;
4210
4211 copy_expr_onside (x, expr);
4212 if (vi != NULL)
4213 {
4214 insn_init_create_new_vinsn_p = false;
4215 change_vinsn_in_expr (x, vi);
4216 }
4217 else
4218 insn_init_create_new_vinsn_p = true;
4219
4220 insn_init_ssid->seqno = seqno;
4221 return x;
4222 }
4223
4224 /* Init data for INSN. */
4225 static void
4226 init_insn_data (insn_t insn)
4227 {
4228 expr_t expr;
4229 sel_insn_data_t ssid = insn_init_ssid;
4230
4231 /* The fields mentioned below are special and hence are not being
4232 propagated to the new insns. */
4233 gcc_assert (!ssid->asm_p && ssid->sched_next == NULL
4234 && !ssid->after_stall_p && ssid->sched_cycle == 0);
4235 gcc_assert (INSN_P (insn) && INSN_LUID (insn) > 0);
4236
4237 expr = INSN_EXPR (insn);
4238 copy_expr (expr, &ssid->expr);
4239 prepare_insn_expr (insn, ssid->seqno);
4240
4241 if (insn_init_create_new_vinsn_p)
4242 change_vinsn_in_expr (expr, vinsn_create (insn, init_insn_force_unique_p));
4243
4244 if (first_time_insn_init (insn))
4245 init_first_time_insn_data (insn);
4246 }
4247
4248 /* This is used to initialize spurious jumps generated by
4249 sel_redirect_edge (). */
4250 static void
4251 init_simplejump_data (insn_t insn)
4252 {
4253 init_expr (INSN_EXPR (insn), vinsn_create (insn, false), 0,
4254 REG_BR_PROB_BASE, 0, 0, 0, 0, 0, 0,
4255 vNULL, true, false, false,
4256 false, true);
4257 INSN_SEQNO (insn) = get_seqno_for_a_jump (insn);
4258 init_first_time_insn_data (insn);
4259 }
4260
4261 /* Perform deferred initialization of insns. This is used to process
4262 a new jump that may be created by redirect_edge. */
4263 void
4264 sel_init_new_insn (insn_t insn, int flags)
4265 {
4266 /* We create data structures for bb when the first insn is emitted in it. */
4267 if (INSN_P (insn)
4268 && INSN_IN_STREAM_P (insn)
4269 && insn_is_the_only_one_in_bb_p (insn))
4270 {
4271 extend_bb_info ();
4272 create_initial_data_sets (BLOCK_FOR_INSN (insn));
4273 }
4274
4275 if (flags & INSN_INIT_TODO_LUID)
4276 {
4277 sched_extend_luids ();
4278 sched_init_insn_luid (insn);
4279 }
4280
4281 if (flags & INSN_INIT_TODO_SSID)
4282 {
4283 extend_insn_data ();
4284 init_insn_data (insn);
4285 clear_expr (&insn_init_ssid->expr);
4286 }
4287
4288 if (flags & INSN_INIT_TODO_SIMPLEJUMP)
4289 {
4290 extend_insn_data ();
4291 init_simplejump_data (insn);
4292 }
4293
4294 gcc_assert (CONTAINING_RGN (BLOCK_NUM (insn))
4295 == CONTAINING_RGN (BB_TO_BLOCK (0)));
4296 }
4297 \f
4298
4299 /* Functions to init/finish work with lv sets. */
4300
4301 /* Init BB_LV_SET of BB from DF_LR_IN set of BB. */
4302 static void
4303 init_lv_set (basic_block bb)
4304 {
4305 gcc_assert (!BB_LV_SET_VALID_P (bb));
4306
4307 BB_LV_SET (bb) = get_regset_from_pool ();
4308 COPY_REG_SET (BB_LV_SET (bb), DF_LR_IN (bb));
4309 BB_LV_SET_VALID_P (bb) = true;
4310 }
4311
4312 /* Copy liveness information to BB from FROM_BB. */
4313 static void
4314 copy_lv_set_from (basic_block bb, basic_block from_bb)
4315 {
4316 gcc_assert (!BB_LV_SET_VALID_P (bb));
4317
4318 COPY_REG_SET (BB_LV_SET (bb), BB_LV_SET (from_bb));
4319 BB_LV_SET_VALID_P (bb) = true;
4320 }
4321
4322 /* Initialize lv set of all bb headers. */
4323 void
4324 init_lv_sets (void)
4325 {
4326 basic_block bb;
4327
4328 /* Initialize of LV sets. */
4329 FOR_EACH_BB (bb)
4330 init_lv_set (bb);
4331
4332 /* Don't forget EXIT_BLOCK. */
4333 init_lv_set (EXIT_BLOCK_PTR);
4334 }
4335
4336 /* Release lv set of HEAD. */
4337 static void
4338 free_lv_set (basic_block bb)
4339 {
4340 gcc_assert (BB_LV_SET (bb) != NULL);
4341
4342 return_regset_to_pool (BB_LV_SET (bb));
4343 BB_LV_SET (bb) = NULL;
4344 BB_LV_SET_VALID_P (bb) = false;
4345 }
4346
4347 /* Finalize lv sets of all bb headers. */
4348 void
4349 free_lv_sets (void)
4350 {
4351 basic_block bb;
4352
4353 /* Don't forget EXIT_BLOCK. */
4354 free_lv_set (EXIT_BLOCK_PTR);
4355
4356 /* Free LV sets. */
4357 FOR_EACH_BB (bb)
4358 if (BB_LV_SET (bb))
4359 free_lv_set (bb);
4360 }
4361
4362 /* Mark AV_SET for BB as invalid, so this set will be updated the next time
4363 compute_av() processes BB. This function is called when creating new basic
4364 blocks, as well as for blocks (either new or existing) where new jumps are
4365 created when the control flow is being updated. */
4366 static void
4367 invalidate_av_set (basic_block bb)
4368 {
4369 BB_AV_LEVEL (bb) = -1;
4370 }
4371
4372 /* Create initial data sets for BB (they will be invalid). */
4373 static void
4374 create_initial_data_sets (basic_block bb)
4375 {
4376 if (BB_LV_SET (bb))
4377 BB_LV_SET_VALID_P (bb) = false;
4378 else
4379 BB_LV_SET (bb) = get_regset_from_pool ();
4380 invalidate_av_set (bb);
4381 }
4382
4383 /* Free av set of BB. */
4384 static void
4385 free_av_set (basic_block bb)
4386 {
4387 av_set_clear (&BB_AV_SET (bb));
4388 BB_AV_LEVEL (bb) = 0;
4389 }
4390
4391 /* Free data sets of BB. */
4392 void
4393 free_data_sets (basic_block bb)
4394 {
4395 free_lv_set (bb);
4396 free_av_set (bb);
4397 }
4398
4399 /* Exchange lv sets of TO and FROM. */
4400 static void
4401 exchange_lv_sets (basic_block to, basic_block from)
4402 {
4403 {
4404 regset to_lv_set = BB_LV_SET (to);
4405
4406 BB_LV_SET (to) = BB_LV_SET (from);
4407 BB_LV_SET (from) = to_lv_set;
4408 }
4409
4410 {
4411 bool to_lv_set_valid_p = BB_LV_SET_VALID_P (to);
4412
4413 BB_LV_SET_VALID_P (to) = BB_LV_SET_VALID_P (from);
4414 BB_LV_SET_VALID_P (from) = to_lv_set_valid_p;
4415 }
4416 }
4417
4418
4419 /* Exchange av sets of TO and FROM. */
4420 static void
4421 exchange_av_sets (basic_block to, basic_block from)
4422 {
4423 {
4424 av_set_t to_av_set = BB_AV_SET (to);
4425
4426 BB_AV_SET (to) = BB_AV_SET (from);
4427 BB_AV_SET (from) = to_av_set;
4428 }
4429
4430 {
4431 int to_av_level = BB_AV_LEVEL (to);
4432
4433 BB_AV_LEVEL (to) = BB_AV_LEVEL (from);
4434 BB_AV_LEVEL (from) = to_av_level;
4435 }
4436 }
4437
4438 /* Exchange data sets of TO and FROM. */
4439 void
4440 exchange_data_sets (basic_block to, basic_block from)
4441 {
4442 exchange_lv_sets (to, from);
4443 exchange_av_sets (to, from);
4444 }
4445
4446 /* Copy data sets of FROM to TO. */
4447 void
4448 copy_data_sets (basic_block to, basic_block from)
4449 {
4450 gcc_assert (!BB_LV_SET_VALID_P (to) && !BB_AV_SET_VALID_P (to));
4451 gcc_assert (BB_AV_SET (to) == NULL);
4452
4453 BB_AV_LEVEL (to) = BB_AV_LEVEL (from);
4454 BB_LV_SET_VALID_P (to) = BB_LV_SET_VALID_P (from);
4455
4456 if (BB_AV_SET_VALID_P (from))
4457 {
4458 BB_AV_SET (to) = av_set_copy (BB_AV_SET (from));
4459 }
4460 if (BB_LV_SET_VALID_P (from))
4461 {
4462 gcc_assert (BB_LV_SET (to) != NULL);
4463 COPY_REG_SET (BB_LV_SET (to), BB_LV_SET (from));
4464 }
4465 }
4466
4467 /* Return an av set for INSN, if any. */
4468 av_set_t
4469 get_av_set (insn_t insn)
4470 {
4471 av_set_t av_set;
4472
4473 gcc_assert (AV_SET_VALID_P (insn));
4474
4475 if (sel_bb_head_p (insn))
4476 av_set = BB_AV_SET (BLOCK_FOR_INSN (insn));
4477 else
4478 av_set = NULL;
4479
4480 return av_set;
4481 }
4482
4483 /* Implementation of AV_LEVEL () macro. Return AV_LEVEL () of INSN. */
4484 int
4485 get_av_level (insn_t insn)
4486 {
4487 int av_level;
4488
4489 gcc_assert (INSN_P (insn));
4490
4491 if (sel_bb_head_p (insn))
4492 av_level = BB_AV_LEVEL (BLOCK_FOR_INSN (insn));
4493 else
4494 av_level = INSN_WS_LEVEL (insn);
4495
4496 return av_level;
4497 }
4498
4499 \f
4500
4501 /* Variables to work with control-flow graph. */
4502
4503 /* The basic block that already has been processed by the sched_data_update (),
4504 but hasn't been in sel_add_bb () yet. */
4505 static vec<basic_block>
4506 last_added_blocks = vNULL;
4507
4508 /* A pool for allocating successor infos. */
4509 static struct
4510 {
4511 /* A stack for saving succs_info structures. */
4512 struct succs_info *stack;
4513
4514 /* Its size. */
4515 int size;
4516
4517 /* Top of the stack. */
4518 int top;
4519
4520 /* Maximal value of the top. */
4521 int max_top;
4522 } succs_info_pool;
4523
4524 /* Functions to work with control-flow graph. */
4525
4526 /* Return basic block note of BB. */
4527 insn_t
4528 sel_bb_head (basic_block bb)
4529 {
4530 insn_t head;
4531
4532 if (bb == EXIT_BLOCK_PTR)
4533 {
4534 gcc_assert (exit_insn != NULL_RTX);
4535 head = exit_insn;
4536 }
4537 else
4538 {
4539 insn_t note;
4540
4541 note = bb_note (bb);
4542 head = next_nonnote_insn (note);
4543
4544 if (head && (BARRIER_P (head) || BLOCK_FOR_INSN (head) != bb))
4545 head = NULL_RTX;
4546 }
4547
4548 return head;
4549 }
4550
4551 /* Return true if INSN is a basic block header. */
4552 bool
4553 sel_bb_head_p (insn_t insn)
4554 {
4555 return sel_bb_head (BLOCK_FOR_INSN (insn)) == insn;
4556 }
4557
4558 /* Return last insn of BB. */
4559 insn_t
4560 sel_bb_end (basic_block bb)
4561 {
4562 if (sel_bb_empty_p (bb))
4563 return NULL_RTX;
4564
4565 gcc_assert (bb != EXIT_BLOCK_PTR);
4566
4567 return BB_END (bb);
4568 }
4569
4570 /* Return true if INSN is the last insn in its basic block. */
4571 bool
4572 sel_bb_end_p (insn_t insn)
4573 {
4574 return insn == sel_bb_end (BLOCK_FOR_INSN (insn));
4575 }
4576
4577 /* Return true if BB consist of single NOTE_INSN_BASIC_BLOCK. */
4578 bool
4579 sel_bb_empty_p (basic_block bb)
4580 {
4581 return sel_bb_head (bb) == NULL;
4582 }
4583
4584 /* True when BB belongs to the current scheduling region. */
4585 bool
4586 in_current_region_p (basic_block bb)
4587 {
4588 if (bb->index < NUM_FIXED_BLOCKS)
4589 return false;
4590
4591 return CONTAINING_RGN (bb->index) == CONTAINING_RGN (BB_TO_BLOCK (0));
4592 }
4593
4594 /* Return the block which is a fallthru bb of a conditional jump JUMP. */
4595 basic_block
4596 fallthru_bb_of_jump (rtx jump)
4597 {
4598 if (!JUMP_P (jump))
4599 return NULL;
4600
4601 if (!any_condjump_p (jump))
4602 return NULL;
4603
4604 /* A basic block that ends with a conditional jump may still have one successor
4605 (and be followed by a barrier), we are not interested. */
4606 if (single_succ_p (BLOCK_FOR_INSN (jump)))
4607 return NULL;
4608
4609 return FALLTHRU_EDGE (BLOCK_FOR_INSN (jump))->dest;
4610 }
4611
4612 /* Remove all notes from BB. */
4613 static void
4614 init_bb (basic_block bb)
4615 {
4616 remove_notes (bb_note (bb), BB_END (bb));
4617 BB_NOTE_LIST (bb) = note_list;
4618 }
4619
4620 void
4621 sel_init_bbs (bb_vec_t bbs)
4622 {
4623 const struct sched_scan_info_def ssi =
4624 {
4625 extend_bb_info, /* extend_bb */
4626 init_bb, /* init_bb */
4627 NULL, /* extend_insn */
4628 NULL /* init_insn */
4629 };
4630
4631 sched_scan (&ssi, bbs);
4632 }
4633
4634 /* Restore notes for the whole region. */
4635 static void
4636 sel_restore_notes (void)
4637 {
4638 int bb;
4639 insn_t insn;
4640
4641 for (bb = 0; bb < current_nr_blocks; bb++)
4642 {
4643 basic_block first, last;
4644
4645 first = EBB_FIRST_BB (bb);
4646 last = EBB_LAST_BB (bb)->next_bb;
4647
4648 do
4649 {
4650 note_list = BB_NOTE_LIST (first);
4651 restore_other_notes (NULL, first);
4652 BB_NOTE_LIST (first) = NULL_RTX;
4653
4654 FOR_BB_INSNS (first, insn)
4655 if (NONDEBUG_INSN_P (insn))
4656 reemit_notes (insn);
4657
4658 first = first->next_bb;
4659 }
4660 while (first != last);
4661 }
4662 }
4663
4664 /* Free per-bb data structures. */
4665 void
4666 sel_finish_bbs (void)
4667 {
4668 sel_restore_notes ();
4669
4670 /* Remove current loop preheader from this loop. */
4671 if (current_loop_nest)
4672 sel_remove_loop_preheader ();
4673
4674 finish_region_bb_info ();
4675 }
4676
4677 /* Return true if INSN has a single successor of type FLAGS. */
4678 bool
4679 sel_insn_has_single_succ_p (insn_t insn, int flags)
4680 {
4681 insn_t succ;
4682 succ_iterator si;
4683 bool first_p = true;
4684
4685 FOR_EACH_SUCC_1 (succ, si, insn, flags)
4686 {
4687 if (first_p)
4688 first_p = false;
4689 else
4690 return false;
4691 }
4692
4693 return true;
4694 }
4695
4696 /* Allocate successor's info. */
4697 static struct succs_info *
4698 alloc_succs_info (void)
4699 {
4700 if (succs_info_pool.top == succs_info_pool.max_top)
4701 {
4702 int i;
4703
4704 if (++succs_info_pool.max_top >= succs_info_pool.size)
4705 gcc_unreachable ();
4706
4707 i = ++succs_info_pool.top;
4708 succs_info_pool.stack[i].succs_ok.create (10);
4709 succs_info_pool.stack[i].succs_other.create (10);
4710 succs_info_pool.stack[i].probs_ok.create (10);
4711 }
4712 else
4713 succs_info_pool.top++;
4714
4715 return &succs_info_pool.stack[succs_info_pool.top];
4716 }
4717
4718 /* Free successor's info. */
4719 void
4720 free_succs_info (struct succs_info * sinfo)
4721 {
4722 gcc_assert (succs_info_pool.top >= 0
4723 && &succs_info_pool.stack[succs_info_pool.top] == sinfo);
4724 succs_info_pool.top--;
4725
4726 /* Clear stale info. */
4727 sinfo->succs_ok.block_remove (0, sinfo->succs_ok.length ());
4728 sinfo->succs_other.block_remove (0, sinfo->succs_other.length ());
4729 sinfo->probs_ok.block_remove (0, sinfo->probs_ok.length ());
4730 sinfo->all_prob = 0;
4731 sinfo->succs_ok_n = 0;
4732 sinfo->all_succs_n = 0;
4733 }
4734
4735 /* Compute successor info for INSN. FLAGS are the flags passed
4736 to the FOR_EACH_SUCC_1 iterator. */
4737 struct succs_info *
4738 compute_succs_info (insn_t insn, short flags)
4739 {
4740 succ_iterator si;
4741 insn_t succ;
4742 struct succs_info *sinfo = alloc_succs_info ();
4743
4744 /* Traverse *all* successors and decide what to do with each. */
4745 FOR_EACH_SUCC_1 (succ, si, insn, SUCCS_ALL)
4746 {
4747 /* FIXME: this doesn't work for skipping to loop exits, as we don't
4748 perform code motion through inner loops. */
4749 short current_flags = si.current_flags & ~SUCCS_SKIP_TO_LOOP_EXITS;
4750
4751 if (current_flags & flags)
4752 {
4753 sinfo->succs_ok.safe_push (succ);
4754 sinfo->probs_ok.safe_push (
4755 /* FIXME: Improve calculation when skipping
4756 inner loop to exits. */
4757 si.bb_end ? si.e1->probability : REG_BR_PROB_BASE);
4758 sinfo->succs_ok_n++;
4759 }
4760 else
4761 sinfo->succs_other.safe_push (succ);
4762
4763 /* Compute all_prob. */
4764 if (!si.bb_end)
4765 sinfo->all_prob = REG_BR_PROB_BASE;
4766 else
4767 sinfo->all_prob += si.e1->probability;
4768
4769 sinfo->all_succs_n++;
4770 }
4771
4772 return sinfo;
4773 }
4774
4775 /* Return the predecessors of BB in PREDS and their number in N.
4776 Empty blocks are skipped. SIZE is used to allocate PREDS. */
4777 static void
4778 cfg_preds_1 (basic_block bb, insn_t **preds, int *n, int *size)
4779 {
4780 edge e;
4781 edge_iterator ei;
4782
4783 gcc_assert (BLOCK_TO_BB (bb->index) != 0);
4784
4785 FOR_EACH_EDGE (e, ei, bb->preds)
4786 {
4787 basic_block pred_bb = e->src;
4788 insn_t bb_end = BB_END (pred_bb);
4789
4790 if (!in_current_region_p (pred_bb))
4791 {
4792 gcc_assert (flag_sel_sched_pipelining_outer_loops
4793 && current_loop_nest);
4794 continue;
4795 }
4796
4797 if (sel_bb_empty_p (pred_bb))
4798 cfg_preds_1 (pred_bb, preds, n, size);
4799 else
4800 {
4801 if (*n == *size)
4802 *preds = XRESIZEVEC (insn_t, *preds,
4803 (*size = 2 * *size + 1));
4804 (*preds)[(*n)++] = bb_end;
4805 }
4806 }
4807
4808 gcc_assert (*n != 0
4809 || (flag_sel_sched_pipelining_outer_loops
4810 && current_loop_nest));
4811 }
4812
4813 /* Find all predecessors of BB and record them in PREDS and their number
4814 in N. Empty blocks are skipped, and only normal (forward in-region)
4815 edges are processed. */
4816 static void
4817 cfg_preds (basic_block bb, insn_t **preds, int *n)
4818 {
4819 int size = 0;
4820
4821 *preds = NULL;
4822 *n = 0;
4823 cfg_preds_1 (bb, preds, n, &size);
4824 }
4825
4826 /* Returns true if we are moving INSN through join point. */
4827 bool
4828 sel_num_cfg_preds_gt_1 (insn_t insn)
4829 {
4830 basic_block bb;
4831
4832 if (!sel_bb_head_p (insn) || INSN_BB (insn) == 0)
4833 return false;
4834
4835 bb = BLOCK_FOR_INSN (insn);
4836
4837 while (1)
4838 {
4839 if (EDGE_COUNT (bb->preds) > 1)
4840 return true;
4841
4842 gcc_assert (EDGE_PRED (bb, 0)->dest == bb);
4843 bb = EDGE_PRED (bb, 0)->src;
4844
4845 if (!sel_bb_empty_p (bb))
4846 break;
4847 }
4848
4849 return false;
4850 }
4851
4852 /* Returns true when BB should be the end of an ebb. Adapted from the
4853 code in sched-ebb.c. */
4854 bool
4855 bb_ends_ebb_p (basic_block bb)
4856 {
4857 basic_block next_bb = bb_next_bb (bb);
4858 edge e;
4859
4860 if (next_bb == EXIT_BLOCK_PTR
4861 || bitmap_bit_p (forced_ebb_heads, next_bb->index)
4862 || (LABEL_P (BB_HEAD (next_bb))
4863 /* NB: LABEL_NUSES () is not maintained outside of jump.c.
4864 Work around that. */
4865 && !single_pred_p (next_bb)))
4866 return true;
4867
4868 if (!in_current_region_p (next_bb))
4869 return true;
4870
4871 e = find_fallthru_edge (bb->succs);
4872 if (e)
4873 {
4874 gcc_assert (e->dest == next_bb);
4875
4876 return false;
4877 }
4878
4879 return true;
4880 }
4881
4882 /* Returns true when INSN and SUCC are in the same EBB, given that SUCC is a
4883 successor of INSN. */
4884 bool
4885 in_same_ebb_p (insn_t insn, insn_t succ)
4886 {
4887 basic_block ptr = BLOCK_FOR_INSN (insn);
4888
4889 for(;;)
4890 {
4891 if (ptr == BLOCK_FOR_INSN (succ))
4892 return true;
4893
4894 if (bb_ends_ebb_p (ptr))
4895 return false;
4896
4897 ptr = bb_next_bb (ptr);
4898 }
4899
4900 gcc_unreachable ();
4901 return false;
4902 }
4903
4904 /* Recomputes the reverse topological order for the function and
4905 saves it in REV_TOP_ORDER_INDEX. REV_TOP_ORDER_INDEX_LEN is also
4906 modified appropriately. */
4907 static void
4908 recompute_rev_top_order (void)
4909 {
4910 int *postorder;
4911 int n_blocks, i;
4912
4913 if (!rev_top_order_index || rev_top_order_index_len < last_basic_block)
4914 {
4915 rev_top_order_index_len = last_basic_block;
4916 rev_top_order_index = XRESIZEVEC (int, rev_top_order_index,
4917 rev_top_order_index_len);
4918 }
4919
4920 postorder = XNEWVEC (int, n_basic_blocks);
4921
4922 n_blocks = post_order_compute (postorder, true, false);
4923 gcc_assert (n_basic_blocks == n_blocks);
4924
4925 /* Build reverse function: for each basic block with BB->INDEX == K
4926 rev_top_order_index[K] is it's reverse topological sort number. */
4927 for (i = 0; i < n_blocks; i++)
4928 {
4929 gcc_assert (postorder[i] < rev_top_order_index_len);
4930 rev_top_order_index[postorder[i]] = i;
4931 }
4932
4933 free (postorder);
4934 }
4935
4936 /* Clear all flags from insns in BB that could spoil its rescheduling. */
4937 void
4938 clear_outdated_rtx_info (basic_block bb)
4939 {
4940 rtx insn;
4941
4942 FOR_BB_INSNS (bb, insn)
4943 if (INSN_P (insn))
4944 {
4945 SCHED_GROUP_P (insn) = 0;
4946 INSN_AFTER_STALL_P (insn) = 0;
4947 INSN_SCHED_TIMES (insn) = 0;
4948 EXPR_PRIORITY_ADJ (INSN_EXPR (insn)) = 0;
4949
4950 /* We cannot use the changed caches, as previously we could ignore
4951 the LHS dependence due to enabled renaming and transform
4952 the expression, and currently we'll be unable to do this. */
4953 htab_empty (INSN_TRANSFORMED_INSNS (insn));
4954 }
4955 }
4956
4957 /* Add BB_NOTE to the pool of available basic block notes. */
4958 static void
4959 return_bb_to_pool (basic_block bb)
4960 {
4961 rtx note = bb_note (bb);
4962
4963 gcc_assert (NOTE_BASIC_BLOCK (note) == bb
4964 && bb->aux == NULL);
4965
4966 /* It turns out that current cfg infrastructure does not support
4967 reuse of basic blocks. Don't bother for now. */
4968 /*bb_note_pool.safe_push (note);*/
4969 }
4970
4971 /* Get a bb_note from pool or return NULL_RTX if pool is empty. */
4972 static rtx
4973 get_bb_note_from_pool (void)
4974 {
4975 if (bb_note_pool.is_empty ())
4976 return NULL_RTX;
4977 else
4978 {
4979 rtx note = bb_note_pool.pop ();
4980
4981 PREV_INSN (note) = NULL_RTX;
4982 NEXT_INSN (note) = NULL_RTX;
4983
4984 return note;
4985 }
4986 }
4987
4988 /* Free bb_note_pool. */
4989 void
4990 free_bb_note_pool (void)
4991 {
4992 bb_note_pool.release ();
4993 }
4994
4995 /* Setup scheduler pool and successor structure. */
4996 void
4997 alloc_sched_pools (void)
4998 {
4999 int succs_size;
5000
5001 succs_size = MAX_WS + 1;
5002 succs_info_pool.stack = XCNEWVEC (struct succs_info, succs_size);
5003 succs_info_pool.size = succs_size;
5004 succs_info_pool.top = -1;
5005 succs_info_pool.max_top = -1;
5006
5007 sched_lists_pool = create_alloc_pool ("sel-sched-lists",
5008 sizeof (struct _list_node), 500);
5009 }
5010
5011 /* Free the pools. */
5012 void
5013 free_sched_pools (void)
5014 {
5015 int i;
5016
5017 free_alloc_pool (sched_lists_pool);
5018 gcc_assert (succs_info_pool.top == -1);
5019 for (i = 0; i < succs_info_pool.max_top; i++)
5020 {
5021 succs_info_pool.stack[i].succs_ok.release ();
5022 succs_info_pool.stack[i].succs_other.release ();
5023 succs_info_pool.stack[i].probs_ok.release ();
5024 }
5025 free (succs_info_pool.stack);
5026 }
5027 \f
5028
5029 /* Returns a position in RGN where BB can be inserted retaining
5030 topological order. */
5031 static int
5032 find_place_to_insert_bb (basic_block bb, int rgn)
5033 {
5034 bool has_preds_outside_rgn = false;
5035 edge e;
5036 edge_iterator ei;
5037
5038 /* Find whether we have preds outside the region. */
5039 FOR_EACH_EDGE (e, ei, bb->preds)
5040 if (!in_current_region_p (e->src))
5041 {
5042 has_preds_outside_rgn = true;
5043 break;
5044 }
5045
5046 /* Recompute the top order -- needed when we have > 1 pred
5047 and in case we don't have preds outside. */
5048 if (flag_sel_sched_pipelining_outer_loops
5049 && (has_preds_outside_rgn || EDGE_COUNT (bb->preds) > 1))
5050 {
5051 int i, bbi = bb->index, cur_bbi;
5052
5053 recompute_rev_top_order ();
5054 for (i = RGN_NR_BLOCKS (rgn) - 1; i >= 0; i--)
5055 {
5056 cur_bbi = BB_TO_BLOCK (i);
5057 if (rev_top_order_index[bbi]
5058 < rev_top_order_index[cur_bbi])
5059 break;
5060 }
5061
5062 /* We skipped the right block, so we increase i. We accommodate
5063 it for increasing by step later, so we decrease i. */
5064 return (i + 1) - 1;
5065 }
5066 else if (has_preds_outside_rgn)
5067 {
5068 /* This is the case when we generate an extra empty block
5069 to serve as region head during pipelining. */
5070 e = EDGE_SUCC (bb, 0);
5071 gcc_assert (EDGE_COUNT (bb->succs) == 1
5072 && in_current_region_p (EDGE_SUCC (bb, 0)->dest)
5073 && (BLOCK_TO_BB (e->dest->index) == 0));
5074 return -1;
5075 }
5076
5077 /* We don't have preds outside the region. We should have
5078 the only pred, because the multiple preds case comes from
5079 the pipelining of outer loops, and that is handled above.
5080 Just take the bbi of this single pred. */
5081 if (EDGE_COUNT (bb->succs) > 0)
5082 {
5083 int pred_bbi;
5084
5085 gcc_assert (EDGE_COUNT (bb->preds) == 1);
5086
5087 pred_bbi = EDGE_PRED (bb, 0)->src->index;
5088 return BLOCK_TO_BB (pred_bbi);
5089 }
5090 else
5091 /* BB has no successors. It is safe to put it in the end. */
5092 return current_nr_blocks - 1;
5093 }
5094
5095 /* Deletes an empty basic block freeing its data. */
5096 static void
5097 delete_and_free_basic_block (basic_block bb)
5098 {
5099 gcc_assert (sel_bb_empty_p (bb));
5100
5101 if (BB_LV_SET (bb))
5102 free_lv_set (bb);
5103
5104 bitmap_clear_bit (blocks_to_reschedule, bb->index);
5105
5106 /* Can't assert av_set properties because we use sel_aremove_bb
5107 when removing loop preheader from the region. At the point of
5108 removing the preheader we already have deallocated sel_region_bb_info. */
5109 gcc_assert (BB_LV_SET (bb) == NULL
5110 && !BB_LV_SET_VALID_P (bb)
5111 && BB_AV_LEVEL (bb) == 0
5112 && BB_AV_SET (bb) == NULL);
5113
5114 delete_basic_block (bb);
5115 }
5116
5117 /* Add BB to the current region and update the region data. */
5118 static void
5119 add_block_to_current_region (basic_block bb)
5120 {
5121 int i, pos, bbi = -2, rgn;
5122
5123 rgn = CONTAINING_RGN (BB_TO_BLOCK (0));
5124 bbi = find_place_to_insert_bb (bb, rgn);
5125 bbi += 1;
5126 pos = RGN_BLOCKS (rgn) + bbi;
5127
5128 gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0
5129 && ebb_head[bbi] == pos);
5130
5131 /* Make a place for the new block. */
5132 extend_regions ();
5133
5134 for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--)
5135 BLOCK_TO_BB (rgn_bb_table[i])++;
5136
5137 memmove (rgn_bb_table + pos + 1,
5138 rgn_bb_table + pos,
5139 (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table));
5140
5141 /* Initialize data for BB. */
5142 rgn_bb_table[pos] = bb->index;
5143 BLOCK_TO_BB (bb->index) = bbi;
5144 CONTAINING_RGN (bb->index) = rgn;
5145
5146 RGN_NR_BLOCKS (rgn)++;
5147
5148 for (i = rgn + 1; i <= nr_regions; i++)
5149 RGN_BLOCKS (i)++;
5150 }
5151
5152 /* Remove BB from the current region and update the region data. */
5153 static void
5154 remove_bb_from_region (basic_block bb)
5155 {
5156 int i, pos, bbi = -2, rgn;
5157
5158 rgn = CONTAINING_RGN (BB_TO_BLOCK (0));
5159 bbi = BLOCK_TO_BB (bb->index);
5160 pos = RGN_BLOCKS (rgn) + bbi;
5161
5162 gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0
5163 && ebb_head[bbi] == pos);
5164
5165 for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--)
5166 BLOCK_TO_BB (rgn_bb_table[i])--;
5167
5168 memmove (rgn_bb_table + pos,
5169 rgn_bb_table + pos + 1,
5170 (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table));
5171
5172 RGN_NR_BLOCKS (rgn)--;
5173 for (i = rgn + 1; i <= nr_regions; i++)
5174 RGN_BLOCKS (i)--;
5175 }
5176
5177 /* Add BB to the current region and update all data. If BB is NULL, add all
5178 blocks from last_added_blocks vector. */
5179 static void
5180 sel_add_bb (basic_block bb)
5181 {
5182 /* Extend luids so that new notes will receive zero luids. */
5183 sched_extend_luids ();
5184 sched_init_bbs ();
5185 sel_init_bbs (last_added_blocks);
5186
5187 /* When bb is passed explicitly, the vector should contain
5188 the only element that equals to bb; otherwise, the vector
5189 should not be NULL. */
5190 gcc_assert (last_added_blocks.exists ());
5191
5192 if (bb != NULL)
5193 {
5194 gcc_assert (last_added_blocks.length () == 1
5195 && last_added_blocks[0] == bb);
5196 add_block_to_current_region (bb);
5197
5198 /* We associate creating/deleting data sets with the first insn
5199 appearing / disappearing in the bb. */
5200 if (!sel_bb_empty_p (bb) && BB_LV_SET (bb) == NULL)
5201 create_initial_data_sets (bb);
5202
5203 last_added_blocks.release ();
5204 }
5205 else
5206 /* BB is NULL - process LAST_ADDED_BLOCKS instead. */
5207 {
5208 int i;
5209 basic_block temp_bb = NULL;
5210
5211 for (i = 0;
5212 last_added_blocks.iterate (i, &bb); i++)
5213 {
5214 add_block_to_current_region (bb);
5215 temp_bb = bb;
5216 }
5217
5218 /* We need to fetch at least one bb so we know the region
5219 to update. */
5220 gcc_assert (temp_bb != NULL);
5221 bb = temp_bb;
5222
5223 last_added_blocks.release ();
5224 }
5225
5226 rgn_setup_region (CONTAINING_RGN (bb->index));
5227 }
5228
5229 /* Remove BB from the current region and update all data.
5230 If REMOVE_FROM_CFG_PBB is true, also remove the block cfom cfg. */
5231 static void
5232 sel_remove_bb (basic_block bb, bool remove_from_cfg_p)
5233 {
5234 unsigned idx = bb->index;
5235
5236 gcc_assert (bb != NULL && BB_NOTE_LIST (bb) == NULL_RTX);
5237
5238 remove_bb_from_region (bb);
5239 return_bb_to_pool (bb);
5240 bitmap_clear_bit (blocks_to_reschedule, idx);
5241
5242 if (remove_from_cfg_p)
5243 {
5244 basic_block succ = single_succ (bb);
5245 delete_and_free_basic_block (bb);
5246 set_immediate_dominator (CDI_DOMINATORS, succ,
5247 recompute_dominator (CDI_DOMINATORS, succ));
5248 }
5249
5250 rgn_setup_region (CONTAINING_RGN (idx));
5251 }
5252
5253 /* Concatenate info of EMPTY_BB to info of MERGE_BB. */
5254 static void
5255 move_bb_info (basic_block merge_bb, basic_block empty_bb)
5256 {
5257 if (in_current_region_p (merge_bb))
5258 concat_note_lists (BB_NOTE_LIST (empty_bb),
5259 &BB_NOTE_LIST (merge_bb));
5260 BB_NOTE_LIST (empty_bb) = NULL_RTX;
5261
5262 }
5263
5264 /* Remove EMPTY_BB. If REMOVE_FROM_CFG_P is false, remove EMPTY_BB from
5265 region, but keep it in CFG. */
5266 static void
5267 remove_empty_bb (basic_block empty_bb, bool remove_from_cfg_p)
5268 {
5269 /* The block should contain just a note or a label.
5270 We try to check whether it is unused below. */
5271 gcc_assert (BB_HEAD (empty_bb) == BB_END (empty_bb)
5272 || LABEL_P (BB_HEAD (empty_bb)));
5273
5274 /* If basic block has predecessors or successors, redirect them. */
5275 if (remove_from_cfg_p
5276 && (EDGE_COUNT (empty_bb->preds) > 0
5277 || EDGE_COUNT (empty_bb->succs) > 0))
5278 {
5279 basic_block pred;
5280 basic_block succ;
5281
5282 /* We need to init PRED and SUCC before redirecting edges. */
5283 if (EDGE_COUNT (empty_bb->preds) > 0)
5284 {
5285 edge e;
5286
5287 gcc_assert (EDGE_COUNT (empty_bb->preds) == 1);
5288
5289 e = EDGE_PRED (empty_bb, 0);
5290 gcc_assert (e->src == empty_bb->prev_bb
5291 && (e->flags & EDGE_FALLTHRU));
5292
5293 pred = empty_bb->prev_bb;
5294 }
5295 else
5296 pred = NULL;
5297
5298 if (EDGE_COUNT (empty_bb->succs) > 0)
5299 {
5300 /* We do not check fallthruness here as above, because
5301 after removing a jump the edge may actually be not fallthru. */
5302 gcc_assert (EDGE_COUNT (empty_bb->succs) == 1);
5303 succ = EDGE_SUCC (empty_bb, 0)->dest;
5304 }
5305 else
5306 succ = NULL;
5307
5308 if (EDGE_COUNT (empty_bb->preds) > 0 && succ != NULL)
5309 {
5310 edge e = EDGE_PRED (empty_bb, 0);
5311
5312 if (e->flags & EDGE_FALLTHRU)
5313 redirect_edge_succ_nodup (e, succ);
5314 else
5315 sel_redirect_edge_and_branch (EDGE_PRED (empty_bb, 0), succ);
5316 }
5317
5318 if (EDGE_COUNT (empty_bb->succs) > 0 && pred != NULL)
5319 {
5320 edge e = EDGE_SUCC (empty_bb, 0);
5321
5322 if (find_edge (pred, e->dest) == NULL)
5323 redirect_edge_pred (e, pred);
5324 }
5325 }
5326
5327 /* Finish removing. */
5328 sel_remove_bb (empty_bb, remove_from_cfg_p);
5329 }
5330
5331 /* An implementation of create_basic_block hook, which additionally updates
5332 per-bb data structures. */
5333 static basic_block
5334 sel_create_basic_block (void *headp, void *endp, basic_block after)
5335 {
5336 basic_block new_bb;
5337 insn_t new_bb_note;
5338
5339 gcc_assert (flag_sel_sched_pipelining_outer_loops
5340 || !last_added_blocks.exists ());
5341
5342 new_bb_note = get_bb_note_from_pool ();
5343
5344 if (new_bb_note == NULL_RTX)
5345 new_bb = orig_cfg_hooks.create_basic_block (headp, endp, after);
5346 else
5347 {
5348 new_bb = create_basic_block_structure ((rtx) headp, (rtx) endp,
5349 new_bb_note, after);
5350 new_bb->aux = NULL;
5351 }
5352
5353 last_added_blocks.safe_push (new_bb);
5354
5355 return new_bb;
5356 }
5357
5358 /* Implement sched_init_only_bb (). */
5359 static void
5360 sel_init_only_bb (basic_block bb, basic_block after)
5361 {
5362 gcc_assert (after == NULL);
5363
5364 extend_regions ();
5365 rgn_make_new_region_out_of_new_block (bb);
5366 }
5367
5368 /* Update the latch when we've splitted or merged it from FROM block to TO.
5369 This should be checked for all outer loops, too. */
5370 static void
5371 change_loops_latches (basic_block from, basic_block to)
5372 {
5373 gcc_assert (from != to);
5374
5375 if (current_loop_nest)
5376 {
5377 struct loop *loop;
5378
5379 for (loop = current_loop_nest; loop; loop = loop_outer (loop))
5380 if (considered_for_pipelining_p (loop) && loop->latch == from)
5381 {
5382 gcc_assert (loop == current_loop_nest);
5383 loop->latch = to;
5384 gcc_assert (loop_latch_edge (loop));
5385 }
5386 }
5387 }
5388
5389 /* Splits BB on two basic blocks, adding it to the region and extending
5390 per-bb data structures. Returns the newly created bb. */
5391 static basic_block
5392 sel_split_block (basic_block bb, rtx after)
5393 {
5394 basic_block new_bb;
5395 insn_t insn;
5396
5397 new_bb = sched_split_block_1 (bb, after);
5398 sel_add_bb (new_bb);
5399
5400 /* This should be called after sel_add_bb, because this uses
5401 CONTAINING_RGN for the new block, which is not yet initialized.
5402 FIXME: this function may be a no-op now. */
5403 change_loops_latches (bb, new_bb);
5404
5405 /* Update ORIG_BB_INDEX for insns moved into the new block. */
5406 FOR_BB_INSNS (new_bb, insn)
5407 if (INSN_P (insn))
5408 EXPR_ORIG_BB_INDEX (INSN_EXPR (insn)) = new_bb->index;
5409
5410 if (sel_bb_empty_p (bb))
5411 {
5412 gcc_assert (!sel_bb_empty_p (new_bb));
5413
5414 /* NEW_BB has data sets that need to be updated and BB holds
5415 data sets that should be removed. Exchange these data sets
5416 so that we won't lose BB's valid data sets. */
5417 exchange_data_sets (new_bb, bb);
5418 free_data_sets (bb);
5419 }
5420
5421 if (!sel_bb_empty_p (new_bb)
5422 && bitmap_bit_p (blocks_to_reschedule, bb->index))
5423 bitmap_set_bit (blocks_to_reschedule, new_bb->index);
5424
5425 return new_bb;
5426 }
5427
5428 /* If BB ends with a jump insn whose ID is bigger then PREV_MAX_UID, return it.
5429 Otherwise returns NULL. */
5430 static rtx
5431 check_for_new_jump (basic_block bb, int prev_max_uid)
5432 {
5433 rtx end;
5434
5435 end = sel_bb_end (bb);
5436 if (end && INSN_UID (end) >= prev_max_uid)
5437 return end;
5438 return NULL;
5439 }
5440
5441 /* Look for a new jump either in FROM_BB block or in newly created JUMP_BB block.
5442 New means having UID at least equal to PREV_MAX_UID. */
5443 static rtx
5444 find_new_jump (basic_block from, basic_block jump_bb, int prev_max_uid)
5445 {
5446 rtx jump;
5447
5448 /* Return immediately if no new insns were emitted. */
5449 if (get_max_uid () == prev_max_uid)
5450 return NULL;
5451
5452 /* Now check both blocks for new jumps. It will ever be only one. */
5453 if ((jump = check_for_new_jump (from, prev_max_uid)))
5454 return jump;
5455
5456 if (jump_bb != NULL
5457 && (jump = check_for_new_jump (jump_bb, prev_max_uid)))
5458 return jump;
5459 return NULL;
5460 }
5461
5462 /* Splits E and adds the newly created basic block to the current region.
5463 Returns this basic block. */
5464 basic_block
5465 sel_split_edge (edge e)
5466 {
5467 basic_block new_bb, src, other_bb = NULL;
5468 int prev_max_uid;
5469 rtx jump;
5470
5471 src = e->src;
5472 prev_max_uid = get_max_uid ();
5473 new_bb = split_edge (e);
5474
5475 if (flag_sel_sched_pipelining_outer_loops
5476 && current_loop_nest)
5477 {
5478 int i;
5479 basic_block bb;
5480
5481 /* Some of the basic blocks might not have been added to the loop.
5482 Add them here, until this is fixed in force_fallthru. */
5483 for (i = 0;
5484 last_added_blocks.iterate (i, &bb); i++)
5485 if (!bb->loop_father)
5486 {
5487 add_bb_to_loop (bb, e->dest->loop_father);
5488
5489 gcc_assert (!other_bb && (new_bb->index != bb->index));
5490 other_bb = bb;
5491 }
5492 }
5493
5494 /* Add all last_added_blocks to the region. */
5495 sel_add_bb (NULL);
5496
5497 jump = find_new_jump (src, new_bb, prev_max_uid);
5498 if (jump)
5499 sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP);
5500
5501 /* Put the correct lv set on this block. */
5502 if (other_bb && !sel_bb_empty_p (other_bb))
5503 compute_live (sel_bb_head (other_bb));
5504
5505 return new_bb;
5506 }
5507
5508 /* Implement sched_create_empty_bb (). */
5509 static basic_block
5510 sel_create_empty_bb (basic_block after)
5511 {
5512 basic_block new_bb;
5513
5514 new_bb = sched_create_empty_bb_1 (after);
5515
5516 /* We'll explicitly initialize NEW_BB via sel_init_only_bb () a bit
5517 later. */
5518 gcc_assert (last_added_blocks.length () == 1
5519 && last_added_blocks[0] == new_bb);
5520
5521 last_added_blocks.release ();
5522 return new_bb;
5523 }
5524
5525 /* Implement sched_create_recovery_block. ORIG_INSN is where block
5526 will be splitted to insert a check. */
5527 basic_block
5528 sel_create_recovery_block (insn_t orig_insn)
5529 {
5530 basic_block first_bb, second_bb, recovery_block;
5531 basic_block before_recovery = NULL;
5532 rtx jump;
5533
5534 first_bb = BLOCK_FOR_INSN (orig_insn);
5535 if (sel_bb_end_p (orig_insn))
5536 {
5537 /* Avoid introducing an empty block while splitting. */
5538 gcc_assert (single_succ_p (first_bb));
5539 second_bb = single_succ (first_bb);
5540 }
5541 else
5542 second_bb = sched_split_block (first_bb, orig_insn);
5543
5544 recovery_block = sched_create_recovery_block (&before_recovery);
5545 if (before_recovery)
5546 copy_lv_set_from (before_recovery, EXIT_BLOCK_PTR);
5547
5548 gcc_assert (sel_bb_empty_p (recovery_block));
5549 sched_create_recovery_edges (first_bb, recovery_block, second_bb);
5550 if (current_loops != NULL)
5551 add_bb_to_loop (recovery_block, first_bb->loop_father);
5552
5553 sel_add_bb (recovery_block);
5554
5555 jump = BB_END (recovery_block);
5556 gcc_assert (sel_bb_head (recovery_block) == jump);
5557 sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP);
5558
5559 return recovery_block;
5560 }
5561
5562 /* Merge basic block B into basic block A. */
5563 static void
5564 sel_merge_blocks (basic_block a, basic_block b)
5565 {
5566 gcc_assert (sel_bb_empty_p (b)
5567 && EDGE_COUNT (b->preds) == 1
5568 && EDGE_PRED (b, 0)->src == b->prev_bb);
5569
5570 move_bb_info (b->prev_bb, b);
5571 remove_empty_bb (b, false);
5572 merge_blocks (a, b);
5573 change_loops_latches (b, a);
5574 }
5575
5576 /* A wrapper for redirect_edge_and_branch_force, which also initializes
5577 data structures for possibly created bb and insns. Returns the newly
5578 added bb or NULL, when a bb was not needed. */
5579 void
5580 sel_redirect_edge_and_branch_force (edge e, basic_block to)
5581 {
5582 basic_block jump_bb, src, orig_dest = e->dest;
5583 int prev_max_uid;
5584 rtx jump;
5585
5586 /* This function is now used only for bookkeeping code creation, where
5587 we'll never get the single pred of orig_dest block and thus will not
5588 hit unreachable blocks when updating dominator info. */
5589 gcc_assert (!sel_bb_empty_p (e->src)
5590 && !single_pred_p (orig_dest));
5591 src = e->src;
5592 prev_max_uid = get_max_uid ();
5593 jump_bb = redirect_edge_and_branch_force (e, to);
5594
5595 if (jump_bb != NULL)
5596 sel_add_bb (jump_bb);
5597
5598 /* This function could not be used to spoil the loop structure by now,
5599 thus we don't care to update anything. But check it to be sure. */
5600 if (current_loop_nest
5601 && pipelining_p)
5602 gcc_assert (loop_latch_edge (current_loop_nest));
5603
5604 jump = find_new_jump (src, jump_bb, prev_max_uid);
5605 if (jump)
5606 sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP);
5607 set_immediate_dominator (CDI_DOMINATORS, to,
5608 recompute_dominator (CDI_DOMINATORS, to));
5609 set_immediate_dominator (CDI_DOMINATORS, orig_dest,
5610 recompute_dominator (CDI_DOMINATORS, orig_dest));
5611 }
5612
5613 /* A wrapper for redirect_edge_and_branch. Return TRUE if blocks connected by
5614 redirected edge are in reverse topological order. */
5615 bool
5616 sel_redirect_edge_and_branch (edge e, basic_block to)
5617 {
5618 bool latch_edge_p;
5619 basic_block src, orig_dest = e->dest;
5620 int prev_max_uid;
5621 rtx jump;
5622 edge redirected;
5623 bool recompute_toporder_p = false;
5624 bool maybe_unreachable = single_pred_p (orig_dest);
5625
5626 latch_edge_p = (pipelining_p
5627 && current_loop_nest
5628 && e == loop_latch_edge (current_loop_nest));
5629
5630 src = e->src;
5631 prev_max_uid = get_max_uid ();
5632
5633 redirected = redirect_edge_and_branch (e, to);
5634
5635 gcc_assert (redirected && !last_added_blocks.exists ());
5636
5637 /* When we've redirected a latch edge, update the header. */
5638 if (latch_edge_p)
5639 {
5640 current_loop_nest->header = to;
5641 gcc_assert (loop_latch_edge (current_loop_nest));
5642 }
5643
5644 /* In rare situations, the topological relation between the blocks connected
5645 by the redirected edge can change (see PR42245 for an example). Update
5646 block_to_bb/bb_to_block. */
5647 if (CONTAINING_RGN (e->src->index) == CONTAINING_RGN (to->index)
5648 && BLOCK_TO_BB (e->src->index) > BLOCK_TO_BB (to->index))
5649 recompute_toporder_p = true;
5650
5651 jump = find_new_jump (src, NULL, prev_max_uid);
5652 if (jump)
5653 sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP);
5654
5655 /* Only update dominator info when we don't have unreachable blocks.
5656 Otherwise we'll update in maybe_tidy_empty_bb. */
5657 if (!maybe_unreachable)
5658 {
5659 set_immediate_dominator (CDI_DOMINATORS, to,
5660 recompute_dominator (CDI_DOMINATORS, to));
5661 set_immediate_dominator (CDI_DOMINATORS, orig_dest,
5662 recompute_dominator (CDI_DOMINATORS, orig_dest));
5663 }
5664 return recompute_toporder_p;
5665 }
5666
5667 /* This variable holds the cfg hooks used by the selective scheduler. */
5668 static struct cfg_hooks sel_cfg_hooks;
5669
5670 /* Register sel-sched cfg hooks. */
5671 void
5672 sel_register_cfg_hooks (void)
5673 {
5674 sched_split_block = sel_split_block;
5675
5676 orig_cfg_hooks = get_cfg_hooks ();
5677 sel_cfg_hooks = orig_cfg_hooks;
5678
5679 sel_cfg_hooks.create_basic_block = sel_create_basic_block;
5680
5681 set_cfg_hooks (sel_cfg_hooks);
5682
5683 sched_init_only_bb = sel_init_only_bb;
5684 sched_split_block = sel_split_block;
5685 sched_create_empty_bb = sel_create_empty_bb;
5686 }
5687
5688 /* Unregister sel-sched cfg hooks. */
5689 void
5690 sel_unregister_cfg_hooks (void)
5691 {
5692 sched_create_empty_bb = NULL;
5693 sched_split_block = NULL;
5694 sched_init_only_bb = NULL;
5695
5696 set_cfg_hooks (orig_cfg_hooks);
5697 }
5698 \f
5699
5700 /* Emit an insn rtx based on PATTERN. If a jump insn is wanted,
5701 LABEL is where this jump should be directed. */
5702 rtx
5703 create_insn_rtx_from_pattern (rtx pattern, rtx label)
5704 {
5705 rtx insn_rtx;
5706
5707 gcc_assert (!INSN_P (pattern));
5708
5709 start_sequence ();
5710
5711 if (label == NULL_RTX)
5712 insn_rtx = emit_insn (pattern);
5713 else if (DEBUG_INSN_P (label))
5714 insn_rtx = emit_debug_insn (pattern);
5715 else
5716 {
5717 insn_rtx = emit_jump_insn (pattern);
5718 JUMP_LABEL (insn_rtx) = label;
5719 ++LABEL_NUSES (label);
5720 }
5721
5722 end_sequence ();
5723
5724 sched_extend_luids ();
5725 sched_extend_target ();
5726 sched_deps_init (false);
5727
5728 /* Initialize INSN_CODE now. */
5729 recog_memoized (insn_rtx);
5730 return insn_rtx;
5731 }
5732
5733 /* Create a new vinsn for INSN_RTX. FORCE_UNIQUE_P is true when the vinsn
5734 must not be clonable. */
5735 vinsn_t
5736 create_vinsn_from_insn_rtx (rtx insn_rtx, bool force_unique_p)
5737 {
5738 gcc_assert (INSN_P (insn_rtx) && !INSN_IN_STREAM_P (insn_rtx));
5739
5740 /* If VINSN_TYPE is not USE, retain its uniqueness. */
5741 return vinsn_create (insn_rtx, force_unique_p);
5742 }
5743
5744 /* Create a copy of INSN_RTX. */
5745 rtx
5746 create_copy_of_insn_rtx (rtx insn_rtx)
5747 {
5748 rtx res, link;
5749
5750 if (DEBUG_INSN_P (insn_rtx))
5751 return create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)),
5752 insn_rtx);
5753
5754 gcc_assert (NONJUMP_INSN_P (insn_rtx));
5755
5756 res = create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)),
5757 NULL_RTX);
5758
5759 /* Copy all REG_NOTES except REG_EQUAL/REG_EQUIV and REG_LABEL_OPERAND
5760 since mark_jump_label will make them. REG_LABEL_TARGETs are created
5761 there too, but are supposed to be sticky, so we copy them. */
5762 for (link = REG_NOTES (insn_rtx); link; link = XEXP (link, 1))
5763 if (REG_NOTE_KIND (link) != REG_LABEL_OPERAND
5764 && REG_NOTE_KIND (link) != REG_EQUAL
5765 && REG_NOTE_KIND (link) != REG_EQUIV)
5766 {
5767 if (GET_CODE (link) == EXPR_LIST)
5768 add_reg_note (res, REG_NOTE_KIND (link),
5769 copy_insn_1 (XEXP (link, 0)));
5770 else
5771 add_reg_note (res, REG_NOTE_KIND (link), XEXP (link, 0));
5772 }
5773
5774 return res;
5775 }
5776
5777 /* Change vinsn field of EXPR to hold NEW_VINSN. */
5778 void
5779 change_vinsn_in_expr (expr_t expr, vinsn_t new_vinsn)
5780 {
5781 vinsn_detach (EXPR_VINSN (expr));
5782
5783 EXPR_VINSN (expr) = new_vinsn;
5784 vinsn_attach (new_vinsn);
5785 }
5786
5787 /* Helpers for global init. */
5788 /* This structure is used to be able to call existing bundling mechanism
5789 and calculate insn priorities. */
5790 static struct haifa_sched_info sched_sel_haifa_sched_info =
5791 {
5792 NULL, /* init_ready_list */
5793 NULL, /* can_schedule_ready_p */
5794 NULL, /* schedule_more_p */
5795 NULL, /* new_ready */
5796 NULL, /* rgn_rank */
5797 sel_print_insn, /* rgn_print_insn */
5798 contributes_to_priority,
5799 NULL, /* insn_finishes_block_p */
5800
5801 NULL, NULL,
5802 NULL, NULL,
5803 0, 0,
5804
5805 NULL, /* add_remove_insn */
5806 NULL, /* begin_schedule_ready */
5807 NULL, /* begin_move_insn */
5808 NULL, /* advance_target_bb */
5809
5810 NULL,
5811 NULL,
5812
5813 SEL_SCHED | NEW_BBS
5814 };
5815
5816 /* Setup special insns used in the scheduler. */
5817 void
5818 setup_nop_and_exit_insns (void)
5819 {
5820 gcc_assert (nop_pattern == NULL_RTX
5821 && exit_insn == NULL_RTX);
5822
5823 nop_pattern = constm1_rtx;
5824
5825 start_sequence ();
5826 emit_insn (nop_pattern);
5827 exit_insn = get_insns ();
5828 end_sequence ();
5829 set_block_for_insn (exit_insn, EXIT_BLOCK_PTR);
5830 }
5831
5832 /* Free special insns used in the scheduler. */
5833 void
5834 free_nop_and_exit_insns (void)
5835 {
5836 exit_insn = NULL_RTX;
5837 nop_pattern = NULL_RTX;
5838 }
5839
5840 /* Setup a special vinsn used in new insns initialization. */
5841 void
5842 setup_nop_vinsn (void)
5843 {
5844 nop_vinsn = vinsn_create (exit_insn, false);
5845 vinsn_attach (nop_vinsn);
5846 }
5847
5848 /* Free a special vinsn used in new insns initialization. */
5849 void
5850 free_nop_vinsn (void)
5851 {
5852 gcc_assert (VINSN_COUNT (nop_vinsn) == 1);
5853 vinsn_detach (nop_vinsn);
5854 nop_vinsn = NULL;
5855 }
5856
5857 /* Call a set_sched_flags hook. */
5858 void
5859 sel_set_sched_flags (void)
5860 {
5861 /* ??? This means that set_sched_flags were called, and we decided to
5862 support speculation. However, set_sched_flags also modifies flags
5863 on current_sched_info, doing this only at global init. And we
5864 sometimes change c_s_i later. So put the correct flags again. */
5865 if (spec_info && targetm.sched.set_sched_flags)
5866 targetm.sched.set_sched_flags (spec_info);
5867 }
5868
5869 /* Setup pointers to global sched info structures. */
5870 void
5871 sel_setup_sched_infos (void)
5872 {
5873 rgn_setup_common_sched_info ();
5874
5875 memcpy (&sel_common_sched_info, common_sched_info,
5876 sizeof (sel_common_sched_info));
5877
5878 sel_common_sched_info.fix_recovery_cfg = NULL;
5879 sel_common_sched_info.add_block = NULL;
5880 sel_common_sched_info.estimate_number_of_insns
5881 = sel_estimate_number_of_insns;
5882 sel_common_sched_info.luid_for_non_insn = sel_luid_for_non_insn;
5883 sel_common_sched_info.sched_pass_id = SCHED_SEL_PASS;
5884
5885 common_sched_info = &sel_common_sched_info;
5886
5887 current_sched_info = &sched_sel_haifa_sched_info;
5888 current_sched_info->sched_max_insns_priority =
5889 get_rgn_sched_max_insns_priority ();
5890
5891 sel_set_sched_flags ();
5892 }
5893 \f
5894
5895 /* Adds basic block BB to region RGN at the position *BB_ORD_INDEX,
5896 *BB_ORD_INDEX after that is increased. */
5897 static void
5898 sel_add_block_to_region (basic_block bb, int *bb_ord_index, int rgn)
5899 {
5900 RGN_NR_BLOCKS (rgn) += 1;
5901 RGN_DONT_CALC_DEPS (rgn) = 0;
5902 RGN_HAS_REAL_EBB (rgn) = 0;
5903 CONTAINING_RGN (bb->index) = rgn;
5904 BLOCK_TO_BB (bb->index) = *bb_ord_index;
5905 rgn_bb_table[RGN_BLOCKS (rgn) + *bb_ord_index] = bb->index;
5906 (*bb_ord_index)++;
5907
5908 /* FIXME: it is true only when not scheduling ebbs. */
5909 RGN_BLOCKS (rgn + 1) = RGN_BLOCKS (rgn) + RGN_NR_BLOCKS (rgn);
5910 }
5911
5912 /* Functions to support pipelining of outer loops. */
5913
5914 /* Creates a new empty region and returns it's number. */
5915 static int
5916 sel_create_new_region (void)
5917 {
5918 int new_rgn_number = nr_regions;
5919
5920 RGN_NR_BLOCKS (new_rgn_number) = 0;
5921
5922 /* FIXME: This will work only when EBBs are not created. */
5923 if (new_rgn_number != 0)
5924 RGN_BLOCKS (new_rgn_number) = RGN_BLOCKS (new_rgn_number - 1) +
5925 RGN_NR_BLOCKS (new_rgn_number - 1);
5926 else
5927 RGN_BLOCKS (new_rgn_number) = 0;
5928
5929 /* Set the blocks of the next region so the other functions may
5930 calculate the number of blocks in the region. */
5931 RGN_BLOCKS (new_rgn_number + 1) = RGN_BLOCKS (new_rgn_number) +
5932 RGN_NR_BLOCKS (new_rgn_number);
5933
5934 nr_regions++;
5935
5936 return new_rgn_number;
5937 }
5938
5939 /* If X has a smaller topological sort number than Y, returns -1;
5940 if greater, returns 1. */
5941 static int
5942 bb_top_order_comparator (const void *x, const void *y)
5943 {
5944 basic_block bb1 = *(const basic_block *) x;
5945 basic_block bb2 = *(const basic_block *) y;
5946
5947 gcc_assert (bb1 == bb2
5948 || rev_top_order_index[bb1->index]
5949 != rev_top_order_index[bb2->index]);
5950
5951 /* It's a reverse topological order in REV_TOP_ORDER_INDEX, so
5952 bbs with greater number should go earlier. */
5953 if (rev_top_order_index[bb1->index] > rev_top_order_index[bb2->index])
5954 return -1;
5955 else
5956 return 1;
5957 }
5958
5959 /* Create a region for LOOP and return its number. If we don't want
5960 to pipeline LOOP, return -1. */
5961 static int
5962 make_region_from_loop (struct loop *loop)
5963 {
5964 unsigned int i;
5965 int new_rgn_number = -1;
5966 struct loop *inner;
5967
5968 /* Basic block index, to be assigned to BLOCK_TO_BB. */
5969 int bb_ord_index = 0;
5970 basic_block *loop_blocks;
5971 basic_block preheader_block;
5972
5973 if (loop->num_nodes
5974 > (unsigned) PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_BLOCKS))
5975 return -1;
5976
5977 /* Don't pipeline loops whose latch belongs to some of its inner loops. */
5978 for (inner = loop->inner; inner; inner = inner->inner)
5979 if (flow_bb_inside_loop_p (inner, loop->latch))
5980 return -1;
5981
5982 loop->ninsns = num_loop_insns (loop);
5983 if ((int) loop->ninsns > PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_INSNS))
5984 return -1;
5985
5986 loop_blocks = get_loop_body_in_custom_order (loop, bb_top_order_comparator);
5987
5988 for (i = 0; i < loop->num_nodes; i++)
5989 if (loop_blocks[i]->flags & BB_IRREDUCIBLE_LOOP)
5990 {
5991 free (loop_blocks);
5992 return -1;
5993 }
5994
5995 preheader_block = loop_preheader_edge (loop)->src;
5996 gcc_assert (preheader_block);
5997 gcc_assert (loop_blocks[0] == loop->header);
5998
5999 new_rgn_number = sel_create_new_region ();
6000
6001 sel_add_block_to_region (preheader_block, &bb_ord_index, new_rgn_number);
6002 bitmap_set_bit (bbs_in_loop_rgns, preheader_block->index);
6003
6004 for (i = 0; i < loop->num_nodes; i++)
6005 {
6006 /* Add only those blocks that haven't been scheduled in the inner loop.
6007 The exception is the basic blocks with bookkeeping code - they should
6008 be added to the region (and they actually don't belong to the loop
6009 body, but to the region containing that loop body). */
6010
6011 gcc_assert (new_rgn_number >= 0);
6012
6013 if (! bitmap_bit_p (bbs_in_loop_rgns, loop_blocks[i]->index))
6014 {
6015 sel_add_block_to_region (loop_blocks[i], &bb_ord_index,
6016 new_rgn_number);
6017 bitmap_set_bit (bbs_in_loop_rgns, loop_blocks[i]->index);
6018 }
6019 }
6020
6021 free (loop_blocks);
6022 MARK_LOOP_FOR_PIPELINING (loop);
6023
6024 return new_rgn_number;
6025 }
6026
6027 /* Create a new region from preheader blocks LOOP_BLOCKS. */
6028 void
6029 make_region_from_loop_preheader (vec<basic_block> *&loop_blocks)
6030 {
6031 unsigned int i;
6032 int new_rgn_number = -1;
6033 basic_block bb;
6034
6035 /* Basic block index, to be assigned to BLOCK_TO_BB. */
6036 int bb_ord_index = 0;
6037
6038 new_rgn_number = sel_create_new_region ();
6039
6040 FOR_EACH_VEC_ELT (*loop_blocks, i, bb)
6041 {
6042 gcc_assert (new_rgn_number >= 0);
6043
6044 sel_add_block_to_region (bb, &bb_ord_index, new_rgn_number);
6045 }
6046
6047 vec_free (loop_blocks);
6048 }
6049
6050
6051 /* Create region(s) from loop nest LOOP, such that inner loops will be
6052 pipelined before outer loops. Returns true when a region for LOOP
6053 is created. */
6054 static bool
6055 make_regions_from_loop_nest (struct loop *loop)
6056 {
6057 struct loop *cur_loop;
6058 int rgn_number;
6059
6060 /* Traverse all inner nodes of the loop. */
6061 for (cur_loop = loop->inner; cur_loop; cur_loop = cur_loop->next)
6062 if (! bitmap_bit_p (bbs_in_loop_rgns, cur_loop->header->index))
6063 return false;
6064
6065 /* At this moment all regular inner loops should have been pipelined.
6066 Try to create a region from this loop. */
6067 rgn_number = make_region_from_loop (loop);
6068
6069 if (rgn_number < 0)
6070 return false;
6071
6072 loop_nests.safe_push (loop);
6073 return true;
6074 }
6075
6076 /* Initalize data structures needed. */
6077 void
6078 sel_init_pipelining (void)
6079 {
6080 /* Collect loop information to be used in outer loops pipelining. */
6081 loop_optimizer_init (LOOPS_HAVE_PREHEADERS
6082 | LOOPS_HAVE_FALLTHRU_PREHEADERS
6083 | LOOPS_HAVE_RECORDED_EXITS
6084 | LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS);
6085 current_loop_nest = NULL;
6086
6087 bbs_in_loop_rgns = sbitmap_alloc (last_basic_block);
6088 bitmap_clear (bbs_in_loop_rgns);
6089
6090 recompute_rev_top_order ();
6091 }
6092
6093 /* Returns a struct loop for region RGN. */
6094 loop_p
6095 get_loop_nest_for_rgn (unsigned int rgn)
6096 {
6097 /* Regions created with extend_rgns don't have corresponding loop nests,
6098 because they don't represent loops. */
6099 if (rgn < loop_nests.length ())
6100 return loop_nests[rgn];
6101 else
6102 return NULL;
6103 }
6104
6105 /* True when LOOP was included into pipelining regions. */
6106 bool
6107 considered_for_pipelining_p (struct loop *loop)
6108 {
6109 if (loop_depth (loop) == 0)
6110 return false;
6111
6112 /* Now, the loop could be too large or irreducible. Check whether its
6113 region is in LOOP_NESTS.
6114 We determine the region number of LOOP as the region number of its
6115 latch. We can't use header here, because this header could be
6116 just removed preheader and it will give us the wrong region number.
6117 Latch can't be used because it could be in the inner loop too. */
6118 if (LOOP_MARKED_FOR_PIPELINING_P (loop))
6119 {
6120 int rgn = CONTAINING_RGN (loop->latch->index);
6121
6122 gcc_assert ((unsigned) rgn < loop_nests.length ());
6123 return true;
6124 }
6125
6126 return false;
6127 }
6128
6129 /* Makes regions from the rest of the blocks, after loops are chosen
6130 for pipelining. */
6131 static void
6132 make_regions_from_the_rest (void)
6133 {
6134 int cur_rgn_blocks;
6135 int *loop_hdr;
6136 int i;
6137
6138 basic_block bb;
6139 edge e;
6140 edge_iterator ei;
6141 int *degree;
6142
6143 /* Index in rgn_bb_table where to start allocating new regions. */
6144 cur_rgn_blocks = nr_regions ? RGN_BLOCKS (nr_regions) : 0;
6145
6146 /* Make regions from all the rest basic blocks - those that don't belong to
6147 any loop or belong to irreducible loops. Prepare the data structures
6148 for extend_rgns. */
6149
6150 /* LOOP_HDR[I] == -1 if I-th bb doesn't belong to any loop,
6151 LOOP_HDR[I] == LOOP_HDR[J] iff basic blocks I and J reside within the same
6152 loop. */
6153 loop_hdr = XNEWVEC (int, last_basic_block);
6154 degree = XCNEWVEC (int, last_basic_block);
6155
6156
6157 /* For each basic block that belongs to some loop assign the number
6158 of innermost loop it belongs to. */
6159 for (i = 0; i < last_basic_block; i++)
6160 loop_hdr[i] = -1;
6161
6162 FOR_EACH_BB (bb)
6163 {
6164 if (bb->loop_father && !bb->loop_father->num == 0
6165 && !(bb->flags & BB_IRREDUCIBLE_LOOP))
6166 loop_hdr[bb->index] = bb->loop_father->num;
6167 }
6168
6169 /* For each basic block degree is calculated as the number of incoming
6170 edges, that are going out of bbs that are not yet scheduled.
6171 The basic blocks that are scheduled have degree value of zero. */
6172 FOR_EACH_BB (bb)
6173 {
6174 degree[bb->index] = 0;
6175
6176 if (!bitmap_bit_p (bbs_in_loop_rgns, bb->index))
6177 {
6178 FOR_EACH_EDGE (e, ei, bb->preds)
6179 if (!bitmap_bit_p (bbs_in_loop_rgns, e->src->index))
6180 degree[bb->index]++;
6181 }
6182 else
6183 degree[bb->index] = -1;
6184 }
6185
6186 extend_rgns (degree, &cur_rgn_blocks, bbs_in_loop_rgns, loop_hdr);
6187
6188 /* Any block that did not end up in a region is placed into a region
6189 by itself. */
6190 FOR_EACH_BB (bb)
6191 if (degree[bb->index] >= 0)
6192 {
6193 rgn_bb_table[cur_rgn_blocks] = bb->index;
6194 RGN_NR_BLOCKS (nr_regions) = 1;
6195 RGN_BLOCKS (nr_regions) = cur_rgn_blocks++;
6196 RGN_DONT_CALC_DEPS (nr_regions) = 0;
6197 RGN_HAS_REAL_EBB (nr_regions) = 0;
6198 CONTAINING_RGN (bb->index) = nr_regions++;
6199 BLOCK_TO_BB (bb->index) = 0;
6200 }
6201
6202 free (degree);
6203 free (loop_hdr);
6204 }
6205
6206 /* Free data structures used in pipelining of loops. */
6207 void sel_finish_pipelining (void)
6208 {
6209 loop_iterator li;
6210 struct loop *loop;
6211
6212 /* Release aux fields so we don't free them later by mistake. */
6213 FOR_EACH_LOOP (li, loop, 0)
6214 loop->aux = NULL;
6215
6216 loop_optimizer_finalize ();
6217
6218 loop_nests.release ();
6219
6220 free (rev_top_order_index);
6221 rev_top_order_index = NULL;
6222 }
6223
6224 /* This function replaces the find_rgns when
6225 FLAG_SEL_SCHED_PIPELINING_OUTER_LOOPS is set. */
6226 void
6227 sel_find_rgns (void)
6228 {
6229 sel_init_pipelining ();
6230 extend_regions ();
6231
6232 if (current_loops)
6233 {
6234 loop_p loop;
6235 loop_iterator li;
6236
6237 FOR_EACH_LOOP (li, loop, (flag_sel_sched_pipelining_outer_loops
6238 ? LI_FROM_INNERMOST
6239 : LI_ONLY_INNERMOST))
6240 make_regions_from_loop_nest (loop);
6241 }
6242
6243 /* Make regions from all the rest basic blocks and schedule them.
6244 These blocks include blocks that don't belong to any loop or belong
6245 to irreducible loops. */
6246 make_regions_from_the_rest ();
6247
6248 /* We don't need bbs_in_loop_rgns anymore. */
6249 sbitmap_free (bbs_in_loop_rgns);
6250 bbs_in_loop_rgns = NULL;
6251 }
6252
6253 /* Add the preheader blocks from previous loop to current region taking
6254 it from LOOP_PREHEADER_BLOCKS (current_loop_nest) and record them in *BBS.
6255 This function is only used with -fsel-sched-pipelining-outer-loops. */
6256 void
6257 sel_add_loop_preheaders (bb_vec_t *bbs)
6258 {
6259 int i;
6260 basic_block bb;
6261 vec<basic_block> *preheader_blocks
6262 = LOOP_PREHEADER_BLOCKS (current_loop_nest);
6263
6264 if (!preheader_blocks)
6265 return;
6266
6267 for (i = 0; preheader_blocks->iterate (i, &bb); i++)
6268 {
6269 bbs->safe_push (bb);
6270 last_added_blocks.safe_push (bb);
6271 sel_add_bb (bb);
6272 }
6273
6274 vec_free (preheader_blocks);
6275 }
6276
6277 /* While pipelining outer loops, returns TRUE if BB is a loop preheader.
6278 Please note that the function should also work when pipelining_p is
6279 false, because it is used when deciding whether we should or should
6280 not reschedule pipelined code. */
6281 bool
6282 sel_is_loop_preheader_p (basic_block bb)
6283 {
6284 if (current_loop_nest)
6285 {
6286 struct loop *outer;
6287
6288 if (preheader_removed)
6289 return false;
6290
6291 /* Preheader is the first block in the region. */
6292 if (BLOCK_TO_BB (bb->index) == 0)
6293 return true;
6294
6295 /* We used to find a preheader with the topological information.
6296 Check that the above code is equivalent to what we did before. */
6297
6298 if (in_current_region_p (current_loop_nest->header))
6299 gcc_assert (!(BLOCK_TO_BB (bb->index)
6300 < BLOCK_TO_BB (current_loop_nest->header->index)));
6301
6302 /* Support the situation when the latch block of outer loop
6303 could be from here. */
6304 for (outer = loop_outer (current_loop_nest);
6305 outer;
6306 outer = loop_outer (outer))
6307 if (considered_for_pipelining_p (outer) && outer->latch == bb)
6308 gcc_unreachable ();
6309 }
6310
6311 return false;
6312 }
6313
6314 /* Check whether JUMP_BB ends with a jump insn that leads only to DEST_BB and
6315 can be removed, making the corresponding edge fallthrough (assuming that
6316 all basic blocks between JUMP_BB and DEST_BB are empty). */
6317 static bool
6318 bb_has_removable_jump_to_p (basic_block jump_bb, basic_block dest_bb)
6319 {
6320 if (!onlyjump_p (BB_END (jump_bb))
6321 || tablejump_p (BB_END (jump_bb), NULL, NULL))
6322 return false;
6323
6324 /* Several outgoing edges, abnormal edge or destination of jump is
6325 not DEST_BB. */
6326 if (EDGE_COUNT (jump_bb->succs) != 1
6327 || EDGE_SUCC (jump_bb, 0)->flags & (EDGE_ABNORMAL | EDGE_CROSSING)
6328 || EDGE_SUCC (jump_bb, 0)->dest != dest_bb)
6329 return false;
6330
6331 /* If not anything of the upper. */
6332 return true;
6333 }
6334
6335 /* Removes the loop preheader from the current region and saves it in
6336 PREHEADER_BLOCKS of the father loop, so they will be added later to
6337 region that represents an outer loop. */
6338 static void
6339 sel_remove_loop_preheader (void)
6340 {
6341 int i, old_len;
6342 int cur_rgn = CONTAINING_RGN (BB_TO_BLOCK (0));
6343 basic_block bb;
6344 bool all_empty_p = true;
6345 vec<basic_block> *preheader_blocks
6346 = LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest));
6347
6348 vec_check_alloc (preheader_blocks, 0);
6349
6350 gcc_assert (current_loop_nest);
6351 old_len = preheader_blocks->length ();
6352
6353 /* Add blocks that aren't within the current loop to PREHEADER_BLOCKS. */
6354 for (i = 0; i < RGN_NR_BLOCKS (cur_rgn); i++)
6355 {
6356 bb = BASIC_BLOCK (BB_TO_BLOCK (i));
6357
6358 /* If the basic block belongs to region, but doesn't belong to
6359 corresponding loop, then it should be a preheader. */
6360 if (sel_is_loop_preheader_p (bb))
6361 {
6362 preheader_blocks->safe_push (bb);
6363 if (BB_END (bb) != bb_note (bb))
6364 all_empty_p = false;
6365 }
6366 }
6367
6368 /* Remove these blocks only after iterating over the whole region. */
6369 for (i = preheader_blocks->length () - 1; i >= old_len; i--)
6370 {
6371 bb = (*preheader_blocks)[i];
6372 sel_remove_bb (bb, false);
6373 }
6374
6375 if (!considered_for_pipelining_p (loop_outer (current_loop_nest)))
6376 {
6377 if (!all_empty_p)
6378 /* Immediately create new region from preheader. */
6379 make_region_from_loop_preheader (preheader_blocks);
6380 else
6381 {
6382 /* If all preheader blocks are empty - dont create new empty region.
6383 Instead, remove them completely. */
6384 FOR_EACH_VEC_ELT (*preheader_blocks, i, bb)
6385 {
6386 edge e;
6387 edge_iterator ei;
6388 basic_block prev_bb = bb->prev_bb, next_bb = bb->next_bb;
6389
6390 /* Redirect all incoming edges to next basic block. */
6391 for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei)); )
6392 {
6393 if (! (e->flags & EDGE_FALLTHRU))
6394 redirect_edge_and_branch (e, bb->next_bb);
6395 else
6396 redirect_edge_succ (e, bb->next_bb);
6397 }
6398 gcc_assert (BB_NOTE_LIST (bb) == NULL);
6399 delete_and_free_basic_block (bb);
6400
6401 /* Check if after deleting preheader there is a nonconditional
6402 jump in PREV_BB that leads to the next basic block NEXT_BB.
6403 If it is so - delete this jump and clear data sets of its
6404 basic block if it becomes empty. */
6405 if (next_bb->prev_bb == prev_bb
6406 && prev_bb != ENTRY_BLOCK_PTR
6407 && bb_has_removable_jump_to_p (prev_bb, next_bb))
6408 {
6409 redirect_edge_and_branch (EDGE_SUCC (prev_bb, 0), next_bb);
6410 if (BB_END (prev_bb) == bb_note (prev_bb))
6411 free_data_sets (prev_bb);
6412 }
6413
6414 set_immediate_dominator (CDI_DOMINATORS, next_bb,
6415 recompute_dominator (CDI_DOMINATORS,
6416 next_bb));
6417 }
6418 }
6419 vec_free (preheader_blocks);
6420 }
6421 else
6422 /* Store preheader within the father's loop structure. */
6423 SET_LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest),
6424 preheader_blocks);
6425 }
6426 #endif