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