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