]> git.ipfire.org Git - thirdparty/gcc.git/blob - gcc/sel-sched-ir.c
2014-10-16 Andrew MacLeod <amacleod@redhat.com>
[thirdparty/gcc.git] / gcc / sel-sched-ir.c
1 /* Instruction scheduling pass. Selective scheduler and pipeliner.
2 Copyright (C) 2006-2014 Free Software Foundation, Inc.
3
4 This file is part of GCC.
5
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
10
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
15
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
19
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "tm.h"
24 #include "diagnostic-core.h"
25 #include "rtl.h"
26 #include "tm_p.h"
27 #include "hard-reg-set.h"
28 #include "regs.h"
29 #include "hashtab.h"
30 #include "hash-set.h"
31 #include "vec.h"
32 #include "machmode.h"
33 #include "input.h"
34 #include "function.h"
35 #include "flags.h"
36 #include "insn-config.h"
37 #include "insn-attr.h"
38 #include "except.h"
39 #include "recog.h"
40 #include "params.h"
41 #include "target.h"
42 #include "sched-int.h"
43 #include "ggc.h"
44 #include "tree.h"
45 #include "langhooks.h"
46 #include "rtlhooks-def.h"
47 #include "emit-rtl.h" /* FIXME: Can go away once crtl is moved to rtl.h. */
48
49 #ifdef INSN_SCHEDULING
50 #include "sel-sched-ir.h"
51 /* We don't have to use it except for sel_print_insn. */
52 #include "sel-sched-dump.h"
53
54 /* A vector holding bb info for whole scheduling pass. */
55 vec<sel_global_bb_info_def>
56 sel_global_bb_info = vNULL;
57
58 /* A vector holding bb info. */
59 vec<sel_region_bb_info_def>
60 sel_region_bb_info = vNULL;
61
62 /* A pool for allocating all lists. */
63 alloc_pool sched_lists_pool;
64
65 /* This contains information about successors for compute_av_set. */
66 struct succs_info current_succs;
67
68 /* Data structure to describe interaction with the generic scheduler utils. */
69 static struct common_sched_info_def sel_common_sched_info;
70
71 /* The loop nest being pipelined. */
72 struct loop *current_loop_nest;
73
74 /* LOOP_NESTS is a vector containing the corresponding loop nest for
75 each region. */
76 static vec<loop_p> loop_nests = vNULL;
77
78 /* Saves blocks already in loop regions, indexed by bb->index. */
79 static sbitmap bbs_in_loop_rgns = NULL;
80
81 /* CFG hooks that are saved before changing create_basic_block hook. */
82 static struct cfg_hooks orig_cfg_hooks;
83 \f
84
85 /* Array containing reverse topological index of function basic blocks,
86 indexed by BB->INDEX. */
87 static int *rev_top_order_index = NULL;
88
89 /* Length of the above array. */
90 static int rev_top_order_index_len = -1;
91
92 /* A regset pool structure. */
93 static struct
94 {
95 /* The stack to which regsets are returned. */
96 regset *v;
97
98 /* Its pointer. */
99 int n;
100
101 /* Its size. */
102 int s;
103
104 /* In VV we save all generated regsets so that, when destructing the
105 pool, we can compare it with V and check that every regset was returned
106 back to pool. */
107 regset *vv;
108
109 /* The pointer of VV stack. */
110 int nn;
111
112 /* Its size. */
113 int ss;
114
115 /* The difference between allocated and returned regsets. */
116 int diff;
117 } regset_pool = { NULL, 0, 0, NULL, 0, 0, 0 };
118
119 /* This represents the nop pool. */
120 static struct
121 {
122 /* The vector which holds previously emitted nops. */
123 insn_t *v;
124
125 /* Its pointer. */
126 int n;
127
128 /* Its size. */
129 int s;
130 } nop_pool = { NULL, 0, 0 };
131
132 /* The pool for basic block notes. */
133 static vec<rtx_note *> bb_note_pool;
134
135 /* A NOP pattern used to emit placeholder insns. */
136 rtx nop_pattern = NULL_RTX;
137 /* A special instruction that resides in EXIT_BLOCK.
138 EXIT_INSN is successor of the insns that lead to EXIT_BLOCK. */
139 rtx_insn *exit_insn = NULL;
140
141 /* TRUE if while scheduling current region, which is loop, its preheader
142 was removed. */
143 bool preheader_removed = false;
144 \f
145
146 /* Forward static declarations. */
147 static void fence_clear (fence_t);
148
149 static void deps_init_id (idata_t, insn_t, bool);
150 static void init_id_from_df (idata_t, insn_t, bool);
151 static expr_t set_insn_init (expr_t, vinsn_t, int);
152
153 static void cfg_preds (basic_block, insn_t **, int *);
154 static void prepare_insn_expr (insn_t, int);
155 static void free_history_vect (vec<expr_history_def> &);
156
157 static void move_bb_info (basic_block, basic_block);
158 static void remove_empty_bb (basic_block, bool);
159 static void sel_merge_blocks (basic_block, basic_block);
160 static void sel_remove_loop_preheader (void);
161 static bool bb_has_removable_jump_to_p (basic_block, basic_block);
162
163 static bool insn_is_the_only_one_in_bb_p (insn_t);
164 static void create_initial_data_sets (basic_block);
165
166 static void free_av_set (basic_block);
167 static void invalidate_av_set (basic_block);
168 static void extend_insn_data (void);
169 static void sel_init_new_insn (insn_t, int, int = -1);
170 static void finish_insns (void);
171 \f
172 /* Various list functions. */
173
174 /* Copy an instruction list L. */
175 ilist_t
176 ilist_copy (ilist_t l)
177 {
178 ilist_t head = NULL, *tailp = &head;
179
180 while (l)
181 {
182 ilist_add (tailp, ILIST_INSN (l));
183 tailp = &ILIST_NEXT (*tailp);
184 l = ILIST_NEXT (l);
185 }
186
187 return head;
188 }
189
190 /* Invert an instruction list L. */
191 ilist_t
192 ilist_invert (ilist_t l)
193 {
194 ilist_t res = NULL;
195
196 while (l)
197 {
198 ilist_add (&res, ILIST_INSN (l));
199 l = ILIST_NEXT (l);
200 }
201
202 return res;
203 }
204
205 /* Add a new boundary to the LP list with parameters TO, PTR, and DC. */
206 void
207 blist_add (blist_t *lp, insn_t to, ilist_t ptr, deps_t dc)
208 {
209 bnd_t bnd;
210
211 _list_add (lp);
212 bnd = BLIST_BND (*lp);
213
214 BND_TO (bnd) = to;
215 BND_PTR (bnd) = ptr;
216 BND_AV (bnd) = NULL;
217 BND_AV1 (bnd) = NULL;
218 BND_DC (bnd) = dc;
219 }
220
221 /* Remove the list note pointed to by LP. */
222 void
223 blist_remove (blist_t *lp)
224 {
225 bnd_t b = BLIST_BND (*lp);
226
227 av_set_clear (&BND_AV (b));
228 av_set_clear (&BND_AV1 (b));
229 ilist_clear (&BND_PTR (b));
230
231 _list_remove (lp);
232 }
233
234 /* Init a fence tail L. */
235 void
236 flist_tail_init (flist_tail_t l)
237 {
238 FLIST_TAIL_HEAD (l) = NULL;
239 FLIST_TAIL_TAILP (l) = &FLIST_TAIL_HEAD (l);
240 }
241
242 /* Try to find fence corresponding to INSN in L. */
243 fence_t
244 flist_lookup (flist_t l, insn_t insn)
245 {
246 while (l)
247 {
248 if (FENCE_INSN (FLIST_FENCE (l)) == insn)
249 return FLIST_FENCE (l);
250
251 l = FLIST_NEXT (l);
252 }
253
254 return NULL;
255 }
256
257 /* Init the fields of F before running fill_insns. */
258 static void
259 init_fence_for_scheduling (fence_t f)
260 {
261 FENCE_BNDS (f) = NULL;
262 FENCE_PROCESSED_P (f) = false;
263 FENCE_SCHEDULED_P (f) = false;
264 }
265
266 /* Add new fence consisting of INSN and STATE to the list pointed to by LP. */
267 static void
268 flist_add (flist_t *lp, insn_t insn, state_t state, deps_t dc, void *tc,
269 insn_t last_scheduled_insn, vec<rtx_insn *, va_gc> *executing_insns,
270 int *ready_ticks, int ready_ticks_size, insn_t sched_next,
271 int cycle, int cycle_issued_insns, int issue_more,
272 bool starts_cycle_p, bool after_stall_p)
273 {
274 fence_t f;
275
276 _list_add (lp);
277 f = FLIST_FENCE (*lp);
278
279 FENCE_INSN (f) = insn;
280
281 gcc_assert (state != NULL);
282 FENCE_STATE (f) = state;
283
284 FENCE_CYCLE (f) = cycle;
285 FENCE_ISSUED_INSNS (f) = cycle_issued_insns;
286 FENCE_STARTS_CYCLE_P (f) = starts_cycle_p;
287 FENCE_AFTER_STALL_P (f) = after_stall_p;
288
289 gcc_assert (dc != NULL);
290 FENCE_DC (f) = dc;
291
292 gcc_assert (tc != NULL || targetm.sched.alloc_sched_context == NULL);
293 FENCE_TC (f) = tc;
294
295 FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn;
296 FENCE_ISSUE_MORE (f) = issue_more;
297 FENCE_EXECUTING_INSNS (f) = executing_insns;
298 FENCE_READY_TICKS (f) = ready_ticks;
299 FENCE_READY_TICKS_SIZE (f) = ready_ticks_size;
300 FENCE_SCHED_NEXT (f) = sched_next;
301
302 init_fence_for_scheduling (f);
303 }
304
305 /* Remove the head node of the list pointed to by LP. */
306 static void
307 flist_remove (flist_t *lp)
308 {
309 if (FENCE_INSN (FLIST_FENCE (*lp)))
310 fence_clear (FLIST_FENCE (*lp));
311 _list_remove (lp);
312 }
313
314 /* Clear the fence list pointed to by LP. */
315 void
316 flist_clear (flist_t *lp)
317 {
318 while (*lp)
319 flist_remove (lp);
320 }
321
322 /* Add ORIGINAL_INSN the def list DL honoring CROSSES_CALL. */
323 void
324 def_list_add (def_list_t *dl, insn_t original_insn, bool crosses_call)
325 {
326 def_t d;
327
328 _list_add (dl);
329 d = DEF_LIST_DEF (*dl);
330
331 d->orig_insn = original_insn;
332 d->crosses_call = crosses_call;
333 }
334 \f
335
336 /* Functions to work with target contexts. */
337
338 /* Bulk target context. It is convenient for debugging purposes to ensure
339 that there are no uninitialized (null) target contexts. */
340 static tc_t bulk_tc = (tc_t) 1;
341
342 /* Target hooks wrappers. In the future we can provide some default
343 implementations for them. */
344
345 /* Allocate a store for the target context. */
346 static tc_t
347 alloc_target_context (void)
348 {
349 return (targetm.sched.alloc_sched_context
350 ? targetm.sched.alloc_sched_context () : bulk_tc);
351 }
352
353 /* Init target context TC.
354 If CLEAN_P is true, then make TC as it is beginning of the scheduler.
355 Overwise, copy current backend context to TC. */
356 static void
357 init_target_context (tc_t tc, bool clean_p)
358 {
359 if (targetm.sched.init_sched_context)
360 targetm.sched.init_sched_context (tc, clean_p);
361 }
362
363 /* Allocate and initialize a target context. Meaning of CLEAN_P is the same as
364 int init_target_context (). */
365 tc_t
366 create_target_context (bool clean_p)
367 {
368 tc_t tc = alloc_target_context ();
369
370 init_target_context (tc, clean_p);
371 return tc;
372 }
373
374 /* Copy TC to the current backend context. */
375 void
376 set_target_context (tc_t tc)
377 {
378 if (targetm.sched.set_sched_context)
379 targetm.sched.set_sched_context (tc);
380 }
381
382 /* TC is about to be destroyed. Free any internal data. */
383 static void
384 clear_target_context (tc_t tc)
385 {
386 if (targetm.sched.clear_sched_context)
387 targetm.sched.clear_sched_context (tc);
388 }
389
390 /* Clear and free it. */
391 static void
392 delete_target_context (tc_t tc)
393 {
394 clear_target_context (tc);
395
396 if (targetm.sched.free_sched_context)
397 targetm.sched.free_sched_context (tc);
398 }
399
400 /* Make a copy of FROM in TO.
401 NB: May be this should be a hook. */
402 static void
403 copy_target_context (tc_t to, tc_t from)
404 {
405 tc_t tmp = create_target_context (false);
406
407 set_target_context (from);
408 init_target_context (to, false);
409
410 set_target_context (tmp);
411 delete_target_context (tmp);
412 }
413
414 /* Create a copy of TC. */
415 static tc_t
416 create_copy_of_target_context (tc_t tc)
417 {
418 tc_t copy = alloc_target_context ();
419
420 copy_target_context (copy, tc);
421
422 return copy;
423 }
424
425 /* Clear TC and initialize it according to CLEAN_P. The meaning of CLEAN_P
426 is the same as in init_target_context (). */
427 void
428 reset_target_context (tc_t tc, bool clean_p)
429 {
430 clear_target_context (tc);
431 init_target_context (tc, clean_p);
432 }
433 \f
434 /* Functions to work with dependence contexts.
435 Dc (aka deps context, aka deps_t, aka struct deps_desc *) is short for dependence
436 context. It accumulates information about processed insns to decide if
437 current insn is dependent on the processed ones. */
438
439 /* Make a copy of FROM in TO. */
440 static void
441 copy_deps_context (deps_t to, deps_t from)
442 {
443 init_deps (to, false);
444 deps_join (to, from);
445 }
446
447 /* Allocate store for dep context. */
448 static deps_t
449 alloc_deps_context (void)
450 {
451 return XNEW (struct deps_desc);
452 }
453
454 /* Allocate and initialize dep context. */
455 static deps_t
456 create_deps_context (void)
457 {
458 deps_t dc = alloc_deps_context ();
459
460 init_deps (dc, false);
461 return dc;
462 }
463
464 /* Create a copy of FROM. */
465 static deps_t
466 create_copy_of_deps_context (deps_t from)
467 {
468 deps_t to = alloc_deps_context ();
469
470 copy_deps_context (to, from);
471 return to;
472 }
473
474 /* Clean up internal data of DC. */
475 static void
476 clear_deps_context (deps_t dc)
477 {
478 free_deps (dc);
479 }
480
481 /* Clear and free DC. */
482 static void
483 delete_deps_context (deps_t dc)
484 {
485 clear_deps_context (dc);
486 free (dc);
487 }
488
489 /* Clear and init DC. */
490 static void
491 reset_deps_context (deps_t dc)
492 {
493 clear_deps_context (dc);
494 init_deps (dc, false);
495 }
496
497 /* This structure describes the dependence analysis hooks for advancing
498 dependence context. */
499 static struct sched_deps_info_def advance_deps_context_sched_deps_info =
500 {
501 NULL,
502
503 NULL, /* start_insn */
504 NULL, /* finish_insn */
505 NULL, /* start_lhs */
506 NULL, /* finish_lhs */
507 NULL, /* start_rhs */
508 NULL, /* finish_rhs */
509 haifa_note_reg_set,
510 haifa_note_reg_clobber,
511 haifa_note_reg_use,
512 NULL, /* note_mem_dep */
513 NULL, /* note_dep */
514
515 0, 0, 0
516 };
517
518 /* Process INSN and add its impact on DC. */
519 void
520 advance_deps_context (deps_t dc, insn_t insn)
521 {
522 sched_deps_info = &advance_deps_context_sched_deps_info;
523 deps_analyze_insn (dc, insn);
524 }
525 \f
526
527 /* Functions to work with DFA states. */
528
529 /* Allocate store for a DFA state. */
530 static state_t
531 state_alloc (void)
532 {
533 return xmalloc (dfa_state_size);
534 }
535
536 /* Allocate and initialize DFA state. */
537 static state_t
538 state_create (void)
539 {
540 state_t state = state_alloc ();
541
542 state_reset (state);
543 advance_state (state);
544 return state;
545 }
546
547 /* Free DFA state. */
548 static void
549 state_free (state_t state)
550 {
551 free (state);
552 }
553
554 /* Make a copy of FROM in TO. */
555 static void
556 state_copy (state_t to, state_t from)
557 {
558 memcpy (to, from, dfa_state_size);
559 }
560
561 /* Create a copy of FROM. */
562 static state_t
563 state_create_copy (state_t from)
564 {
565 state_t to = state_alloc ();
566
567 state_copy (to, from);
568 return to;
569 }
570 \f
571
572 /* Functions to work with fences. */
573
574 /* Clear the fence. */
575 static void
576 fence_clear (fence_t f)
577 {
578 state_t s = FENCE_STATE (f);
579 deps_t dc = FENCE_DC (f);
580 void *tc = FENCE_TC (f);
581
582 ilist_clear (&FENCE_BNDS (f));
583
584 gcc_assert ((s != NULL && dc != NULL && tc != NULL)
585 || (s == NULL && dc == NULL && tc == NULL));
586
587 free (s);
588
589 if (dc != NULL)
590 delete_deps_context (dc);
591
592 if (tc != NULL)
593 delete_target_context (tc);
594 vec_free (FENCE_EXECUTING_INSNS (f));
595 free (FENCE_READY_TICKS (f));
596 FENCE_READY_TICKS (f) = NULL;
597 }
598
599 /* Init a list of fences with successors of OLD_FENCE. */
600 void
601 init_fences (insn_t old_fence)
602 {
603 insn_t succ;
604 succ_iterator si;
605 bool first = true;
606 int ready_ticks_size = get_max_uid () + 1;
607
608 FOR_EACH_SUCC_1 (succ, si, old_fence,
609 SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS)
610 {
611
612 if (first)
613 first = false;
614 else
615 gcc_assert (flag_sel_sched_pipelining_outer_loops);
616
617 flist_add (&fences, succ,
618 state_create (),
619 create_deps_context () /* dc */,
620 create_target_context (true) /* tc */,
621 NULL /* last_scheduled_insn */,
622 NULL, /* executing_insns */
623 XCNEWVEC (int, ready_ticks_size), /* ready_ticks */
624 ready_ticks_size,
625 NULL /* sched_next */,
626 1 /* cycle */, 0 /* cycle_issued_insns */,
627 issue_rate, /* issue_more */
628 1 /* starts_cycle_p */, 0 /* after_stall_p */);
629 }
630 }
631
632 /* Merges two fences (filling fields of fence F with resulting values) by
633 following rules: 1) state, target context and last scheduled insn are
634 propagated from fallthrough edge if it is available;
635 2) deps context and cycle is propagated from more probable edge;
636 3) all other fields are set to corresponding constant values.
637
638 INSN, STATE, DC, TC, LAST_SCHEDULED_INSN, EXECUTING_INSNS,
639 READY_TICKS, READY_TICKS_SIZE, SCHED_NEXT, CYCLE, ISSUE_MORE
640 and AFTER_STALL_P are the corresponding fields of the second fence. */
641 static void
642 merge_fences (fence_t f, insn_t insn,
643 state_t state, deps_t dc, void *tc,
644 rtx_insn *last_scheduled_insn,
645 vec<rtx_insn *, va_gc> *executing_insns,
646 int *ready_ticks, int ready_ticks_size,
647 rtx sched_next, int cycle, int issue_more, bool after_stall_p)
648 {
649 insn_t last_scheduled_insn_old = FENCE_LAST_SCHEDULED_INSN (f);
650
651 gcc_assert (sel_bb_head_p (FENCE_INSN (f))
652 && !sched_next && !FENCE_SCHED_NEXT (f));
653
654 /* Check if we can decide which path fences came.
655 If we can't (or don't want to) - reset all. */
656 if (last_scheduled_insn == NULL
657 || last_scheduled_insn_old == NULL
658 /* This is a case when INSN is reachable on several paths from
659 one insn (this can happen when pipelining of outer loops is on and
660 there are two edges: one going around of inner loop and the other -
661 right through it; in such case just reset everything). */
662 || last_scheduled_insn == last_scheduled_insn_old)
663 {
664 state_reset (FENCE_STATE (f));
665 state_free (state);
666
667 reset_deps_context (FENCE_DC (f));
668 delete_deps_context (dc);
669
670 reset_target_context (FENCE_TC (f), true);
671 delete_target_context (tc);
672
673 if (cycle > FENCE_CYCLE (f))
674 FENCE_CYCLE (f) = cycle;
675
676 FENCE_LAST_SCHEDULED_INSN (f) = NULL;
677 FENCE_ISSUE_MORE (f) = issue_rate;
678 vec_free (executing_insns);
679 free (ready_ticks);
680 if (FENCE_EXECUTING_INSNS (f))
681 FENCE_EXECUTING_INSNS (f)->block_remove (0,
682 FENCE_EXECUTING_INSNS (f)->length ());
683 if (FENCE_READY_TICKS (f))
684 memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f));
685 }
686 else
687 {
688 edge edge_old = NULL, edge_new = NULL;
689 edge candidate;
690 succ_iterator si;
691 insn_t succ;
692
693 /* Find fallthrough edge. */
694 gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb);
695 candidate = find_fallthru_edge_from (BLOCK_FOR_INSN (insn)->prev_bb);
696
697 if (!candidate
698 || (candidate->src != BLOCK_FOR_INSN (last_scheduled_insn)
699 && candidate->src != BLOCK_FOR_INSN (last_scheduled_insn_old)))
700 {
701 /* No fallthrough edge leading to basic block of INSN. */
702 state_reset (FENCE_STATE (f));
703 state_free (state);
704
705 reset_target_context (FENCE_TC (f), true);
706 delete_target_context (tc);
707
708 FENCE_LAST_SCHEDULED_INSN (f) = NULL;
709 FENCE_ISSUE_MORE (f) = issue_rate;
710 }
711 else
712 if (candidate->src == BLOCK_FOR_INSN (last_scheduled_insn))
713 {
714 /* Would be weird if same insn is successor of several fallthrough
715 edges. */
716 gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb
717 != BLOCK_FOR_INSN (last_scheduled_insn_old));
718
719 state_free (FENCE_STATE (f));
720 FENCE_STATE (f) = state;
721
722 delete_target_context (FENCE_TC (f));
723 FENCE_TC (f) = tc;
724
725 FENCE_LAST_SCHEDULED_INSN (f) = last_scheduled_insn;
726 FENCE_ISSUE_MORE (f) = issue_more;
727 }
728 else
729 {
730 /* Leave STATE, TC and LAST_SCHEDULED_INSN fields untouched. */
731 state_free (state);
732 delete_target_context (tc);
733
734 gcc_assert (BLOCK_FOR_INSN (insn)->prev_bb
735 != BLOCK_FOR_INSN (last_scheduled_insn));
736 }
737
738 /* Find edge of first predecessor (last_scheduled_insn_old->insn). */
739 FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn_old,
740 SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS)
741 {
742 if (succ == insn)
743 {
744 /* No same successor allowed from several edges. */
745 gcc_assert (!edge_old);
746 edge_old = si.e1;
747 }
748 }
749 /* Find edge of second predecessor (last_scheduled_insn->insn). */
750 FOR_EACH_SUCC_1 (succ, si, last_scheduled_insn,
751 SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS)
752 {
753 if (succ == insn)
754 {
755 /* No same successor allowed from several edges. */
756 gcc_assert (!edge_new);
757 edge_new = si.e1;
758 }
759 }
760
761 /* Check if we can choose most probable predecessor. */
762 if (edge_old == NULL || edge_new == NULL)
763 {
764 reset_deps_context (FENCE_DC (f));
765 delete_deps_context (dc);
766 vec_free (executing_insns);
767 free (ready_ticks);
768
769 FENCE_CYCLE (f) = MAX (FENCE_CYCLE (f), cycle);
770 if (FENCE_EXECUTING_INSNS (f))
771 FENCE_EXECUTING_INSNS (f)->block_remove (0,
772 FENCE_EXECUTING_INSNS (f)->length ());
773 if (FENCE_READY_TICKS (f))
774 memset (FENCE_READY_TICKS (f), 0, FENCE_READY_TICKS_SIZE (f));
775 }
776 else
777 if (edge_new->probability > edge_old->probability)
778 {
779 delete_deps_context (FENCE_DC (f));
780 FENCE_DC (f) = dc;
781 vec_free (FENCE_EXECUTING_INSNS (f));
782 FENCE_EXECUTING_INSNS (f) = executing_insns;
783 free (FENCE_READY_TICKS (f));
784 FENCE_READY_TICKS (f) = ready_ticks;
785 FENCE_READY_TICKS_SIZE (f) = ready_ticks_size;
786 FENCE_CYCLE (f) = cycle;
787 }
788 else
789 {
790 /* Leave DC and CYCLE untouched. */
791 delete_deps_context (dc);
792 vec_free (executing_insns);
793 free (ready_ticks);
794 }
795 }
796
797 /* Fill remaining invariant fields. */
798 if (after_stall_p)
799 FENCE_AFTER_STALL_P (f) = 1;
800
801 FENCE_ISSUED_INSNS (f) = 0;
802 FENCE_STARTS_CYCLE_P (f) = 1;
803 FENCE_SCHED_NEXT (f) = NULL;
804 }
805
806 /* Add a new fence to NEW_FENCES list, initializing it from all
807 other parameters. */
808 static void
809 add_to_fences (flist_tail_t new_fences, insn_t insn,
810 state_t state, deps_t dc, void *tc,
811 rtx_insn *last_scheduled_insn,
812 vec<rtx_insn *, va_gc> *executing_insns, int *ready_ticks,
813 int ready_ticks_size, rtx_insn *sched_next, int cycle,
814 int cycle_issued_insns, int issue_rate,
815 bool starts_cycle_p, bool after_stall_p)
816 {
817 fence_t f = flist_lookup (FLIST_TAIL_HEAD (new_fences), insn);
818
819 if (! f)
820 {
821 flist_add (FLIST_TAIL_TAILP (new_fences), insn, state, dc, tc,
822 last_scheduled_insn, executing_insns, ready_ticks,
823 ready_ticks_size, sched_next, cycle, cycle_issued_insns,
824 issue_rate, starts_cycle_p, after_stall_p);
825
826 FLIST_TAIL_TAILP (new_fences)
827 = &FLIST_NEXT (*FLIST_TAIL_TAILP (new_fences));
828 }
829 else
830 {
831 merge_fences (f, insn, state, dc, tc, last_scheduled_insn,
832 executing_insns, ready_ticks, ready_ticks_size,
833 sched_next, cycle, issue_rate, after_stall_p);
834 }
835 }
836
837 /* Move the first fence in the OLD_FENCES list to NEW_FENCES. */
838 void
839 move_fence_to_fences (flist_t old_fences, flist_tail_t new_fences)
840 {
841 fence_t f, old;
842 flist_t *tailp = FLIST_TAIL_TAILP (new_fences);
843
844 old = FLIST_FENCE (old_fences);
845 f = flist_lookup (FLIST_TAIL_HEAD (new_fences),
846 FENCE_INSN (FLIST_FENCE (old_fences)));
847 if (f)
848 {
849 merge_fences (f, old->insn, old->state, old->dc, old->tc,
850 old->last_scheduled_insn, old->executing_insns,
851 old->ready_ticks, old->ready_ticks_size,
852 old->sched_next, old->cycle, old->issue_more,
853 old->after_stall_p);
854 }
855 else
856 {
857 _list_add (tailp);
858 FLIST_TAIL_TAILP (new_fences) = &FLIST_NEXT (*tailp);
859 *FLIST_FENCE (*tailp) = *old;
860 init_fence_for_scheduling (FLIST_FENCE (*tailp));
861 }
862 FENCE_INSN (old) = NULL;
863 }
864
865 /* Add a new fence to NEW_FENCES list and initialize most of its data
866 as a clean one. */
867 void
868 add_clean_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence)
869 {
870 int ready_ticks_size = get_max_uid () + 1;
871
872 add_to_fences (new_fences,
873 succ, state_create (), create_deps_context (),
874 create_target_context (true),
875 NULL, NULL,
876 XCNEWVEC (int, ready_ticks_size), ready_ticks_size,
877 NULL, FENCE_CYCLE (fence) + 1,
878 0, issue_rate, 1, FENCE_AFTER_STALL_P (fence));
879 }
880
881 /* Add a new fence to NEW_FENCES list and initialize all of its data
882 from FENCE and SUCC. */
883 void
884 add_dirty_fence_to_fences (flist_tail_t new_fences, insn_t succ, fence_t fence)
885 {
886 int * new_ready_ticks
887 = XNEWVEC (int, FENCE_READY_TICKS_SIZE (fence));
888
889 memcpy (new_ready_ticks, FENCE_READY_TICKS (fence),
890 FENCE_READY_TICKS_SIZE (fence) * sizeof (int));
891 add_to_fences (new_fences,
892 succ, state_create_copy (FENCE_STATE (fence)),
893 create_copy_of_deps_context (FENCE_DC (fence)),
894 create_copy_of_target_context (FENCE_TC (fence)),
895 FENCE_LAST_SCHEDULED_INSN (fence),
896 vec_safe_copy (FENCE_EXECUTING_INSNS (fence)),
897 new_ready_ticks,
898 FENCE_READY_TICKS_SIZE (fence),
899 FENCE_SCHED_NEXT (fence),
900 FENCE_CYCLE (fence),
901 FENCE_ISSUED_INSNS (fence),
902 FENCE_ISSUE_MORE (fence),
903 FENCE_STARTS_CYCLE_P (fence),
904 FENCE_AFTER_STALL_P (fence));
905 }
906 \f
907
908 /* Functions to work with regset and nop pools. */
909
910 /* Returns the new regset from pool. It might have some of the bits set
911 from the previous usage. */
912 regset
913 get_regset_from_pool (void)
914 {
915 regset rs;
916
917 if (regset_pool.n != 0)
918 rs = regset_pool.v[--regset_pool.n];
919 else
920 /* We need to create the regset. */
921 {
922 rs = ALLOC_REG_SET (&reg_obstack);
923
924 if (regset_pool.nn == regset_pool.ss)
925 regset_pool.vv = XRESIZEVEC (regset, regset_pool.vv,
926 (regset_pool.ss = 2 * regset_pool.ss + 1));
927 regset_pool.vv[regset_pool.nn++] = rs;
928 }
929
930 regset_pool.diff++;
931
932 return rs;
933 }
934
935 /* Same as above, but returns the empty regset. */
936 regset
937 get_clear_regset_from_pool (void)
938 {
939 regset rs = get_regset_from_pool ();
940
941 CLEAR_REG_SET (rs);
942 return rs;
943 }
944
945 /* Return regset RS to the pool for future use. */
946 void
947 return_regset_to_pool (regset rs)
948 {
949 gcc_assert (rs);
950 regset_pool.diff--;
951
952 if (regset_pool.n == regset_pool.s)
953 regset_pool.v = XRESIZEVEC (regset, regset_pool.v,
954 (regset_pool.s = 2 * regset_pool.s + 1));
955 regset_pool.v[regset_pool.n++] = rs;
956 }
957
958 #ifdef ENABLE_CHECKING
959 /* This is used as a qsort callback for sorting regset pool stacks.
960 X and XX are addresses of two regsets. They are never equal. */
961 static int
962 cmp_v_in_regset_pool (const void *x, const void *xx)
963 {
964 uintptr_t r1 = (uintptr_t) *((const regset *) x);
965 uintptr_t r2 = (uintptr_t) *((const regset *) xx);
966 if (r1 > r2)
967 return 1;
968 else if (r1 < r2)
969 return -1;
970 gcc_unreachable ();
971 }
972 #endif
973
974 /* Free the regset pool possibly checking for memory leaks. */
975 void
976 free_regset_pool (void)
977 {
978 #ifdef ENABLE_CHECKING
979 {
980 regset *v = regset_pool.v;
981 int i = 0;
982 int n = regset_pool.n;
983
984 regset *vv = regset_pool.vv;
985 int ii = 0;
986 int nn = regset_pool.nn;
987
988 int diff = 0;
989
990 gcc_assert (n <= nn);
991
992 /* Sort both vectors so it will be possible to compare them. */
993 qsort (v, n, sizeof (*v), cmp_v_in_regset_pool);
994 qsort (vv, nn, sizeof (*vv), cmp_v_in_regset_pool);
995
996 while (ii < nn)
997 {
998 if (v[i] == vv[ii])
999 i++;
1000 else
1001 /* VV[II] was lost. */
1002 diff++;
1003
1004 ii++;
1005 }
1006
1007 gcc_assert (diff == regset_pool.diff);
1008 }
1009 #endif
1010
1011 /* If not true - we have a memory leak. */
1012 gcc_assert (regset_pool.diff == 0);
1013
1014 while (regset_pool.n)
1015 {
1016 --regset_pool.n;
1017 FREE_REG_SET (regset_pool.v[regset_pool.n]);
1018 }
1019
1020 free (regset_pool.v);
1021 regset_pool.v = NULL;
1022 regset_pool.s = 0;
1023
1024 free (regset_pool.vv);
1025 regset_pool.vv = NULL;
1026 regset_pool.nn = 0;
1027 regset_pool.ss = 0;
1028
1029 regset_pool.diff = 0;
1030 }
1031 \f
1032
1033 /* Functions to work with nop pools. NOP insns are used as temporary
1034 placeholders of the insns being scheduled to allow correct update of
1035 the data sets. When update is finished, NOPs are deleted. */
1036
1037 /* A vinsn that is used to represent a nop. This vinsn is shared among all
1038 nops sel-sched generates. */
1039 static vinsn_t nop_vinsn = NULL;
1040
1041 /* Emit a nop before INSN, taking it from pool. */
1042 insn_t
1043 get_nop_from_pool (insn_t insn)
1044 {
1045 rtx nop_pat;
1046 insn_t nop;
1047 bool old_p = nop_pool.n != 0;
1048 int flags;
1049
1050 if (old_p)
1051 nop_pat = nop_pool.v[--nop_pool.n];
1052 else
1053 nop_pat = nop_pattern;
1054
1055 nop = emit_insn_before (nop_pat, insn);
1056
1057 if (old_p)
1058 flags = INSN_INIT_TODO_SSID;
1059 else
1060 flags = INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID;
1061
1062 set_insn_init (INSN_EXPR (insn), nop_vinsn, INSN_SEQNO (insn));
1063 sel_init_new_insn (nop, flags);
1064
1065 return nop;
1066 }
1067
1068 /* Remove NOP from the instruction stream and return it to the pool. */
1069 void
1070 return_nop_to_pool (insn_t nop, bool full_tidying)
1071 {
1072 gcc_assert (INSN_IN_STREAM_P (nop));
1073 sel_remove_insn (nop, false, full_tidying);
1074
1075 /* We'll recycle this nop. */
1076 nop->set_undeleted ();
1077
1078 if (nop_pool.n == nop_pool.s)
1079 nop_pool.v = XRESIZEVEC (rtx_insn *, nop_pool.v,
1080 (nop_pool.s = 2 * nop_pool.s + 1));
1081 nop_pool.v[nop_pool.n++] = nop;
1082 }
1083
1084 /* Free the nop pool. */
1085 void
1086 free_nop_pool (void)
1087 {
1088 nop_pool.n = 0;
1089 nop_pool.s = 0;
1090 free (nop_pool.v);
1091 nop_pool.v = NULL;
1092 }
1093 \f
1094
1095 /* Skip unspec to support ia64 speculation. Called from rtx_equal_p_cb.
1096 The callback is given two rtxes XX and YY and writes the new rtxes
1097 to NX and NY in case some needs to be skipped. */
1098 static int
1099 skip_unspecs_callback (const_rtx *xx, const_rtx *yy, rtx *nx, rtx* ny)
1100 {
1101 const_rtx x = *xx;
1102 const_rtx y = *yy;
1103
1104 if (GET_CODE (x) == UNSPEC
1105 && (targetm.sched.skip_rtx_p == NULL
1106 || targetm.sched.skip_rtx_p (x)))
1107 {
1108 *nx = XVECEXP (x, 0, 0);
1109 *ny = CONST_CAST_RTX (y);
1110 return 1;
1111 }
1112
1113 if (GET_CODE (y) == UNSPEC
1114 && (targetm.sched.skip_rtx_p == NULL
1115 || targetm.sched.skip_rtx_p (y)))
1116 {
1117 *nx = CONST_CAST_RTX (x);
1118 *ny = XVECEXP (y, 0, 0);
1119 return 1;
1120 }
1121
1122 return 0;
1123 }
1124
1125 /* Callback, called from hash_rtx_cb. Helps to hash UNSPEC rtx X in a correct way
1126 to support ia64 speculation. When changes are needed, new rtx X and new mode
1127 NMODE are written, and the callback returns true. */
1128 static int
1129 hash_with_unspec_callback (const_rtx x, enum machine_mode mode ATTRIBUTE_UNUSED,
1130 rtx *nx, enum machine_mode* nmode)
1131 {
1132 if (GET_CODE (x) == UNSPEC
1133 && targetm.sched.skip_rtx_p
1134 && targetm.sched.skip_rtx_p (x))
1135 {
1136 *nx = XVECEXP (x, 0 ,0);
1137 *nmode = VOIDmode;
1138 return 1;
1139 }
1140
1141 return 0;
1142 }
1143
1144 /* Returns LHS and RHS are ok to be scheduled separately. */
1145 static bool
1146 lhs_and_rhs_separable_p (rtx lhs, rtx rhs)
1147 {
1148 if (lhs == NULL || rhs == NULL)
1149 return false;
1150
1151 /* Do not schedule constants as rhs: no point to use reg, if const
1152 can be used. Moreover, scheduling const as rhs may lead to mode
1153 mismatch cause consts don't have modes but they could be merged
1154 from branches where the same const used in different modes. */
1155 if (CONSTANT_P (rhs))
1156 return false;
1157
1158 /* ??? Do not rename predicate registers to avoid ICEs in bundling. */
1159 if (COMPARISON_P (rhs))
1160 return false;
1161
1162 /* Do not allow single REG to be an rhs. */
1163 if (REG_P (rhs))
1164 return false;
1165
1166 /* See comment at find_used_regs_1 (*1) for explanation of this
1167 restriction. */
1168 /* FIXME: remove this later. */
1169 if (MEM_P (lhs))
1170 return false;
1171
1172 /* This will filter all tricky things like ZERO_EXTRACT etc.
1173 For now we don't handle it. */
1174 if (!REG_P (lhs) && !MEM_P (lhs))
1175 return false;
1176
1177 return true;
1178 }
1179
1180 /* Initialize vinsn VI for INSN. Only for use from vinsn_create (). When
1181 FORCE_UNIQUE_P is true, the resulting vinsn will not be clonable. This is
1182 used e.g. for insns from recovery blocks. */
1183 static void
1184 vinsn_init (vinsn_t vi, insn_t insn, bool force_unique_p)
1185 {
1186 hash_rtx_callback_function hrcf;
1187 int insn_class;
1188
1189 VINSN_INSN_RTX (vi) = insn;
1190 VINSN_COUNT (vi) = 0;
1191 vi->cost = -1;
1192
1193 if (INSN_NOP_P (insn))
1194 return;
1195
1196 if (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL)
1197 init_id_from_df (VINSN_ID (vi), insn, force_unique_p);
1198 else
1199 deps_init_id (VINSN_ID (vi), insn, force_unique_p);
1200
1201 /* Hash vinsn depending on whether it is separable or not. */
1202 hrcf = targetm.sched.skip_rtx_p ? hash_with_unspec_callback : NULL;
1203 if (VINSN_SEPARABLE_P (vi))
1204 {
1205 rtx rhs = VINSN_RHS (vi);
1206
1207 VINSN_HASH (vi) = hash_rtx_cb (rhs, GET_MODE (rhs),
1208 NULL, NULL, false, hrcf);
1209 VINSN_HASH_RTX (vi) = hash_rtx_cb (VINSN_PATTERN (vi),
1210 VOIDmode, NULL, NULL,
1211 false, hrcf);
1212 }
1213 else
1214 {
1215 VINSN_HASH (vi) = hash_rtx_cb (VINSN_PATTERN (vi), VOIDmode,
1216 NULL, NULL, false, hrcf);
1217 VINSN_HASH_RTX (vi) = VINSN_HASH (vi);
1218 }
1219
1220 insn_class = haifa_classify_insn (insn);
1221 if (insn_class >= 2
1222 && (!targetm.sched.get_insn_spec_ds
1223 || ((targetm.sched.get_insn_spec_ds (insn) & BEGIN_CONTROL)
1224 == 0)))
1225 VINSN_MAY_TRAP_P (vi) = true;
1226 else
1227 VINSN_MAY_TRAP_P (vi) = false;
1228 }
1229
1230 /* Indicate that VI has become the part of an rtx object. */
1231 void
1232 vinsn_attach (vinsn_t vi)
1233 {
1234 /* Assert that VI is not pending for deletion. */
1235 gcc_assert (VINSN_INSN_RTX (vi));
1236
1237 VINSN_COUNT (vi)++;
1238 }
1239
1240 /* Create and init VI from the INSN. Use UNIQUE_P for determining the correct
1241 VINSN_TYPE (VI). */
1242 static vinsn_t
1243 vinsn_create (insn_t insn, bool force_unique_p)
1244 {
1245 vinsn_t vi = XCNEW (struct vinsn_def);
1246
1247 vinsn_init (vi, insn, force_unique_p);
1248 return vi;
1249 }
1250
1251 /* Return a copy of VI. When REATTACH_P is true, detach VI and attach
1252 the copy. */
1253 vinsn_t
1254 vinsn_copy (vinsn_t vi, bool reattach_p)
1255 {
1256 rtx_insn *copy;
1257 bool unique = VINSN_UNIQUE_P (vi);
1258 vinsn_t new_vi;
1259
1260 copy = create_copy_of_insn_rtx (VINSN_INSN_RTX (vi));
1261 new_vi = create_vinsn_from_insn_rtx (copy, unique);
1262 if (reattach_p)
1263 {
1264 vinsn_detach (vi);
1265 vinsn_attach (new_vi);
1266 }
1267
1268 return new_vi;
1269 }
1270
1271 /* Delete the VI vinsn and free its data. */
1272 static void
1273 vinsn_delete (vinsn_t vi)
1274 {
1275 gcc_assert (VINSN_COUNT (vi) == 0);
1276
1277 if (!INSN_NOP_P (VINSN_INSN_RTX (vi)))
1278 {
1279 return_regset_to_pool (VINSN_REG_SETS (vi));
1280 return_regset_to_pool (VINSN_REG_USES (vi));
1281 return_regset_to_pool (VINSN_REG_CLOBBERS (vi));
1282 }
1283
1284 free (vi);
1285 }
1286
1287 /* Indicate that VI is no longer a part of some rtx object.
1288 Remove VI if it is no longer needed. */
1289 void
1290 vinsn_detach (vinsn_t vi)
1291 {
1292 gcc_assert (VINSN_COUNT (vi) > 0);
1293
1294 if (--VINSN_COUNT (vi) == 0)
1295 vinsn_delete (vi);
1296 }
1297
1298 /* Returns TRUE if VI is a branch. */
1299 bool
1300 vinsn_cond_branch_p (vinsn_t vi)
1301 {
1302 insn_t insn;
1303
1304 if (!VINSN_UNIQUE_P (vi))
1305 return false;
1306
1307 insn = VINSN_INSN_RTX (vi);
1308 if (BB_END (BLOCK_FOR_INSN (insn)) != insn)
1309 return false;
1310
1311 return control_flow_insn_p (insn);
1312 }
1313
1314 /* Return latency of INSN. */
1315 static int
1316 sel_insn_rtx_cost (rtx_insn *insn)
1317 {
1318 int cost;
1319
1320 /* A USE insn, or something else we don't need to
1321 understand. We can't pass these directly to
1322 result_ready_cost or insn_default_latency because it will
1323 trigger a fatal error for unrecognizable insns. */
1324 if (recog_memoized (insn) < 0)
1325 cost = 0;
1326 else
1327 {
1328 cost = insn_default_latency (insn);
1329
1330 if (cost < 0)
1331 cost = 0;
1332 }
1333
1334 return cost;
1335 }
1336
1337 /* Return the cost of the VI.
1338 !!! FIXME: Unify with haifa-sched.c: insn_cost (). */
1339 int
1340 sel_vinsn_cost (vinsn_t vi)
1341 {
1342 int cost = vi->cost;
1343
1344 if (cost < 0)
1345 {
1346 cost = sel_insn_rtx_cost (VINSN_INSN_RTX (vi));
1347 vi->cost = cost;
1348 }
1349
1350 return cost;
1351 }
1352 \f
1353
1354 /* Functions for insn emitting. */
1355
1356 /* Emit new insn after AFTER based on PATTERN and initialize its data from
1357 EXPR and SEQNO. */
1358 insn_t
1359 sel_gen_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno, insn_t after)
1360 {
1361 insn_t new_insn;
1362
1363 gcc_assert (EXPR_TARGET_AVAILABLE (expr) == true);
1364
1365 new_insn = emit_insn_after (pattern, after);
1366 set_insn_init (expr, NULL, seqno);
1367 sel_init_new_insn (new_insn, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SSID);
1368
1369 return new_insn;
1370 }
1371
1372 /* Force newly generated vinsns to be unique. */
1373 static bool init_insn_force_unique_p = false;
1374
1375 /* Emit new speculation recovery insn after AFTER based on PATTERN and
1376 initialize its data from EXPR and SEQNO. */
1377 insn_t
1378 sel_gen_recovery_insn_from_rtx_after (rtx pattern, expr_t expr, int seqno,
1379 insn_t after)
1380 {
1381 insn_t insn;
1382
1383 gcc_assert (!init_insn_force_unique_p);
1384
1385 init_insn_force_unique_p = true;
1386 insn = sel_gen_insn_from_rtx_after (pattern, expr, seqno, after);
1387 CANT_MOVE (insn) = 1;
1388 init_insn_force_unique_p = false;
1389
1390 return insn;
1391 }
1392
1393 /* Emit new insn after AFTER based on EXPR and SEQNO. If VINSN is not NULL,
1394 take it as a new vinsn instead of EXPR's vinsn.
1395 We simplify insns later, after scheduling region in
1396 simplify_changed_insns. */
1397 insn_t
1398 sel_gen_insn_from_expr_after (expr_t expr, vinsn_t vinsn, int seqno,
1399 insn_t after)
1400 {
1401 expr_t emit_expr;
1402 insn_t insn;
1403 int flags;
1404
1405 emit_expr = set_insn_init (expr, vinsn ? vinsn : EXPR_VINSN (expr),
1406 seqno);
1407 insn = EXPR_INSN_RTX (emit_expr);
1408
1409 /* The insn may come from the transformation cache, which may hold already
1410 deleted insns, so mark it as not deleted. */
1411 insn->set_undeleted ();
1412
1413 add_insn_after (insn, after, BLOCK_FOR_INSN (insn));
1414
1415 flags = INSN_INIT_TODO_SSID;
1416 if (INSN_LUID (insn) == 0)
1417 flags |= INSN_INIT_TODO_LUID;
1418 sel_init_new_insn (insn, flags);
1419
1420 return insn;
1421 }
1422
1423 /* Move insn from EXPR after AFTER. */
1424 insn_t
1425 sel_move_insn (expr_t expr, int seqno, insn_t after)
1426 {
1427 insn_t insn = EXPR_INSN_RTX (expr);
1428 basic_block bb = BLOCK_FOR_INSN (after);
1429 insn_t next = NEXT_INSN (after);
1430
1431 /* Assert that in move_op we disconnected this insn properly. */
1432 gcc_assert (EXPR_VINSN (INSN_EXPR (insn)) != NULL);
1433 SET_PREV_INSN (insn) = after;
1434 SET_NEXT_INSN (insn) = next;
1435
1436 SET_NEXT_INSN (after) = insn;
1437 SET_PREV_INSN (next) = insn;
1438
1439 /* Update links from insn to bb and vice versa. */
1440 df_insn_change_bb (insn, bb);
1441 if (BB_END (bb) == after)
1442 BB_END (bb) = insn;
1443
1444 prepare_insn_expr (insn, seqno);
1445 return insn;
1446 }
1447
1448 \f
1449 /* Functions to work with right-hand sides. */
1450
1451 /* Search for a hash value determined by UID/NEW_VINSN in a sorted vector
1452 VECT and return true when found. Use NEW_VINSN for comparison only when
1453 COMPARE_VINSNS is true. Write to INDP the index on which
1454 the search has stopped, such that inserting the new element at INDP will
1455 retain VECT's sort order. */
1456 static bool
1457 find_in_history_vect_1 (vec<expr_history_def> vect,
1458 unsigned uid, vinsn_t new_vinsn,
1459 bool compare_vinsns, int *indp)
1460 {
1461 expr_history_def *arr;
1462 int i, j, len = vect.length ();
1463
1464 if (len == 0)
1465 {
1466 *indp = 0;
1467 return false;
1468 }
1469
1470 arr = vect.address ();
1471 i = 0, j = len - 1;
1472
1473 while (i <= j)
1474 {
1475 unsigned auid = arr[i].uid;
1476 vinsn_t avinsn = arr[i].new_expr_vinsn;
1477
1478 if (auid == uid
1479 /* When undoing transformation on a bookkeeping copy, the new vinsn
1480 may not be exactly equal to the one that is saved in the vector.
1481 This is because the insn whose copy we're checking was possibly
1482 substituted itself. */
1483 && (! compare_vinsns
1484 || vinsn_equal_p (avinsn, new_vinsn)))
1485 {
1486 *indp = i;
1487 return true;
1488 }
1489 else if (auid > uid)
1490 break;
1491 i++;
1492 }
1493
1494 *indp = i;
1495 return false;
1496 }
1497
1498 /* Search for a uid of INSN and NEW_VINSN in a sorted vector VECT. Return
1499 the position found or -1, if no such value is in vector.
1500 Search also for UIDs of insn's originators, if ORIGINATORS_P is true. */
1501 int
1502 find_in_history_vect (vec<expr_history_def> vect, rtx insn,
1503 vinsn_t new_vinsn, bool originators_p)
1504 {
1505 int ind;
1506
1507 if (find_in_history_vect_1 (vect, INSN_UID (insn), new_vinsn,
1508 false, &ind))
1509 return ind;
1510
1511 if (INSN_ORIGINATORS (insn) && originators_p)
1512 {
1513 unsigned uid;
1514 bitmap_iterator bi;
1515
1516 EXECUTE_IF_SET_IN_BITMAP (INSN_ORIGINATORS (insn), 0, uid, bi)
1517 if (find_in_history_vect_1 (vect, uid, new_vinsn, false, &ind))
1518 return ind;
1519 }
1520
1521 return -1;
1522 }
1523
1524 /* Insert new element in a sorted history vector pointed to by PVECT,
1525 if it is not there already. The element is searched using
1526 UID/NEW_EXPR_VINSN pair. TYPE, OLD_EXPR_VINSN and SPEC_DS save
1527 the history of a transformation. */
1528 void
1529 insert_in_history_vect (vec<expr_history_def> *pvect,
1530 unsigned uid, enum local_trans_type type,
1531 vinsn_t old_expr_vinsn, vinsn_t new_expr_vinsn,
1532 ds_t spec_ds)
1533 {
1534 vec<expr_history_def> vect = *pvect;
1535 expr_history_def temp;
1536 bool res;
1537 int ind;
1538
1539 res = find_in_history_vect_1 (vect, uid, new_expr_vinsn, true, &ind);
1540
1541 if (res)
1542 {
1543 expr_history_def *phist = &vect[ind];
1544
1545 /* It is possible that speculation types of expressions that were
1546 propagated through different paths will be different here. In this
1547 case, merge the status to get the correct check later. */
1548 if (phist->spec_ds != spec_ds)
1549 phist->spec_ds = ds_max_merge (phist->spec_ds, spec_ds);
1550 return;
1551 }
1552
1553 temp.uid = uid;
1554 temp.old_expr_vinsn = old_expr_vinsn;
1555 temp.new_expr_vinsn = new_expr_vinsn;
1556 temp.spec_ds = spec_ds;
1557 temp.type = type;
1558
1559 vinsn_attach (old_expr_vinsn);
1560 vinsn_attach (new_expr_vinsn);
1561 vect.safe_insert (ind, temp);
1562 *pvect = vect;
1563 }
1564
1565 /* Free history vector PVECT. */
1566 static void
1567 free_history_vect (vec<expr_history_def> &pvect)
1568 {
1569 unsigned i;
1570 expr_history_def *phist;
1571
1572 if (! pvect.exists ())
1573 return;
1574
1575 for (i = 0; pvect.iterate (i, &phist); i++)
1576 {
1577 vinsn_detach (phist->old_expr_vinsn);
1578 vinsn_detach (phist->new_expr_vinsn);
1579 }
1580
1581 pvect.release ();
1582 }
1583
1584 /* Merge vector FROM to PVECT. */
1585 static void
1586 merge_history_vect (vec<expr_history_def> *pvect,
1587 vec<expr_history_def> from)
1588 {
1589 expr_history_def *phist;
1590 int i;
1591
1592 /* We keep this vector sorted. */
1593 for (i = 0; from.iterate (i, &phist); i++)
1594 insert_in_history_vect (pvect, phist->uid, phist->type,
1595 phist->old_expr_vinsn, phist->new_expr_vinsn,
1596 phist->spec_ds);
1597 }
1598
1599 /* Compare two vinsns as rhses if possible and as vinsns otherwise. */
1600 bool
1601 vinsn_equal_p (vinsn_t x, vinsn_t y)
1602 {
1603 rtx_equal_p_callback_function repcf;
1604
1605 if (x == y)
1606 return true;
1607
1608 if (VINSN_TYPE (x) != VINSN_TYPE (y))
1609 return false;
1610
1611 if (VINSN_HASH (x) != VINSN_HASH (y))
1612 return false;
1613
1614 repcf = targetm.sched.skip_rtx_p ? skip_unspecs_callback : NULL;
1615 if (VINSN_SEPARABLE_P (x))
1616 {
1617 /* Compare RHSes of VINSNs. */
1618 gcc_assert (VINSN_RHS (x));
1619 gcc_assert (VINSN_RHS (y));
1620
1621 return rtx_equal_p_cb (VINSN_RHS (x), VINSN_RHS (y), repcf);
1622 }
1623
1624 return rtx_equal_p_cb (VINSN_PATTERN (x), VINSN_PATTERN (y), repcf);
1625 }
1626 \f
1627
1628 /* Functions for working with expressions. */
1629
1630 /* Initialize EXPR. */
1631 static void
1632 init_expr (expr_t expr, vinsn_t vi, int spec, int use, int priority,
1633 int sched_times, int orig_bb_index, ds_t spec_done_ds,
1634 ds_t spec_to_check_ds, int orig_sched_cycle,
1635 vec<expr_history_def> history,
1636 signed char target_available,
1637 bool was_substituted, bool was_renamed, bool needs_spec_check_p,
1638 bool cant_move)
1639 {
1640 vinsn_attach (vi);
1641
1642 EXPR_VINSN (expr) = vi;
1643 EXPR_SPEC (expr) = spec;
1644 EXPR_USEFULNESS (expr) = use;
1645 EXPR_PRIORITY (expr) = priority;
1646 EXPR_PRIORITY_ADJ (expr) = 0;
1647 EXPR_SCHED_TIMES (expr) = sched_times;
1648 EXPR_ORIG_BB_INDEX (expr) = orig_bb_index;
1649 EXPR_ORIG_SCHED_CYCLE (expr) = orig_sched_cycle;
1650 EXPR_SPEC_DONE_DS (expr) = spec_done_ds;
1651 EXPR_SPEC_TO_CHECK_DS (expr) = spec_to_check_ds;
1652
1653 if (history.exists ())
1654 EXPR_HISTORY_OF_CHANGES (expr) = history;
1655 else
1656 EXPR_HISTORY_OF_CHANGES (expr).create (0);
1657
1658 EXPR_TARGET_AVAILABLE (expr) = target_available;
1659 EXPR_WAS_SUBSTITUTED (expr) = was_substituted;
1660 EXPR_WAS_RENAMED (expr) = was_renamed;
1661 EXPR_NEEDS_SPEC_CHECK_P (expr) = needs_spec_check_p;
1662 EXPR_CANT_MOVE (expr) = cant_move;
1663 }
1664
1665 /* Make a copy of the expr FROM into the expr TO. */
1666 void
1667 copy_expr (expr_t to, expr_t from)
1668 {
1669 vec<expr_history_def> temp = vNULL;
1670
1671 if (EXPR_HISTORY_OF_CHANGES (from).exists ())
1672 {
1673 unsigned i;
1674 expr_history_def *phist;
1675
1676 temp = EXPR_HISTORY_OF_CHANGES (from).copy ();
1677 for (i = 0;
1678 temp.iterate (i, &phist);
1679 i++)
1680 {
1681 vinsn_attach (phist->old_expr_vinsn);
1682 vinsn_attach (phist->new_expr_vinsn);
1683 }
1684 }
1685
1686 init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from),
1687 EXPR_USEFULNESS (from), EXPR_PRIORITY (from),
1688 EXPR_SCHED_TIMES (from), EXPR_ORIG_BB_INDEX (from),
1689 EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from),
1690 EXPR_ORIG_SCHED_CYCLE (from), temp,
1691 EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from),
1692 EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from),
1693 EXPR_CANT_MOVE (from));
1694 }
1695
1696 /* Same, but the final expr will not ever be in av sets, so don't copy
1697 "uninteresting" data such as bitmap cache. */
1698 void
1699 copy_expr_onside (expr_t to, expr_t from)
1700 {
1701 init_expr (to, EXPR_VINSN (from), EXPR_SPEC (from), EXPR_USEFULNESS (from),
1702 EXPR_PRIORITY (from), EXPR_SCHED_TIMES (from), 0,
1703 EXPR_SPEC_DONE_DS (from), EXPR_SPEC_TO_CHECK_DS (from), 0,
1704 vNULL,
1705 EXPR_TARGET_AVAILABLE (from), EXPR_WAS_SUBSTITUTED (from),
1706 EXPR_WAS_RENAMED (from), EXPR_NEEDS_SPEC_CHECK_P (from),
1707 EXPR_CANT_MOVE (from));
1708 }
1709
1710 /* Prepare the expr of INSN for scheduling. Used when moving insn and when
1711 initializing new insns. */
1712 static void
1713 prepare_insn_expr (insn_t insn, int seqno)
1714 {
1715 expr_t expr = INSN_EXPR (insn);
1716 ds_t ds;
1717
1718 INSN_SEQNO (insn) = seqno;
1719 EXPR_ORIG_BB_INDEX (expr) = BLOCK_NUM (insn);
1720 EXPR_SPEC (expr) = 0;
1721 EXPR_ORIG_SCHED_CYCLE (expr) = 0;
1722 EXPR_WAS_SUBSTITUTED (expr) = 0;
1723 EXPR_WAS_RENAMED (expr) = 0;
1724 EXPR_TARGET_AVAILABLE (expr) = 1;
1725 INSN_LIVE_VALID_P (insn) = false;
1726
1727 /* ??? If this expression is speculative, make its dependence
1728 as weak as possible. We can filter this expression later
1729 in process_spec_exprs, because we do not distinguish
1730 between the status we got during compute_av_set and the
1731 existing status. To be fixed. */
1732 ds = EXPR_SPEC_DONE_DS (expr);
1733 if (ds)
1734 EXPR_SPEC_DONE_DS (expr) = ds_get_max_dep_weak (ds);
1735
1736 free_history_vect (EXPR_HISTORY_OF_CHANGES (expr));
1737 }
1738
1739 /* Update target_available bits when merging exprs TO and FROM. SPLIT_POINT
1740 is non-null when expressions are merged from different successors at
1741 a split point. */
1742 static void
1743 update_target_availability (expr_t to, expr_t from, insn_t split_point)
1744 {
1745 if (EXPR_TARGET_AVAILABLE (to) < 0
1746 || EXPR_TARGET_AVAILABLE (from) < 0)
1747 EXPR_TARGET_AVAILABLE (to) = -1;
1748 else
1749 {
1750 /* We try to detect the case when one of the expressions
1751 can only be reached through another one. In this case,
1752 we can do better. */
1753 if (split_point == NULL)
1754 {
1755 int toind, fromind;
1756
1757 toind = EXPR_ORIG_BB_INDEX (to);
1758 fromind = EXPR_ORIG_BB_INDEX (from);
1759
1760 if (toind && toind == fromind)
1761 /* Do nothing -- everything is done in
1762 merge_with_other_exprs. */
1763 ;
1764 else
1765 EXPR_TARGET_AVAILABLE (to) = -1;
1766 }
1767 else if (EXPR_TARGET_AVAILABLE (from) == 0
1768 && EXPR_LHS (from)
1769 && REG_P (EXPR_LHS (from))
1770 && REGNO (EXPR_LHS (to)) != REGNO (EXPR_LHS (from)))
1771 EXPR_TARGET_AVAILABLE (to) = -1;
1772 else
1773 EXPR_TARGET_AVAILABLE (to) &= EXPR_TARGET_AVAILABLE (from);
1774 }
1775 }
1776
1777 /* Update speculation bits when merging exprs TO and FROM. SPLIT_POINT
1778 is non-null when expressions are merged from different successors at
1779 a split point. */
1780 static void
1781 update_speculative_bits (expr_t to, expr_t from, insn_t split_point)
1782 {
1783 ds_t old_to_ds, old_from_ds;
1784
1785 old_to_ds = EXPR_SPEC_DONE_DS (to);
1786 old_from_ds = EXPR_SPEC_DONE_DS (from);
1787
1788 EXPR_SPEC_DONE_DS (to) = ds_max_merge (old_to_ds, old_from_ds);
1789 EXPR_SPEC_TO_CHECK_DS (to) |= EXPR_SPEC_TO_CHECK_DS (from);
1790 EXPR_NEEDS_SPEC_CHECK_P (to) |= EXPR_NEEDS_SPEC_CHECK_P (from);
1791
1792 /* When merging e.g. control & data speculative exprs, or a control
1793 speculative with a control&data speculative one, we really have
1794 to change vinsn too. Also, when speculative status is changed,
1795 we also need to record this as a transformation in expr's history. */
1796 if ((old_to_ds & SPECULATIVE) || (old_from_ds & SPECULATIVE))
1797 {
1798 old_to_ds = ds_get_speculation_types (old_to_ds);
1799 old_from_ds = ds_get_speculation_types (old_from_ds);
1800
1801 if (old_to_ds != old_from_ds)
1802 {
1803 ds_t record_ds;
1804
1805 /* When both expressions are speculative, we need to change
1806 the vinsn first. */
1807 if ((old_to_ds & SPECULATIVE) && (old_from_ds & SPECULATIVE))
1808 {
1809 int res;
1810
1811 res = speculate_expr (to, EXPR_SPEC_DONE_DS (to));
1812 gcc_assert (res >= 0);
1813 }
1814
1815 if (split_point != NULL)
1816 {
1817 /* Record the change with proper status. */
1818 record_ds = EXPR_SPEC_DONE_DS (to) & SPECULATIVE;
1819 record_ds &= ~(old_to_ds & SPECULATIVE);
1820 record_ds &= ~(old_from_ds & SPECULATIVE);
1821
1822 insert_in_history_vect (&EXPR_HISTORY_OF_CHANGES (to),
1823 INSN_UID (split_point), TRANS_SPECULATION,
1824 EXPR_VINSN (from), EXPR_VINSN (to),
1825 record_ds);
1826 }
1827 }
1828 }
1829 }
1830
1831
1832 /* Merge bits of FROM expr to TO expr. When SPLIT_POINT is not NULL,
1833 this is done along different paths. */
1834 void
1835 merge_expr_data (expr_t to, expr_t from, insn_t split_point)
1836 {
1837 /* Choose the maximum of the specs of merged exprs. This is required
1838 for correctness of bookkeeping. */
1839 if (EXPR_SPEC (to) < EXPR_SPEC (from))
1840 EXPR_SPEC (to) = EXPR_SPEC (from);
1841
1842 if (split_point)
1843 EXPR_USEFULNESS (to) += EXPR_USEFULNESS (from);
1844 else
1845 EXPR_USEFULNESS (to) = MAX (EXPR_USEFULNESS (to),
1846 EXPR_USEFULNESS (from));
1847
1848 if (EXPR_PRIORITY (to) < EXPR_PRIORITY (from))
1849 EXPR_PRIORITY (to) = EXPR_PRIORITY (from);
1850
1851 if (EXPR_SCHED_TIMES (to) > EXPR_SCHED_TIMES (from))
1852 EXPR_SCHED_TIMES (to) = EXPR_SCHED_TIMES (from);
1853
1854 if (EXPR_ORIG_BB_INDEX (to) != EXPR_ORIG_BB_INDEX (from))
1855 EXPR_ORIG_BB_INDEX (to) = 0;
1856
1857 EXPR_ORIG_SCHED_CYCLE (to) = MIN (EXPR_ORIG_SCHED_CYCLE (to),
1858 EXPR_ORIG_SCHED_CYCLE (from));
1859
1860 EXPR_WAS_SUBSTITUTED (to) |= EXPR_WAS_SUBSTITUTED (from);
1861 EXPR_WAS_RENAMED (to) |= EXPR_WAS_RENAMED (from);
1862 EXPR_CANT_MOVE (to) |= EXPR_CANT_MOVE (from);
1863
1864 merge_history_vect (&EXPR_HISTORY_OF_CHANGES (to),
1865 EXPR_HISTORY_OF_CHANGES (from));
1866 update_target_availability (to, from, split_point);
1867 update_speculative_bits (to, from, split_point);
1868 }
1869
1870 /* Merge bits of FROM expr to TO expr. Vinsns in the exprs should be equal
1871 in terms of vinsn_equal_p. SPLIT_POINT is non-null when expressions
1872 are merged from different successors at a split point. */
1873 void
1874 merge_expr (expr_t to, expr_t from, insn_t split_point)
1875 {
1876 vinsn_t to_vi = EXPR_VINSN (to);
1877 vinsn_t from_vi = EXPR_VINSN (from);
1878
1879 gcc_assert (vinsn_equal_p (to_vi, from_vi));
1880
1881 /* Make sure that speculative pattern is propagated into exprs that
1882 have non-speculative one. This will provide us with consistent
1883 speculative bits and speculative patterns inside expr. */
1884 if ((EXPR_SPEC_DONE_DS (from) != 0
1885 && EXPR_SPEC_DONE_DS (to) == 0)
1886 /* Do likewise for volatile insns, so that we always retain
1887 the may_trap_p bit on the resulting expression. */
1888 || (VINSN_MAY_TRAP_P (EXPR_VINSN (from))
1889 && !VINSN_MAY_TRAP_P (EXPR_VINSN (to))))
1890 change_vinsn_in_expr (to, EXPR_VINSN (from));
1891
1892 merge_expr_data (to, from, split_point);
1893 gcc_assert (EXPR_USEFULNESS (to) <= REG_BR_PROB_BASE);
1894 }
1895
1896 /* Clear the information of this EXPR. */
1897 void
1898 clear_expr (expr_t expr)
1899 {
1900
1901 vinsn_detach (EXPR_VINSN (expr));
1902 EXPR_VINSN (expr) = NULL;
1903
1904 free_history_vect (EXPR_HISTORY_OF_CHANGES (expr));
1905 }
1906
1907 /* For a given LV_SET, mark EXPR having unavailable target register. */
1908 static void
1909 set_unavailable_target_for_expr (expr_t expr, regset lv_set)
1910 {
1911 if (EXPR_SEPARABLE_P (expr))
1912 {
1913 if (REG_P (EXPR_LHS (expr))
1914 && register_unavailable_p (lv_set, EXPR_LHS (expr)))
1915 {
1916 /* If it's an insn like r1 = use (r1, ...), and it exists in
1917 different forms in each of the av_sets being merged, we can't say
1918 whether original destination register is available or not.
1919 However, this still works if destination register is not used
1920 in the original expression: if the branch at which LV_SET we're
1921 looking here is not actually 'other branch' in sense that same
1922 expression is available through it (but it can't be determined
1923 at computation stage because of transformations on one of the
1924 branches), it still won't affect the availability.
1925 Liveness of a register somewhere on a code motion path means
1926 it's either read somewhere on a codemotion path, live on
1927 'other' branch, live at the point immediately following
1928 the original operation, or is read by the original operation.
1929 The latter case is filtered out in the condition below.
1930 It still doesn't cover the case when register is defined and used
1931 somewhere within the code motion path, and in this case we could
1932 miss a unifying code motion along both branches using a renamed
1933 register, but it won't affect a code correctness since upon
1934 an actual code motion a bookkeeping code would be generated. */
1935 if (register_unavailable_p (VINSN_REG_USES (EXPR_VINSN (expr)),
1936 EXPR_LHS (expr)))
1937 EXPR_TARGET_AVAILABLE (expr) = -1;
1938 else
1939 EXPR_TARGET_AVAILABLE (expr) = false;
1940 }
1941 }
1942 else
1943 {
1944 unsigned regno;
1945 reg_set_iterator rsi;
1946
1947 EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_SETS (EXPR_VINSN (expr)),
1948 0, regno, rsi)
1949 if (bitmap_bit_p (lv_set, regno))
1950 {
1951 EXPR_TARGET_AVAILABLE (expr) = false;
1952 break;
1953 }
1954
1955 EXECUTE_IF_SET_IN_REG_SET (VINSN_REG_CLOBBERS (EXPR_VINSN (expr)),
1956 0, regno, rsi)
1957 if (bitmap_bit_p (lv_set, regno))
1958 {
1959 EXPR_TARGET_AVAILABLE (expr) = false;
1960 break;
1961 }
1962 }
1963 }
1964
1965 /* Try to make EXPR speculative. Return 1 when EXPR's pattern
1966 or dependence status have changed, 2 when also the target register
1967 became unavailable, 0 if nothing had to be changed. */
1968 int
1969 speculate_expr (expr_t expr, ds_t ds)
1970 {
1971 int res;
1972 rtx_insn *orig_insn_rtx;
1973 rtx spec_pat;
1974 ds_t target_ds, current_ds;
1975
1976 /* Obtain the status we need to put on EXPR. */
1977 target_ds = (ds & SPECULATIVE);
1978 current_ds = EXPR_SPEC_DONE_DS (expr);
1979 ds = ds_full_merge (current_ds, target_ds, NULL_RTX, NULL_RTX);
1980
1981 orig_insn_rtx = EXPR_INSN_RTX (expr);
1982
1983 res = sched_speculate_insn (orig_insn_rtx, ds, &spec_pat);
1984
1985 switch (res)
1986 {
1987 case 0:
1988 EXPR_SPEC_DONE_DS (expr) = ds;
1989 return current_ds != ds ? 1 : 0;
1990
1991 case 1:
1992 {
1993 rtx_insn *spec_insn_rtx =
1994 create_insn_rtx_from_pattern (spec_pat, NULL_RTX);
1995 vinsn_t spec_vinsn = create_vinsn_from_insn_rtx (spec_insn_rtx, false);
1996
1997 change_vinsn_in_expr (expr, spec_vinsn);
1998 EXPR_SPEC_DONE_DS (expr) = ds;
1999 EXPR_NEEDS_SPEC_CHECK_P (expr) = true;
2000
2001 /* Do not allow clobbering the address register of speculative
2002 insns. */
2003 if (register_unavailable_p (VINSN_REG_USES (EXPR_VINSN (expr)),
2004 expr_dest_reg (expr)))
2005 {
2006 EXPR_TARGET_AVAILABLE (expr) = false;
2007 return 2;
2008 }
2009
2010 return 1;
2011 }
2012
2013 case -1:
2014 return -1;
2015
2016 default:
2017 gcc_unreachable ();
2018 return -1;
2019 }
2020 }
2021
2022 /* Return a destination register, if any, of EXPR. */
2023 rtx
2024 expr_dest_reg (expr_t expr)
2025 {
2026 rtx dest = VINSN_LHS (EXPR_VINSN (expr));
2027
2028 if (dest != NULL_RTX && REG_P (dest))
2029 return dest;
2030
2031 return NULL_RTX;
2032 }
2033
2034 /* Returns the REGNO of the R's destination. */
2035 unsigned
2036 expr_dest_regno (expr_t expr)
2037 {
2038 rtx dest = expr_dest_reg (expr);
2039
2040 gcc_assert (dest != NULL_RTX);
2041 return REGNO (dest);
2042 }
2043
2044 /* For a given LV_SET, mark all expressions in JOIN_SET, but not present in
2045 AV_SET having unavailable target register. */
2046 void
2047 mark_unavailable_targets (av_set_t join_set, av_set_t av_set, regset lv_set)
2048 {
2049 expr_t expr;
2050 av_set_iterator avi;
2051
2052 FOR_EACH_EXPR (expr, avi, join_set)
2053 if (av_set_lookup (av_set, EXPR_VINSN (expr)) == NULL)
2054 set_unavailable_target_for_expr (expr, lv_set);
2055 }
2056 \f
2057
2058 /* Returns true if REG (at least partially) is present in REGS. */
2059 bool
2060 register_unavailable_p (regset regs, rtx reg)
2061 {
2062 unsigned regno, end_regno;
2063
2064 regno = REGNO (reg);
2065 if (bitmap_bit_p (regs, regno))
2066 return true;
2067
2068 end_regno = END_REGNO (reg);
2069
2070 while (++regno < end_regno)
2071 if (bitmap_bit_p (regs, regno))
2072 return true;
2073
2074 return false;
2075 }
2076
2077 /* Av set functions. */
2078
2079 /* Add a new element to av set SETP.
2080 Return the element added. */
2081 static av_set_t
2082 av_set_add_element (av_set_t *setp)
2083 {
2084 /* Insert at the beginning of the list. */
2085 _list_add (setp);
2086 return *setp;
2087 }
2088
2089 /* Add EXPR to SETP. */
2090 void
2091 av_set_add (av_set_t *setp, expr_t expr)
2092 {
2093 av_set_t elem;
2094
2095 gcc_assert (!INSN_NOP_P (EXPR_INSN_RTX (expr)));
2096 elem = av_set_add_element (setp);
2097 copy_expr (_AV_SET_EXPR (elem), expr);
2098 }
2099
2100 /* Same, but do not copy EXPR. */
2101 static void
2102 av_set_add_nocopy (av_set_t *setp, expr_t expr)
2103 {
2104 av_set_t elem;
2105
2106 elem = av_set_add_element (setp);
2107 *_AV_SET_EXPR (elem) = *expr;
2108 }
2109
2110 /* Remove expr pointed to by IP from the av_set. */
2111 void
2112 av_set_iter_remove (av_set_iterator *ip)
2113 {
2114 clear_expr (_AV_SET_EXPR (*ip->lp));
2115 _list_iter_remove (ip);
2116 }
2117
2118 /* Search for an expr in SET, such that it's equivalent to SOUGHT_VINSN in the
2119 sense of vinsn_equal_p function. Return NULL if no such expr is
2120 in SET was found. */
2121 expr_t
2122 av_set_lookup (av_set_t set, vinsn_t sought_vinsn)
2123 {
2124 expr_t expr;
2125 av_set_iterator i;
2126
2127 FOR_EACH_EXPR (expr, i, set)
2128 if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn))
2129 return expr;
2130 return NULL;
2131 }
2132
2133 /* Same, but also remove the EXPR found. */
2134 static expr_t
2135 av_set_lookup_and_remove (av_set_t *setp, vinsn_t sought_vinsn)
2136 {
2137 expr_t expr;
2138 av_set_iterator i;
2139
2140 FOR_EACH_EXPR_1 (expr, i, setp)
2141 if (vinsn_equal_p (EXPR_VINSN (expr), sought_vinsn))
2142 {
2143 _list_iter_remove_nofree (&i);
2144 return expr;
2145 }
2146 return NULL;
2147 }
2148
2149 /* Search for an expr in SET, such that it's equivalent to EXPR in the
2150 sense of vinsn_equal_p function of their vinsns, but not EXPR itself.
2151 Returns NULL if no such expr is in SET was found. */
2152 static expr_t
2153 av_set_lookup_other_equiv_expr (av_set_t set, expr_t expr)
2154 {
2155 expr_t cur_expr;
2156 av_set_iterator i;
2157
2158 FOR_EACH_EXPR (cur_expr, i, set)
2159 {
2160 if (cur_expr == expr)
2161 continue;
2162 if (vinsn_equal_p (EXPR_VINSN (cur_expr), EXPR_VINSN (expr)))
2163 return cur_expr;
2164 }
2165
2166 return NULL;
2167 }
2168
2169 /* If other expression is already in AVP, remove one of them. */
2170 expr_t
2171 merge_with_other_exprs (av_set_t *avp, av_set_iterator *ip, expr_t expr)
2172 {
2173 expr_t expr2;
2174
2175 expr2 = av_set_lookup_other_equiv_expr (*avp, expr);
2176 if (expr2 != NULL)
2177 {
2178 /* Reset target availability on merge, since taking it only from one
2179 of the exprs would be controversial for different code. */
2180 EXPR_TARGET_AVAILABLE (expr2) = -1;
2181 EXPR_USEFULNESS (expr2) = 0;
2182
2183 merge_expr (expr2, expr, NULL);
2184
2185 /* Fix usefulness as it should be now REG_BR_PROB_BASE. */
2186 EXPR_USEFULNESS (expr2) = REG_BR_PROB_BASE;
2187
2188 av_set_iter_remove (ip);
2189 return expr2;
2190 }
2191
2192 return expr;
2193 }
2194
2195 /* Return true if there is an expr that correlates to VI in SET. */
2196 bool
2197 av_set_is_in_p (av_set_t set, vinsn_t vi)
2198 {
2199 return av_set_lookup (set, vi) != NULL;
2200 }
2201
2202 /* Return a copy of SET. */
2203 av_set_t
2204 av_set_copy (av_set_t set)
2205 {
2206 expr_t expr;
2207 av_set_iterator i;
2208 av_set_t res = NULL;
2209
2210 FOR_EACH_EXPR (expr, i, set)
2211 av_set_add (&res, expr);
2212
2213 return res;
2214 }
2215
2216 /* Join two av sets that do not have common elements by attaching second set
2217 (pointed to by FROMP) to the end of first set (TO_TAILP must point to
2218 _AV_SET_NEXT of first set's last element). */
2219 static void
2220 join_distinct_sets (av_set_t *to_tailp, av_set_t *fromp)
2221 {
2222 gcc_assert (*to_tailp == NULL);
2223 *to_tailp = *fromp;
2224 *fromp = NULL;
2225 }
2226
2227 /* Makes set pointed to by TO to be the union of TO and FROM. Clear av_set
2228 pointed to by FROMP afterwards. */
2229 void
2230 av_set_union_and_clear (av_set_t *top, av_set_t *fromp, insn_t insn)
2231 {
2232 expr_t expr1;
2233 av_set_iterator i;
2234
2235 /* Delete from TOP all exprs, that present in FROMP. */
2236 FOR_EACH_EXPR_1 (expr1, i, top)
2237 {
2238 expr_t expr2 = av_set_lookup (*fromp, EXPR_VINSN (expr1));
2239
2240 if (expr2)
2241 {
2242 merge_expr (expr2, expr1, insn);
2243 av_set_iter_remove (&i);
2244 }
2245 }
2246
2247 join_distinct_sets (i.lp, fromp);
2248 }
2249
2250 /* Same as above, but also update availability of target register in
2251 TOP judging by TO_LV_SET and FROM_LV_SET. */
2252 void
2253 av_set_union_and_live (av_set_t *top, av_set_t *fromp, regset to_lv_set,
2254 regset from_lv_set, insn_t insn)
2255 {
2256 expr_t expr1;
2257 av_set_iterator i;
2258 av_set_t *to_tailp, in_both_set = NULL;
2259
2260 /* Delete from TOP all expres, that present in FROMP. */
2261 FOR_EACH_EXPR_1 (expr1, i, top)
2262 {
2263 expr_t expr2 = av_set_lookup_and_remove (fromp, EXPR_VINSN (expr1));
2264
2265 if (expr2)
2266 {
2267 /* It may be that the expressions have different destination
2268 registers, in which case we need to check liveness here. */
2269 if (EXPR_SEPARABLE_P (expr1))
2270 {
2271 int regno1 = (REG_P (EXPR_LHS (expr1))
2272 ? (int) expr_dest_regno (expr1) : -1);
2273 int regno2 = (REG_P (EXPR_LHS (expr2))
2274 ? (int) expr_dest_regno (expr2) : -1);
2275
2276 /* ??? We don't have a way to check restrictions for
2277 *other* register on the current path, we did it only
2278 for the current target register. Give up. */
2279 if (regno1 != regno2)
2280 EXPR_TARGET_AVAILABLE (expr2) = -1;
2281 }
2282 else if (EXPR_INSN_RTX (expr1) != EXPR_INSN_RTX (expr2))
2283 EXPR_TARGET_AVAILABLE (expr2) = -1;
2284
2285 merge_expr (expr2, expr1, insn);
2286 av_set_add_nocopy (&in_both_set, expr2);
2287 av_set_iter_remove (&i);
2288 }
2289 else
2290 /* EXPR1 is present in TOP, but not in FROMP. Check it on
2291 FROM_LV_SET. */
2292 set_unavailable_target_for_expr (expr1, from_lv_set);
2293 }
2294 to_tailp = i.lp;
2295
2296 /* These expressions are not present in TOP. Check liveness
2297 restrictions on TO_LV_SET. */
2298 FOR_EACH_EXPR (expr1, i, *fromp)
2299 set_unavailable_target_for_expr (expr1, to_lv_set);
2300
2301 join_distinct_sets (i.lp, &in_both_set);
2302 join_distinct_sets (to_tailp, fromp);
2303 }
2304
2305 /* Clear av_set pointed to by SETP. */
2306 void
2307 av_set_clear (av_set_t *setp)
2308 {
2309 expr_t expr;
2310 av_set_iterator i;
2311
2312 FOR_EACH_EXPR_1 (expr, i, setp)
2313 av_set_iter_remove (&i);
2314
2315 gcc_assert (*setp == NULL);
2316 }
2317
2318 /* Leave only one non-speculative element in the SETP. */
2319 void
2320 av_set_leave_one_nonspec (av_set_t *setp)
2321 {
2322 expr_t expr;
2323 av_set_iterator i;
2324 bool has_one_nonspec = false;
2325
2326 /* Keep all speculative exprs, and leave one non-speculative
2327 (the first one). */
2328 FOR_EACH_EXPR_1 (expr, i, setp)
2329 {
2330 if (!EXPR_SPEC_DONE_DS (expr))
2331 {
2332 if (has_one_nonspec)
2333 av_set_iter_remove (&i);
2334 else
2335 has_one_nonspec = true;
2336 }
2337 }
2338 }
2339
2340 /* Return the N'th element of the SET. */
2341 expr_t
2342 av_set_element (av_set_t set, int n)
2343 {
2344 expr_t expr;
2345 av_set_iterator i;
2346
2347 FOR_EACH_EXPR (expr, i, set)
2348 if (n-- == 0)
2349 return expr;
2350
2351 gcc_unreachable ();
2352 return NULL;
2353 }
2354
2355 /* Deletes all expressions from AVP that are conditional branches (IFs). */
2356 void
2357 av_set_substract_cond_branches (av_set_t *avp)
2358 {
2359 av_set_iterator i;
2360 expr_t expr;
2361
2362 FOR_EACH_EXPR_1 (expr, i, avp)
2363 if (vinsn_cond_branch_p (EXPR_VINSN (expr)))
2364 av_set_iter_remove (&i);
2365 }
2366
2367 /* Multiplies usefulness attribute of each member of av-set *AVP by
2368 value PROB / ALL_PROB. */
2369 void
2370 av_set_split_usefulness (av_set_t av, int prob, int all_prob)
2371 {
2372 av_set_iterator i;
2373 expr_t expr;
2374
2375 FOR_EACH_EXPR (expr, i, av)
2376 EXPR_USEFULNESS (expr) = (all_prob
2377 ? (EXPR_USEFULNESS (expr) * prob) / all_prob
2378 : 0);
2379 }
2380
2381 /* Leave in AVP only those expressions, which are present in AV,
2382 and return it, merging history expressions. */
2383 void
2384 av_set_code_motion_filter (av_set_t *avp, av_set_t av)
2385 {
2386 av_set_iterator i;
2387 expr_t expr, expr2;
2388
2389 FOR_EACH_EXPR_1 (expr, i, avp)
2390 if ((expr2 = av_set_lookup (av, EXPR_VINSN (expr))) == NULL)
2391 av_set_iter_remove (&i);
2392 else
2393 /* When updating av sets in bookkeeping blocks, we can add more insns
2394 there which will be transformed but the upper av sets will not
2395 reflect those transformations. We then fail to undo those
2396 when searching for such insns. So merge the history saved
2397 in the av set of the block we are processing. */
2398 merge_history_vect (&EXPR_HISTORY_OF_CHANGES (expr),
2399 EXPR_HISTORY_OF_CHANGES (expr2));
2400 }
2401
2402 \f
2403
2404 /* Dependence hooks to initialize insn data. */
2405
2406 /* This is used in hooks callable from dependence analysis when initializing
2407 instruction's data. */
2408 static struct
2409 {
2410 /* Where the dependence was found (lhs/rhs). */
2411 deps_where_t where;
2412
2413 /* The actual data object to initialize. */
2414 idata_t id;
2415
2416 /* True when the insn should not be made clonable. */
2417 bool force_unique_p;
2418
2419 /* True when insn should be treated as of type USE, i.e. never renamed. */
2420 bool force_use_p;
2421 } deps_init_id_data;
2422
2423
2424 /* Setup ID for INSN. FORCE_UNIQUE_P is true when INSN should not be
2425 clonable. */
2426 static void
2427 setup_id_for_insn (idata_t id, insn_t insn, bool force_unique_p)
2428 {
2429 int type;
2430
2431 /* Determine whether INSN could be cloned and return appropriate vinsn type.
2432 That clonable insns which can be separated into lhs and rhs have type SET.
2433 Other clonable insns have type USE. */
2434 type = GET_CODE (insn);
2435
2436 /* Only regular insns could be cloned. */
2437 if (type == INSN && !force_unique_p)
2438 type = SET;
2439 else if (type == JUMP_INSN && simplejump_p (insn))
2440 type = PC;
2441 else if (type == DEBUG_INSN)
2442 type = !force_unique_p ? USE : INSN;
2443
2444 IDATA_TYPE (id) = type;
2445 IDATA_REG_SETS (id) = get_clear_regset_from_pool ();
2446 IDATA_REG_USES (id) = get_clear_regset_from_pool ();
2447 IDATA_REG_CLOBBERS (id) = get_clear_regset_from_pool ();
2448 }
2449
2450 /* Start initializing insn data. */
2451 static void
2452 deps_init_id_start_insn (insn_t insn)
2453 {
2454 gcc_assert (deps_init_id_data.where == DEPS_IN_NOWHERE);
2455
2456 setup_id_for_insn (deps_init_id_data.id, insn,
2457 deps_init_id_data.force_unique_p);
2458 deps_init_id_data.where = DEPS_IN_INSN;
2459 }
2460
2461 /* Start initializing lhs data. */
2462 static void
2463 deps_init_id_start_lhs (rtx lhs)
2464 {
2465 gcc_assert (deps_init_id_data.where == DEPS_IN_INSN);
2466 gcc_assert (IDATA_LHS (deps_init_id_data.id) == NULL);
2467
2468 if (IDATA_TYPE (deps_init_id_data.id) == SET)
2469 {
2470 IDATA_LHS (deps_init_id_data.id) = lhs;
2471 deps_init_id_data.where = DEPS_IN_LHS;
2472 }
2473 }
2474
2475 /* Finish initializing lhs data. */
2476 static void
2477 deps_init_id_finish_lhs (void)
2478 {
2479 deps_init_id_data.where = DEPS_IN_INSN;
2480 }
2481
2482 /* Note a set of REGNO. */
2483 static void
2484 deps_init_id_note_reg_set (int regno)
2485 {
2486 haifa_note_reg_set (regno);
2487
2488 if (deps_init_id_data.where == DEPS_IN_RHS)
2489 deps_init_id_data.force_use_p = true;
2490
2491 if (IDATA_TYPE (deps_init_id_data.id) != PC)
2492 SET_REGNO_REG_SET (IDATA_REG_SETS (deps_init_id_data.id), regno);
2493
2494 #ifdef STACK_REGS
2495 /* Make instructions that set stack registers to be ineligible for
2496 renaming to avoid issues with find_used_regs. */
2497 if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG))
2498 deps_init_id_data.force_use_p = true;
2499 #endif
2500 }
2501
2502 /* Note a clobber of REGNO. */
2503 static void
2504 deps_init_id_note_reg_clobber (int regno)
2505 {
2506 haifa_note_reg_clobber (regno);
2507
2508 if (deps_init_id_data.where == DEPS_IN_RHS)
2509 deps_init_id_data.force_use_p = true;
2510
2511 if (IDATA_TYPE (deps_init_id_data.id) != PC)
2512 SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (deps_init_id_data.id), regno);
2513 }
2514
2515 /* Note a use of REGNO. */
2516 static void
2517 deps_init_id_note_reg_use (int regno)
2518 {
2519 haifa_note_reg_use (regno);
2520
2521 if (IDATA_TYPE (deps_init_id_data.id) != PC)
2522 SET_REGNO_REG_SET (IDATA_REG_USES (deps_init_id_data.id), regno);
2523 }
2524
2525 /* Start initializing rhs data. */
2526 static void
2527 deps_init_id_start_rhs (rtx rhs)
2528 {
2529 gcc_assert (deps_init_id_data.where == DEPS_IN_INSN);
2530
2531 /* And there was no sel_deps_reset_to_insn (). */
2532 if (IDATA_LHS (deps_init_id_data.id) != NULL)
2533 {
2534 IDATA_RHS (deps_init_id_data.id) = rhs;
2535 deps_init_id_data.where = DEPS_IN_RHS;
2536 }
2537 }
2538
2539 /* Finish initializing rhs data. */
2540 static void
2541 deps_init_id_finish_rhs (void)
2542 {
2543 gcc_assert (deps_init_id_data.where == DEPS_IN_RHS
2544 || deps_init_id_data.where == DEPS_IN_INSN);
2545 deps_init_id_data.where = DEPS_IN_INSN;
2546 }
2547
2548 /* Finish initializing insn data. */
2549 static void
2550 deps_init_id_finish_insn (void)
2551 {
2552 gcc_assert (deps_init_id_data.where == DEPS_IN_INSN);
2553
2554 if (IDATA_TYPE (deps_init_id_data.id) == SET)
2555 {
2556 rtx lhs = IDATA_LHS (deps_init_id_data.id);
2557 rtx rhs = IDATA_RHS (deps_init_id_data.id);
2558
2559 if (lhs == NULL || rhs == NULL || !lhs_and_rhs_separable_p (lhs, rhs)
2560 || deps_init_id_data.force_use_p)
2561 {
2562 /* This should be a USE, as we don't want to schedule its RHS
2563 separately. However, we still want to have them recorded
2564 for the purposes of substitution. That's why we don't
2565 simply call downgrade_to_use () here. */
2566 gcc_assert (IDATA_TYPE (deps_init_id_data.id) == SET);
2567 gcc_assert (!lhs == !rhs);
2568
2569 IDATA_TYPE (deps_init_id_data.id) = USE;
2570 }
2571 }
2572
2573 deps_init_id_data.where = DEPS_IN_NOWHERE;
2574 }
2575
2576 /* This is dependence info used for initializing insn's data. */
2577 static struct sched_deps_info_def deps_init_id_sched_deps_info;
2578
2579 /* This initializes most of the static part of the above structure. */
2580 static const struct sched_deps_info_def const_deps_init_id_sched_deps_info =
2581 {
2582 NULL,
2583
2584 deps_init_id_start_insn,
2585 deps_init_id_finish_insn,
2586 deps_init_id_start_lhs,
2587 deps_init_id_finish_lhs,
2588 deps_init_id_start_rhs,
2589 deps_init_id_finish_rhs,
2590 deps_init_id_note_reg_set,
2591 deps_init_id_note_reg_clobber,
2592 deps_init_id_note_reg_use,
2593 NULL, /* note_mem_dep */
2594 NULL, /* note_dep */
2595
2596 0, /* use_cselib */
2597 0, /* use_deps_list */
2598 0 /* generate_spec_deps */
2599 };
2600
2601 /* Initialize INSN's lhs and rhs in ID. When FORCE_UNIQUE_P is true,
2602 we don't actually need information about lhs and rhs. */
2603 static void
2604 setup_id_lhs_rhs (idata_t id, insn_t insn, bool force_unique_p)
2605 {
2606 rtx pat = PATTERN (insn);
2607
2608 if (NONJUMP_INSN_P (insn)
2609 && GET_CODE (pat) == SET
2610 && !force_unique_p)
2611 {
2612 IDATA_RHS (id) = SET_SRC (pat);
2613 IDATA_LHS (id) = SET_DEST (pat);
2614 }
2615 else
2616 IDATA_LHS (id) = IDATA_RHS (id) = NULL;
2617 }
2618
2619 /* Possibly downgrade INSN to USE. */
2620 static void
2621 maybe_downgrade_id_to_use (idata_t id, insn_t insn)
2622 {
2623 bool must_be_use = false;
2624 df_ref def;
2625 rtx lhs = IDATA_LHS (id);
2626 rtx rhs = IDATA_RHS (id);
2627
2628 /* We downgrade only SETs. */
2629 if (IDATA_TYPE (id) != SET)
2630 return;
2631
2632 if (!lhs || !lhs_and_rhs_separable_p (lhs, rhs))
2633 {
2634 IDATA_TYPE (id) = USE;
2635 return;
2636 }
2637
2638 FOR_EACH_INSN_DEF (def, insn)
2639 {
2640 if (DF_REF_INSN (def)
2641 && DF_REF_FLAGS_IS_SET (def, DF_REF_PRE_POST_MODIFY)
2642 && loc_mentioned_in_p (DF_REF_LOC (def), IDATA_RHS (id)))
2643 {
2644 must_be_use = true;
2645 break;
2646 }
2647
2648 #ifdef STACK_REGS
2649 /* Make instructions that set stack registers to be ineligible for
2650 renaming to avoid issues with find_used_regs. */
2651 if (IN_RANGE (DF_REF_REGNO (def), FIRST_STACK_REG, LAST_STACK_REG))
2652 {
2653 must_be_use = true;
2654 break;
2655 }
2656 #endif
2657 }
2658
2659 if (must_be_use)
2660 IDATA_TYPE (id) = USE;
2661 }
2662
2663 /* Setup register sets describing INSN in ID. */
2664 static void
2665 setup_id_reg_sets (idata_t id, insn_t insn)
2666 {
2667 struct df_insn_info *insn_info = DF_INSN_INFO_GET (insn);
2668 df_ref def, use;
2669 regset tmp = get_clear_regset_from_pool ();
2670
2671 FOR_EACH_INSN_INFO_DEF (def, insn_info)
2672 {
2673 unsigned int regno = DF_REF_REGNO (def);
2674
2675 /* Post modifies are treated like clobbers by sched-deps.c. */
2676 if (DF_REF_FLAGS_IS_SET (def, (DF_REF_MUST_CLOBBER
2677 | DF_REF_PRE_POST_MODIFY)))
2678 SET_REGNO_REG_SET (IDATA_REG_CLOBBERS (id), regno);
2679 else if (! DF_REF_FLAGS_IS_SET (def, DF_REF_MAY_CLOBBER))
2680 {
2681 SET_REGNO_REG_SET (IDATA_REG_SETS (id), regno);
2682
2683 #ifdef STACK_REGS
2684 /* For stack registers, treat writes to them as writes
2685 to the first one to be consistent with sched-deps.c. */
2686 if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG))
2687 SET_REGNO_REG_SET (IDATA_REG_SETS (id), FIRST_STACK_REG);
2688 #endif
2689 }
2690 /* Mark special refs that generate read/write def pair. */
2691 if (DF_REF_FLAGS_IS_SET (def, DF_REF_CONDITIONAL)
2692 || regno == STACK_POINTER_REGNUM)
2693 bitmap_set_bit (tmp, regno);
2694 }
2695
2696 FOR_EACH_INSN_INFO_USE (use, insn_info)
2697 {
2698 unsigned int regno = DF_REF_REGNO (use);
2699
2700 /* When these refs are met for the first time, skip them, as
2701 these uses are just counterparts of some defs. */
2702 if (bitmap_bit_p (tmp, regno))
2703 bitmap_clear_bit (tmp, regno);
2704 else if (! DF_REF_FLAGS_IS_SET (use, DF_REF_CALL_STACK_USAGE))
2705 {
2706 SET_REGNO_REG_SET (IDATA_REG_USES (id), regno);
2707
2708 #ifdef STACK_REGS
2709 /* For stack registers, treat reads from them as reads from
2710 the first one to be consistent with sched-deps.c. */
2711 if (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG))
2712 SET_REGNO_REG_SET (IDATA_REG_USES (id), FIRST_STACK_REG);
2713 #endif
2714 }
2715 }
2716
2717 return_regset_to_pool (tmp);
2718 }
2719
2720 /* Initialize instruction data for INSN in ID using DF's data. */
2721 static void
2722 init_id_from_df (idata_t id, insn_t insn, bool force_unique_p)
2723 {
2724 gcc_assert (DF_INSN_UID_SAFE_GET (INSN_UID (insn)) != NULL);
2725
2726 setup_id_for_insn (id, insn, force_unique_p);
2727 setup_id_lhs_rhs (id, insn, force_unique_p);
2728
2729 if (INSN_NOP_P (insn))
2730 return;
2731
2732 maybe_downgrade_id_to_use (id, insn);
2733 setup_id_reg_sets (id, insn);
2734 }
2735
2736 /* Initialize instruction data for INSN in ID. */
2737 static void
2738 deps_init_id (idata_t id, insn_t insn, bool force_unique_p)
2739 {
2740 struct deps_desc _dc, *dc = &_dc;
2741
2742 deps_init_id_data.where = DEPS_IN_NOWHERE;
2743 deps_init_id_data.id = id;
2744 deps_init_id_data.force_unique_p = force_unique_p;
2745 deps_init_id_data.force_use_p = false;
2746
2747 init_deps (dc, false);
2748
2749 memcpy (&deps_init_id_sched_deps_info,
2750 &const_deps_init_id_sched_deps_info,
2751 sizeof (deps_init_id_sched_deps_info));
2752
2753 if (spec_info != NULL)
2754 deps_init_id_sched_deps_info.generate_spec_deps = 1;
2755
2756 sched_deps_info = &deps_init_id_sched_deps_info;
2757
2758 deps_analyze_insn (dc, insn);
2759
2760 free_deps (dc);
2761
2762 deps_init_id_data.id = NULL;
2763 }
2764
2765 \f
2766 struct sched_scan_info_def
2767 {
2768 /* This hook notifies scheduler frontend to extend its internal per basic
2769 block data structures. This hook should be called once before a series of
2770 calls to bb_init (). */
2771 void (*extend_bb) (void);
2772
2773 /* This hook makes scheduler frontend to initialize its internal data
2774 structures for the passed basic block. */
2775 void (*init_bb) (basic_block);
2776
2777 /* This hook notifies scheduler frontend to extend its internal per insn data
2778 structures. This hook should be called once before a series of calls to
2779 insn_init (). */
2780 void (*extend_insn) (void);
2781
2782 /* This hook makes scheduler frontend to initialize its internal data
2783 structures for the passed insn. */
2784 void (*init_insn) (insn_t);
2785 };
2786
2787 /* A driver function to add a set of basic blocks (BBS) to the
2788 scheduling region. */
2789 static void
2790 sched_scan (const struct sched_scan_info_def *ssi, bb_vec_t bbs)
2791 {
2792 unsigned i;
2793 basic_block bb;
2794
2795 if (ssi->extend_bb)
2796 ssi->extend_bb ();
2797
2798 if (ssi->init_bb)
2799 FOR_EACH_VEC_ELT (bbs, i, bb)
2800 ssi->init_bb (bb);
2801
2802 if (ssi->extend_insn)
2803 ssi->extend_insn ();
2804
2805 if (ssi->init_insn)
2806 FOR_EACH_VEC_ELT (bbs, i, bb)
2807 {
2808 rtx_insn *insn;
2809
2810 FOR_BB_INSNS (bb, insn)
2811 ssi->init_insn (insn);
2812 }
2813 }
2814
2815 /* Implement hooks for collecting fundamental insn properties like if insn is
2816 an ASM or is within a SCHED_GROUP. */
2817
2818 /* True when a "one-time init" data for INSN was already inited. */
2819 static bool
2820 first_time_insn_init (insn_t insn)
2821 {
2822 return INSN_LIVE (insn) == NULL;
2823 }
2824
2825 /* Hash an entry in a transformed_insns hashtable. */
2826 static hashval_t
2827 hash_transformed_insns (const void *p)
2828 {
2829 return VINSN_HASH_RTX (((const struct transformed_insns *) p)->vinsn_old);
2830 }
2831
2832 /* Compare the entries in a transformed_insns hashtable. */
2833 static int
2834 eq_transformed_insns (const void *p, const void *q)
2835 {
2836 rtx_insn *i1 =
2837 VINSN_INSN_RTX (((const struct transformed_insns *) p)->vinsn_old);
2838 rtx_insn *i2 =
2839 VINSN_INSN_RTX (((const struct transformed_insns *) q)->vinsn_old);
2840
2841 if (INSN_UID (i1) == INSN_UID (i2))
2842 return 1;
2843 return rtx_equal_p (PATTERN (i1), PATTERN (i2));
2844 }
2845
2846 /* Free an entry in a transformed_insns hashtable. */
2847 static void
2848 free_transformed_insns (void *p)
2849 {
2850 struct transformed_insns *pti = (struct transformed_insns *) p;
2851
2852 vinsn_detach (pti->vinsn_old);
2853 vinsn_detach (pti->vinsn_new);
2854 free (pti);
2855 }
2856
2857 /* Init the s_i_d data for INSN which should be inited just once, when
2858 we first see the insn. */
2859 static void
2860 init_first_time_insn_data (insn_t insn)
2861 {
2862 /* This should not be set if this is the first time we init data for
2863 insn. */
2864 gcc_assert (first_time_insn_init (insn));
2865
2866 /* These are needed for nops too. */
2867 INSN_LIVE (insn) = get_regset_from_pool ();
2868 INSN_LIVE_VALID_P (insn) = false;
2869
2870 if (!INSN_NOP_P (insn))
2871 {
2872 INSN_ANALYZED_DEPS (insn) = BITMAP_ALLOC (NULL);
2873 INSN_FOUND_DEPS (insn) = BITMAP_ALLOC (NULL);
2874 INSN_TRANSFORMED_INSNS (insn)
2875 = htab_create (16, hash_transformed_insns,
2876 eq_transformed_insns, free_transformed_insns);
2877 init_deps (&INSN_DEPS_CONTEXT (insn), true);
2878 }
2879 }
2880
2881 /* Free almost all above data for INSN that is scheduled already.
2882 Used for extra-large basic blocks. */
2883 void
2884 free_data_for_scheduled_insn (insn_t insn)
2885 {
2886 gcc_assert (! first_time_insn_init (insn));
2887
2888 if (! INSN_ANALYZED_DEPS (insn))
2889 return;
2890
2891 BITMAP_FREE (INSN_ANALYZED_DEPS (insn));
2892 BITMAP_FREE (INSN_FOUND_DEPS (insn));
2893 htab_delete (INSN_TRANSFORMED_INSNS (insn));
2894
2895 /* This is allocated only for bookkeeping insns. */
2896 if (INSN_ORIGINATORS (insn))
2897 BITMAP_FREE (INSN_ORIGINATORS (insn));
2898 free_deps (&INSN_DEPS_CONTEXT (insn));
2899
2900 INSN_ANALYZED_DEPS (insn) = NULL;
2901
2902 /* Clear the readonly flag so we would ICE when trying to recalculate
2903 the deps context (as we believe that it should not happen). */
2904 (&INSN_DEPS_CONTEXT (insn))->readonly = 0;
2905 }
2906
2907 /* Free the same data as above for INSN. */
2908 static void
2909 free_first_time_insn_data (insn_t insn)
2910 {
2911 gcc_assert (! first_time_insn_init (insn));
2912
2913 free_data_for_scheduled_insn (insn);
2914 return_regset_to_pool (INSN_LIVE (insn));
2915 INSN_LIVE (insn) = NULL;
2916 INSN_LIVE_VALID_P (insn) = false;
2917 }
2918
2919 /* Initialize region-scope data structures for basic blocks. */
2920 static void
2921 init_global_and_expr_for_bb (basic_block bb)
2922 {
2923 if (sel_bb_empty_p (bb))
2924 return;
2925
2926 invalidate_av_set (bb);
2927 }
2928
2929 /* Data for global dependency analysis (to initialize CANT_MOVE and
2930 SCHED_GROUP_P). */
2931 static struct
2932 {
2933 /* Previous insn. */
2934 insn_t prev_insn;
2935 } init_global_data;
2936
2937 /* Determine if INSN is in the sched_group, is an asm or should not be
2938 cloned. After that initialize its expr. */
2939 static void
2940 init_global_and_expr_for_insn (insn_t insn)
2941 {
2942 if (LABEL_P (insn))
2943 return;
2944
2945 if (NOTE_INSN_BASIC_BLOCK_P (insn))
2946 {
2947 init_global_data.prev_insn = NULL;
2948 return;
2949 }
2950
2951 gcc_assert (INSN_P (insn));
2952
2953 if (SCHED_GROUP_P (insn))
2954 /* Setup a sched_group. */
2955 {
2956 insn_t prev_insn = init_global_data.prev_insn;
2957
2958 if (prev_insn)
2959 INSN_SCHED_NEXT (prev_insn) = insn;
2960
2961 init_global_data.prev_insn = insn;
2962 }
2963 else
2964 init_global_data.prev_insn = NULL;
2965
2966 if (GET_CODE (PATTERN (insn)) == ASM_INPUT
2967 || asm_noperands (PATTERN (insn)) >= 0)
2968 /* Mark INSN as an asm. */
2969 INSN_ASM_P (insn) = true;
2970
2971 {
2972 bool force_unique_p;
2973 ds_t spec_done_ds;
2974
2975 /* Certain instructions cannot be cloned, and frame related insns and
2976 the insn adjacent to NOTE_INSN_EPILOGUE_BEG cannot be moved out of
2977 their block. */
2978 if (prologue_epilogue_contains (insn))
2979 {
2980 if (RTX_FRAME_RELATED_P (insn))
2981 CANT_MOVE (insn) = 1;
2982 else
2983 {
2984 rtx note;
2985 for (note = REG_NOTES (insn); note; note = XEXP (note, 1))
2986 if (REG_NOTE_KIND (note) == REG_SAVE_NOTE
2987 && ((enum insn_note) INTVAL (XEXP (note, 0))
2988 == NOTE_INSN_EPILOGUE_BEG))
2989 {
2990 CANT_MOVE (insn) = 1;
2991 break;
2992 }
2993 }
2994 force_unique_p = true;
2995 }
2996 else
2997 if (CANT_MOVE (insn)
2998 || INSN_ASM_P (insn)
2999 || SCHED_GROUP_P (insn)
3000 || CALL_P (insn)
3001 /* Exception handling insns are always unique. */
3002 || (cfun->can_throw_non_call_exceptions && can_throw_internal (insn))
3003 /* TRAP_IF though have an INSN code is control_flow_insn_p (). */
3004 || control_flow_insn_p (insn)
3005 || volatile_insn_p (PATTERN (insn))
3006 || (targetm.cannot_copy_insn_p
3007 && targetm.cannot_copy_insn_p (insn)))
3008 force_unique_p = true;
3009 else
3010 force_unique_p = false;
3011
3012 if (targetm.sched.get_insn_spec_ds)
3013 {
3014 spec_done_ds = targetm.sched.get_insn_spec_ds (insn);
3015 spec_done_ds = ds_get_max_dep_weak (spec_done_ds);
3016 }
3017 else
3018 spec_done_ds = 0;
3019
3020 /* Initialize INSN's expr. */
3021 init_expr (INSN_EXPR (insn), vinsn_create (insn, force_unique_p), 0,
3022 REG_BR_PROB_BASE, INSN_PRIORITY (insn), 0, BLOCK_NUM (insn),
3023 spec_done_ds, 0, 0, vNULL, true,
3024 false, false, false, CANT_MOVE (insn));
3025 }
3026
3027 init_first_time_insn_data (insn);
3028 }
3029
3030 /* Scan the region and initialize instruction data for basic blocks BBS. */
3031 void
3032 sel_init_global_and_expr (bb_vec_t bbs)
3033 {
3034 /* ??? It would be nice to implement push / pop scheme for sched_infos. */
3035 const struct sched_scan_info_def ssi =
3036 {
3037 NULL, /* extend_bb */
3038 init_global_and_expr_for_bb, /* init_bb */
3039 extend_insn_data, /* extend_insn */
3040 init_global_and_expr_for_insn /* init_insn */
3041 };
3042
3043 sched_scan (&ssi, bbs);
3044 }
3045
3046 /* Finalize region-scope data structures for basic blocks. */
3047 static void
3048 finish_global_and_expr_for_bb (basic_block bb)
3049 {
3050 av_set_clear (&BB_AV_SET (bb));
3051 BB_AV_LEVEL (bb) = 0;
3052 }
3053
3054 /* Finalize INSN's data. */
3055 static void
3056 finish_global_and_expr_insn (insn_t insn)
3057 {
3058 if (LABEL_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
3059 return;
3060
3061 gcc_assert (INSN_P (insn));
3062
3063 if (INSN_LUID (insn) > 0)
3064 {
3065 free_first_time_insn_data (insn);
3066 INSN_WS_LEVEL (insn) = 0;
3067 CANT_MOVE (insn) = 0;
3068
3069 /* We can no longer assert this, as vinsns of this insn could be
3070 easily live in other insn's caches. This should be changed to
3071 a counter-like approach among all vinsns. */
3072 gcc_assert (true || VINSN_COUNT (INSN_VINSN (insn)) == 1);
3073 clear_expr (INSN_EXPR (insn));
3074 }
3075 }
3076
3077 /* Finalize per instruction data for the whole region. */
3078 void
3079 sel_finish_global_and_expr (void)
3080 {
3081 {
3082 bb_vec_t bbs;
3083 int i;
3084
3085 bbs.create (current_nr_blocks);
3086
3087 for (i = 0; i < current_nr_blocks; i++)
3088 bbs.quick_push (BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i)));
3089
3090 /* Clear AV_SETs and INSN_EXPRs. */
3091 {
3092 const struct sched_scan_info_def ssi =
3093 {
3094 NULL, /* extend_bb */
3095 finish_global_and_expr_for_bb, /* init_bb */
3096 NULL, /* extend_insn */
3097 finish_global_and_expr_insn /* init_insn */
3098 };
3099
3100 sched_scan (&ssi, bbs);
3101 }
3102
3103 bbs.release ();
3104 }
3105
3106 finish_insns ();
3107 }
3108 \f
3109
3110 /* In the below hooks, we merely calculate whether or not a dependence
3111 exists, and in what part of insn. However, we will need more data
3112 when we'll start caching dependence requests. */
3113
3114 /* Container to hold information for dependency analysis. */
3115 static struct
3116 {
3117 deps_t dc;
3118
3119 /* A variable to track which part of rtx we are scanning in
3120 sched-deps.c: sched_analyze_insn (). */
3121 deps_where_t where;
3122
3123 /* Current producer. */
3124 insn_t pro;
3125
3126 /* Current consumer. */
3127 vinsn_t con;
3128
3129 /* Is SEL_DEPS_HAS_DEP_P[DEPS_IN_X] is true, then X has a dependence.
3130 X is from { INSN, LHS, RHS }. */
3131 ds_t has_dep_p[DEPS_IN_NOWHERE];
3132 } has_dependence_data;
3133
3134 /* Start analyzing dependencies of INSN. */
3135 static void
3136 has_dependence_start_insn (insn_t insn ATTRIBUTE_UNUSED)
3137 {
3138 gcc_assert (has_dependence_data.where == DEPS_IN_NOWHERE);
3139
3140 has_dependence_data.where = DEPS_IN_INSN;
3141 }
3142
3143 /* Finish analyzing dependencies of an insn. */
3144 static void
3145 has_dependence_finish_insn (void)
3146 {
3147 gcc_assert (has_dependence_data.where == DEPS_IN_INSN);
3148
3149 has_dependence_data.where = DEPS_IN_NOWHERE;
3150 }
3151
3152 /* Start analyzing dependencies of LHS. */
3153 static void
3154 has_dependence_start_lhs (rtx lhs ATTRIBUTE_UNUSED)
3155 {
3156 gcc_assert (has_dependence_data.where == DEPS_IN_INSN);
3157
3158 if (VINSN_LHS (has_dependence_data.con) != NULL)
3159 has_dependence_data.where = DEPS_IN_LHS;
3160 }
3161
3162 /* Finish analyzing dependencies of an lhs. */
3163 static void
3164 has_dependence_finish_lhs (void)
3165 {
3166 has_dependence_data.where = DEPS_IN_INSN;
3167 }
3168
3169 /* Start analyzing dependencies of RHS. */
3170 static void
3171 has_dependence_start_rhs (rtx rhs ATTRIBUTE_UNUSED)
3172 {
3173 gcc_assert (has_dependence_data.where == DEPS_IN_INSN);
3174
3175 if (VINSN_RHS (has_dependence_data.con) != NULL)
3176 has_dependence_data.where = DEPS_IN_RHS;
3177 }
3178
3179 /* Start analyzing dependencies of an rhs. */
3180 static void
3181 has_dependence_finish_rhs (void)
3182 {
3183 gcc_assert (has_dependence_data.where == DEPS_IN_RHS
3184 || has_dependence_data.where == DEPS_IN_INSN);
3185
3186 has_dependence_data.where = DEPS_IN_INSN;
3187 }
3188
3189 /* Note a set of REGNO. */
3190 static void
3191 has_dependence_note_reg_set (int regno)
3192 {
3193 struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno];
3194
3195 if (!sched_insns_conditions_mutex_p (has_dependence_data.pro,
3196 VINSN_INSN_RTX
3197 (has_dependence_data.con)))
3198 {
3199 ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where];
3200
3201 if (reg_last->sets != NULL
3202 || reg_last->clobbers != NULL)
3203 *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT;
3204
3205 if (reg_last->uses || reg_last->implicit_sets)
3206 *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI;
3207 }
3208 }
3209
3210 /* Note a clobber of REGNO. */
3211 static void
3212 has_dependence_note_reg_clobber (int regno)
3213 {
3214 struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno];
3215
3216 if (!sched_insns_conditions_mutex_p (has_dependence_data.pro,
3217 VINSN_INSN_RTX
3218 (has_dependence_data.con)))
3219 {
3220 ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where];
3221
3222 if (reg_last->sets)
3223 *dsp = (*dsp & ~SPECULATIVE) | DEP_OUTPUT;
3224
3225 if (reg_last->uses || reg_last->implicit_sets)
3226 *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI;
3227 }
3228 }
3229
3230 /* Note a use of REGNO. */
3231 static void
3232 has_dependence_note_reg_use (int regno)
3233 {
3234 struct deps_reg *reg_last = &has_dependence_data.dc->reg_last[regno];
3235
3236 if (!sched_insns_conditions_mutex_p (has_dependence_data.pro,
3237 VINSN_INSN_RTX
3238 (has_dependence_data.con)))
3239 {
3240 ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where];
3241
3242 if (reg_last->sets)
3243 *dsp = (*dsp & ~SPECULATIVE) | DEP_TRUE;
3244
3245 if (reg_last->clobbers || reg_last->implicit_sets)
3246 *dsp = (*dsp & ~SPECULATIVE) | DEP_ANTI;
3247
3248 /* Merge BE_IN_SPEC bits into *DSP when the dependency producer
3249 is actually a check insn. We need to do this for any register
3250 read-read dependency with the check unless we track properly
3251 all registers written by BE_IN_SPEC-speculated insns, as
3252 we don't have explicit dependence lists. See PR 53975. */
3253 if (reg_last->uses)
3254 {
3255 ds_t pro_spec_checked_ds;
3256
3257 pro_spec_checked_ds = INSN_SPEC_CHECKED_DS (has_dependence_data.pro);
3258 pro_spec_checked_ds = ds_get_max_dep_weak (pro_spec_checked_ds);
3259
3260 if (pro_spec_checked_ds != 0)
3261 *dsp = ds_full_merge (*dsp, pro_spec_checked_ds,
3262 NULL_RTX, NULL_RTX);
3263 }
3264 }
3265 }
3266
3267 /* Note a memory dependence. */
3268 static void
3269 has_dependence_note_mem_dep (rtx mem ATTRIBUTE_UNUSED,
3270 rtx pending_mem ATTRIBUTE_UNUSED,
3271 insn_t pending_insn ATTRIBUTE_UNUSED,
3272 ds_t ds ATTRIBUTE_UNUSED)
3273 {
3274 if (!sched_insns_conditions_mutex_p (has_dependence_data.pro,
3275 VINSN_INSN_RTX (has_dependence_data.con)))
3276 {
3277 ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where];
3278
3279 *dsp = ds_full_merge (ds, *dsp, pending_mem, mem);
3280 }
3281 }
3282
3283 /* Note a dependence. */
3284 static void
3285 has_dependence_note_dep (insn_t pro ATTRIBUTE_UNUSED,
3286 ds_t ds ATTRIBUTE_UNUSED)
3287 {
3288 if (!sched_insns_conditions_mutex_p (has_dependence_data.pro,
3289 VINSN_INSN_RTX (has_dependence_data.con)))
3290 {
3291 ds_t *dsp = &has_dependence_data.has_dep_p[has_dependence_data.where];
3292
3293 *dsp = ds_full_merge (ds, *dsp, NULL_RTX, NULL_RTX);
3294 }
3295 }
3296
3297 /* Mark the insn as having a hard dependence that prevents speculation. */
3298 void
3299 sel_mark_hard_insn (rtx insn)
3300 {
3301 int i;
3302
3303 /* Only work when we're in has_dependence_p mode.
3304 ??? This is a hack, this should actually be a hook. */
3305 if (!has_dependence_data.dc || !has_dependence_data.pro)
3306 return;
3307
3308 gcc_assert (insn == VINSN_INSN_RTX (has_dependence_data.con));
3309 gcc_assert (has_dependence_data.where == DEPS_IN_INSN);
3310
3311 for (i = 0; i < DEPS_IN_NOWHERE; i++)
3312 has_dependence_data.has_dep_p[i] &= ~SPECULATIVE;
3313 }
3314
3315 /* This structure holds the hooks for the dependency analysis used when
3316 actually processing dependencies in the scheduler. */
3317 static struct sched_deps_info_def has_dependence_sched_deps_info;
3318
3319 /* This initializes most of the fields of the above structure. */
3320 static const struct sched_deps_info_def const_has_dependence_sched_deps_info =
3321 {
3322 NULL,
3323
3324 has_dependence_start_insn,
3325 has_dependence_finish_insn,
3326 has_dependence_start_lhs,
3327 has_dependence_finish_lhs,
3328 has_dependence_start_rhs,
3329 has_dependence_finish_rhs,
3330 has_dependence_note_reg_set,
3331 has_dependence_note_reg_clobber,
3332 has_dependence_note_reg_use,
3333 has_dependence_note_mem_dep,
3334 has_dependence_note_dep,
3335
3336 0, /* use_cselib */
3337 0, /* use_deps_list */
3338 0 /* generate_spec_deps */
3339 };
3340
3341 /* Initialize has_dependence_sched_deps_info with extra spec field. */
3342 static void
3343 setup_has_dependence_sched_deps_info (void)
3344 {
3345 memcpy (&has_dependence_sched_deps_info,
3346 &const_has_dependence_sched_deps_info,
3347 sizeof (has_dependence_sched_deps_info));
3348
3349 if (spec_info != NULL)
3350 has_dependence_sched_deps_info.generate_spec_deps = 1;
3351
3352 sched_deps_info = &has_dependence_sched_deps_info;
3353 }
3354
3355 /* Remove all dependences found and recorded in has_dependence_data array. */
3356 void
3357 sel_clear_has_dependence (void)
3358 {
3359 int i;
3360
3361 for (i = 0; i < DEPS_IN_NOWHERE; i++)
3362 has_dependence_data.has_dep_p[i] = 0;
3363 }
3364
3365 /* Return nonzero if EXPR has is dependent upon PRED. Return the pointer
3366 to the dependence information array in HAS_DEP_PP. */
3367 ds_t
3368 has_dependence_p (expr_t expr, insn_t pred, ds_t **has_dep_pp)
3369 {
3370 int i;
3371 ds_t ds;
3372 struct deps_desc *dc;
3373
3374 if (INSN_SIMPLEJUMP_P (pred))
3375 /* Unconditional jump is just a transfer of control flow.
3376 Ignore it. */
3377 return false;
3378
3379 dc = &INSN_DEPS_CONTEXT (pred);
3380
3381 /* We init this field lazily. */
3382 if (dc->reg_last == NULL)
3383 init_deps_reg_last (dc);
3384
3385 if (!dc->readonly)
3386 {
3387 has_dependence_data.pro = NULL;
3388 /* Initialize empty dep context with information about PRED. */
3389 advance_deps_context (dc, pred);
3390 dc->readonly = 1;
3391 }
3392
3393 has_dependence_data.where = DEPS_IN_NOWHERE;
3394 has_dependence_data.pro = pred;
3395 has_dependence_data.con = EXPR_VINSN (expr);
3396 has_dependence_data.dc = dc;
3397
3398 sel_clear_has_dependence ();
3399
3400 /* Now catch all dependencies that would be generated between PRED and
3401 INSN. */
3402 setup_has_dependence_sched_deps_info ();
3403 deps_analyze_insn (dc, EXPR_INSN_RTX (expr));
3404 has_dependence_data.dc = NULL;
3405
3406 /* When a barrier was found, set DEPS_IN_INSN bits. */
3407 if (dc->last_reg_pending_barrier == TRUE_BARRIER)
3408 has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_TRUE;
3409 else if (dc->last_reg_pending_barrier == MOVE_BARRIER)
3410 has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI;
3411
3412 /* Do not allow stores to memory to move through checks. Currently
3413 we don't move this to sched-deps.c as the check doesn't have
3414 obvious places to which this dependence can be attached.
3415 FIMXE: this should go to a hook. */
3416 if (EXPR_LHS (expr)
3417 && MEM_P (EXPR_LHS (expr))
3418 && sel_insn_is_speculation_check (pred))
3419 has_dependence_data.has_dep_p[DEPS_IN_INSN] = DEP_ANTI;
3420
3421 *has_dep_pp = has_dependence_data.has_dep_p;
3422 ds = 0;
3423 for (i = 0; i < DEPS_IN_NOWHERE; i++)
3424 ds = ds_full_merge (ds, has_dependence_data.has_dep_p[i],
3425 NULL_RTX, NULL_RTX);
3426
3427 return ds;
3428 }
3429 \f
3430
3431 /* Dependence hooks implementation that checks dependence latency constraints
3432 on the insns being scheduled. The entry point for these routines is
3433 tick_check_p predicate. */
3434
3435 static struct
3436 {
3437 /* An expr we are currently checking. */
3438 expr_t expr;
3439
3440 /* A minimal cycle for its scheduling. */
3441 int cycle;
3442
3443 /* Whether we have seen a true dependence while checking. */
3444 bool seen_true_dep_p;
3445 } tick_check_data;
3446
3447 /* Update minimal scheduling cycle for tick_check_insn given that it depends
3448 on PRO with status DS and weight DW. */
3449 static void
3450 tick_check_dep_with_dw (insn_t pro_insn, ds_t ds, dw_t dw)
3451 {
3452 expr_t con_expr = tick_check_data.expr;
3453 insn_t con_insn = EXPR_INSN_RTX (con_expr);
3454
3455 if (con_insn != pro_insn)
3456 {
3457 enum reg_note dt;
3458 int tick;
3459
3460 if (/* PROducer was removed from above due to pipelining. */
3461 !INSN_IN_STREAM_P (pro_insn)
3462 /* Or PROducer was originally on the next iteration regarding the
3463 CONsumer. */
3464 || (INSN_SCHED_TIMES (pro_insn)
3465 - EXPR_SCHED_TIMES (con_expr)) > 1)
3466 /* Don't count this dependence. */
3467 return;
3468
3469 dt = ds_to_dt (ds);
3470 if (dt == REG_DEP_TRUE)
3471 tick_check_data.seen_true_dep_p = true;
3472
3473 gcc_assert (INSN_SCHED_CYCLE (pro_insn) > 0);
3474
3475 {
3476 dep_def _dep, *dep = &_dep;
3477
3478 init_dep (dep, pro_insn, con_insn, dt);
3479
3480 tick = INSN_SCHED_CYCLE (pro_insn) + dep_cost_1 (dep, dw);
3481 }
3482
3483 /* When there are several kinds of dependencies between pro and con,
3484 only REG_DEP_TRUE should be taken into account. */
3485 if (tick > tick_check_data.cycle
3486 && (dt == REG_DEP_TRUE || !tick_check_data.seen_true_dep_p))
3487 tick_check_data.cycle = tick;
3488 }
3489 }
3490
3491 /* An implementation of note_dep hook. */
3492 static void
3493 tick_check_note_dep (insn_t pro, ds_t ds)
3494 {
3495 tick_check_dep_with_dw (pro, ds, 0);
3496 }
3497
3498 /* An implementation of note_mem_dep hook. */
3499 static void
3500 tick_check_note_mem_dep (rtx mem1, rtx mem2, insn_t pro, ds_t ds)
3501 {
3502 dw_t dw;
3503
3504 dw = (ds_to_dt (ds) == REG_DEP_TRUE
3505 ? estimate_dep_weak (mem1, mem2)
3506 : 0);
3507
3508 tick_check_dep_with_dw (pro, ds, dw);
3509 }
3510
3511 /* This structure contains hooks for dependence analysis used when determining
3512 whether an insn is ready for scheduling. */
3513 static struct sched_deps_info_def tick_check_sched_deps_info =
3514 {
3515 NULL,
3516
3517 NULL,
3518 NULL,
3519 NULL,
3520 NULL,
3521 NULL,
3522 NULL,
3523 haifa_note_reg_set,
3524 haifa_note_reg_clobber,
3525 haifa_note_reg_use,
3526 tick_check_note_mem_dep,
3527 tick_check_note_dep,
3528
3529 0, 0, 0
3530 };
3531
3532 /* Estimate number of cycles from the current cycle of FENCE until EXPR can be
3533 scheduled. Return 0 if all data from producers in DC is ready. */
3534 int
3535 tick_check_p (expr_t expr, deps_t dc, fence_t fence)
3536 {
3537 int cycles_left;
3538 /* Initialize variables. */
3539 tick_check_data.expr = expr;
3540 tick_check_data.cycle = 0;
3541 tick_check_data.seen_true_dep_p = false;
3542 sched_deps_info = &tick_check_sched_deps_info;
3543
3544 gcc_assert (!dc->readonly);
3545 dc->readonly = 1;
3546 deps_analyze_insn (dc, EXPR_INSN_RTX (expr));
3547 dc->readonly = 0;
3548
3549 cycles_left = tick_check_data.cycle - FENCE_CYCLE (fence);
3550
3551 return cycles_left >= 0 ? cycles_left : 0;
3552 }
3553 \f
3554
3555 /* Functions to work with insns. */
3556
3557 /* Returns true if LHS of INSN is the same as DEST of an insn
3558 being moved. */
3559 bool
3560 lhs_of_insn_equals_to_dest_p (insn_t insn, rtx dest)
3561 {
3562 rtx lhs = INSN_LHS (insn);
3563
3564 if (lhs == NULL || dest == NULL)
3565 return false;
3566
3567 return rtx_equal_p (lhs, dest);
3568 }
3569
3570 /* Return s_i_d entry of INSN. Callable from debugger. */
3571 sel_insn_data_def
3572 insn_sid (insn_t insn)
3573 {
3574 return *SID (insn);
3575 }
3576
3577 /* True when INSN is a speculative check. We can tell this by looking
3578 at the data structures of the selective scheduler, not by examining
3579 the pattern. */
3580 bool
3581 sel_insn_is_speculation_check (rtx insn)
3582 {
3583 return s_i_d.exists () && !! INSN_SPEC_CHECKED_DS (insn);
3584 }
3585
3586 /* Extracts machine mode MODE and destination location DST_LOC
3587 for given INSN. */
3588 void
3589 get_dest_and_mode (rtx insn, rtx *dst_loc, enum machine_mode *mode)
3590 {
3591 rtx pat = PATTERN (insn);
3592
3593 gcc_assert (dst_loc);
3594 gcc_assert (GET_CODE (pat) == SET);
3595
3596 *dst_loc = SET_DEST (pat);
3597
3598 gcc_assert (*dst_loc);
3599 gcc_assert (MEM_P (*dst_loc) || REG_P (*dst_loc));
3600
3601 if (mode)
3602 *mode = GET_MODE (*dst_loc);
3603 }
3604
3605 /* Returns true when moving through JUMP will result in bookkeeping
3606 creation. */
3607 bool
3608 bookkeeping_can_be_created_if_moved_through_p (insn_t jump)
3609 {
3610 insn_t succ;
3611 succ_iterator si;
3612
3613 FOR_EACH_SUCC (succ, si, jump)
3614 if (sel_num_cfg_preds_gt_1 (succ))
3615 return true;
3616
3617 return false;
3618 }
3619
3620 /* Return 'true' if INSN is the only one in its basic block. */
3621 static bool
3622 insn_is_the_only_one_in_bb_p (insn_t insn)
3623 {
3624 return sel_bb_head_p (insn) && sel_bb_end_p (insn);
3625 }
3626
3627 #ifdef ENABLE_CHECKING
3628 /* Check that the region we're scheduling still has at most one
3629 backedge. */
3630 static void
3631 verify_backedges (void)
3632 {
3633 if (pipelining_p)
3634 {
3635 int i, n = 0;
3636 edge e;
3637 edge_iterator ei;
3638
3639 for (i = 0; i < current_nr_blocks; i++)
3640 FOR_EACH_EDGE (e, ei, BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i))->succs)
3641 if (in_current_region_p (e->dest)
3642 && BLOCK_TO_BB (e->dest->index) < i)
3643 n++;
3644
3645 gcc_assert (n <= 1);
3646 }
3647 }
3648 #endif
3649 \f
3650
3651 /* Functions to work with control flow. */
3652
3653 /* Recompute BLOCK_TO_BB and BB_FOR_BLOCK for current region so that blocks
3654 are sorted in topological order (it might have been invalidated by
3655 redirecting an edge). */
3656 static void
3657 sel_recompute_toporder (void)
3658 {
3659 int i, n, rgn;
3660 int *postorder, n_blocks;
3661
3662 postorder = XALLOCAVEC (int, n_basic_blocks_for_fn (cfun));
3663 n_blocks = post_order_compute (postorder, false, false);
3664
3665 rgn = CONTAINING_RGN (BB_TO_BLOCK (0));
3666 for (n = 0, i = n_blocks - 1; i >= 0; i--)
3667 if (CONTAINING_RGN (postorder[i]) == rgn)
3668 {
3669 BLOCK_TO_BB (postorder[i]) = n;
3670 BB_TO_BLOCK (n) = postorder[i];
3671 n++;
3672 }
3673
3674 /* Assert that we updated info for all blocks. We may miss some blocks if
3675 this function is called when redirecting an edge made a block
3676 unreachable, but that block is not deleted yet. */
3677 gcc_assert (n == RGN_NR_BLOCKS (rgn));
3678 }
3679
3680 /* Tidy the possibly empty block BB. */
3681 static bool
3682 maybe_tidy_empty_bb (basic_block bb)
3683 {
3684 basic_block succ_bb, pred_bb, note_bb;
3685 vec<basic_block> dom_bbs;
3686 edge e;
3687 edge_iterator ei;
3688 bool rescan_p;
3689
3690 /* Keep empty bb only if this block immediately precedes EXIT and
3691 has incoming non-fallthrough edge, or it has no predecessors or
3692 successors. Otherwise remove it. */
3693 if (!sel_bb_empty_p (bb)
3694 || (single_succ_p (bb)
3695 && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun)
3696 && (!single_pred_p (bb)
3697 || !(single_pred_edge (bb)->flags & EDGE_FALLTHRU)))
3698 || EDGE_COUNT (bb->preds) == 0
3699 || EDGE_COUNT (bb->succs) == 0)
3700 return false;
3701
3702 /* Do not attempt to redirect complex edges. */
3703 FOR_EACH_EDGE (e, ei, bb->preds)
3704 if (e->flags & EDGE_COMPLEX)
3705 return false;
3706 else if (e->flags & EDGE_FALLTHRU)
3707 {
3708 rtx note;
3709 /* If prev bb ends with asm goto, see if any of the
3710 ASM_OPERANDS_LABELs don't point to the fallthru
3711 label. Do not attempt to redirect it in that case. */
3712 if (JUMP_P (BB_END (e->src))
3713 && (note = extract_asm_operands (PATTERN (BB_END (e->src)))))
3714 {
3715 int i, n = ASM_OPERANDS_LABEL_LENGTH (note);
3716
3717 for (i = 0; i < n; ++i)
3718 if (XEXP (ASM_OPERANDS_LABEL (note, i), 0) == BB_HEAD (bb))
3719 return false;
3720 }
3721 }
3722
3723 free_data_sets (bb);
3724
3725 /* Do not delete BB if it has more than one successor.
3726 That can occur when we moving a jump. */
3727 if (!single_succ_p (bb))
3728 {
3729 gcc_assert (can_merge_blocks_p (bb->prev_bb, bb));
3730 sel_merge_blocks (bb->prev_bb, bb);
3731 return true;
3732 }
3733
3734 succ_bb = single_succ (bb);
3735 rescan_p = true;
3736 pred_bb = NULL;
3737 dom_bbs.create (0);
3738
3739 /* Save a pred/succ from the current region to attach the notes to. */
3740 note_bb = NULL;
3741 FOR_EACH_EDGE (e, ei, bb->preds)
3742 if (in_current_region_p (e->src))
3743 {
3744 note_bb = e->src;
3745 break;
3746 }
3747 if (note_bb == NULL)
3748 note_bb = succ_bb;
3749
3750 /* Redirect all non-fallthru edges to the next bb. */
3751 while (rescan_p)
3752 {
3753 rescan_p = false;
3754
3755 FOR_EACH_EDGE (e, ei, bb->preds)
3756 {
3757 pred_bb = e->src;
3758
3759 if (!(e->flags & EDGE_FALLTHRU))
3760 {
3761 /* We can not invalidate computed topological order by moving
3762 the edge destination block (E->SUCC) along a fallthru edge.
3763
3764 We will update dominators here only when we'll get
3765 an unreachable block when redirecting, otherwise
3766 sel_redirect_edge_and_branch will take care of it. */
3767 if (e->dest != bb
3768 && single_pred_p (e->dest))
3769 dom_bbs.safe_push (e->dest);
3770 sel_redirect_edge_and_branch (e, succ_bb);
3771 rescan_p = true;
3772 break;
3773 }
3774 /* If the edge is fallthru, but PRED_BB ends in a conditional jump
3775 to BB (so there is no non-fallthru edge from PRED_BB to BB), we
3776 still have to adjust it. */
3777 else if (single_succ_p (pred_bb) && any_condjump_p (BB_END (pred_bb)))
3778 {
3779 /* If possible, try to remove the unneeded conditional jump. */
3780 if (INSN_SCHED_TIMES (BB_END (pred_bb)) == 0
3781 && !IN_CURRENT_FENCE_P (BB_END (pred_bb)))
3782 {
3783 if (!sel_remove_insn (BB_END (pred_bb), false, false))
3784 tidy_fallthru_edge (e);
3785 }
3786 else
3787 sel_redirect_edge_and_branch (e, succ_bb);
3788 rescan_p = true;
3789 break;
3790 }
3791 }
3792 }
3793
3794 if (can_merge_blocks_p (bb->prev_bb, bb))
3795 sel_merge_blocks (bb->prev_bb, bb);
3796 else
3797 {
3798 /* This is a block without fallthru predecessor. Just delete it. */
3799 gcc_assert (note_bb);
3800 move_bb_info (note_bb, bb);
3801 remove_empty_bb (bb, true);
3802 }
3803
3804 if (!dom_bbs.is_empty ())
3805 {
3806 dom_bbs.safe_push (succ_bb);
3807 iterate_fix_dominators (CDI_DOMINATORS, dom_bbs, false);
3808 dom_bbs.release ();
3809 }
3810
3811 return true;
3812 }
3813
3814 /* Tidy the control flow after we have removed original insn from
3815 XBB. Return true if we have removed some blocks. When FULL_TIDYING
3816 is true, also try to optimize control flow on non-empty blocks. */
3817 bool
3818 tidy_control_flow (basic_block xbb, bool full_tidying)
3819 {
3820 bool changed = true;
3821 insn_t first, last;
3822
3823 /* First check whether XBB is empty. */
3824 changed = maybe_tidy_empty_bb (xbb);
3825 if (changed || !full_tidying)
3826 return changed;
3827
3828 /* Check if there is a unnecessary jump after insn left. */
3829 if (bb_has_removable_jump_to_p (xbb, xbb->next_bb)
3830 && INSN_SCHED_TIMES (BB_END (xbb)) == 0
3831 && !IN_CURRENT_FENCE_P (BB_END (xbb)))
3832 {
3833 if (sel_remove_insn (BB_END (xbb), false, false))
3834 return true;
3835 tidy_fallthru_edge (EDGE_SUCC (xbb, 0));
3836 }
3837
3838 first = sel_bb_head (xbb);
3839 last = sel_bb_end (xbb);
3840 if (MAY_HAVE_DEBUG_INSNS)
3841 {
3842 if (first != last && DEBUG_INSN_P (first))
3843 do
3844 first = NEXT_INSN (first);
3845 while (first != last && (DEBUG_INSN_P (first) || NOTE_P (first)));
3846
3847 if (first != last && DEBUG_INSN_P (last))
3848 do
3849 last = PREV_INSN (last);
3850 while (first != last && (DEBUG_INSN_P (last) || NOTE_P (last)));
3851 }
3852 /* Check if there is an unnecessary jump in previous basic block leading
3853 to next basic block left after removing INSN from stream.
3854 If it is so, remove that jump and redirect edge to current
3855 basic block (where there was INSN before deletion). This way
3856 when NOP will be deleted several instructions later with its
3857 basic block we will not get a jump to next instruction, which
3858 can be harmful. */
3859 if (first == last
3860 && !sel_bb_empty_p (xbb)
3861 && INSN_NOP_P (last)
3862 /* Flow goes fallthru from current block to the next. */
3863 && EDGE_COUNT (xbb->succs) == 1
3864 && (EDGE_SUCC (xbb, 0)->flags & EDGE_FALLTHRU)
3865 /* When successor is an EXIT block, it may not be the next block. */
3866 && single_succ (xbb) != EXIT_BLOCK_PTR_FOR_FN (cfun)
3867 /* And unconditional jump in previous basic block leads to
3868 next basic block of XBB and this jump can be safely removed. */
3869 && in_current_region_p (xbb->prev_bb)
3870 && bb_has_removable_jump_to_p (xbb->prev_bb, xbb->next_bb)
3871 && INSN_SCHED_TIMES (BB_END (xbb->prev_bb)) == 0
3872 /* Also this jump is not at the scheduling boundary. */
3873 && !IN_CURRENT_FENCE_P (BB_END (xbb->prev_bb)))
3874 {
3875 bool recompute_toporder_p;
3876 /* Clear data structures of jump - jump itself will be removed
3877 by sel_redirect_edge_and_branch. */
3878 clear_expr (INSN_EXPR (BB_END (xbb->prev_bb)));
3879 recompute_toporder_p
3880 = sel_redirect_edge_and_branch (EDGE_SUCC (xbb->prev_bb, 0), xbb);
3881
3882 gcc_assert (EDGE_SUCC (xbb->prev_bb, 0)->flags & EDGE_FALLTHRU);
3883
3884 /* It can turn out that after removing unused jump, basic block
3885 that contained that jump, becomes empty too. In such case
3886 remove it too. */
3887 if (sel_bb_empty_p (xbb->prev_bb))
3888 changed = maybe_tidy_empty_bb (xbb->prev_bb);
3889 if (recompute_toporder_p)
3890 sel_recompute_toporder ();
3891 }
3892
3893 #ifdef ENABLE_CHECKING
3894 verify_backedges ();
3895 verify_dominators (CDI_DOMINATORS);
3896 #endif
3897
3898 return changed;
3899 }
3900
3901 /* Purge meaningless empty blocks in the middle of a region. */
3902 void
3903 purge_empty_blocks (void)
3904 {
3905 int i;
3906
3907 /* Do not attempt to delete the first basic block in the region. */
3908 for (i = 1; i < current_nr_blocks; )
3909 {
3910 basic_block b = BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i));
3911
3912 if (maybe_tidy_empty_bb (b))
3913 continue;
3914
3915 i++;
3916 }
3917 }
3918
3919 /* Rip-off INSN from the insn stream. When ONLY_DISCONNECT is true,
3920 do not delete insn's data, because it will be later re-emitted.
3921 Return true if we have removed some blocks afterwards. */
3922 bool
3923 sel_remove_insn (insn_t insn, bool only_disconnect, bool full_tidying)
3924 {
3925 basic_block bb = BLOCK_FOR_INSN (insn);
3926
3927 gcc_assert (INSN_IN_STREAM_P (insn));
3928
3929 if (DEBUG_INSN_P (insn) && BB_AV_SET_VALID_P (bb))
3930 {
3931 expr_t expr;
3932 av_set_iterator i;
3933
3934 /* When we remove a debug insn that is head of a BB, it remains
3935 in the AV_SET of the block, but it shouldn't. */
3936 FOR_EACH_EXPR_1 (expr, i, &BB_AV_SET (bb))
3937 if (EXPR_INSN_RTX (expr) == insn)
3938 {
3939 av_set_iter_remove (&i);
3940 break;
3941 }
3942 }
3943
3944 if (only_disconnect)
3945 remove_insn (insn);
3946 else
3947 {
3948 delete_insn (insn);
3949 clear_expr (INSN_EXPR (insn));
3950 }
3951
3952 /* It is necessary to NULL these fields in case we are going to re-insert
3953 INSN into the insns stream, as will usually happen in the ONLY_DISCONNECT
3954 case, but also for NOPs that we will return to the nop pool. */
3955 SET_PREV_INSN (insn) = NULL_RTX;
3956 SET_NEXT_INSN (insn) = NULL_RTX;
3957 set_block_for_insn (insn, NULL);
3958
3959 return tidy_control_flow (bb, full_tidying);
3960 }
3961
3962 /* Estimate number of the insns in BB. */
3963 static int
3964 sel_estimate_number_of_insns (basic_block bb)
3965 {
3966 int res = 0;
3967 insn_t insn = NEXT_INSN (BB_HEAD (bb)), next_tail = NEXT_INSN (BB_END (bb));
3968
3969 for (; insn != next_tail; insn = NEXT_INSN (insn))
3970 if (NONDEBUG_INSN_P (insn))
3971 res++;
3972
3973 return res;
3974 }
3975
3976 /* We don't need separate luids for notes or labels. */
3977 static int
3978 sel_luid_for_non_insn (rtx x)
3979 {
3980 gcc_assert (NOTE_P (x) || LABEL_P (x));
3981
3982 return -1;
3983 }
3984
3985 /* Find the proper seqno for inserting at INSN by successors.
3986 Return -1 if no successors with positive seqno exist. */
3987 static int
3988 get_seqno_by_succs (rtx_insn *insn)
3989 {
3990 basic_block bb = BLOCK_FOR_INSN (insn);
3991 rtx_insn *tmp = insn, *end = BB_END (bb);
3992 int seqno;
3993 insn_t succ = NULL;
3994 succ_iterator si;
3995
3996 while (tmp != end)
3997 {
3998 tmp = NEXT_INSN (tmp);
3999 if (INSN_P (tmp))
4000 return INSN_SEQNO (tmp);
4001 }
4002
4003 seqno = INT_MAX;
4004
4005 FOR_EACH_SUCC_1 (succ, si, end, SUCCS_NORMAL)
4006 if (INSN_SEQNO (succ) > 0)
4007 seqno = MIN (seqno, INSN_SEQNO (succ));
4008
4009 if (seqno == INT_MAX)
4010 return -1;
4011
4012 return seqno;
4013 }
4014
4015 /* Compute seqno for INSN by its preds or succs. Use OLD_SEQNO to compute
4016 seqno in corner cases. */
4017 static int
4018 get_seqno_for_a_jump (insn_t insn, int old_seqno)
4019 {
4020 int seqno;
4021
4022 gcc_assert (INSN_SIMPLEJUMP_P (insn));
4023
4024 if (!sel_bb_head_p (insn))
4025 seqno = INSN_SEQNO (PREV_INSN (insn));
4026 else
4027 {
4028 basic_block bb = BLOCK_FOR_INSN (insn);
4029
4030 if (single_pred_p (bb)
4031 && !in_current_region_p (single_pred (bb)))
4032 {
4033 /* We can have preds outside a region when splitting edges
4034 for pipelining of an outer loop. Use succ instead.
4035 There should be only one of them. */
4036 insn_t succ = NULL;
4037 succ_iterator si;
4038 bool first = true;
4039
4040 gcc_assert (flag_sel_sched_pipelining_outer_loops
4041 && current_loop_nest);
4042 FOR_EACH_SUCC_1 (succ, si, insn,
4043 SUCCS_NORMAL | SUCCS_SKIP_TO_LOOP_EXITS)
4044 {
4045 gcc_assert (first);
4046 first = false;
4047 }
4048
4049 gcc_assert (succ != NULL);
4050 seqno = INSN_SEQNO (succ);
4051 }
4052 else
4053 {
4054 insn_t *preds;
4055 int n;
4056
4057 cfg_preds (BLOCK_FOR_INSN (insn), &preds, &n);
4058
4059 gcc_assert (n > 0);
4060 /* For one predecessor, use simple method. */
4061 if (n == 1)
4062 seqno = INSN_SEQNO (preds[0]);
4063 else
4064 seqno = get_seqno_by_preds (insn);
4065
4066 free (preds);
4067 }
4068 }
4069
4070 /* We were unable to find a good seqno among preds. */
4071 if (seqno < 0)
4072 seqno = get_seqno_by_succs (insn);
4073
4074 if (seqno < 0)
4075 {
4076 /* The only case where this could be here legally is that the only
4077 unscheduled insn was a conditional jump that got removed and turned
4078 into this unconditional one. Initialize from the old seqno
4079 of that jump passed down to here. */
4080 seqno = old_seqno;
4081 }
4082
4083 gcc_assert (seqno >= 0);
4084 return seqno;
4085 }
4086
4087 /* Find the proper seqno for inserting at INSN. Returns -1 if no predecessors
4088 with positive seqno exist. */
4089 int
4090 get_seqno_by_preds (rtx_insn *insn)
4091 {
4092 basic_block bb = BLOCK_FOR_INSN (insn);
4093 rtx_insn *tmp = insn, *head = BB_HEAD (bb);
4094 insn_t *preds;
4095 int n, i, seqno;
4096
4097 while (tmp != head)
4098 {
4099 tmp = PREV_INSN (tmp);
4100 if (INSN_P (tmp))
4101 return INSN_SEQNO (tmp);
4102 }
4103
4104 cfg_preds (bb, &preds, &n);
4105 for (i = 0, seqno = -1; i < n; i++)
4106 seqno = MAX (seqno, INSN_SEQNO (preds[i]));
4107
4108 return seqno;
4109 }
4110
4111 \f
4112
4113 /* Extend pass-scope data structures for basic blocks. */
4114 void
4115 sel_extend_global_bb_info (void)
4116 {
4117 sel_global_bb_info.safe_grow_cleared (last_basic_block_for_fn (cfun));
4118 }
4119
4120 /* Extend region-scope data structures for basic blocks. */
4121 static void
4122 extend_region_bb_info (void)
4123 {
4124 sel_region_bb_info.safe_grow_cleared (last_basic_block_for_fn (cfun));
4125 }
4126
4127 /* Extend all data structures to fit for all basic blocks. */
4128 static void
4129 extend_bb_info (void)
4130 {
4131 sel_extend_global_bb_info ();
4132 extend_region_bb_info ();
4133 }
4134
4135 /* Finalize pass-scope data structures for basic blocks. */
4136 void
4137 sel_finish_global_bb_info (void)
4138 {
4139 sel_global_bb_info.release ();
4140 }
4141
4142 /* Finalize region-scope data structures for basic blocks. */
4143 static void
4144 finish_region_bb_info (void)
4145 {
4146 sel_region_bb_info.release ();
4147 }
4148 \f
4149
4150 /* Data for each insn in current region. */
4151 vec<sel_insn_data_def> s_i_d = vNULL;
4152
4153 /* Extend data structures for insns from current region. */
4154 static void
4155 extend_insn_data (void)
4156 {
4157 int reserve;
4158
4159 sched_extend_target ();
4160 sched_deps_init (false);
4161
4162 /* Extend data structures for insns from current region. */
4163 reserve = (sched_max_luid + 1 - s_i_d.length ());
4164 if (reserve > 0 && ! s_i_d.space (reserve))
4165 {
4166 int size;
4167
4168 if (sched_max_luid / 2 > 1024)
4169 size = sched_max_luid + 1024;
4170 else
4171 size = 3 * sched_max_luid / 2;
4172
4173
4174 s_i_d.safe_grow_cleared (size);
4175 }
4176 }
4177
4178 /* Finalize data structures for insns from current region. */
4179 static void
4180 finish_insns (void)
4181 {
4182 unsigned i;
4183
4184 /* Clear here all dependence contexts that may have left from insns that were
4185 removed during the scheduling. */
4186 for (i = 0; i < s_i_d.length (); i++)
4187 {
4188 sel_insn_data_def *sid_entry = &s_i_d[i];
4189
4190 if (sid_entry->live)
4191 return_regset_to_pool (sid_entry->live);
4192 if (sid_entry->analyzed_deps)
4193 {
4194 BITMAP_FREE (sid_entry->analyzed_deps);
4195 BITMAP_FREE (sid_entry->found_deps);
4196 htab_delete (sid_entry->transformed_insns);
4197 free_deps (&sid_entry->deps_context);
4198 }
4199 if (EXPR_VINSN (&sid_entry->expr))
4200 {
4201 clear_expr (&sid_entry->expr);
4202
4203 /* Also, clear CANT_MOVE bit here, because we really don't want it
4204 to be passed to the next region. */
4205 CANT_MOVE_BY_LUID (i) = 0;
4206 }
4207 }
4208
4209 s_i_d.release ();
4210 }
4211
4212 /* A proxy to pass initialization data to init_insn (). */
4213 static sel_insn_data_def _insn_init_ssid;
4214 static sel_insn_data_t insn_init_ssid = &_insn_init_ssid;
4215
4216 /* If true create a new vinsn. Otherwise use the one from EXPR. */
4217 static bool insn_init_create_new_vinsn_p;
4218
4219 /* Set all necessary data for initialization of the new insn[s]. */
4220 static expr_t
4221 set_insn_init (expr_t expr, vinsn_t vi, int seqno)
4222 {
4223 expr_t x = &insn_init_ssid->expr;
4224
4225 copy_expr_onside (x, expr);
4226 if (vi != NULL)
4227 {
4228 insn_init_create_new_vinsn_p = false;
4229 change_vinsn_in_expr (x, vi);
4230 }
4231 else
4232 insn_init_create_new_vinsn_p = true;
4233
4234 insn_init_ssid->seqno = seqno;
4235 return x;
4236 }
4237
4238 /* Init data for INSN. */
4239 static void
4240 init_insn_data (insn_t insn)
4241 {
4242 expr_t expr;
4243 sel_insn_data_t ssid = insn_init_ssid;
4244
4245 /* The fields mentioned below are special and hence are not being
4246 propagated to the new insns. */
4247 gcc_assert (!ssid->asm_p && ssid->sched_next == NULL
4248 && !ssid->after_stall_p && ssid->sched_cycle == 0);
4249 gcc_assert (INSN_P (insn) && INSN_LUID (insn) > 0);
4250
4251 expr = INSN_EXPR (insn);
4252 copy_expr (expr, &ssid->expr);
4253 prepare_insn_expr (insn, ssid->seqno);
4254
4255 if (insn_init_create_new_vinsn_p)
4256 change_vinsn_in_expr (expr, vinsn_create (insn, init_insn_force_unique_p));
4257
4258 if (first_time_insn_init (insn))
4259 init_first_time_insn_data (insn);
4260 }
4261
4262 /* This is used to initialize spurious jumps generated by
4263 sel_redirect_edge (). OLD_SEQNO is used for initializing seqnos
4264 in corner cases within get_seqno_for_a_jump. */
4265 static void
4266 init_simplejump_data (insn_t insn, int old_seqno)
4267 {
4268 init_expr (INSN_EXPR (insn), vinsn_create (insn, false), 0,
4269 REG_BR_PROB_BASE, 0, 0, 0, 0, 0, 0,
4270 vNULL, true, false, false,
4271 false, true);
4272 INSN_SEQNO (insn) = get_seqno_for_a_jump (insn, old_seqno);
4273 init_first_time_insn_data (insn);
4274 }
4275
4276 /* Perform deferred initialization of insns. This is used to process
4277 a new jump that may be created by redirect_edge. OLD_SEQNO is used
4278 for initializing simplejumps in init_simplejump_data. */
4279 static void
4280 sel_init_new_insn (insn_t insn, int flags, int old_seqno)
4281 {
4282 /* We create data structures for bb when the first insn is emitted in it. */
4283 if (INSN_P (insn)
4284 && INSN_IN_STREAM_P (insn)
4285 && insn_is_the_only_one_in_bb_p (insn))
4286 {
4287 extend_bb_info ();
4288 create_initial_data_sets (BLOCK_FOR_INSN (insn));
4289 }
4290
4291 if (flags & INSN_INIT_TODO_LUID)
4292 {
4293 sched_extend_luids ();
4294 sched_init_insn_luid (insn);
4295 }
4296
4297 if (flags & INSN_INIT_TODO_SSID)
4298 {
4299 extend_insn_data ();
4300 init_insn_data (insn);
4301 clear_expr (&insn_init_ssid->expr);
4302 }
4303
4304 if (flags & INSN_INIT_TODO_SIMPLEJUMP)
4305 {
4306 extend_insn_data ();
4307 init_simplejump_data (insn, old_seqno);
4308 }
4309
4310 gcc_assert (CONTAINING_RGN (BLOCK_NUM (insn))
4311 == CONTAINING_RGN (BB_TO_BLOCK (0)));
4312 }
4313 \f
4314
4315 /* Functions to init/finish work with lv sets. */
4316
4317 /* Init BB_LV_SET of BB from DF_LR_IN set of BB. */
4318 static void
4319 init_lv_set (basic_block bb)
4320 {
4321 gcc_assert (!BB_LV_SET_VALID_P (bb));
4322
4323 BB_LV_SET (bb) = get_regset_from_pool ();
4324 COPY_REG_SET (BB_LV_SET (bb), DF_LR_IN (bb));
4325 BB_LV_SET_VALID_P (bb) = true;
4326 }
4327
4328 /* Copy liveness information to BB from FROM_BB. */
4329 static void
4330 copy_lv_set_from (basic_block bb, basic_block from_bb)
4331 {
4332 gcc_assert (!BB_LV_SET_VALID_P (bb));
4333
4334 COPY_REG_SET (BB_LV_SET (bb), BB_LV_SET (from_bb));
4335 BB_LV_SET_VALID_P (bb) = true;
4336 }
4337
4338 /* Initialize lv set of all bb headers. */
4339 void
4340 init_lv_sets (void)
4341 {
4342 basic_block bb;
4343
4344 /* Initialize of LV sets. */
4345 FOR_EACH_BB_FN (bb, cfun)
4346 init_lv_set (bb);
4347
4348 /* Don't forget EXIT_BLOCK. */
4349 init_lv_set (EXIT_BLOCK_PTR_FOR_FN (cfun));
4350 }
4351
4352 /* Release lv set of HEAD. */
4353 static void
4354 free_lv_set (basic_block bb)
4355 {
4356 gcc_assert (BB_LV_SET (bb) != NULL);
4357
4358 return_regset_to_pool (BB_LV_SET (bb));
4359 BB_LV_SET (bb) = NULL;
4360 BB_LV_SET_VALID_P (bb) = false;
4361 }
4362
4363 /* Finalize lv sets of all bb headers. */
4364 void
4365 free_lv_sets (void)
4366 {
4367 basic_block bb;
4368
4369 /* Don't forget EXIT_BLOCK. */
4370 free_lv_set (EXIT_BLOCK_PTR_FOR_FN (cfun));
4371
4372 /* Free LV sets. */
4373 FOR_EACH_BB_FN (bb, cfun)
4374 if (BB_LV_SET (bb))
4375 free_lv_set (bb);
4376 }
4377
4378 /* Mark AV_SET for BB as invalid, so this set will be updated the next time
4379 compute_av() processes BB. This function is called when creating new basic
4380 blocks, as well as for blocks (either new or existing) where new jumps are
4381 created when the control flow is being updated. */
4382 static void
4383 invalidate_av_set (basic_block bb)
4384 {
4385 BB_AV_LEVEL (bb) = -1;
4386 }
4387
4388 /* Create initial data sets for BB (they will be invalid). */
4389 static void
4390 create_initial_data_sets (basic_block bb)
4391 {
4392 if (BB_LV_SET (bb))
4393 BB_LV_SET_VALID_P (bb) = false;
4394 else
4395 BB_LV_SET (bb) = get_regset_from_pool ();
4396 invalidate_av_set (bb);
4397 }
4398
4399 /* Free av set of BB. */
4400 static void
4401 free_av_set (basic_block bb)
4402 {
4403 av_set_clear (&BB_AV_SET (bb));
4404 BB_AV_LEVEL (bb) = 0;
4405 }
4406
4407 /* Free data sets of BB. */
4408 void
4409 free_data_sets (basic_block bb)
4410 {
4411 free_lv_set (bb);
4412 free_av_set (bb);
4413 }
4414
4415 /* Exchange lv sets of TO and FROM. */
4416 static void
4417 exchange_lv_sets (basic_block to, basic_block from)
4418 {
4419 {
4420 regset to_lv_set = BB_LV_SET (to);
4421
4422 BB_LV_SET (to) = BB_LV_SET (from);
4423 BB_LV_SET (from) = to_lv_set;
4424 }
4425
4426 {
4427 bool to_lv_set_valid_p = BB_LV_SET_VALID_P (to);
4428
4429 BB_LV_SET_VALID_P (to) = BB_LV_SET_VALID_P (from);
4430 BB_LV_SET_VALID_P (from) = to_lv_set_valid_p;
4431 }
4432 }
4433
4434
4435 /* Exchange av sets of TO and FROM. */
4436 static void
4437 exchange_av_sets (basic_block to, basic_block from)
4438 {
4439 {
4440 av_set_t to_av_set = BB_AV_SET (to);
4441
4442 BB_AV_SET (to) = BB_AV_SET (from);
4443 BB_AV_SET (from) = to_av_set;
4444 }
4445
4446 {
4447 int to_av_level = BB_AV_LEVEL (to);
4448
4449 BB_AV_LEVEL (to) = BB_AV_LEVEL (from);
4450 BB_AV_LEVEL (from) = to_av_level;
4451 }
4452 }
4453
4454 /* Exchange data sets of TO and FROM. */
4455 void
4456 exchange_data_sets (basic_block to, basic_block from)
4457 {
4458 exchange_lv_sets (to, from);
4459 exchange_av_sets (to, from);
4460 }
4461
4462 /* Copy data sets of FROM to TO. */
4463 void
4464 copy_data_sets (basic_block to, basic_block from)
4465 {
4466 gcc_assert (!BB_LV_SET_VALID_P (to) && !BB_AV_SET_VALID_P (to));
4467 gcc_assert (BB_AV_SET (to) == NULL);
4468
4469 BB_AV_LEVEL (to) = BB_AV_LEVEL (from);
4470 BB_LV_SET_VALID_P (to) = BB_LV_SET_VALID_P (from);
4471
4472 if (BB_AV_SET_VALID_P (from))
4473 {
4474 BB_AV_SET (to) = av_set_copy (BB_AV_SET (from));
4475 }
4476 if (BB_LV_SET_VALID_P (from))
4477 {
4478 gcc_assert (BB_LV_SET (to) != NULL);
4479 COPY_REG_SET (BB_LV_SET (to), BB_LV_SET (from));
4480 }
4481 }
4482
4483 /* Return an av set for INSN, if any. */
4484 av_set_t
4485 get_av_set (insn_t insn)
4486 {
4487 av_set_t av_set;
4488
4489 gcc_assert (AV_SET_VALID_P (insn));
4490
4491 if (sel_bb_head_p (insn))
4492 av_set = BB_AV_SET (BLOCK_FOR_INSN (insn));
4493 else
4494 av_set = NULL;
4495
4496 return av_set;
4497 }
4498
4499 /* Implementation of AV_LEVEL () macro. Return AV_LEVEL () of INSN. */
4500 int
4501 get_av_level (insn_t insn)
4502 {
4503 int av_level;
4504
4505 gcc_assert (INSN_P (insn));
4506
4507 if (sel_bb_head_p (insn))
4508 av_level = BB_AV_LEVEL (BLOCK_FOR_INSN (insn));
4509 else
4510 av_level = INSN_WS_LEVEL (insn);
4511
4512 return av_level;
4513 }
4514
4515 \f
4516
4517 /* Variables to work with control-flow graph. */
4518
4519 /* The basic block that already has been processed by the sched_data_update (),
4520 but hasn't been in sel_add_bb () yet. */
4521 static vec<basic_block>
4522 last_added_blocks = vNULL;
4523
4524 /* A pool for allocating successor infos. */
4525 static struct
4526 {
4527 /* A stack for saving succs_info structures. */
4528 struct succs_info *stack;
4529
4530 /* Its size. */
4531 int size;
4532
4533 /* Top of the stack. */
4534 int top;
4535
4536 /* Maximal value of the top. */
4537 int max_top;
4538 } succs_info_pool;
4539
4540 /* Functions to work with control-flow graph. */
4541
4542 /* Return basic block note of BB. */
4543 rtx_insn *
4544 sel_bb_head (basic_block bb)
4545 {
4546 rtx_insn *head;
4547
4548 if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
4549 {
4550 gcc_assert (exit_insn != NULL_RTX);
4551 head = exit_insn;
4552 }
4553 else
4554 {
4555 insn_t note;
4556
4557 note = bb_note (bb);
4558 head = next_nonnote_insn (note);
4559
4560 if (head && (BARRIER_P (head) || BLOCK_FOR_INSN (head) != bb))
4561 head = NULL;
4562 }
4563
4564 return head;
4565 }
4566
4567 /* Return true if INSN is a basic block header. */
4568 bool
4569 sel_bb_head_p (insn_t insn)
4570 {
4571 return sel_bb_head (BLOCK_FOR_INSN (insn)) == insn;
4572 }
4573
4574 /* Return last insn of BB. */
4575 rtx_insn *
4576 sel_bb_end (basic_block bb)
4577 {
4578 if (sel_bb_empty_p (bb))
4579 return NULL;
4580
4581 gcc_assert (bb != EXIT_BLOCK_PTR_FOR_FN (cfun));
4582
4583 return BB_END (bb);
4584 }
4585
4586 /* Return true if INSN is the last insn in its basic block. */
4587 bool
4588 sel_bb_end_p (insn_t insn)
4589 {
4590 return insn == sel_bb_end (BLOCK_FOR_INSN (insn));
4591 }
4592
4593 /* Return true if BB consist of single NOTE_INSN_BASIC_BLOCK. */
4594 bool
4595 sel_bb_empty_p (basic_block bb)
4596 {
4597 return sel_bb_head (bb) == NULL;
4598 }
4599
4600 /* True when BB belongs to the current scheduling region. */
4601 bool
4602 in_current_region_p (basic_block bb)
4603 {
4604 if (bb->index < NUM_FIXED_BLOCKS)
4605 return false;
4606
4607 return CONTAINING_RGN (bb->index) == CONTAINING_RGN (BB_TO_BLOCK (0));
4608 }
4609
4610 /* Return the block which is a fallthru bb of a conditional jump JUMP. */
4611 basic_block
4612 fallthru_bb_of_jump (const rtx_insn *jump)
4613 {
4614 if (!JUMP_P (jump))
4615 return NULL;
4616
4617 if (!any_condjump_p (jump))
4618 return NULL;
4619
4620 /* A basic block that ends with a conditional jump may still have one successor
4621 (and be followed by a barrier), we are not interested. */
4622 if (single_succ_p (BLOCK_FOR_INSN (jump)))
4623 return NULL;
4624
4625 return FALLTHRU_EDGE (BLOCK_FOR_INSN (jump))->dest;
4626 }
4627
4628 /* Remove all notes from BB. */
4629 static void
4630 init_bb (basic_block bb)
4631 {
4632 remove_notes (bb_note (bb), BB_END (bb));
4633 BB_NOTE_LIST (bb) = note_list;
4634 }
4635
4636 void
4637 sel_init_bbs (bb_vec_t bbs)
4638 {
4639 const struct sched_scan_info_def ssi =
4640 {
4641 extend_bb_info, /* extend_bb */
4642 init_bb, /* init_bb */
4643 NULL, /* extend_insn */
4644 NULL /* init_insn */
4645 };
4646
4647 sched_scan (&ssi, bbs);
4648 }
4649
4650 /* Restore notes for the whole region. */
4651 static void
4652 sel_restore_notes (void)
4653 {
4654 int bb;
4655 insn_t insn;
4656
4657 for (bb = 0; bb < current_nr_blocks; bb++)
4658 {
4659 basic_block first, last;
4660
4661 first = EBB_FIRST_BB (bb);
4662 last = EBB_LAST_BB (bb)->next_bb;
4663
4664 do
4665 {
4666 note_list = BB_NOTE_LIST (first);
4667 restore_other_notes (NULL, first);
4668 BB_NOTE_LIST (first) = NULL;
4669
4670 FOR_BB_INSNS (first, insn)
4671 if (NONDEBUG_INSN_P (insn))
4672 reemit_notes (insn);
4673
4674 first = first->next_bb;
4675 }
4676 while (first != last);
4677 }
4678 }
4679
4680 /* Free per-bb data structures. */
4681 void
4682 sel_finish_bbs (void)
4683 {
4684 sel_restore_notes ();
4685
4686 /* Remove current loop preheader from this loop. */
4687 if (current_loop_nest)
4688 sel_remove_loop_preheader ();
4689
4690 finish_region_bb_info ();
4691 }
4692
4693 /* Return true if INSN has a single successor of type FLAGS. */
4694 bool
4695 sel_insn_has_single_succ_p (insn_t insn, int flags)
4696 {
4697 insn_t succ;
4698 succ_iterator si;
4699 bool first_p = true;
4700
4701 FOR_EACH_SUCC_1 (succ, si, insn, flags)
4702 {
4703 if (first_p)
4704 first_p = false;
4705 else
4706 return false;
4707 }
4708
4709 return true;
4710 }
4711
4712 /* Allocate successor's info. */
4713 static struct succs_info *
4714 alloc_succs_info (void)
4715 {
4716 if (succs_info_pool.top == succs_info_pool.max_top)
4717 {
4718 int i;
4719
4720 if (++succs_info_pool.max_top >= succs_info_pool.size)
4721 gcc_unreachable ();
4722
4723 i = ++succs_info_pool.top;
4724 succs_info_pool.stack[i].succs_ok.create (10);
4725 succs_info_pool.stack[i].succs_other.create (10);
4726 succs_info_pool.stack[i].probs_ok.create (10);
4727 }
4728 else
4729 succs_info_pool.top++;
4730
4731 return &succs_info_pool.stack[succs_info_pool.top];
4732 }
4733
4734 /* Free successor's info. */
4735 void
4736 free_succs_info (struct succs_info * sinfo)
4737 {
4738 gcc_assert (succs_info_pool.top >= 0
4739 && &succs_info_pool.stack[succs_info_pool.top] == sinfo);
4740 succs_info_pool.top--;
4741
4742 /* Clear stale info. */
4743 sinfo->succs_ok.block_remove (0, sinfo->succs_ok.length ());
4744 sinfo->succs_other.block_remove (0, sinfo->succs_other.length ());
4745 sinfo->probs_ok.block_remove (0, sinfo->probs_ok.length ());
4746 sinfo->all_prob = 0;
4747 sinfo->succs_ok_n = 0;
4748 sinfo->all_succs_n = 0;
4749 }
4750
4751 /* Compute successor info for INSN. FLAGS are the flags passed
4752 to the FOR_EACH_SUCC_1 iterator. */
4753 struct succs_info *
4754 compute_succs_info (insn_t insn, short flags)
4755 {
4756 succ_iterator si;
4757 insn_t succ;
4758 struct succs_info *sinfo = alloc_succs_info ();
4759
4760 /* Traverse *all* successors and decide what to do with each. */
4761 FOR_EACH_SUCC_1 (succ, si, insn, SUCCS_ALL)
4762 {
4763 /* FIXME: this doesn't work for skipping to loop exits, as we don't
4764 perform code motion through inner loops. */
4765 short current_flags = si.current_flags & ~SUCCS_SKIP_TO_LOOP_EXITS;
4766
4767 if (current_flags & flags)
4768 {
4769 sinfo->succs_ok.safe_push (succ);
4770 sinfo->probs_ok.safe_push (
4771 /* FIXME: Improve calculation when skipping
4772 inner loop to exits. */
4773 si.bb_end ? si.e1->probability : REG_BR_PROB_BASE);
4774 sinfo->succs_ok_n++;
4775 }
4776 else
4777 sinfo->succs_other.safe_push (succ);
4778
4779 /* Compute all_prob. */
4780 if (!si.bb_end)
4781 sinfo->all_prob = REG_BR_PROB_BASE;
4782 else
4783 sinfo->all_prob += si.e1->probability;
4784
4785 sinfo->all_succs_n++;
4786 }
4787
4788 return sinfo;
4789 }
4790
4791 /* Return the predecessors of BB in PREDS and their number in N.
4792 Empty blocks are skipped. SIZE is used to allocate PREDS. */
4793 static void
4794 cfg_preds_1 (basic_block bb, insn_t **preds, int *n, int *size)
4795 {
4796 edge e;
4797 edge_iterator ei;
4798
4799 gcc_assert (BLOCK_TO_BB (bb->index) != 0);
4800
4801 FOR_EACH_EDGE (e, ei, bb->preds)
4802 {
4803 basic_block pred_bb = e->src;
4804 insn_t bb_end = BB_END (pred_bb);
4805
4806 if (!in_current_region_p (pred_bb))
4807 {
4808 gcc_assert (flag_sel_sched_pipelining_outer_loops
4809 && current_loop_nest);
4810 continue;
4811 }
4812
4813 if (sel_bb_empty_p (pred_bb))
4814 cfg_preds_1 (pred_bb, preds, n, size);
4815 else
4816 {
4817 if (*n == *size)
4818 *preds = XRESIZEVEC (insn_t, *preds,
4819 (*size = 2 * *size + 1));
4820 (*preds)[(*n)++] = bb_end;
4821 }
4822 }
4823
4824 gcc_assert (*n != 0
4825 || (flag_sel_sched_pipelining_outer_loops
4826 && current_loop_nest));
4827 }
4828
4829 /* Find all predecessors of BB and record them in PREDS and their number
4830 in N. Empty blocks are skipped, and only normal (forward in-region)
4831 edges are processed. */
4832 static void
4833 cfg_preds (basic_block bb, insn_t **preds, int *n)
4834 {
4835 int size = 0;
4836
4837 *preds = NULL;
4838 *n = 0;
4839 cfg_preds_1 (bb, preds, n, &size);
4840 }
4841
4842 /* Returns true if we are moving INSN through join point. */
4843 bool
4844 sel_num_cfg_preds_gt_1 (insn_t insn)
4845 {
4846 basic_block bb;
4847
4848 if (!sel_bb_head_p (insn) || INSN_BB (insn) == 0)
4849 return false;
4850
4851 bb = BLOCK_FOR_INSN (insn);
4852
4853 while (1)
4854 {
4855 if (EDGE_COUNT (bb->preds) > 1)
4856 return true;
4857
4858 gcc_assert (EDGE_PRED (bb, 0)->dest == bb);
4859 bb = EDGE_PRED (bb, 0)->src;
4860
4861 if (!sel_bb_empty_p (bb))
4862 break;
4863 }
4864
4865 return false;
4866 }
4867
4868 /* Returns true when BB should be the end of an ebb. Adapted from the
4869 code in sched-ebb.c. */
4870 bool
4871 bb_ends_ebb_p (basic_block bb)
4872 {
4873 basic_block next_bb = bb_next_bb (bb);
4874 edge e;
4875
4876 if (next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
4877 || bitmap_bit_p (forced_ebb_heads, next_bb->index)
4878 || (LABEL_P (BB_HEAD (next_bb))
4879 /* NB: LABEL_NUSES () is not maintained outside of jump.c.
4880 Work around that. */
4881 && !single_pred_p (next_bb)))
4882 return true;
4883
4884 if (!in_current_region_p (next_bb))
4885 return true;
4886
4887 e = find_fallthru_edge (bb->succs);
4888 if (e)
4889 {
4890 gcc_assert (e->dest == next_bb);
4891
4892 return false;
4893 }
4894
4895 return true;
4896 }
4897
4898 /* Returns true when INSN and SUCC are in the same EBB, given that SUCC is a
4899 successor of INSN. */
4900 bool
4901 in_same_ebb_p (insn_t insn, insn_t succ)
4902 {
4903 basic_block ptr = BLOCK_FOR_INSN (insn);
4904
4905 for (;;)
4906 {
4907 if (ptr == BLOCK_FOR_INSN (succ))
4908 return true;
4909
4910 if (bb_ends_ebb_p (ptr))
4911 return false;
4912
4913 ptr = bb_next_bb (ptr);
4914 }
4915
4916 gcc_unreachable ();
4917 return false;
4918 }
4919
4920 /* Recomputes the reverse topological order for the function and
4921 saves it in REV_TOP_ORDER_INDEX. REV_TOP_ORDER_INDEX_LEN is also
4922 modified appropriately. */
4923 static void
4924 recompute_rev_top_order (void)
4925 {
4926 int *postorder;
4927 int n_blocks, i;
4928
4929 if (!rev_top_order_index
4930 || rev_top_order_index_len < last_basic_block_for_fn (cfun))
4931 {
4932 rev_top_order_index_len = last_basic_block_for_fn (cfun);
4933 rev_top_order_index = XRESIZEVEC (int, rev_top_order_index,
4934 rev_top_order_index_len);
4935 }
4936
4937 postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
4938
4939 n_blocks = post_order_compute (postorder, true, false);
4940 gcc_assert (n_basic_blocks_for_fn (cfun) == n_blocks);
4941
4942 /* Build reverse function: for each basic block with BB->INDEX == K
4943 rev_top_order_index[K] is it's reverse topological sort number. */
4944 for (i = 0; i < n_blocks; i++)
4945 {
4946 gcc_assert (postorder[i] < rev_top_order_index_len);
4947 rev_top_order_index[postorder[i]] = i;
4948 }
4949
4950 free (postorder);
4951 }
4952
4953 /* Clear all flags from insns in BB that could spoil its rescheduling. */
4954 void
4955 clear_outdated_rtx_info (basic_block bb)
4956 {
4957 rtx_insn *insn;
4958
4959 FOR_BB_INSNS (bb, insn)
4960 if (INSN_P (insn))
4961 {
4962 SCHED_GROUP_P (insn) = 0;
4963 INSN_AFTER_STALL_P (insn) = 0;
4964 INSN_SCHED_TIMES (insn) = 0;
4965 EXPR_PRIORITY_ADJ (INSN_EXPR (insn)) = 0;
4966
4967 /* We cannot use the changed caches, as previously we could ignore
4968 the LHS dependence due to enabled renaming and transform
4969 the expression, and currently we'll be unable to do this. */
4970 htab_empty (INSN_TRANSFORMED_INSNS (insn));
4971 }
4972 }
4973
4974 /* Add BB_NOTE to the pool of available basic block notes. */
4975 static void
4976 return_bb_to_pool (basic_block bb)
4977 {
4978 rtx note = bb_note (bb);
4979
4980 gcc_assert (NOTE_BASIC_BLOCK (note) == bb
4981 && bb->aux == NULL);
4982
4983 /* It turns out that current cfg infrastructure does not support
4984 reuse of basic blocks. Don't bother for now. */
4985 /*bb_note_pool.safe_push (note);*/
4986 }
4987
4988 /* Get a bb_note from pool or return NULL_RTX if pool is empty. */
4989 static rtx_note *
4990 get_bb_note_from_pool (void)
4991 {
4992 if (bb_note_pool.is_empty ())
4993 return NULL;
4994 else
4995 {
4996 rtx_note *note = bb_note_pool.pop ();
4997
4998 SET_PREV_INSN (note) = NULL_RTX;
4999 SET_NEXT_INSN (note) = NULL_RTX;
5000
5001 return note;
5002 }
5003 }
5004
5005 /* Free bb_note_pool. */
5006 void
5007 free_bb_note_pool (void)
5008 {
5009 bb_note_pool.release ();
5010 }
5011
5012 /* Setup scheduler pool and successor structure. */
5013 void
5014 alloc_sched_pools (void)
5015 {
5016 int succs_size;
5017
5018 succs_size = MAX_WS + 1;
5019 succs_info_pool.stack = XCNEWVEC (struct succs_info, succs_size);
5020 succs_info_pool.size = succs_size;
5021 succs_info_pool.top = -1;
5022 succs_info_pool.max_top = -1;
5023
5024 sched_lists_pool = create_alloc_pool ("sel-sched-lists",
5025 sizeof (struct _list_node), 500);
5026 }
5027
5028 /* Free the pools. */
5029 void
5030 free_sched_pools (void)
5031 {
5032 int i;
5033
5034 free_alloc_pool (sched_lists_pool);
5035 gcc_assert (succs_info_pool.top == -1);
5036 for (i = 0; i <= succs_info_pool.max_top; i++)
5037 {
5038 succs_info_pool.stack[i].succs_ok.release ();
5039 succs_info_pool.stack[i].succs_other.release ();
5040 succs_info_pool.stack[i].probs_ok.release ();
5041 }
5042 free (succs_info_pool.stack);
5043 }
5044 \f
5045
5046 /* Returns a position in RGN where BB can be inserted retaining
5047 topological order. */
5048 static int
5049 find_place_to_insert_bb (basic_block bb, int rgn)
5050 {
5051 bool has_preds_outside_rgn = false;
5052 edge e;
5053 edge_iterator ei;
5054
5055 /* Find whether we have preds outside the region. */
5056 FOR_EACH_EDGE (e, ei, bb->preds)
5057 if (!in_current_region_p (e->src))
5058 {
5059 has_preds_outside_rgn = true;
5060 break;
5061 }
5062
5063 /* Recompute the top order -- needed when we have > 1 pred
5064 and in case we don't have preds outside. */
5065 if (flag_sel_sched_pipelining_outer_loops
5066 && (has_preds_outside_rgn || EDGE_COUNT (bb->preds) > 1))
5067 {
5068 int i, bbi = bb->index, cur_bbi;
5069
5070 recompute_rev_top_order ();
5071 for (i = RGN_NR_BLOCKS (rgn) - 1; i >= 0; i--)
5072 {
5073 cur_bbi = BB_TO_BLOCK (i);
5074 if (rev_top_order_index[bbi]
5075 < rev_top_order_index[cur_bbi])
5076 break;
5077 }
5078
5079 /* We skipped the right block, so we increase i. We accommodate
5080 it for increasing by step later, so we decrease i. */
5081 return (i + 1) - 1;
5082 }
5083 else if (has_preds_outside_rgn)
5084 {
5085 /* This is the case when we generate an extra empty block
5086 to serve as region head during pipelining. */
5087 e = EDGE_SUCC (bb, 0);
5088 gcc_assert (EDGE_COUNT (bb->succs) == 1
5089 && in_current_region_p (EDGE_SUCC (bb, 0)->dest)
5090 && (BLOCK_TO_BB (e->dest->index) == 0));
5091 return -1;
5092 }
5093
5094 /* We don't have preds outside the region. We should have
5095 the only pred, because the multiple preds case comes from
5096 the pipelining of outer loops, and that is handled above.
5097 Just take the bbi of this single pred. */
5098 if (EDGE_COUNT (bb->succs) > 0)
5099 {
5100 int pred_bbi;
5101
5102 gcc_assert (EDGE_COUNT (bb->preds) == 1);
5103
5104 pred_bbi = EDGE_PRED (bb, 0)->src->index;
5105 return BLOCK_TO_BB (pred_bbi);
5106 }
5107 else
5108 /* BB has no successors. It is safe to put it in the end. */
5109 return current_nr_blocks - 1;
5110 }
5111
5112 /* Deletes an empty basic block freeing its data. */
5113 static void
5114 delete_and_free_basic_block (basic_block bb)
5115 {
5116 gcc_assert (sel_bb_empty_p (bb));
5117
5118 if (BB_LV_SET (bb))
5119 free_lv_set (bb);
5120
5121 bitmap_clear_bit (blocks_to_reschedule, bb->index);
5122
5123 /* Can't assert av_set properties because we use sel_aremove_bb
5124 when removing loop preheader from the region. At the point of
5125 removing the preheader we already have deallocated sel_region_bb_info. */
5126 gcc_assert (BB_LV_SET (bb) == NULL
5127 && !BB_LV_SET_VALID_P (bb)
5128 && BB_AV_LEVEL (bb) == 0
5129 && BB_AV_SET (bb) == NULL);
5130
5131 delete_basic_block (bb);
5132 }
5133
5134 /* Add BB to the current region and update the region data. */
5135 static void
5136 add_block_to_current_region (basic_block bb)
5137 {
5138 int i, pos, bbi = -2, rgn;
5139
5140 rgn = CONTAINING_RGN (BB_TO_BLOCK (0));
5141 bbi = find_place_to_insert_bb (bb, rgn);
5142 bbi += 1;
5143 pos = RGN_BLOCKS (rgn) + bbi;
5144
5145 gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0
5146 && ebb_head[bbi] == pos);
5147
5148 /* Make a place for the new block. */
5149 extend_regions ();
5150
5151 for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--)
5152 BLOCK_TO_BB (rgn_bb_table[i])++;
5153
5154 memmove (rgn_bb_table + pos + 1,
5155 rgn_bb_table + pos,
5156 (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table));
5157
5158 /* Initialize data for BB. */
5159 rgn_bb_table[pos] = bb->index;
5160 BLOCK_TO_BB (bb->index) = bbi;
5161 CONTAINING_RGN (bb->index) = rgn;
5162
5163 RGN_NR_BLOCKS (rgn)++;
5164
5165 for (i = rgn + 1; i <= nr_regions; i++)
5166 RGN_BLOCKS (i)++;
5167 }
5168
5169 /* Remove BB from the current region and update the region data. */
5170 static void
5171 remove_bb_from_region (basic_block bb)
5172 {
5173 int i, pos, bbi = -2, rgn;
5174
5175 rgn = CONTAINING_RGN (BB_TO_BLOCK (0));
5176 bbi = BLOCK_TO_BB (bb->index);
5177 pos = RGN_BLOCKS (rgn) + bbi;
5178
5179 gcc_assert (RGN_HAS_REAL_EBB (rgn) == 0
5180 && ebb_head[bbi] == pos);
5181
5182 for (i = RGN_BLOCKS (rgn + 1) - 1; i >= pos; i--)
5183 BLOCK_TO_BB (rgn_bb_table[i])--;
5184
5185 memmove (rgn_bb_table + pos,
5186 rgn_bb_table + pos + 1,
5187 (RGN_BLOCKS (nr_regions) - pos) * sizeof (*rgn_bb_table));
5188
5189 RGN_NR_BLOCKS (rgn)--;
5190 for (i = rgn + 1; i <= nr_regions; i++)
5191 RGN_BLOCKS (i)--;
5192 }
5193
5194 /* Add BB to the current region and update all data. If BB is NULL, add all
5195 blocks from last_added_blocks vector. */
5196 static void
5197 sel_add_bb (basic_block bb)
5198 {
5199 /* Extend luids so that new notes will receive zero luids. */
5200 sched_extend_luids ();
5201 sched_init_bbs ();
5202 sel_init_bbs (last_added_blocks);
5203
5204 /* When bb is passed explicitly, the vector should contain
5205 the only element that equals to bb; otherwise, the vector
5206 should not be NULL. */
5207 gcc_assert (last_added_blocks.exists ());
5208
5209 if (bb != NULL)
5210 {
5211 gcc_assert (last_added_blocks.length () == 1
5212 && last_added_blocks[0] == bb);
5213 add_block_to_current_region (bb);
5214
5215 /* We associate creating/deleting data sets with the first insn
5216 appearing / disappearing in the bb. */
5217 if (!sel_bb_empty_p (bb) && BB_LV_SET (bb) == NULL)
5218 create_initial_data_sets (bb);
5219
5220 last_added_blocks.release ();
5221 }
5222 else
5223 /* BB is NULL - process LAST_ADDED_BLOCKS instead. */
5224 {
5225 int i;
5226 basic_block temp_bb = NULL;
5227
5228 for (i = 0;
5229 last_added_blocks.iterate (i, &bb); i++)
5230 {
5231 add_block_to_current_region (bb);
5232 temp_bb = bb;
5233 }
5234
5235 /* We need to fetch at least one bb so we know the region
5236 to update. */
5237 gcc_assert (temp_bb != NULL);
5238 bb = temp_bb;
5239
5240 last_added_blocks.release ();
5241 }
5242
5243 rgn_setup_region (CONTAINING_RGN (bb->index));
5244 }
5245
5246 /* Remove BB from the current region and update all data.
5247 If REMOVE_FROM_CFG_PBB is true, also remove the block cfom cfg. */
5248 static void
5249 sel_remove_bb (basic_block bb, bool remove_from_cfg_p)
5250 {
5251 unsigned idx = bb->index;
5252
5253 gcc_assert (bb != NULL && BB_NOTE_LIST (bb) == NULL_RTX);
5254
5255 remove_bb_from_region (bb);
5256 return_bb_to_pool (bb);
5257 bitmap_clear_bit (blocks_to_reschedule, idx);
5258
5259 if (remove_from_cfg_p)
5260 {
5261 basic_block succ = single_succ (bb);
5262 delete_and_free_basic_block (bb);
5263 set_immediate_dominator (CDI_DOMINATORS, succ,
5264 recompute_dominator (CDI_DOMINATORS, succ));
5265 }
5266
5267 rgn_setup_region (CONTAINING_RGN (idx));
5268 }
5269
5270 /* Concatenate info of EMPTY_BB to info of MERGE_BB. */
5271 static void
5272 move_bb_info (basic_block merge_bb, basic_block empty_bb)
5273 {
5274 if (in_current_region_p (merge_bb))
5275 concat_note_lists (BB_NOTE_LIST (empty_bb),
5276 &BB_NOTE_LIST (merge_bb));
5277 BB_NOTE_LIST (empty_bb) = NULL;
5278
5279 }
5280
5281 /* Remove EMPTY_BB. If REMOVE_FROM_CFG_P is false, remove EMPTY_BB from
5282 region, but keep it in CFG. */
5283 static void
5284 remove_empty_bb (basic_block empty_bb, bool remove_from_cfg_p)
5285 {
5286 /* The block should contain just a note or a label.
5287 We try to check whether it is unused below. */
5288 gcc_assert (BB_HEAD (empty_bb) == BB_END (empty_bb)
5289 || LABEL_P (BB_HEAD (empty_bb)));
5290
5291 /* If basic block has predecessors or successors, redirect them. */
5292 if (remove_from_cfg_p
5293 && (EDGE_COUNT (empty_bb->preds) > 0
5294 || EDGE_COUNT (empty_bb->succs) > 0))
5295 {
5296 basic_block pred;
5297 basic_block succ;
5298
5299 /* We need to init PRED and SUCC before redirecting edges. */
5300 if (EDGE_COUNT (empty_bb->preds) > 0)
5301 {
5302 edge e;
5303
5304 gcc_assert (EDGE_COUNT (empty_bb->preds) == 1);
5305
5306 e = EDGE_PRED (empty_bb, 0);
5307 gcc_assert (e->src == empty_bb->prev_bb
5308 && (e->flags & EDGE_FALLTHRU));
5309
5310 pred = empty_bb->prev_bb;
5311 }
5312 else
5313 pred = NULL;
5314
5315 if (EDGE_COUNT (empty_bb->succs) > 0)
5316 {
5317 /* We do not check fallthruness here as above, because
5318 after removing a jump the edge may actually be not fallthru. */
5319 gcc_assert (EDGE_COUNT (empty_bb->succs) == 1);
5320 succ = EDGE_SUCC (empty_bb, 0)->dest;
5321 }
5322 else
5323 succ = NULL;
5324
5325 if (EDGE_COUNT (empty_bb->preds) > 0 && succ != NULL)
5326 {
5327 edge e = EDGE_PRED (empty_bb, 0);
5328
5329 if (e->flags & EDGE_FALLTHRU)
5330 redirect_edge_succ_nodup (e, succ);
5331 else
5332 sel_redirect_edge_and_branch (EDGE_PRED (empty_bb, 0), succ);
5333 }
5334
5335 if (EDGE_COUNT (empty_bb->succs) > 0 && pred != NULL)
5336 {
5337 edge e = EDGE_SUCC (empty_bb, 0);
5338
5339 if (find_edge (pred, e->dest) == NULL)
5340 redirect_edge_pred (e, pred);
5341 }
5342 }
5343
5344 /* Finish removing. */
5345 sel_remove_bb (empty_bb, remove_from_cfg_p);
5346 }
5347
5348 /* An implementation of create_basic_block hook, which additionally updates
5349 per-bb data structures. */
5350 static basic_block
5351 sel_create_basic_block (void *headp, void *endp, basic_block after)
5352 {
5353 basic_block new_bb;
5354 rtx_note *new_bb_note;
5355
5356 gcc_assert (flag_sel_sched_pipelining_outer_loops
5357 || !last_added_blocks.exists ());
5358
5359 new_bb_note = get_bb_note_from_pool ();
5360
5361 if (new_bb_note == NULL_RTX)
5362 new_bb = orig_cfg_hooks.create_basic_block (headp, endp, after);
5363 else
5364 {
5365 new_bb = create_basic_block_structure ((rtx_insn *) headp,
5366 (rtx_insn *) endp,
5367 new_bb_note, after);
5368 new_bb->aux = NULL;
5369 }
5370
5371 last_added_blocks.safe_push (new_bb);
5372
5373 return new_bb;
5374 }
5375
5376 /* Implement sched_init_only_bb (). */
5377 static void
5378 sel_init_only_bb (basic_block bb, basic_block after)
5379 {
5380 gcc_assert (after == NULL);
5381
5382 extend_regions ();
5383 rgn_make_new_region_out_of_new_block (bb);
5384 }
5385
5386 /* Update the latch when we've splitted or merged it from FROM block to TO.
5387 This should be checked for all outer loops, too. */
5388 static void
5389 change_loops_latches (basic_block from, basic_block to)
5390 {
5391 gcc_assert (from != to);
5392
5393 if (current_loop_nest)
5394 {
5395 struct loop *loop;
5396
5397 for (loop = current_loop_nest; loop; loop = loop_outer (loop))
5398 if (considered_for_pipelining_p (loop) && loop->latch == from)
5399 {
5400 gcc_assert (loop == current_loop_nest);
5401 loop->latch = to;
5402 gcc_assert (loop_latch_edge (loop));
5403 }
5404 }
5405 }
5406
5407 /* Splits BB on two basic blocks, adding it to the region and extending
5408 per-bb data structures. Returns the newly created bb. */
5409 static basic_block
5410 sel_split_block (basic_block bb, rtx after)
5411 {
5412 basic_block new_bb;
5413 insn_t insn;
5414
5415 new_bb = sched_split_block_1 (bb, after);
5416 sel_add_bb (new_bb);
5417
5418 /* This should be called after sel_add_bb, because this uses
5419 CONTAINING_RGN for the new block, which is not yet initialized.
5420 FIXME: this function may be a no-op now. */
5421 change_loops_latches (bb, new_bb);
5422
5423 /* Update ORIG_BB_INDEX for insns moved into the new block. */
5424 FOR_BB_INSNS (new_bb, insn)
5425 if (INSN_P (insn))
5426 EXPR_ORIG_BB_INDEX (INSN_EXPR (insn)) = new_bb->index;
5427
5428 if (sel_bb_empty_p (bb))
5429 {
5430 gcc_assert (!sel_bb_empty_p (new_bb));
5431
5432 /* NEW_BB has data sets that need to be updated and BB holds
5433 data sets that should be removed. Exchange these data sets
5434 so that we won't lose BB's valid data sets. */
5435 exchange_data_sets (new_bb, bb);
5436 free_data_sets (bb);
5437 }
5438
5439 if (!sel_bb_empty_p (new_bb)
5440 && bitmap_bit_p (blocks_to_reschedule, bb->index))
5441 bitmap_set_bit (blocks_to_reschedule, new_bb->index);
5442
5443 return new_bb;
5444 }
5445
5446 /* If BB ends with a jump insn whose ID is bigger then PREV_MAX_UID, return it.
5447 Otherwise returns NULL. */
5448 static rtx_insn *
5449 check_for_new_jump (basic_block bb, int prev_max_uid)
5450 {
5451 rtx_insn *end;
5452
5453 end = sel_bb_end (bb);
5454 if (end && INSN_UID (end) >= prev_max_uid)
5455 return end;
5456 return NULL;
5457 }
5458
5459 /* Look for a new jump either in FROM_BB block or in newly created JUMP_BB block.
5460 New means having UID at least equal to PREV_MAX_UID. */
5461 static rtx_insn *
5462 find_new_jump (basic_block from, basic_block jump_bb, int prev_max_uid)
5463 {
5464 rtx_insn *jump;
5465
5466 /* Return immediately if no new insns were emitted. */
5467 if (get_max_uid () == prev_max_uid)
5468 return NULL;
5469
5470 /* Now check both blocks for new jumps. It will ever be only one. */
5471 if ((jump = check_for_new_jump (from, prev_max_uid)))
5472 return jump;
5473
5474 if (jump_bb != NULL
5475 && (jump = check_for_new_jump (jump_bb, prev_max_uid)))
5476 return jump;
5477 return NULL;
5478 }
5479
5480 /* Splits E and adds the newly created basic block to the current region.
5481 Returns this basic block. */
5482 basic_block
5483 sel_split_edge (edge e)
5484 {
5485 basic_block new_bb, src, other_bb = NULL;
5486 int prev_max_uid;
5487 rtx_insn *jump;
5488
5489 src = e->src;
5490 prev_max_uid = get_max_uid ();
5491 new_bb = split_edge (e);
5492
5493 if (flag_sel_sched_pipelining_outer_loops
5494 && current_loop_nest)
5495 {
5496 int i;
5497 basic_block bb;
5498
5499 /* Some of the basic blocks might not have been added to the loop.
5500 Add them here, until this is fixed in force_fallthru. */
5501 for (i = 0;
5502 last_added_blocks.iterate (i, &bb); i++)
5503 if (!bb->loop_father)
5504 {
5505 add_bb_to_loop (bb, e->dest->loop_father);
5506
5507 gcc_assert (!other_bb && (new_bb->index != bb->index));
5508 other_bb = bb;
5509 }
5510 }
5511
5512 /* Add all last_added_blocks to the region. */
5513 sel_add_bb (NULL);
5514
5515 jump = find_new_jump (src, new_bb, prev_max_uid);
5516 if (jump)
5517 sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP);
5518
5519 /* Put the correct lv set on this block. */
5520 if (other_bb && !sel_bb_empty_p (other_bb))
5521 compute_live (sel_bb_head (other_bb));
5522
5523 return new_bb;
5524 }
5525
5526 /* Implement sched_create_empty_bb (). */
5527 static basic_block
5528 sel_create_empty_bb (basic_block after)
5529 {
5530 basic_block new_bb;
5531
5532 new_bb = sched_create_empty_bb_1 (after);
5533
5534 /* We'll explicitly initialize NEW_BB via sel_init_only_bb () a bit
5535 later. */
5536 gcc_assert (last_added_blocks.length () == 1
5537 && last_added_blocks[0] == new_bb);
5538
5539 last_added_blocks.release ();
5540 return new_bb;
5541 }
5542
5543 /* Implement sched_create_recovery_block. ORIG_INSN is where block
5544 will be splitted to insert a check. */
5545 basic_block
5546 sel_create_recovery_block (insn_t orig_insn)
5547 {
5548 basic_block first_bb, second_bb, recovery_block;
5549 basic_block before_recovery = NULL;
5550 rtx_insn *jump;
5551
5552 first_bb = BLOCK_FOR_INSN (orig_insn);
5553 if (sel_bb_end_p (orig_insn))
5554 {
5555 /* Avoid introducing an empty block while splitting. */
5556 gcc_assert (single_succ_p (first_bb));
5557 second_bb = single_succ (first_bb);
5558 }
5559 else
5560 second_bb = sched_split_block (first_bb, orig_insn);
5561
5562 recovery_block = sched_create_recovery_block (&before_recovery);
5563 if (before_recovery)
5564 copy_lv_set_from (before_recovery, EXIT_BLOCK_PTR_FOR_FN (cfun));
5565
5566 gcc_assert (sel_bb_empty_p (recovery_block));
5567 sched_create_recovery_edges (first_bb, recovery_block, second_bb);
5568 if (current_loops != NULL)
5569 add_bb_to_loop (recovery_block, first_bb->loop_father);
5570
5571 sel_add_bb (recovery_block);
5572
5573 jump = BB_END (recovery_block);
5574 gcc_assert (sel_bb_head (recovery_block) == jump);
5575 sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP);
5576
5577 return recovery_block;
5578 }
5579
5580 /* Merge basic block B into basic block A. */
5581 static void
5582 sel_merge_blocks (basic_block a, basic_block b)
5583 {
5584 gcc_assert (sel_bb_empty_p (b)
5585 && EDGE_COUNT (b->preds) == 1
5586 && EDGE_PRED (b, 0)->src == b->prev_bb);
5587
5588 move_bb_info (b->prev_bb, b);
5589 remove_empty_bb (b, false);
5590 merge_blocks (a, b);
5591 change_loops_latches (b, a);
5592 }
5593
5594 /* A wrapper for redirect_edge_and_branch_force, which also initializes
5595 data structures for possibly created bb and insns. */
5596 void
5597 sel_redirect_edge_and_branch_force (edge e, basic_block to)
5598 {
5599 basic_block jump_bb, src, orig_dest = e->dest;
5600 int prev_max_uid;
5601 rtx_insn *jump;
5602 int old_seqno = -1;
5603
5604 /* This function is now used only for bookkeeping code creation, where
5605 we'll never get the single pred of orig_dest block and thus will not
5606 hit unreachable blocks when updating dominator info. */
5607 gcc_assert (!sel_bb_empty_p (e->src)
5608 && !single_pred_p (orig_dest));
5609 src = e->src;
5610 prev_max_uid = get_max_uid ();
5611 /* Compute and pass old_seqno down to sel_init_new_insn only for the case
5612 when the conditional jump being redirected may become unconditional. */
5613 if (any_condjump_p (BB_END (src))
5614 && INSN_SEQNO (BB_END (src)) >= 0)
5615 old_seqno = INSN_SEQNO (BB_END (src));
5616
5617 jump_bb = redirect_edge_and_branch_force (e, to);
5618 if (jump_bb != NULL)
5619 sel_add_bb (jump_bb);
5620
5621 /* This function could not be used to spoil the loop structure by now,
5622 thus we don't care to update anything. But check it to be sure. */
5623 if (current_loop_nest
5624 && pipelining_p)
5625 gcc_assert (loop_latch_edge (current_loop_nest));
5626
5627 jump = find_new_jump (src, jump_bb, prev_max_uid);
5628 if (jump)
5629 sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP,
5630 old_seqno);
5631 set_immediate_dominator (CDI_DOMINATORS, to,
5632 recompute_dominator (CDI_DOMINATORS, to));
5633 set_immediate_dominator (CDI_DOMINATORS, orig_dest,
5634 recompute_dominator (CDI_DOMINATORS, orig_dest));
5635 }
5636
5637 /* A wrapper for redirect_edge_and_branch. Return TRUE if blocks connected by
5638 redirected edge are in reverse topological order. */
5639 bool
5640 sel_redirect_edge_and_branch (edge e, basic_block to)
5641 {
5642 bool latch_edge_p;
5643 basic_block src, orig_dest = e->dest;
5644 int prev_max_uid;
5645 rtx_insn *jump;
5646 edge redirected;
5647 bool recompute_toporder_p = false;
5648 bool maybe_unreachable = single_pred_p (orig_dest);
5649 int old_seqno = -1;
5650
5651 latch_edge_p = (pipelining_p
5652 && current_loop_nest
5653 && e == loop_latch_edge (current_loop_nest));
5654
5655 src = e->src;
5656 prev_max_uid = get_max_uid ();
5657
5658 /* Compute and pass old_seqno down to sel_init_new_insn only for the case
5659 when the conditional jump being redirected may become unconditional. */
5660 if (any_condjump_p (BB_END (src))
5661 && INSN_SEQNO (BB_END (src)) >= 0)
5662 old_seqno = INSN_SEQNO (BB_END (src));
5663
5664 redirected = redirect_edge_and_branch (e, to);
5665
5666 gcc_assert (redirected && !last_added_blocks.exists ());
5667
5668 /* When we've redirected a latch edge, update the header. */
5669 if (latch_edge_p)
5670 {
5671 current_loop_nest->header = to;
5672 gcc_assert (loop_latch_edge (current_loop_nest));
5673 }
5674
5675 /* In rare situations, the topological relation between the blocks connected
5676 by the redirected edge can change (see PR42245 for an example). Update
5677 block_to_bb/bb_to_block. */
5678 if (CONTAINING_RGN (e->src->index) == CONTAINING_RGN (to->index)
5679 && BLOCK_TO_BB (e->src->index) > BLOCK_TO_BB (to->index))
5680 recompute_toporder_p = true;
5681
5682 jump = find_new_jump (src, NULL, prev_max_uid);
5683 if (jump)
5684 sel_init_new_insn (jump, INSN_INIT_TODO_LUID | INSN_INIT_TODO_SIMPLEJUMP, old_seqno);
5685
5686 /* Only update dominator info when we don't have unreachable blocks.
5687 Otherwise we'll update in maybe_tidy_empty_bb. */
5688 if (!maybe_unreachable)
5689 {
5690 set_immediate_dominator (CDI_DOMINATORS, to,
5691 recompute_dominator (CDI_DOMINATORS, to));
5692 set_immediate_dominator (CDI_DOMINATORS, orig_dest,
5693 recompute_dominator (CDI_DOMINATORS, orig_dest));
5694 }
5695 return recompute_toporder_p;
5696 }
5697
5698 /* This variable holds the cfg hooks used by the selective scheduler. */
5699 static struct cfg_hooks sel_cfg_hooks;
5700
5701 /* Register sel-sched cfg hooks. */
5702 void
5703 sel_register_cfg_hooks (void)
5704 {
5705 sched_split_block = sel_split_block;
5706
5707 orig_cfg_hooks = get_cfg_hooks ();
5708 sel_cfg_hooks = orig_cfg_hooks;
5709
5710 sel_cfg_hooks.create_basic_block = sel_create_basic_block;
5711
5712 set_cfg_hooks (sel_cfg_hooks);
5713
5714 sched_init_only_bb = sel_init_only_bb;
5715 sched_split_block = sel_split_block;
5716 sched_create_empty_bb = sel_create_empty_bb;
5717 }
5718
5719 /* Unregister sel-sched cfg hooks. */
5720 void
5721 sel_unregister_cfg_hooks (void)
5722 {
5723 sched_create_empty_bb = NULL;
5724 sched_split_block = NULL;
5725 sched_init_only_bb = NULL;
5726
5727 set_cfg_hooks (orig_cfg_hooks);
5728 }
5729 \f
5730
5731 /* Emit an insn rtx based on PATTERN. If a jump insn is wanted,
5732 LABEL is where this jump should be directed. */
5733 rtx_insn *
5734 create_insn_rtx_from_pattern (rtx pattern, rtx label)
5735 {
5736 rtx_insn *insn_rtx;
5737
5738 gcc_assert (!INSN_P (pattern));
5739
5740 start_sequence ();
5741
5742 if (label == NULL_RTX)
5743 insn_rtx = emit_insn (pattern);
5744 else if (DEBUG_INSN_P (label))
5745 insn_rtx = emit_debug_insn (pattern);
5746 else
5747 {
5748 insn_rtx = emit_jump_insn (pattern);
5749 JUMP_LABEL (insn_rtx) = label;
5750 ++LABEL_NUSES (label);
5751 }
5752
5753 end_sequence ();
5754
5755 sched_extend_luids ();
5756 sched_extend_target ();
5757 sched_deps_init (false);
5758
5759 /* Initialize INSN_CODE now. */
5760 recog_memoized (insn_rtx);
5761 return insn_rtx;
5762 }
5763
5764 /* Create a new vinsn for INSN_RTX. FORCE_UNIQUE_P is true when the vinsn
5765 must not be clonable. */
5766 vinsn_t
5767 create_vinsn_from_insn_rtx (rtx_insn *insn_rtx, bool force_unique_p)
5768 {
5769 gcc_assert (INSN_P (insn_rtx) && !INSN_IN_STREAM_P (insn_rtx));
5770
5771 /* If VINSN_TYPE is not USE, retain its uniqueness. */
5772 return vinsn_create (insn_rtx, force_unique_p);
5773 }
5774
5775 /* Create a copy of INSN_RTX. */
5776 rtx_insn *
5777 create_copy_of_insn_rtx (rtx insn_rtx)
5778 {
5779 rtx_insn *res;
5780 rtx link;
5781
5782 if (DEBUG_INSN_P (insn_rtx))
5783 return create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)),
5784 insn_rtx);
5785
5786 gcc_assert (NONJUMP_INSN_P (insn_rtx));
5787
5788 res = create_insn_rtx_from_pattern (copy_rtx (PATTERN (insn_rtx)),
5789 NULL_RTX);
5790
5791 /* Copy all REG_NOTES except REG_EQUAL/REG_EQUIV and REG_LABEL_OPERAND
5792 since mark_jump_label will make them. REG_LABEL_TARGETs are created
5793 there too, but are supposed to be sticky, so we copy them. */
5794 for (link = REG_NOTES (insn_rtx); link; link = XEXP (link, 1))
5795 if (REG_NOTE_KIND (link) != REG_LABEL_OPERAND
5796 && REG_NOTE_KIND (link) != REG_EQUAL
5797 && REG_NOTE_KIND (link) != REG_EQUIV)
5798 {
5799 if (GET_CODE (link) == EXPR_LIST)
5800 add_reg_note (res, REG_NOTE_KIND (link),
5801 copy_insn_1 (XEXP (link, 0)));
5802 else
5803 add_reg_note (res, REG_NOTE_KIND (link), XEXP (link, 0));
5804 }
5805
5806 return res;
5807 }
5808
5809 /* Change vinsn field of EXPR to hold NEW_VINSN. */
5810 void
5811 change_vinsn_in_expr (expr_t expr, vinsn_t new_vinsn)
5812 {
5813 vinsn_detach (EXPR_VINSN (expr));
5814
5815 EXPR_VINSN (expr) = new_vinsn;
5816 vinsn_attach (new_vinsn);
5817 }
5818
5819 /* Helpers for global init. */
5820 /* This structure is used to be able to call existing bundling mechanism
5821 and calculate insn priorities. */
5822 static struct haifa_sched_info sched_sel_haifa_sched_info =
5823 {
5824 NULL, /* init_ready_list */
5825 NULL, /* can_schedule_ready_p */
5826 NULL, /* schedule_more_p */
5827 NULL, /* new_ready */
5828 NULL, /* rgn_rank */
5829 sel_print_insn, /* rgn_print_insn */
5830 contributes_to_priority,
5831 NULL, /* insn_finishes_block_p */
5832
5833 NULL, NULL,
5834 NULL, NULL,
5835 0, 0,
5836
5837 NULL, /* add_remove_insn */
5838 NULL, /* begin_schedule_ready */
5839 NULL, /* begin_move_insn */
5840 NULL, /* advance_target_bb */
5841
5842 NULL,
5843 NULL,
5844
5845 SEL_SCHED | NEW_BBS
5846 };
5847
5848 /* Setup special insns used in the scheduler. */
5849 void
5850 setup_nop_and_exit_insns (void)
5851 {
5852 gcc_assert (nop_pattern == NULL_RTX
5853 && exit_insn == NULL_RTX);
5854
5855 nop_pattern = constm1_rtx;
5856
5857 start_sequence ();
5858 emit_insn (nop_pattern);
5859 exit_insn = get_insns ();
5860 end_sequence ();
5861 set_block_for_insn (exit_insn, EXIT_BLOCK_PTR_FOR_FN (cfun));
5862 }
5863
5864 /* Free special insns used in the scheduler. */
5865 void
5866 free_nop_and_exit_insns (void)
5867 {
5868 exit_insn = NULL;
5869 nop_pattern = NULL_RTX;
5870 }
5871
5872 /* Setup a special vinsn used in new insns initialization. */
5873 void
5874 setup_nop_vinsn (void)
5875 {
5876 nop_vinsn = vinsn_create (exit_insn, false);
5877 vinsn_attach (nop_vinsn);
5878 }
5879
5880 /* Free a special vinsn used in new insns initialization. */
5881 void
5882 free_nop_vinsn (void)
5883 {
5884 gcc_assert (VINSN_COUNT (nop_vinsn) == 1);
5885 vinsn_detach (nop_vinsn);
5886 nop_vinsn = NULL;
5887 }
5888
5889 /* Call a set_sched_flags hook. */
5890 void
5891 sel_set_sched_flags (void)
5892 {
5893 /* ??? This means that set_sched_flags were called, and we decided to
5894 support speculation. However, set_sched_flags also modifies flags
5895 on current_sched_info, doing this only at global init. And we
5896 sometimes change c_s_i later. So put the correct flags again. */
5897 if (spec_info && targetm.sched.set_sched_flags)
5898 targetm.sched.set_sched_flags (spec_info);
5899 }
5900
5901 /* Setup pointers to global sched info structures. */
5902 void
5903 sel_setup_sched_infos (void)
5904 {
5905 rgn_setup_common_sched_info ();
5906
5907 memcpy (&sel_common_sched_info, common_sched_info,
5908 sizeof (sel_common_sched_info));
5909
5910 sel_common_sched_info.fix_recovery_cfg = NULL;
5911 sel_common_sched_info.add_block = NULL;
5912 sel_common_sched_info.estimate_number_of_insns
5913 = sel_estimate_number_of_insns;
5914 sel_common_sched_info.luid_for_non_insn = sel_luid_for_non_insn;
5915 sel_common_sched_info.sched_pass_id = SCHED_SEL_PASS;
5916
5917 common_sched_info = &sel_common_sched_info;
5918
5919 current_sched_info = &sched_sel_haifa_sched_info;
5920 current_sched_info->sched_max_insns_priority =
5921 get_rgn_sched_max_insns_priority ();
5922
5923 sel_set_sched_flags ();
5924 }
5925 \f
5926
5927 /* Adds basic block BB to region RGN at the position *BB_ORD_INDEX,
5928 *BB_ORD_INDEX after that is increased. */
5929 static void
5930 sel_add_block_to_region (basic_block bb, int *bb_ord_index, int rgn)
5931 {
5932 RGN_NR_BLOCKS (rgn) += 1;
5933 RGN_DONT_CALC_DEPS (rgn) = 0;
5934 RGN_HAS_REAL_EBB (rgn) = 0;
5935 CONTAINING_RGN (bb->index) = rgn;
5936 BLOCK_TO_BB (bb->index) = *bb_ord_index;
5937 rgn_bb_table[RGN_BLOCKS (rgn) + *bb_ord_index] = bb->index;
5938 (*bb_ord_index)++;
5939
5940 /* FIXME: it is true only when not scheduling ebbs. */
5941 RGN_BLOCKS (rgn + 1) = RGN_BLOCKS (rgn) + RGN_NR_BLOCKS (rgn);
5942 }
5943
5944 /* Functions to support pipelining of outer loops. */
5945
5946 /* Creates a new empty region and returns it's number. */
5947 static int
5948 sel_create_new_region (void)
5949 {
5950 int new_rgn_number = nr_regions;
5951
5952 RGN_NR_BLOCKS (new_rgn_number) = 0;
5953
5954 /* FIXME: This will work only when EBBs are not created. */
5955 if (new_rgn_number != 0)
5956 RGN_BLOCKS (new_rgn_number) = RGN_BLOCKS (new_rgn_number - 1) +
5957 RGN_NR_BLOCKS (new_rgn_number - 1);
5958 else
5959 RGN_BLOCKS (new_rgn_number) = 0;
5960
5961 /* Set the blocks of the next region so the other functions may
5962 calculate the number of blocks in the region. */
5963 RGN_BLOCKS (new_rgn_number + 1) = RGN_BLOCKS (new_rgn_number) +
5964 RGN_NR_BLOCKS (new_rgn_number);
5965
5966 nr_regions++;
5967
5968 return new_rgn_number;
5969 }
5970
5971 /* If X has a smaller topological sort number than Y, returns -1;
5972 if greater, returns 1. */
5973 static int
5974 bb_top_order_comparator (const void *x, const void *y)
5975 {
5976 basic_block bb1 = *(const basic_block *) x;
5977 basic_block bb2 = *(const basic_block *) y;
5978
5979 gcc_assert (bb1 == bb2
5980 || rev_top_order_index[bb1->index]
5981 != rev_top_order_index[bb2->index]);
5982
5983 /* It's a reverse topological order in REV_TOP_ORDER_INDEX, so
5984 bbs with greater number should go earlier. */
5985 if (rev_top_order_index[bb1->index] > rev_top_order_index[bb2->index])
5986 return -1;
5987 else
5988 return 1;
5989 }
5990
5991 /* Create a region for LOOP and return its number. If we don't want
5992 to pipeline LOOP, return -1. */
5993 static int
5994 make_region_from_loop (struct loop *loop)
5995 {
5996 unsigned int i;
5997 int new_rgn_number = -1;
5998 struct loop *inner;
5999
6000 /* Basic block index, to be assigned to BLOCK_TO_BB. */
6001 int bb_ord_index = 0;
6002 basic_block *loop_blocks;
6003 basic_block preheader_block;
6004
6005 if (loop->num_nodes
6006 > (unsigned) PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_BLOCKS))
6007 return -1;
6008
6009 /* Don't pipeline loops whose latch belongs to some of its inner loops. */
6010 for (inner = loop->inner; inner; inner = inner->inner)
6011 if (flow_bb_inside_loop_p (inner, loop->latch))
6012 return -1;
6013
6014 loop->ninsns = num_loop_insns (loop);
6015 if ((int) loop->ninsns > PARAM_VALUE (PARAM_MAX_PIPELINE_REGION_INSNS))
6016 return -1;
6017
6018 loop_blocks = get_loop_body_in_custom_order (loop, bb_top_order_comparator);
6019
6020 for (i = 0; i < loop->num_nodes; i++)
6021 if (loop_blocks[i]->flags & BB_IRREDUCIBLE_LOOP)
6022 {
6023 free (loop_blocks);
6024 return -1;
6025 }
6026
6027 preheader_block = loop_preheader_edge (loop)->src;
6028 gcc_assert (preheader_block);
6029 gcc_assert (loop_blocks[0] == loop->header);
6030
6031 new_rgn_number = sel_create_new_region ();
6032
6033 sel_add_block_to_region (preheader_block, &bb_ord_index, new_rgn_number);
6034 bitmap_set_bit (bbs_in_loop_rgns, preheader_block->index);
6035
6036 for (i = 0; i < loop->num_nodes; i++)
6037 {
6038 /* Add only those blocks that haven't been scheduled in the inner loop.
6039 The exception is the basic blocks with bookkeeping code - they should
6040 be added to the region (and they actually don't belong to the loop
6041 body, but to the region containing that loop body). */
6042
6043 gcc_assert (new_rgn_number >= 0);
6044
6045 if (! bitmap_bit_p (bbs_in_loop_rgns, loop_blocks[i]->index))
6046 {
6047 sel_add_block_to_region (loop_blocks[i], &bb_ord_index,
6048 new_rgn_number);
6049 bitmap_set_bit (bbs_in_loop_rgns, loop_blocks[i]->index);
6050 }
6051 }
6052
6053 free (loop_blocks);
6054 MARK_LOOP_FOR_PIPELINING (loop);
6055
6056 return new_rgn_number;
6057 }
6058
6059 /* Create a new region from preheader blocks LOOP_BLOCKS. */
6060 void
6061 make_region_from_loop_preheader (vec<basic_block> *&loop_blocks)
6062 {
6063 unsigned int i;
6064 int new_rgn_number = -1;
6065 basic_block bb;
6066
6067 /* Basic block index, to be assigned to BLOCK_TO_BB. */
6068 int bb_ord_index = 0;
6069
6070 new_rgn_number = sel_create_new_region ();
6071
6072 FOR_EACH_VEC_ELT (*loop_blocks, i, bb)
6073 {
6074 gcc_assert (new_rgn_number >= 0);
6075
6076 sel_add_block_to_region (bb, &bb_ord_index, new_rgn_number);
6077 }
6078
6079 vec_free (loop_blocks);
6080 }
6081
6082
6083 /* Create region(s) from loop nest LOOP, such that inner loops will be
6084 pipelined before outer loops. Returns true when a region for LOOP
6085 is created. */
6086 static bool
6087 make_regions_from_loop_nest (struct loop *loop)
6088 {
6089 struct loop *cur_loop;
6090 int rgn_number;
6091
6092 /* Traverse all inner nodes of the loop. */
6093 for (cur_loop = loop->inner; cur_loop; cur_loop = cur_loop->next)
6094 if (! bitmap_bit_p (bbs_in_loop_rgns, cur_loop->header->index))
6095 return false;
6096
6097 /* At this moment all regular inner loops should have been pipelined.
6098 Try to create a region from this loop. */
6099 rgn_number = make_region_from_loop (loop);
6100
6101 if (rgn_number < 0)
6102 return false;
6103
6104 loop_nests.safe_push (loop);
6105 return true;
6106 }
6107
6108 /* Initalize data structures needed. */
6109 void
6110 sel_init_pipelining (void)
6111 {
6112 /* Collect loop information to be used in outer loops pipelining. */
6113 loop_optimizer_init (LOOPS_HAVE_PREHEADERS
6114 | LOOPS_HAVE_FALLTHRU_PREHEADERS
6115 | LOOPS_HAVE_RECORDED_EXITS
6116 | LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS);
6117 current_loop_nest = NULL;
6118
6119 bbs_in_loop_rgns = sbitmap_alloc (last_basic_block_for_fn (cfun));
6120 bitmap_clear (bbs_in_loop_rgns);
6121
6122 recompute_rev_top_order ();
6123 }
6124
6125 /* Returns a struct loop for region RGN. */
6126 loop_p
6127 get_loop_nest_for_rgn (unsigned int rgn)
6128 {
6129 /* Regions created with extend_rgns don't have corresponding loop nests,
6130 because they don't represent loops. */
6131 if (rgn < loop_nests.length ())
6132 return loop_nests[rgn];
6133 else
6134 return NULL;
6135 }
6136
6137 /* True when LOOP was included into pipelining regions. */
6138 bool
6139 considered_for_pipelining_p (struct loop *loop)
6140 {
6141 if (loop_depth (loop) == 0)
6142 return false;
6143
6144 /* Now, the loop could be too large or irreducible. Check whether its
6145 region is in LOOP_NESTS.
6146 We determine the region number of LOOP as the region number of its
6147 latch. We can't use header here, because this header could be
6148 just removed preheader and it will give us the wrong region number.
6149 Latch can't be used because it could be in the inner loop too. */
6150 if (LOOP_MARKED_FOR_PIPELINING_P (loop))
6151 {
6152 int rgn = CONTAINING_RGN (loop->latch->index);
6153
6154 gcc_assert ((unsigned) rgn < loop_nests.length ());
6155 return true;
6156 }
6157
6158 return false;
6159 }
6160
6161 /* Makes regions from the rest of the blocks, after loops are chosen
6162 for pipelining. */
6163 static void
6164 make_regions_from_the_rest (void)
6165 {
6166 int cur_rgn_blocks;
6167 int *loop_hdr;
6168 int i;
6169
6170 basic_block bb;
6171 edge e;
6172 edge_iterator ei;
6173 int *degree;
6174
6175 /* Index in rgn_bb_table where to start allocating new regions. */
6176 cur_rgn_blocks = nr_regions ? RGN_BLOCKS (nr_regions) : 0;
6177
6178 /* Make regions from all the rest basic blocks - those that don't belong to
6179 any loop or belong to irreducible loops. Prepare the data structures
6180 for extend_rgns. */
6181
6182 /* LOOP_HDR[I] == -1 if I-th bb doesn't belong to any loop,
6183 LOOP_HDR[I] == LOOP_HDR[J] iff basic blocks I and J reside within the same
6184 loop. */
6185 loop_hdr = XNEWVEC (int, last_basic_block_for_fn (cfun));
6186 degree = XCNEWVEC (int, last_basic_block_for_fn (cfun));
6187
6188
6189 /* For each basic block that belongs to some loop assign the number
6190 of innermost loop it belongs to. */
6191 for (i = 0; i < last_basic_block_for_fn (cfun); i++)
6192 loop_hdr[i] = -1;
6193
6194 FOR_EACH_BB_FN (bb, cfun)
6195 {
6196 if (bb->loop_father && bb->loop_father->num != 0
6197 && !(bb->flags & BB_IRREDUCIBLE_LOOP))
6198 loop_hdr[bb->index] = bb->loop_father->num;
6199 }
6200
6201 /* For each basic block degree is calculated as the number of incoming
6202 edges, that are going out of bbs that are not yet scheduled.
6203 The basic blocks that are scheduled have degree value of zero. */
6204 FOR_EACH_BB_FN (bb, cfun)
6205 {
6206 degree[bb->index] = 0;
6207
6208 if (!bitmap_bit_p (bbs_in_loop_rgns, bb->index))
6209 {
6210 FOR_EACH_EDGE (e, ei, bb->preds)
6211 if (!bitmap_bit_p (bbs_in_loop_rgns, e->src->index))
6212 degree[bb->index]++;
6213 }
6214 else
6215 degree[bb->index] = -1;
6216 }
6217
6218 extend_rgns (degree, &cur_rgn_blocks, bbs_in_loop_rgns, loop_hdr);
6219
6220 /* Any block that did not end up in a region is placed into a region
6221 by itself. */
6222 FOR_EACH_BB_FN (bb, cfun)
6223 if (degree[bb->index] >= 0)
6224 {
6225 rgn_bb_table[cur_rgn_blocks] = bb->index;
6226 RGN_NR_BLOCKS (nr_regions) = 1;
6227 RGN_BLOCKS (nr_regions) = cur_rgn_blocks++;
6228 RGN_DONT_CALC_DEPS (nr_regions) = 0;
6229 RGN_HAS_REAL_EBB (nr_regions) = 0;
6230 CONTAINING_RGN (bb->index) = nr_regions++;
6231 BLOCK_TO_BB (bb->index) = 0;
6232 }
6233
6234 free (degree);
6235 free (loop_hdr);
6236 }
6237
6238 /* Free data structures used in pipelining of loops. */
6239 void sel_finish_pipelining (void)
6240 {
6241 struct loop *loop;
6242
6243 /* Release aux fields so we don't free them later by mistake. */
6244 FOR_EACH_LOOP (loop, 0)
6245 loop->aux = NULL;
6246
6247 loop_optimizer_finalize ();
6248
6249 loop_nests.release ();
6250
6251 free (rev_top_order_index);
6252 rev_top_order_index = NULL;
6253 }
6254
6255 /* This function replaces the find_rgns when
6256 FLAG_SEL_SCHED_PIPELINING_OUTER_LOOPS is set. */
6257 void
6258 sel_find_rgns (void)
6259 {
6260 sel_init_pipelining ();
6261 extend_regions ();
6262
6263 if (current_loops)
6264 {
6265 loop_p loop;
6266
6267 FOR_EACH_LOOP (loop, (flag_sel_sched_pipelining_outer_loops
6268 ? LI_FROM_INNERMOST
6269 : LI_ONLY_INNERMOST))
6270 make_regions_from_loop_nest (loop);
6271 }
6272
6273 /* Make regions from all the rest basic blocks and schedule them.
6274 These blocks include blocks that don't belong to any loop or belong
6275 to irreducible loops. */
6276 make_regions_from_the_rest ();
6277
6278 /* We don't need bbs_in_loop_rgns anymore. */
6279 sbitmap_free (bbs_in_loop_rgns);
6280 bbs_in_loop_rgns = NULL;
6281 }
6282
6283 /* Add the preheader blocks from previous loop to current region taking
6284 it from LOOP_PREHEADER_BLOCKS (current_loop_nest) and record them in *BBS.
6285 This function is only used with -fsel-sched-pipelining-outer-loops. */
6286 void
6287 sel_add_loop_preheaders (bb_vec_t *bbs)
6288 {
6289 int i;
6290 basic_block bb;
6291 vec<basic_block> *preheader_blocks
6292 = LOOP_PREHEADER_BLOCKS (current_loop_nest);
6293
6294 if (!preheader_blocks)
6295 return;
6296
6297 for (i = 0; preheader_blocks->iterate (i, &bb); i++)
6298 {
6299 bbs->safe_push (bb);
6300 last_added_blocks.safe_push (bb);
6301 sel_add_bb (bb);
6302 }
6303
6304 vec_free (preheader_blocks);
6305 }
6306
6307 /* While pipelining outer loops, returns TRUE if BB is a loop preheader.
6308 Please note that the function should also work when pipelining_p is
6309 false, because it is used when deciding whether we should or should
6310 not reschedule pipelined code. */
6311 bool
6312 sel_is_loop_preheader_p (basic_block bb)
6313 {
6314 if (current_loop_nest)
6315 {
6316 struct loop *outer;
6317
6318 if (preheader_removed)
6319 return false;
6320
6321 /* Preheader is the first block in the region. */
6322 if (BLOCK_TO_BB (bb->index) == 0)
6323 return true;
6324
6325 /* We used to find a preheader with the topological information.
6326 Check that the above code is equivalent to what we did before. */
6327
6328 if (in_current_region_p (current_loop_nest->header))
6329 gcc_assert (!(BLOCK_TO_BB (bb->index)
6330 < BLOCK_TO_BB (current_loop_nest->header->index)));
6331
6332 /* Support the situation when the latch block of outer loop
6333 could be from here. */
6334 for (outer = loop_outer (current_loop_nest);
6335 outer;
6336 outer = loop_outer (outer))
6337 if (considered_for_pipelining_p (outer) && outer->latch == bb)
6338 gcc_unreachable ();
6339 }
6340
6341 return false;
6342 }
6343
6344 /* Check whether JUMP_BB ends with a jump insn that leads only to DEST_BB and
6345 can be removed, making the corresponding edge fallthrough (assuming that
6346 all basic blocks between JUMP_BB and DEST_BB are empty). */
6347 static bool
6348 bb_has_removable_jump_to_p (basic_block jump_bb, basic_block dest_bb)
6349 {
6350 if (!onlyjump_p (BB_END (jump_bb))
6351 || tablejump_p (BB_END (jump_bb), NULL, NULL))
6352 return false;
6353
6354 /* Several outgoing edges, abnormal edge or destination of jump is
6355 not DEST_BB. */
6356 if (EDGE_COUNT (jump_bb->succs) != 1
6357 || EDGE_SUCC (jump_bb, 0)->flags & (EDGE_ABNORMAL | EDGE_CROSSING)
6358 || EDGE_SUCC (jump_bb, 0)->dest != dest_bb)
6359 return false;
6360
6361 /* If not anything of the upper. */
6362 return true;
6363 }
6364
6365 /* Removes the loop preheader from the current region and saves it in
6366 PREHEADER_BLOCKS of the father loop, so they will be added later to
6367 region that represents an outer loop. */
6368 static void
6369 sel_remove_loop_preheader (void)
6370 {
6371 int i, old_len;
6372 int cur_rgn = CONTAINING_RGN (BB_TO_BLOCK (0));
6373 basic_block bb;
6374 bool all_empty_p = true;
6375 vec<basic_block> *preheader_blocks
6376 = LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest));
6377
6378 vec_check_alloc (preheader_blocks, 0);
6379
6380 gcc_assert (current_loop_nest);
6381 old_len = preheader_blocks->length ();
6382
6383 /* Add blocks that aren't within the current loop to PREHEADER_BLOCKS. */
6384 for (i = 0; i < RGN_NR_BLOCKS (cur_rgn); i++)
6385 {
6386 bb = BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (i));
6387
6388 /* If the basic block belongs to region, but doesn't belong to
6389 corresponding loop, then it should be a preheader. */
6390 if (sel_is_loop_preheader_p (bb))
6391 {
6392 preheader_blocks->safe_push (bb);
6393 if (BB_END (bb) != bb_note (bb))
6394 all_empty_p = false;
6395 }
6396 }
6397
6398 /* Remove these blocks only after iterating over the whole region. */
6399 for (i = preheader_blocks->length () - 1; i >= old_len; i--)
6400 {
6401 bb = (*preheader_blocks)[i];
6402 sel_remove_bb (bb, false);
6403 }
6404
6405 if (!considered_for_pipelining_p (loop_outer (current_loop_nest)))
6406 {
6407 if (!all_empty_p)
6408 /* Immediately create new region from preheader. */
6409 make_region_from_loop_preheader (preheader_blocks);
6410 else
6411 {
6412 /* If all preheader blocks are empty - dont create new empty region.
6413 Instead, remove them completely. */
6414 FOR_EACH_VEC_ELT (*preheader_blocks, i, bb)
6415 {
6416 edge e;
6417 edge_iterator ei;
6418 basic_block prev_bb = bb->prev_bb, next_bb = bb->next_bb;
6419
6420 /* Redirect all incoming edges to next basic block. */
6421 for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei)); )
6422 {
6423 if (! (e->flags & EDGE_FALLTHRU))
6424 redirect_edge_and_branch (e, bb->next_bb);
6425 else
6426 redirect_edge_succ (e, bb->next_bb);
6427 }
6428 gcc_assert (BB_NOTE_LIST (bb) == NULL);
6429 delete_and_free_basic_block (bb);
6430
6431 /* Check if after deleting preheader there is a nonconditional
6432 jump in PREV_BB that leads to the next basic block NEXT_BB.
6433 If it is so - delete this jump and clear data sets of its
6434 basic block if it becomes empty. */
6435 if (next_bb->prev_bb == prev_bb
6436 && prev_bb != ENTRY_BLOCK_PTR_FOR_FN (cfun)
6437 && bb_has_removable_jump_to_p (prev_bb, next_bb))
6438 {
6439 redirect_edge_and_branch (EDGE_SUCC (prev_bb, 0), next_bb);
6440 if (BB_END (prev_bb) == bb_note (prev_bb))
6441 free_data_sets (prev_bb);
6442 }
6443
6444 set_immediate_dominator (CDI_DOMINATORS, next_bb,
6445 recompute_dominator (CDI_DOMINATORS,
6446 next_bb));
6447 }
6448 }
6449 vec_free (preheader_blocks);
6450 }
6451 else
6452 /* Store preheader within the father's loop structure. */
6453 SET_LOOP_PREHEADER_BLOCKS (loop_outer (current_loop_nest),
6454 preheader_blocks);
6455 }
6456
6457 #endif