]> git.ipfire.org Git - thirdparty/gcc.git/blob - gcc/cfgcleanup.c
2014-10-16 Andrew MacLeod <amacleod@redhat.com>
[thirdparty/gcc.git] / gcc / cfgcleanup.c
1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987-2014 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 /* This file contains optimizer of the control flow. The main entry point is
21 cleanup_cfg. Following optimizations are performed:
22
23 - Unreachable blocks removal
24 - Edge forwarding (edge to the forwarder block is forwarded to its
25 successor. Simplification of the branch instruction is performed by
26 underlying infrastructure so branch can be converted to simplejump or
27 eliminated).
28 - Cross jumping (tail merging)
29 - Conditional jump-around-simplejump simplification
30 - Basic block merging. */
31
32 #include "config.h"
33 #include "system.h"
34 #include "coretypes.h"
35 #include "tm.h"
36 #include "rtl.h"
37 #include "tree.h"
38 #include "hard-reg-set.h"
39 #include "regs.h"
40 #include "insn-config.h"
41 #include "flags.h"
42 #include "recog.h"
43 #include "diagnostic-core.h"
44 #include "cselib.h"
45 #include "params.h"
46 #include "tm_p.h"
47 #include "target.h"
48 #include "hashtab.h"
49 #include "hash-set.h"
50 #include "vec.h"
51 #include "machmode.h"
52 #include "input.h"
53 #include "function.h" /* For inline functions in emit-rtl.h they need crtl. */
54 #include "emit-rtl.h"
55 #include "tree-pass.h"
56 #include "cfgloop.h"
57 #include "expr.h"
58 #include "df.h"
59 #include "dce.h"
60 #include "dbgcnt.h"
61 #include "emit-rtl.h"
62 #include "rtl-iter.h"
63
64 #define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
65
66 /* Set to true when we are running first pass of try_optimize_cfg loop. */
67 static bool first_pass;
68
69 /* Set to true if crossjumps occurred in the latest run of try_optimize_cfg. */
70 static bool crossjumps_occured;
71
72 /* Set to true if we couldn't run an optimization due to stale liveness
73 information; we should run df_analyze to enable more opportunities. */
74 static bool block_was_dirty;
75
76 static bool try_crossjump_to_edge (int, edge, edge, enum replace_direction);
77 static bool try_crossjump_bb (int, basic_block);
78 static bool outgoing_edges_match (int, basic_block, basic_block);
79 static enum replace_direction old_insns_match_p (int, rtx_insn *, rtx_insn *);
80
81 static void merge_blocks_move_predecessor_nojumps (basic_block, basic_block);
82 static void merge_blocks_move_successor_nojumps (basic_block, basic_block);
83 static bool try_optimize_cfg (int);
84 static bool try_simplify_condjump (basic_block);
85 static bool try_forward_edges (int, basic_block);
86 static edge thread_jump (edge, basic_block);
87 static bool mark_effect (rtx, bitmap);
88 static void notice_new_block (basic_block);
89 static void update_forwarder_flag (basic_block);
90 static void merge_memattrs (rtx, rtx);
91 \f
92 /* Set flags for newly created block. */
93
94 static void
95 notice_new_block (basic_block bb)
96 {
97 if (!bb)
98 return;
99
100 if (forwarder_block_p (bb))
101 bb->flags |= BB_FORWARDER_BLOCK;
102 }
103
104 /* Recompute forwarder flag after block has been modified. */
105
106 static void
107 update_forwarder_flag (basic_block bb)
108 {
109 if (forwarder_block_p (bb))
110 bb->flags |= BB_FORWARDER_BLOCK;
111 else
112 bb->flags &= ~BB_FORWARDER_BLOCK;
113 }
114 \f
115 /* Simplify a conditional jump around an unconditional jump.
116 Return true if something changed. */
117
118 static bool
119 try_simplify_condjump (basic_block cbranch_block)
120 {
121 basic_block jump_block, jump_dest_block, cbranch_dest_block;
122 edge cbranch_jump_edge, cbranch_fallthru_edge;
123 rtx_insn *cbranch_insn;
124
125 /* Verify that there are exactly two successors. */
126 if (EDGE_COUNT (cbranch_block->succs) != 2)
127 return false;
128
129 /* Verify that we've got a normal conditional branch at the end
130 of the block. */
131 cbranch_insn = BB_END (cbranch_block);
132 if (!any_condjump_p (cbranch_insn))
133 return false;
134
135 cbranch_fallthru_edge = FALLTHRU_EDGE (cbranch_block);
136 cbranch_jump_edge = BRANCH_EDGE (cbranch_block);
137
138 /* The next block must not have multiple predecessors, must not
139 be the last block in the function, and must contain just the
140 unconditional jump. */
141 jump_block = cbranch_fallthru_edge->dest;
142 if (!single_pred_p (jump_block)
143 || jump_block->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
144 || !FORWARDER_BLOCK_P (jump_block))
145 return false;
146 jump_dest_block = single_succ (jump_block);
147
148 /* If we are partitioning hot/cold basic blocks, we don't want to
149 mess up unconditional or indirect jumps that cross between hot
150 and cold sections.
151
152 Basic block partitioning may result in some jumps that appear to
153 be optimizable (or blocks that appear to be mergeable), but which really
154 must be left untouched (they are required to make it safely across
155 partition boundaries). See the comments at the top of
156 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
157
158 if (BB_PARTITION (jump_block) != BB_PARTITION (jump_dest_block)
159 || (cbranch_jump_edge->flags & EDGE_CROSSING))
160 return false;
161
162 /* The conditional branch must target the block after the
163 unconditional branch. */
164 cbranch_dest_block = cbranch_jump_edge->dest;
165
166 if (cbranch_dest_block == EXIT_BLOCK_PTR_FOR_FN (cfun)
167 || !can_fallthru (jump_block, cbranch_dest_block))
168 return false;
169
170 /* Invert the conditional branch. */
171 if (!invert_jump (cbranch_insn, block_label (jump_dest_block), 0))
172 return false;
173
174 if (dump_file)
175 fprintf (dump_file, "Simplifying condjump %i around jump %i\n",
176 INSN_UID (cbranch_insn), INSN_UID (BB_END (jump_block)));
177
178 /* Success. Update the CFG to match. Note that after this point
179 the edge variable names appear backwards; the redirection is done
180 this way to preserve edge profile data. */
181 cbranch_jump_edge = redirect_edge_succ_nodup (cbranch_jump_edge,
182 cbranch_dest_block);
183 cbranch_fallthru_edge = redirect_edge_succ_nodup (cbranch_fallthru_edge,
184 jump_dest_block);
185 cbranch_jump_edge->flags |= EDGE_FALLTHRU;
186 cbranch_fallthru_edge->flags &= ~EDGE_FALLTHRU;
187 update_br_prob_note (cbranch_block);
188
189 /* Delete the block with the unconditional jump, and clean up the mess. */
190 delete_basic_block (jump_block);
191 tidy_fallthru_edge (cbranch_jump_edge);
192 update_forwarder_flag (cbranch_block);
193
194 return true;
195 }
196 \f
197 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
198 on register. Used by jump threading. */
199
200 static bool
201 mark_effect (rtx exp, regset nonequal)
202 {
203 int regno;
204 rtx dest;
205 switch (GET_CODE (exp))
206 {
207 /* In case we do clobber the register, mark it as equal, as we know the
208 value is dead so it don't have to match. */
209 case CLOBBER:
210 if (REG_P (XEXP (exp, 0)))
211 {
212 dest = XEXP (exp, 0);
213 regno = REGNO (dest);
214 if (HARD_REGISTER_NUM_P (regno))
215 bitmap_clear_range (nonequal, regno,
216 hard_regno_nregs[regno][GET_MODE (dest)]);
217 else
218 bitmap_clear_bit (nonequal, regno);
219 }
220 return false;
221
222 case SET:
223 if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
224 return false;
225 dest = SET_DEST (exp);
226 if (dest == pc_rtx)
227 return false;
228 if (!REG_P (dest))
229 return true;
230 regno = REGNO (dest);
231 if (HARD_REGISTER_NUM_P (regno))
232 bitmap_set_range (nonequal, regno,
233 hard_regno_nregs[regno][GET_MODE (dest)]);
234 else
235 bitmap_set_bit (nonequal, regno);
236 return false;
237
238 default:
239 return false;
240 }
241 }
242
243 /* Return true if X contains a register in NONEQUAL. */
244 static bool
245 mentions_nonequal_regs (const_rtx x, regset nonequal)
246 {
247 subrtx_iterator::array_type array;
248 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
249 {
250 const_rtx x = *iter;
251 if (REG_P (x))
252 {
253 unsigned int regno = REGNO (x);
254 if (REGNO_REG_SET_P (nonequal, regno))
255 return true;
256 if (regno < FIRST_PSEUDO_REGISTER)
257 {
258 int n = hard_regno_nregs[regno][GET_MODE (x)];
259 while (--n > 0)
260 if (REGNO_REG_SET_P (nonequal, regno + n))
261 return true;
262 }
263 }
264 }
265 return false;
266 }
267
268 /* Attempt to prove that the basic block B will have no side effects and
269 always continues in the same edge if reached via E. Return the edge
270 if exist, NULL otherwise. */
271
272 static edge
273 thread_jump (edge e, basic_block b)
274 {
275 rtx set1, set2, cond1, cond2;
276 rtx_insn *insn;
277 enum rtx_code code1, code2, reversed_code2;
278 bool reverse1 = false;
279 unsigned i;
280 regset nonequal;
281 bool failed = false;
282 reg_set_iterator rsi;
283
284 if (b->flags & BB_NONTHREADABLE_BLOCK)
285 return NULL;
286
287 /* At the moment, we do handle only conditional jumps, but later we may
288 want to extend this code to tablejumps and others. */
289 if (EDGE_COUNT (e->src->succs) != 2)
290 return NULL;
291 if (EDGE_COUNT (b->succs) != 2)
292 {
293 b->flags |= BB_NONTHREADABLE_BLOCK;
294 return NULL;
295 }
296
297 /* Second branch must end with onlyjump, as we will eliminate the jump. */
298 if (!any_condjump_p (BB_END (e->src)))
299 return NULL;
300
301 if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
302 {
303 b->flags |= BB_NONTHREADABLE_BLOCK;
304 return NULL;
305 }
306
307 set1 = pc_set (BB_END (e->src));
308 set2 = pc_set (BB_END (b));
309 if (((e->flags & EDGE_FALLTHRU) != 0)
310 != (XEXP (SET_SRC (set1), 1) == pc_rtx))
311 reverse1 = true;
312
313 cond1 = XEXP (SET_SRC (set1), 0);
314 cond2 = XEXP (SET_SRC (set2), 0);
315 if (reverse1)
316 code1 = reversed_comparison_code (cond1, BB_END (e->src));
317 else
318 code1 = GET_CODE (cond1);
319
320 code2 = GET_CODE (cond2);
321 reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
322
323 if (!comparison_dominates_p (code1, code2)
324 && !comparison_dominates_p (code1, reversed_code2))
325 return NULL;
326
327 /* Ensure that the comparison operators are equivalent.
328 ??? This is far too pessimistic. We should allow swapped operands,
329 different CCmodes, or for example comparisons for interval, that
330 dominate even when operands are not equivalent. */
331 if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
332 || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
333 return NULL;
334
335 /* Short circuit cases where block B contains some side effects, as we can't
336 safely bypass it. */
337 for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
338 insn = NEXT_INSN (insn))
339 if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
340 {
341 b->flags |= BB_NONTHREADABLE_BLOCK;
342 return NULL;
343 }
344
345 cselib_init (0);
346
347 /* First process all values computed in the source basic block. */
348 for (insn = NEXT_INSN (BB_HEAD (e->src));
349 insn != NEXT_INSN (BB_END (e->src));
350 insn = NEXT_INSN (insn))
351 if (INSN_P (insn))
352 cselib_process_insn (insn);
353
354 nonequal = BITMAP_ALLOC (NULL);
355 CLEAR_REG_SET (nonequal);
356
357 /* Now assume that we've continued by the edge E to B and continue
358 processing as if it were same basic block.
359 Our goal is to prove that whole block is an NOOP. */
360
361 for (insn = NEXT_INSN (BB_HEAD (b));
362 insn != NEXT_INSN (BB_END (b)) && !failed;
363 insn = NEXT_INSN (insn))
364 {
365 if (INSN_P (insn))
366 {
367 rtx pat = PATTERN (insn);
368
369 if (GET_CODE (pat) == PARALLEL)
370 {
371 for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
372 failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
373 }
374 else
375 failed |= mark_effect (pat, nonequal);
376 }
377
378 cselib_process_insn (insn);
379 }
380
381 /* Later we should clear nonequal of dead registers. So far we don't
382 have life information in cfg_cleanup. */
383 if (failed)
384 {
385 b->flags |= BB_NONTHREADABLE_BLOCK;
386 goto failed_exit;
387 }
388
389 /* cond2 must not mention any register that is not equal to the
390 former block. */
391 if (mentions_nonequal_regs (cond2, nonequal))
392 goto failed_exit;
393
394 EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
395 goto failed_exit;
396
397 BITMAP_FREE (nonequal);
398 cselib_finish ();
399 if ((comparison_dominates_p (code1, code2) != 0)
400 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
401 return BRANCH_EDGE (b);
402 else
403 return FALLTHRU_EDGE (b);
404
405 failed_exit:
406 BITMAP_FREE (nonequal);
407 cselib_finish ();
408 return NULL;
409 }
410 \f
411 /* Attempt to forward edges leaving basic block B.
412 Return true if successful. */
413
414 static bool
415 try_forward_edges (int mode, basic_block b)
416 {
417 bool changed = false;
418 edge_iterator ei;
419 edge e, *threaded_edges = NULL;
420
421 /* If we are partitioning hot/cold basic blocks, we don't want to
422 mess up unconditional or indirect jumps that cross between hot
423 and cold sections.
424
425 Basic block partitioning may result in some jumps that appear to
426 be optimizable (or blocks that appear to be mergeable), but which really
427 must be left untouched (they are required to make it safely across
428 partition boundaries). See the comments at the top of
429 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
430
431 if (JUMP_P (BB_END (b)) && CROSSING_JUMP_P (BB_END (b)))
432 return false;
433
434 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
435 {
436 basic_block target, first;
437 location_t goto_locus;
438 int counter;
439 bool threaded = false;
440 int nthreaded_edges = 0;
441 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
442
443 /* Skip complex edges because we don't know how to update them.
444
445 Still handle fallthru edges, as we can succeed to forward fallthru
446 edge to the same place as the branch edge of conditional branch
447 and turn conditional branch to an unconditional branch. */
448 if (e->flags & EDGE_COMPLEX)
449 {
450 ei_next (&ei);
451 continue;
452 }
453
454 target = first = e->dest;
455 counter = NUM_FIXED_BLOCKS;
456 goto_locus = e->goto_locus;
457
458 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
459 up jumps that cross between hot/cold sections.
460
461 Basic block partitioning may result in some jumps that appear
462 to be optimizable (or blocks that appear to be mergeable), but which
463 really must be left untouched (they are required to make it safely
464 across partition boundaries). See the comments at the top of
465 bb-reorder.c:partition_hot_cold_basic_blocks for complete
466 details. */
467
468 if (first != EXIT_BLOCK_PTR_FOR_FN (cfun)
469 && JUMP_P (BB_END (first))
470 && CROSSING_JUMP_P (BB_END (first)))
471 return changed;
472
473 while (counter < n_basic_blocks_for_fn (cfun))
474 {
475 basic_block new_target = NULL;
476 bool new_target_threaded = false;
477 may_thread |= (target->flags & BB_MODIFIED) != 0;
478
479 if (FORWARDER_BLOCK_P (target)
480 && !(single_succ_edge (target)->flags & EDGE_CROSSING)
481 && single_succ (target) != EXIT_BLOCK_PTR_FOR_FN (cfun))
482 {
483 /* Bypass trivial infinite loops. */
484 new_target = single_succ (target);
485 if (target == new_target)
486 counter = n_basic_blocks_for_fn (cfun);
487 else if (!optimize)
488 {
489 /* When not optimizing, ensure that edges or forwarder
490 blocks with different locus are not optimized out. */
491 location_t new_locus = single_succ_edge (target)->goto_locus;
492 location_t locus = goto_locus;
493
494 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
495 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
496 && new_locus != locus)
497 new_target = NULL;
498 else
499 {
500 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
501 locus = new_locus;
502
503 rtx_insn *last = BB_END (target);
504 if (DEBUG_INSN_P (last))
505 last = prev_nondebug_insn (last);
506 if (last && INSN_P (last))
507 new_locus = INSN_LOCATION (last);
508 else
509 new_locus = UNKNOWN_LOCATION;
510
511 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
512 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
513 && new_locus != locus)
514 new_target = NULL;
515 else
516 {
517 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
518 locus = new_locus;
519
520 goto_locus = locus;
521 }
522 }
523 }
524 }
525
526 /* Allow to thread only over one edge at time to simplify updating
527 of probabilities. */
528 else if ((mode & CLEANUP_THREADING) && may_thread)
529 {
530 edge t = thread_jump (e, target);
531 if (t)
532 {
533 if (!threaded_edges)
534 threaded_edges = XNEWVEC (edge,
535 n_basic_blocks_for_fn (cfun));
536 else
537 {
538 int i;
539
540 /* Detect an infinite loop across blocks not
541 including the start block. */
542 for (i = 0; i < nthreaded_edges; ++i)
543 if (threaded_edges[i] == t)
544 break;
545 if (i < nthreaded_edges)
546 {
547 counter = n_basic_blocks_for_fn (cfun);
548 break;
549 }
550 }
551
552 /* Detect an infinite loop across the start block. */
553 if (t->dest == b)
554 break;
555
556 gcc_assert (nthreaded_edges
557 < (n_basic_blocks_for_fn (cfun)
558 - NUM_FIXED_BLOCKS));
559 threaded_edges[nthreaded_edges++] = t;
560
561 new_target = t->dest;
562 new_target_threaded = true;
563 }
564 }
565
566 if (!new_target)
567 break;
568
569 counter++;
570 target = new_target;
571 threaded |= new_target_threaded;
572 }
573
574 if (counter >= n_basic_blocks_for_fn (cfun))
575 {
576 if (dump_file)
577 fprintf (dump_file, "Infinite loop in BB %i.\n",
578 target->index);
579 }
580 else if (target == first)
581 ; /* We didn't do anything. */
582 else
583 {
584 /* Save the values now, as the edge may get removed. */
585 gcov_type edge_count = e->count;
586 int edge_probability = e->probability;
587 int edge_frequency;
588 int n = 0;
589
590 e->goto_locus = goto_locus;
591
592 /* Don't force if target is exit block. */
593 if (threaded && target != EXIT_BLOCK_PTR_FOR_FN (cfun))
594 {
595 notice_new_block (redirect_edge_and_branch_force (e, target));
596 if (dump_file)
597 fprintf (dump_file, "Conditionals threaded.\n");
598 }
599 else if (!redirect_edge_and_branch (e, target))
600 {
601 if (dump_file)
602 fprintf (dump_file,
603 "Forwarding edge %i->%i to %i failed.\n",
604 b->index, e->dest->index, target->index);
605 ei_next (&ei);
606 continue;
607 }
608
609 /* We successfully forwarded the edge. Now update profile
610 data: for each edge we traversed in the chain, remove
611 the original edge's execution count. */
612 edge_frequency = apply_probability (b->frequency, edge_probability);
613
614 do
615 {
616 edge t;
617
618 if (!single_succ_p (first))
619 {
620 gcc_assert (n < nthreaded_edges);
621 t = threaded_edges [n++];
622 gcc_assert (t->src == first);
623 update_bb_profile_for_threading (first, edge_frequency,
624 edge_count, t);
625 update_br_prob_note (first);
626 }
627 else
628 {
629 first->count -= edge_count;
630 if (first->count < 0)
631 first->count = 0;
632 first->frequency -= edge_frequency;
633 if (first->frequency < 0)
634 first->frequency = 0;
635 /* It is possible that as the result of
636 threading we've removed edge as it is
637 threaded to the fallthru edge. Avoid
638 getting out of sync. */
639 if (n < nthreaded_edges
640 && first == threaded_edges [n]->src)
641 n++;
642 t = single_succ_edge (first);
643 }
644
645 t->count -= edge_count;
646 if (t->count < 0)
647 t->count = 0;
648 first = t->dest;
649 }
650 while (first != target);
651
652 changed = true;
653 continue;
654 }
655 ei_next (&ei);
656 }
657
658 free (threaded_edges);
659 return changed;
660 }
661 \f
662
663 /* Blocks A and B are to be merged into a single block. A has no incoming
664 fallthru edge, so it can be moved before B without adding or modifying
665 any jumps (aside from the jump from A to B). */
666
667 static void
668 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
669 {
670 rtx_insn *barrier;
671
672 /* If we are partitioning hot/cold basic blocks, we don't want to
673 mess up unconditional or indirect jumps that cross between hot
674 and cold sections.
675
676 Basic block partitioning may result in some jumps that appear to
677 be optimizable (or blocks that appear to be mergeable), but which really
678 must be left untouched (they are required to make it safely across
679 partition boundaries). See the comments at the top of
680 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
681
682 if (BB_PARTITION (a) != BB_PARTITION (b))
683 return;
684
685 barrier = next_nonnote_insn (BB_END (a));
686 gcc_assert (BARRIER_P (barrier));
687 delete_insn (barrier);
688
689 /* Scramble the insn chain. */
690 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
691 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
692 df_set_bb_dirty (a);
693
694 if (dump_file)
695 fprintf (dump_file, "Moved block %d before %d and merged.\n",
696 a->index, b->index);
697
698 /* Swap the records for the two blocks around. */
699
700 unlink_block (a);
701 link_block (a, b->prev_bb);
702
703 /* Now blocks A and B are contiguous. Merge them. */
704 merge_blocks (a, b);
705 }
706
707 /* Blocks A and B are to be merged into a single block. B has no outgoing
708 fallthru edge, so it can be moved after A without adding or modifying
709 any jumps (aside from the jump from A to B). */
710
711 static void
712 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
713 {
714 rtx_insn *barrier, *real_b_end;
715 rtx label;
716 rtx_jump_table_data *table;
717
718 /* If we are partitioning hot/cold basic blocks, we don't want to
719 mess up unconditional or indirect jumps that cross between hot
720 and cold sections.
721
722 Basic block partitioning may result in some jumps that appear to
723 be optimizable (or blocks that appear to be mergeable), but which really
724 must be left untouched (they are required to make it safely across
725 partition boundaries). See the comments at the top of
726 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
727
728 if (BB_PARTITION (a) != BB_PARTITION (b))
729 return;
730
731 real_b_end = BB_END (b);
732
733 /* If there is a jump table following block B temporarily add the jump table
734 to block B so that it will also be moved to the correct location. */
735 if (tablejump_p (BB_END (b), &label, &table)
736 && prev_active_insn (label) == BB_END (b))
737 {
738 BB_END (b) = table;
739 }
740
741 /* There had better have been a barrier there. Delete it. */
742 barrier = NEXT_INSN (BB_END (b));
743 if (barrier && BARRIER_P (barrier))
744 delete_insn (barrier);
745
746
747 /* Scramble the insn chain. */
748 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
749
750 /* Restore the real end of b. */
751 BB_END (b) = real_b_end;
752
753 if (dump_file)
754 fprintf (dump_file, "Moved block %d after %d and merged.\n",
755 b->index, a->index);
756
757 /* Now blocks A and B are contiguous. Merge them. */
758 merge_blocks (a, b);
759 }
760
761 /* Attempt to merge basic blocks that are potentially non-adjacent.
762 Return NULL iff the attempt failed, otherwise return basic block
763 where cleanup_cfg should continue. Because the merging commonly
764 moves basic block away or introduces another optimization
765 possibility, return basic block just before B so cleanup_cfg don't
766 need to iterate.
767
768 It may be good idea to return basic block before C in the case
769 C has been moved after B and originally appeared earlier in the
770 insn sequence, but we have no information available about the
771 relative ordering of these two. Hopefully it is not too common. */
772
773 static basic_block
774 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
775 {
776 basic_block next;
777
778 /* If we are partitioning hot/cold basic blocks, we don't want to
779 mess up unconditional or indirect jumps that cross between hot
780 and cold sections.
781
782 Basic block partitioning may result in some jumps that appear to
783 be optimizable (or blocks that appear to be mergeable), but which really
784 must be left untouched (they are required to make it safely across
785 partition boundaries). See the comments at the top of
786 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
787
788 if (BB_PARTITION (b) != BB_PARTITION (c))
789 return NULL;
790
791 /* If B has a fallthru edge to C, no need to move anything. */
792 if (e->flags & EDGE_FALLTHRU)
793 {
794 int b_index = b->index, c_index = c->index;
795
796 /* Protect the loop latches. */
797 if (current_loops && c->loop_father->latch == c)
798 return NULL;
799
800 merge_blocks (b, c);
801 update_forwarder_flag (b);
802
803 if (dump_file)
804 fprintf (dump_file, "Merged %d and %d without moving.\n",
805 b_index, c_index);
806
807 return b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? b : b->prev_bb;
808 }
809
810 /* Otherwise we will need to move code around. Do that only if expensive
811 transformations are allowed. */
812 else if (mode & CLEANUP_EXPENSIVE)
813 {
814 edge tmp_edge, b_fallthru_edge;
815 bool c_has_outgoing_fallthru;
816 bool b_has_incoming_fallthru;
817
818 /* Avoid overactive code motion, as the forwarder blocks should be
819 eliminated by edge redirection instead. One exception might have
820 been if B is a forwarder block and C has no fallthru edge, but
821 that should be cleaned up by bb-reorder instead. */
822 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
823 return NULL;
824
825 /* We must make sure to not munge nesting of lexical blocks,
826 and loop notes. This is done by squeezing out all the notes
827 and leaving them there to lie. Not ideal, but functional. */
828
829 tmp_edge = find_fallthru_edge (c->succs);
830 c_has_outgoing_fallthru = (tmp_edge != NULL);
831
832 tmp_edge = find_fallthru_edge (b->preds);
833 b_has_incoming_fallthru = (tmp_edge != NULL);
834 b_fallthru_edge = tmp_edge;
835 next = b->prev_bb;
836 if (next == c)
837 next = next->prev_bb;
838
839 /* Otherwise, we're going to try to move C after B. If C does
840 not have an outgoing fallthru, then it can be moved
841 immediately after B without introducing or modifying jumps. */
842 if (! c_has_outgoing_fallthru)
843 {
844 merge_blocks_move_successor_nojumps (b, c);
845 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
846 }
847
848 /* If B does not have an incoming fallthru, then it can be moved
849 immediately before C without introducing or modifying jumps.
850 C cannot be the first block, so we do not have to worry about
851 accessing a non-existent block. */
852
853 if (b_has_incoming_fallthru)
854 {
855 basic_block bb;
856
857 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
858 return NULL;
859 bb = force_nonfallthru (b_fallthru_edge);
860 if (bb)
861 notice_new_block (bb);
862 }
863
864 merge_blocks_move_predecessor_nojumps (b, c);
865 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
866 }
867
868 return NULL;
869 }
870 \f
871
872 /* Removes the memory attributes of MEM expression
873 if they are not equal. */
874
875 void
876 merge_memattrs (rtx x, rtx y)
877 {
878 int i;
879 int j;
880 enum rtx_code code;
881 const char *fmt;
882
883 if (x == y)
884 return;
885 if (x == 0 || y == 0)
886 return;
887
888 code = GET_CODE (x);
889
890 if (code != GET_CODE (y))
891 return;
892
893 if (GET_MODE (x) != GET_MODE (y))
894 return;
895
896 if (code == MEM && !mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
897 {
898 if (! MEM_ATTRS (x))
899 MEM_ATTRS (y) = 0;
900 else if (! MEM_ATTRS (y))
901 MEM_ATTRS (x) = 0;
902 else
903 {
904 HOST_WIDE_INT mem_size;
905
906 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
907 {
908 set_mem_alias_set (x, 0);
909 set_mem_alias_set (y, 0);
910 }
911
912 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
913 {
914 set_mem_expr (x, 0);
915 set_mem_expr (y, 0);
916 clear_mem_offset (x);
917 clear_mem_offset (y);
918 }
919 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
920 || (MEM_OFFSET_KNOWN_P (x)
921 && MEM_OFFSET (x) != MEM_OFFSET (y)))
922 {
923 clear_mem_offset (x);
924 clear_mem_offset (y);
925 }
926
927 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
928 {
929 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
930 set_mem_size (x, mem_size);
931 set_mem_size (y, mem_size);
932 }
933 else
934 {
935 clear_mem_size (x);
936 clear_mem_size (y);
937 }
938
939 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
940 set_mem_align (y, MEM_ALIGN (x));
941 }
942 }
943 if (code == MEM)
944 {
945 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
946 {
947 MEM_READONLY_P (x) = 0;
948 MEM_READONLY_P (y) = 0;
949 }
950 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
951 {
952 MEM_NOTRAP_P (x) = 0;
953 MEM_NOTRAP_P (y) = 0;
954 }
955 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
956 {
957 MEM_VOLATILE_P (x) = 1;
958 MEM_VOLATILE_P (y) = 1;
959 }
960 }
961
962 fmt = GET_RTX_FORMAT (code);
963 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
964 {
965 switch (fmt[i])
966 {
967 case 'E':
968 /* Two vectors must have the same length. */
969 if (XVECLEN (x, i) != XVECLEN (y, i))
970 return;
971
972 for (j = 0; j < XVECLEN (x, i); j++)
973 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
974
975 break;
976
977 case 'e':
978 merge_memattrs (XEXP (x, i), XEXP (y, i));
979 }
980 }
981 return;
982 }
983
984
985 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
986 different single sets S1 and S2. */
987
988 static bool
989 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
990 {
991 int i;
992 rtx e1, e2;
993
994 if (p1 == s1 && p2 == s2)
995 return true;
996
997 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
998 return false;
999
1000 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
1001 return false;
1002
1003 for (i = 0; i < XVECLEN (p1, 0); i++)
1004 {
1005 e1 = XVECEXP (p1, 0, i);
1006 e2 = XVECEXP (p2, 0, i);
1007 if (e1 == s1 && e2 == s2)
1008 continue;
1009 if (reload_completed
1010 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
1011 continue;
1012
1013 return false;
1014 }
1015
1016 return true;
1017 }
1018
1019 /* Examine register notes on I1 and I2 and return:
1020 - dir_forward if I1 can be replaced by I2, or
1021 - dir_backward if I2 can be replaced by I1, or
1022 - dir_both if both are the case. */
1023
1024 static enum replace_direction
1025 can_replace_by (rtx_insn *i1, rtx_insn *i2)
1026 {
1027 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1028 bool c1, c2;
1029
1030 /* Check for 2 sets. */
1031 s1 = single_set (i1);
1032 s2 = single_set (i2);
1033 if (s1 == NULL_RTX || s2 == NULL_RTX)
1034 return dir_none;
1035
1036 /* Check that the 2 sets set the same dest. */
1037 d1 = SET_DEST (s1);
1038 d2 = SET_DEST (s2);
1039 if (!(reload_completed
1040 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1041 return dir_none;
1042
1043 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1044 set dest to the same value. */
1045 note1 = find_reg_equal_equiv_note (i1);
1046 note2 = find_reg_equal_equiv_note (i2);
1047 if (!note1 || !note2 || !rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0))
1048 || !CONST_INT_P (XEXP (note1, 0)))
1049 return dir_none;
1050
1051 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1052 return dir_none;
1053
1054 /* Although the 2 sets set dest to the same value, we cannot replace
1055 (set (dest) (const_int))
1056 by
1057 (set (dest) (reg))
1058 because we don't know if the reg is live and has the same value at the
1059 location of replacement. */
1060 src1 = SET_SRC (s1);
1061 src2 = SET_SRC (s2);
1062 c1 = CONST_INT_P (src1);
1063 c2 = CONST_INT_P (src2);
1064 if (c1 && c2)
1065 return dir_both;
1066 else if (c2)
1067 return dir_forward;
1068 else if (c1)
1069 return dir_backward;
1070
1071 return dir_none;
1072 }
1073
1074 /* Merges directions A and B. */
1075
1076 static enum replace_direction
1077 merge_dir (enum replace_direction a, enum replace_direction b)
1078 {
1079 /* Implements the following table:
1080 |bo fw bw no
1081 ---+-----------
1082 bo |bo fw bw no
1083 fw |-- fw no no
1084 bw |-- -- bw no
1085 no |-- -- -- no. */
1086
1087 if (a == b)
1088 return a;
1089
1090 if (a == dir_both)
1091 return b;
1092 if (b == dir_both)
1093 return a;
1094
1095 return dir_none;
1096 }
1097
1098 /* Examine I1 and I2 and return:
1099 - dir_forward if I1 can be replaced by I2, or
1100 - dir_backward if I2 can be replaced by I1, or
1101 - dir_both if both are the case. */
1102
1103 static enum replace_direction
1104 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx_insn *i1, rtx_insn *i2)
1105 {
1106 rtx p1, p2;
1107
1108 /* Verify that I1 and I2 are equivalent. */
1109 if (GET_CODE (i1) != GET_CODE (i2))
1110 return dir_none;
1111
1112 /* __builtin_unreachable() may lead to empty blocks (ending with
1113 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1114 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1115 return dir_both;
1116
1117 /* ??? Do not allow cross-jumping between different stack levels. */
1118 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1119 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1120 if (p1 && p2)
1121 {
1122 p1 = XEXP (p1, 0);
1123 p2 = XEXP (p2, 0);
1124 if (!rtx_equal_p (p1, p2))
1125 return dir_none;
1126
1127 /* ??? Worse, this adjustment had better be constant lest we
1128 have differing incoming stack levels. */
1129 if (!frame_pointer_needed
1130 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1131 return dir_none;
1132 }
1133 else if (p1 || p2)
1134 return dir_none;
1135
1136 p1 = PATTERN (i1);
1137 p2 = PATTERN (i2);
1138
1139 if (GET_CODE (p1) != GET_CODE (p2))
1140 return dir_none;
1141
1142 /* If this is a CALL_INSN, compare register usage information.
1143 If we don't check this on stack register machines, the two
1144 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1145 numbers of stack registers in the same basic block.
1146 If we don't check this on machines with delay slots, a delay slot may
1147 be filled that clobbers a parameter expected by the subroutine.
1148
1149 ??? We take the simple route for now and assume that if they're
1150 equal, they were constructed identically.
1151
1152 Also check for identical exception regions. */
1153
1154 if (CALL_P (i1))
1155 {
1156 /* Ensure the same EH region. */
1157 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1158 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1159
1160 if (!n1 && n2)
1161 return dir_none;
1162
1163 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1164 return dir_none;
1165
1166 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1167 CALL_INSN_FUNCTION_USAGE (i2))
1168 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1169 return dir_none;
1170
1171 /* For address sanitizer, never crossjump __asan_report_* builtins,
1172 otherwise errors might be reported on incorrect lines. */
1173 if (flag_sanitize & SANITIZE_ADDRESS)
1174 {
1175 rtx call = get_call_rtx_from (i1);
1176 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1177 {
1178 rtx symbol = XEXP (XEXP (call, 0), 0);
1179 if (SYMBOL_REF_DECL (symbol)
1180 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1181 {
1182 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1183 == BUILT_IN_NORMAL)
1184 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1185 >= BUILT_IN_ASAN_REPORT_LOAD1
1186 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1187 <= BUILT_IN_ASAN_STOREN)
1188 return dir_none;
1189 }
1190 }
1191 }
1192 }
1193
1194 #ifdef STACK_REGS
1195 /* If cross_jump_death_matters is not 0, the insn's mode
1196 indicates whether or not the insn contains any stack-like
1197 regs. */
1198
1199 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1200 {
1201 /* If register stack conversion has already been done, then
1202 death notes must also be compared before it is certain that
1203 the two instruction streams match. */
1204
1205 rtx note;
1206 HARD_REG_SET i1_regset, i2_regset;
1207
1208 CLEAR_HARD_REG_SET (i1_regset);
1209 CLEAR_HARD_REG_SET (i2_regset);
1210
1211 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1212 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1213 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1214
1215 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1216 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1217 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1218
1219 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1220 return dir_none;
1221 }
1222 #endif
1223
1224 if (reload_completed
1225 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1226 return dir_both;
1227
1228 return can_replace_by (i1, i2);
1229 }
1230 \f
1231 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1232 flow_find_head_matching_sequence, ensure the notes match. */
1233
1234 static void
1235 merge_notes (rtx_insn *i1, rtx_insn *i2)
1236 {
1237 /* If the merged insns have different REG_EQUAL notes, then
1238 remove them. */
1239 rtx equiv1 = find_reg_equal_equiv_note (i1);
1240 rtx equiv2 = find_reg_equal_equiv_note (i2);
1241
1242 if (equiv1 && !equiv2)
1243 remove_note (i1, equiv1);
1244 else if (!equiv1 && equiv2)
1245 remove_note (i2, equiv2);
1246 else if (equiv1 && equiv2
1247 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1248 {
1249 remove_note (i1, equiv1);
1250 remove_note (i2, equiv2);
1251 }
1252 }
1253
1254 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1255 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1256 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1257 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1258 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1259
1260 static void
1261 walk_to_nondebug_insn (rtx_insn **i1, basic_block *bb1, bool follow_fallthru,
1262 bool *did_fallthru)
1263 {
1264 edge fallthru;
1265
1266 *did_fallthru = false;
1267
1268 /* Ignore notes. */
1269 while (!NONDEBUG_INSN_P (*i1))
1270 {
1271 if (*i1 != BB_HEAD (*bb1))
1272 {
1273 *i1 = PREV_INSN (*i1);
1274 continue;
1275 }
1276
1277 if (!follow_fallthru)
1278 return;
1279
1280 fallthru = find_fallthru_edge ((*bb1)->preds);
1281 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)
1282 || !single_succ_p (fallthru->src))
1283 return;
1284
1285 *bb1 = fallthru->src;
1286 *i1 = BB_END (*bb1);
1287 *did_fallthru = true;
1288 }
1289 }
1290
1291 /* Look through the insns at the end of BB1 and BB2 and find the longest
1292 sequence that are either equivalent, or allow forward or backward
1293 replacement. Store the first insns for that sequence in *F1 and *F2 and
1294 return the sequence length.
1295
1296 DIR_P indicates the allowed replacement direction on function entry, and
1297 the actual replacement direction on function exit. If NULL, only equivalent
1298 sequences are allowed.
1299
1300 To simplify callers of this function, if the blocks match exactly,
1301 store the head of the blocks in *F1 and *F2. */
1302
1303 int
1304 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx_insn **f1,
1305 rtx_insn **f2, enum replace_direction *dir_p)
1306 {
1307 rtx_insn *i1, *i2, *last1, *last2, *afterlast1, *afterlast2;
1308 int ninsns = 0;
1309 enum replace_direction dir, last_dir, afterlast_dir;
1310 bool follow_fallthru, did_fallthru;
1311
1312 if (dir_p)
1313 dir = *dir_p;
1314 else
1315 dir = dir_both;
1316 afterlast_dir = dir;
1317 last_dir = afterlast_dir;
1318
1319 /* Skip simple jumps at the end of the blocks. Complex jumps still
1320 need to be compared for equivalence, which we'll do below. */
1321
1322 i1 = BB_END (bb1);
1323 last1 = afterlast1 = last2 = afterlast2 = NULL;
1324 if (onlyjump_p (i1)
1325 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1326 {
1327 last1 = i1;
1328 i1 = PREV_INSN (i1);
1329 }
1330
1331 i2 = BB_END (bb2);
1332 if (onlyjump_p (i2)
1333 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1334 {
1335 last2 = i2;
1336 /* Count everything except for unconditional jump as insn.
1337 Don't count any jumps if dir_p is NULL. */
1338 if (!simplejump_p (i2) && !returnjump_p (i2) && last1 && dir_p)
1339 ninsns++;
1340 i2 = PREV_INSN (i2);
1341 }
1342
1343 while (true)
1344 {
1345 /* In the following example, we can replace all jumps to C by jumps to A.
1346
1347 This removes 4 duplicate insns.
1348 [bb A] insn1 [bb C] insn1
1349 insn2 insn2
1350 [bb B] insn3 insn3
1351 insn4 insn4
1352 jump_insn jump_insn
1353
1354 We could also replace all jumps to A by jumps to C, but that leaves B
1355 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1356 step, all jumps to B would be replaced with jumps to the middle of C,
1357 achieving the same result with more effort.
1358 So we allow only the first possibility, which means that we don't allow
1359 fallthru in the block that's being replaced. */
1360
1361 follow_fallthru = dir_p && dir != dir_forward;
1362 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1363 if (did_fallthru)
1364 dir = dir_backward;
1365
1366 follow_fallthru = dir_p && dir != dir_backward;
1367 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1368 if (did_fallthru)
1369 dir = dir_forward;
1370
1371 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1372 break;
1373
1374 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1375 if (dir == dir_none || (!dir_p && dir != dir_both))
1376 break;
1377
1378 merge_memattrs (i1, i2);
1379
1380 /* Don't begin a cross-jump with a NOTE insn. */
1381 if (INSN_P (i1))
1382 {
1383 merge_notes (i1, i2);
1384
1385 afterlast1 = last1, afterlast2 = last2;
1386 last1 = i1, last2 = i2;
1387 afterlast_dir = last_dir;
1388 last_dir = dir;
1389 if (active_insn_p (i1))
1390 ninsns++;
1391 }
1392
1393 i1 = PREV_INSN (i1);
1394 i2 = PREV_INSN (i2);
1395 }
1396
1397 #ifdef HAVE_cc0
1398 /* Don't allow the insn after a compare to be shared by
1399 cross-jumping unless the compare is also shared. */
1400 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && ! sets_cc0_p (last1))
1401 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1402 #endif
1403
1404 /* Include preceding notes and labels in the cross-jump. One,
1405 this may bring us to the head of the blocks as requested above.
1406 Two, it keeps line number notes as matched as may be. */
1407 if (ninsns)
1408 {
1409 bb1 = BLOCK_FOR_INSN (last1);
1410 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1411 last1 = PREV_INSN (last1);
1412
1413 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1414 last1 = PREV_INSN (last1);
1415
1416 bb2 = BLOCK_FOR_INSN (last2);
1417 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1418 last2 = PREV_INSN (last2);
1419
1420 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1421 last2 = PREV_INSN (last2);
1422
1423 *f1 = last1;
1424 *f2 = last2;
1425 }
1426
1427 if (dir_p)
1428 *dir_p = last_dir;
1429 return ninsns;
1430 }
1431
1432 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1433 the head of the two blocks. Do not include jumps at the end.
1434 If STOP_AFTER is nonzero, stop after finding that many matching
1435 instructions. If STOP_AFTER is zero, count all INSN_P insns, if it is
1436 non-zero, only count active insns. */
1437
1438 int
1439 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx_insn **f1,
1440 rtx_insn **f2, int stop_after)
1441 {
1442 rtx_insn *i1, *i2, *last1, *last2, *beforelast1, *beforelast2;
1443 int ninsns = 0;
1444 edge e;
1445 edge_iterator ei;
1446 int nehedges1 = 0, nehedges2 = 0;
1447
1448 FOR_EACH_EDGE (e, ei, bb1->succs)
1449 if (e->flags & EDGE_EH)
1450 nehedges1++;
1451 FOR_EACH_EDGE (e, ei, bb2->succs)
1452 if (e->flags & EDGE_EH)
1453 nehedges2++;
1454
1455 i1 = BB_HEAD (bb1);
1456 i2 = BB_HEAD (bb2);
1457 last1 = beforelast1 = last2 = beforelast2 = NULL;
1458
1459 while (true)
1460 {
1461 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1462 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1463 {
1464 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1465 break;
1466 i1 = NEXT_INSN (i1);
1467 }
1468
1469 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1470 {
1471 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1472 break;
1473 i2 = NEXT_INSN (i2);
1474 }
1475
1476 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1477 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1478 break;
1479
1480 if (NOTE_P (i1) || NOTE_P (i2)
1481 || JUMP_P (i1) || JUMP_P (i2))
1482 break;
1483
1484 /* A sanity check to make sure we're not merging insns with different
1485 effects on EH. If only one of them ends a basic block, it shouldn't
1486 have an EH edge; if both end a basic block, there should be the same
1487 number of EH edges. */
1488 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1489 && nehedges1 > 0)
1490 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1491 && nehedges2 > 0)
1492 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1493 && nehedges1 != nehedges2))
1494 break;
1495
1496 if (old_insns_match_p (0, i1, i2) != dir_both)
1497 break;
1498
1499 merge_memattrs (i1, i2);
1500
1501 /* Don't begin a cross-jump with a NOTE insn. */
1502 if (INSN_P (i1))
1503 {
1504 merge_notes (i1, i2);
1505
1506 beforelast1 = last1, beforelast2 = last2;
1507 last1 = i1, last2 = i2;
1508 if (!stop_after || active_insn_p (i1))
1509 ninsns++;
1510 }
1511
1512 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1513 || (stop_after > 0 && ninsns == stop_after))
1514 break;
1515
1516 i1 = NEXT_INSN (i1);
1517 i2 = NEXT_INSN (i2);
1518 }
1519
1520 #ifdef HAVE_cc0
1521 /* Don't allow a compare to be shared by cross-jumping unless the insn
1522 after the compare is also shared. */
1523 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && sets_cc0_p (last1))
1524 last1 = beforelast1, last2 = beforelast2, ninsns--;
1525 #endif
1526
1527 if (ninsns)
1528 {
1529 *f1 = last1;
1530 *f2 = last2;
1531 }
1532
1533 return ninsns;
1534 }
1535
1536 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1537 the branch instruction. This means that if we commonize the control
1538 flow before end of the basic block, the semantic remains unchanged.
1539
1540 We may assume that there exists one edge with a common destination. */
1541
1542 static bool
1543 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1544 {
1545 int nehedges1 = 0, nehedges2 = 0;
1546 edge fallthru1 = 0, fallthru2 = 0;
1547 edge e1, e2;
1548 edge_iterator ei;
1549
1550 /* If we performed shrink-wrapping, edges to the exit block can
1551 only be distinguished for JUMP_INSNs. The two paths may differ in
1552 whether they went through the prologue. Sibcalls are fine, we know
1553 that we either didn't need or inserted an epilogue before them. */
1554 if (crtl->shrink_wrapped
1555 && single_succ_p (bb1)
1556 && single_succ (bb1) == EXIT_BLOCK_PTR_FOR_FN (cfun)
1557 && !JUMP_P (BB_END (bb1))
1558 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1559 return false;
1560
1561 /* If BB1 has only one successor, we may be looking at either an
1562 unconditional jump, or a fake edge to exit. */
1563 if (single_succ_p (bb1)
1564 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1565 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1566 return (single_succ_p (bb2)
1567 && (single_succ_edge (bb2)->flags
1568 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1569 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1570
1571 /* Match conditional jumps - this may get tricky when fallthru and branch
1572 edges are crossed. */
1573 if (EDGE_COUNT (bb1->succs) == 2
1574 && any_condjump_p (BB_END (bb1))
1575 && onlyjump_p (BB_END (bb1)))
1576 {
1577 edge b1, f1, b2, f2;
1578 bool reverse, match;
1579 rtx set1, set2, cond1, cond2;
1580 enum rtx_code code1, code2;
1581
1582 if (EDGE_COUNT (bb2->succs) != 2
1583 || !any_condjump_p (BB_END (bb2))
1584 || !onlyjump_p (BB_END (bb2)))
1585 return false;
1586
1587 b1 = BRANCH_EDGE (bb1);
1588 b2 = BRANCH_EDGE (bb2);
1589 f1 = FALLTHRU_EDGE (bb1);
1590 f2 = FALLTHRU_EDGE (bb2);
1591
1592 /* Get around possible forwarders on fallthru edges. Other cases
1593 should be optimized out already. */
1594 if (FORWARDER_BLOCK_P (f1->dest))
1595 f1 = single_succ_edge (f1->dest);
1596
1597 if (FORWARDER_BLOCK_P (f2->dest))
1598 f2 = single_succ_edge (f2->dest);
1599
1600 /* To simplify use of this function, return false if there are
1601 unneeded forwarder blocks. These will get eliminated later
1602 during cleanup_cfg. */
1603 if (FORWARDER_BLOCK_P (f1->dest)
1604 || FORWARDER_BLOCK_P (f2->dest)
1605 || FORWARDER_BLOCK_P (b1->dest)
1606 || FORWARDER_BLOCK_P (b2->dest))
1607 return false;
1608
1609 if (f1->dest == f2->dest && b1->dest == b2->dest)
1610 reverse = false;
1611 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1612 reverse = true;
1613 else
1614 return false;
1615
1616 set1 = pc_set (BB_END (bb1));
1617 set2 = pc_set (BB_END (bb2));
1618 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1619 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1620 reverse = !reverse;
1621
1622 cond1 = XEXP (SET_SRC (set1), 0);
1623 cond2 = XEXP (SET_SRC (set2), 0);
1624 code1 = GET_CODE (cond1);
1625 if (reverse)
1626 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1627 else
1628 code2 = GET_CODE (cond2);
1629
1630 if (code2 == UNKNOWN)
1631 return false;
1632
1633 /* Verify codes and operands match. */
1634 match = ((code1 == code2
1635 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1636 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1637 || (code1 == swap_condition (code2)
1638 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1639 XEXP (cond2, 0))
1640 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1641 XEXP (cond2, 1))));
1642
1643 /* If we return true, we will join the blocks. Which means that
1644 we will only have one branch prediction bit to work with. Thus
1645 we require the existing branches to have probabilities that are
1646 roughly similar. */
1647 if (match
1648 && optimize_bb_for_speed_p (bb1)
1649 && optimize_bb_for_speed_p (bb2))
1650 {
1651 int prob2;
1652
1653 if (b1->dest == b2->dest)
1654 prob2 = b2->probability;
1655 else
1656 /* Do not use f2 probability as f2 may be forwarded. */
1657 prob2 = REG_BR_PROB_BASE - b2->probability;
1658
1659 /* Fail if the difference in probabilities is greater than 50%.
1660 This rules out two well-predicted branches with opposite
1661 outcomes. */
1662 if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
1663 {
1664 if (dump_file)
1665 fprintf (dump_file,
1666 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1667 bb1->index, bb2->index, b1->probability, prob2);
1668
1669 return false;
1670 }
1671 }
1672
1673 if (dump_file && match)
1674 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1675 bb1->index, bb2->index);
1676
1677 return match;
1678 }
1679
1680 /* Generic case - we are seeing a computed jump, table jump or trapping
1681 instruction. */
1682
1683 /* Check whether there are tablejumps in the end of BB1 and BB2.
1684 Return true if they are identical. */
1685 {
1686 rtx label1, label2;
1687 rtx_jump_table_data *table1, *table2;
1688
1689 if (tablejump_p (BB_END (bb1), &label1, &table1)
1690 && tablejump_p (BB_END (bb2), &label2, &table2)
1691 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1692 {
1693 /* The labels should never be the same rtx. If they really are same
1694 the jump tables are same too. So disable crossjumping of blocks BB1
1695 and BB2 because when deleting the common insns in the end of BB1
1696 by delete_basic_block () the jump table would be deleted too. */
1697 /* If LABEL2 is referenced in BB1->END do not do anything
1698 because we would loose information when replacing
1699 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1700 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1701 {
1702 /* Set IDENTICAL to true when the tables are identical. */
1703 bool identical = false;
1704 rtx p1, p2;
1705
1706 p1 = PATTERN (table1);
1707 p2 = PATTERN (table2);
1708 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1709 {
1710 identical = true;
1711 }
1712 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1713 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1714 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1715 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1716 {
1717 int i;
1718
1719 identical = true;
1720 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1721 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1722 identical = false;
1723 }
1724
1725 if (identical)
1726 {
1727 bool match;
1728
1729 /* Temporarily replace references to LABEL1 with LABEL2
1730 in BB1->END so that we could compare the instructions. */
1731 replace_label_in_insn (BB_END (bb1), label1, label2, false);
1732
1733 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1734 == dir_both);
1735 if (dump_file && match)
1736 fprintf (dump_file,
1737 "Tablejumps in bb %i and %i match.\n",
1738 bb1->index, bb2->index);
1739
1740 /* Set the original label in BB1->END because when deleting
1741 a block whose end is a tablejump, the tablejump referenced
1742 from the instruction is deleted too. */
1743 replace_label_in_insn (BB_END (bb1), label2, label1, false);
1744
1745 return match;
1746 }
1747 }
1748 return false;
1749 }
1750 }
1751
1752 /* Find the last non-debug non-note instruction in each bb, except
1753 stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
1754 handles that case specially. old_insns_match_p does not handle
1755 other types of instruction notes. */
1756 rtx_insn *last1 = BB_END (bb1);
1757 rtx_insn *last2 = BB_END (bb2);
1758 while (!NOTE_INSN_BASIC_BLOCK_P (last1) &&
1759 (DEBUG_INSN_P (last1) || NOTE_P (last1)))
1760 last1 = PREV_INSN (last1);
1761 while (!NOTE_INSN_BASIC_BLOCK_P (last2) &&
1762 (DEBUG_INSN_P (last2) || NOTE_P (last2)))
1763 last2 = PREV_INSN (last2);
1764 gcc_assert (last1 && last2);
1765
1766 /* First ensure that the instructions match. There may be many outgoing
1767 edges so this test is generally cheaper. */
1768 if (old_insns_match_p (mode, last1, last2) != dir_both)
1769 return false;
1770
1771 /* Search the outgoing edges, ensure that the counts do match, find possible
1772 fallthru and exception handling edges since these needs more
1773 validation. */
1774 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1775 return false;
1776
1777 bool nonfakeedges = false;
1778 FOR_EACH_EDGE (e1, ei, bb1->succs)
1779 {
1780 e2 = EDGE_SUCC (bb2, ei.index);
1781
1782 if ((e1->flags & EDGE_FAKE) == 0)
1783 nonfakeedges = true;
1784
1785 if (e1->flags & EDGE_EH)
1786 nehedges1++;
1787
1788 if (e2->flags & EDGE_EH)
1789 nehedges2++;
1790
1791 if (e1->flags & EDGE_FALLTHRU)
1792 fallthru1 = e1;
1793 if (e2->flags & EDGE_FALLTHRU)
1794 fallthru2 = e2;
1795 }
1796
1797 /* If number of edges of various types does not match, fail. */
1798 if (nehedges1 != nehedges2
1799 || (fallthru1 != 0) != (fallthru2 != 0))
1800 return false;
1801
1802 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1803 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1804 attempt to optimize, as the two basic blocks might have different
1805 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1806 traps there should be REG_ARG_SIZE notes, they could be missing
1807 for __builtin_unreachable () uses though. */
1808 if (!nonfakeedges
1809 && !ACCUMULATE_OUTGOING_ARGS
1810 && (!INSN_P (last1)
1811 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1812 return false;
1813
1814 /* fallthru edges must be forwarded to the same destination. */
1815 if (fallthru1)
1816 {
1817 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1818 ? single_succ (fallthru1->dest): fallthru1->dest);
1819 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1820 ? single_succ (fallthru2->dest): fallthru2->dest);
1821
1822 if (d1 != d2)
1823 return false;
1824 }
1825
1826 /* Ensure the same EH region. */
1827 {
1828 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1829 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1830
1831 if (!n1 && n2)
1832 return false;
1833
1834 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1835 return false;
1836 }
1837
1838 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1839 version of sequence abstraction. */
1840 FOR_EACH_EDGE (e1, ei, bb2->succs)
1841 {
1842 edge e2;
1843 edge_iterator ei;
1844 basic_block d1 = e1->dest;
1845
1846 if (FORWARDER_BLOCK_P (d1))
1847 d1 = EDGE_SUCC (d1, 0)->dest;
1848
1849 FOR_EACH_EDGE (e2, ei, bb1->succs)
1850 {
1851 basic_block d2 = e2->dest;
1852 if (FORWARDER_BLOCK_P (d2))
1853 d2 = EDGE_SUCC (d2, 0)->dest;
1854 if (d1 == d2)
1855 break;
1856 }
1857
1858 if (!e2)
1859 return false;
1860 }
1861
1862 return true;
1863 }
1864
1865 /* Returns true if BB basic block has a preserve label. */
1866
1867 static bool
1868 block_has_preserve_label (basic_block bb)
1869 {
1870 return (bb
1871 && block_label (bb)
1872 && LABEL_PRESERVE_P (block_label (bb)));
1873 }
1874
1875 /* E1 and E2 are edges with the same destination block. Search their
1876 predecessors for common code. If found, redirect control flow from
1877 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1878 or the other way around (dir_backward). DIR specifies the allowed
1879 replacement direction. */
1880
1881 static bool
1882 try_crossjump_to_edge (int mode, edge e1, edge e2,
1883 enum replace_direction dir)
1884 {
1885 int nmatch;
1886 basic_block src1 = e1->src, src2 = e2->src;
1887 basic_block redirect_to, redirect_from, to_remove;
1888 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1889 rtx_insn *newpos1, *newpos2;
1890 edge s;
1891 edge_iterator ei;
1892
1893 newpos1 = newpos2 = NULL;
1894
1895 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1896 to try this optimization.
1897
1898 Basic block partitioning may result in some jumps that appear to
1899 be optimizable (or blocks that appear to be mergeable), but which really
1900 must be left untouched (they are required to make it safely across
1901 partition boundaries). See the comments at the top of
1902 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1903
1904 if (crtl->has_bb_partition && reload_completed)
1905 return false;
1906
1907 /* Search backward through forwarder blocks. We don't need to worry
1908 about multiple entry or chained forwarders, as they will be optimized
1909 away. We do this to look past the unconditional jump following a
1910 conditional jump that is required due to the current CFG shape. */
1911 if (single_pred_p (src1)
1912 && FORWARDER_BLOCK_P (src1))
1913 e1 = single_pred_edge (src1), src1 = e1->src;
1914
1915 if (single_pred_p (src2)
1916 && FORWARDER_BLOCK_P (src2))
1917 e2 = single_pred_edge (src2), src2 = e2->src;
1918
1919 /* Nothing to do if we reach ENTRY, or a common source block. */
1920 if (src1 == ENTRY_BLOCK_PTR_FOR_FN (cfun) || src2
1921 == ENTRY_BLOCK_PTR_FOR_FN (cfun))
1922 return false;
1923 if (src1 == src2)
1924 return false;
1925
1926 /* Seeing more than 1 forwarder blocks would confuse us later... */
1927 if (FORWARDER_BLOCK_P (e1->dest)
1928 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1929 return false;
1930
1931 if (FORWARDER_BLOCK_P (e2->dest)
1932 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1933 return false;
1934
1935 /* Likewise with dead code (possibly newly created by the other optimizations
1936 of cfg_cleanup). */
1937 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1938 return false;
1939
1940 /* Look for the common insn sequence, part the first ... */
1941 if (!outgoing_edges_match (mode, src1, src2))
1942 return false;
1943
1944 /* ... and part the second. */
1945 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
1946
1947 osrc1 = src1;
1948 osrc2 = src2;
1949 if (newpos1 != NULL_RTX)
1950 src1 = BLOCK_FOR_INSN (newpos1);
1951 if (newpos2 != NULL_RTX)
1952 src2 = BLOCK_FOR_INSN (newpos2);
1953
1954 if (dir == dir_backward)
1955 {
1956 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
1957 SWAP (basic_block, osrc1, osrc2);
1958 SWAP (basic_block, src1, src2);
1959 SWAP (edge, e1, e2);
1960 SWAP (rtx_insn *, newpos1, newpos2);
1961 #undef SWAP
1962 }
1963
1964 /* Don't proceed with the crossjump unless we found a sufficient number
1965 of matching instructions or the 'from' block was totally matched
1966 (such that its predecessors will hopefully be redirected and the
1967 block removed). */
1968 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
1969 && (newpos1 != BB_HEAD (src1)))
1970 return false;
1971
1972 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
1973 if (block_has_preserve_label (e1->dest)
1974 && (e1->flags & EDGE_ABNORMAL))
1975 return false;
1976
1977 /* Here we know that the insns in the end of SRC1 which are common with SRC2
1978 will be deleted.
1979 If we have tablejumps in the end of SRC1 and SRC2
1980 they have been already compared for equivalence in outgoing_edges_match ()
1981 so replace the references to TABLE1 by references to TABLE2. */
1982 {
1983 rtx label1, label2;
1984 rtx_jump_table_data *table1, *table2;
1985
1986 if (tablejump_p (BB_END (osrc1), &label1, &table1)
1987 && tablejump_p (BB_END (osrc2), &label2, &table2)
1988 && label1 != label2)
1989 {
1990 rtx_insn *insn;
1991
1992 /* Replace references to LABEL1 with LABEL2. */
1993 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
1994 {
1995 /* Do not replace the label in SRC1->END because when deleting
1996 a block whose end is a tablejump, the tablejump referenced
1997 from the instruction is deleted too. */
1998 if (insn != BB_END (osrc1))
1999 replace_label_in_insn (insn, label1, label2, true);
2000 }
2001 }
2002 }
2003
2004 /* Avoid splitting if possible. We must always split when SRC2 has
2005 EH predecessor edges, or we may end up with basic blocks with both
2006 normal and EH predecessor edges. */
2007 if (newpos2 == BB_HEAD (src2)
2008 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
2009 redirect_to = src2;
2010 else
2011 {
2012 if (newpos2 == BB_HEAD (src2))
2013 {
2014 /* Skip possible basic block header. */
2015 if (LABEL_P (newpos2))
2016 newpos2 = NEXT_INSN (newpos2);
2017 while (DEBUG_INSN_P (newpos2))
2018 newpos2 = NEXT_INSN (newpos2);
2019 if (NOTE_P (newpos2))
2020 newpos2 = NEXT_INSN (newpos2);
2021 while (DEBUG_INSN_P (newpos2))
2022 newpos2 = NEXT_INSN (newpos2);
2023 }
2024
2025 if (dump_file)
2026 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2027 src2->index, nmatch);
2028 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2029 }
2030
2031 if (dump_file)
2032 fprintf (dump_file,
2033 "Cross jumping from bb %i to bb %i; %i common insns\n",
2034 src1->index, src2->index, nmatch);
2035
2036 /* We may have some registers visible through the block. */
2037 df_set_bb_dirty (redirect_to);
2038
2039 if (osrc2 == src2)
2040 redirect_edges_to = redirect_to;
2041 else
2042 redirect_edges_to = osrc2;
2043
2044 /* Recompute the frequencies and counts of outgoing edges. */
2045 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2046 {
2047 edge s2;
2048 edge_iterator ei;
2049 basic_block d = s->dest;
2050
2051 if (FORWARDER_BLOCK_P (d))
2052 d = single_succ (d);
2053
2054 FOR_EACH_EDGE (s2, ei, src1->succs)
2055 {
2056 basic_block d2 = s2->dest;
2057 if (FORWARDER_BLOCK_P (d2))
2058 d2 = single_succ (d2);
2059 if (d == d2)
2060 break;
2061 }
2062
2063 s->count += s2->count;
2064
2065 /* Take care to update possible forwarder blocks. We verified
2066 that there is no more than one in the chain, so we can't run
2067 into infinite loop. */
2068 if (FORWARDER_BLOCK_P (s->dest))
2069 {
2070 single_succ_edge (s->dest)->count += s2->count;
2071 s->dest->count += s2->count;
2072 s->dest->frequency += EDGE_FREQUENCY (s);
2073 }
2074
2075 if (FORWARDER_BLOCK_P (s2->dest))
2076 {
2077 single_succ_edge (s2->dest)->count -= s2->count;
2078 if (single_succ_edge (s2->dest)->count < 0)
2079 single_succ_edge (s2->dest)->count = 0;
2080 s2->dest->count -= s2->count;
2081 s2->dest->frequency -= EDGE_FREQUENCY (s);
2082 if (s2->dest->frequency < 0)
2083 s2->dest->frequency = 0;
2084 if (s2->dest->count < 0)
2085 s2->dest->count = 0;
2086 }
2087
2088 if (!redirect_edges_to->frequency && !src1->frequency)
2089 s->probability = (s->probability + s2->probability) / 2;
2090 else
2091 s->probability
2092 = ((s->probability * redirect_edges_to->frequency +
2093 s2->probability * src1->frequency)
2094 / (redirect_edges_to->frequency + src1->frequency));
2095 }
2096
2097 /* Adjust count and frequency for the block. An earlier jump
2098 threading pass may have left the profile in an inconsistent
2099 state (see update_bb_profile_for_threading) so we must be
2100 prepared for overflows. */
2101 tmp = redirect_to;
2102 do
2103 {
2104 tmp->count += src1->count;
2105 tmp->frequency += src1->frequency;
2106 if (tmp->frequency > BB_FREQ_MAX)
2107 tmp->frequency = BB_FREQ_MAX;
2108 if (tmp == redirect_edges_to)
2109 break;
2110 tmp = find_fallthru_edge (tmp->succs)->dest;
2111 }
2112 while (true);
2113 update_br_prob_note (redirect_edges_to);
2114
2115 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2116
2117 /* Skip possible basic block header. */
2118 if (LABEL_P (newpos1))
2119 newpos1 = NEXT_INSN (newpos1);
2120
2121 while (DEBUG_INSN_P (newpos1))
2122 newpos1 = NEXT_INSN (newpos1);
2123
2124 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2125 newpos1 = NEXT_INSN (newpos1);
2126
2127 while (DEBUG_INSN_P (newpos1))
2128 newpos1 = NEXT_INSN (newpos1);
2129
2130 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2131 to_remove = single_succ (redirect_from);
2132
2133 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2134 delete_basic_block (to_remove);
2135
2136 update_forwarder_flag (redirect_from);
2137 if (redirect_to != src2)
2138 update_forwarder_flag (src2);
2139
2140 return true;
2141 }
2142
2143 /* Search the predecessors of BB for common insn sequences. When found,
2144 share code between them by redirecting control flow. Return true if
2145 any changes made. */
2146
2147 static bool
2148 try_crossjump_bb (int mode, basic_block bb)
2149 {
2150 edge e, e2, fallthru;
2151 bool changed;
2152 unsigned max, ix, ix2;
2153
2154 /* Nothing to do if there is not at least two incoming edges. */
2155 if (EDGE_COUNT (bb->preds) < 2)
2156 return false;
2157
2158 /* Don't crossjump if this block ends in a computed jump,
2159 unless we are optimizing for size. */
2160 if (optimize_bb_for_size_p (bb)
2161 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2162 && computed_jump_p (BB_END (bb)))
2163 return false;
2164
2165 /* If we are partitioning hot/cold basic blocks, we don't want to
2166 mess up unconditional or indirect jumps that cross between hot
2167 and cold sections.
2168
2169 Basic block partitioning may result in some jumps that appear to
2170 be optimizable (or blocks that appear to be mergeable), but which really
2171 must be left untouched (they are required to make it safely across
2172 partition boundaries). See the comments at the top of
2173 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2174
2175 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2176 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2177 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2178 return false;
2179
2180 /* It is always cheapest to redirect a block that ends in a branch to
2181 a block that falls through into BB, as that adds no branches to the
2182 program. We'll try that combination first. */
2183 fallthru = NULL;
2184 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2185
2186 if (EDGE_COUNT (bb->preds) > max)
2187 return false;
2188
2189 fallthru = find_fallthru_edge (bb->preds);
2190
2191 changed = false;
2192 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2193 {
2194 e = EDGE_PRED (bb, ix);
2195 ix++;
2196
2197 /* As noted above, first try with the fallthru predecessor (or, a
2198 fallthru predecessor if we are in cfglayout mode). */
2199 if (fallthru)
2200 {
2201 /* Don't combine the fallthru edge into anything else.
2202 If there is a match, we'll do it the other way around. */
2203 if (e == fallthru)
2204 continue;
2205 /* If nothing changed since the last attempt, there is nothing
2206 we can do. */
2207 if (!first_pass
2208 && !((e->src->flags & BB_MODIFIED)
2209 || (fallthru->src->flags & BB_MODIFIED)))
2210 continue;
2211
2212 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2213 {
2214 changed = true;
2215 ix = 0;
2216 continue;
2217 }
2218 }
2219
2220 /* Non-obvious work limiting check: Recognize that we're going
2221 to call try_crossjump_bb on every basic block. So if we have
2222 two blocks with lots of outgoing edges (a switch) and they
2223 share lots of common destinations, then we would do the
2224 cross-jump check once for each common destination.
2225
2226 Now, if the blocks actually are cross-jump candidates, then
2227 all of their destinations will be shared. Which means that
2228 we only need check them for cross-jump candidacy once. We
2229 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2230 choosing to do the check from the block for which the edge
2231 in question is the first successor of A. */
2232 if (EDGE_SUCC (e->src, 0) != e)
2233 continue;
2234
2235 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2236 {
2237 e2 = EDGE_PRED (bb, ix2);
2238
2239 if (e2 == e)
2240 continue;
2241
2242 /* We've already checked the fallthru edge above. */
2243 if (e2 == fallthru)
2244 continue;
2245
2246 /* The "first successor" check above only prevents multiple
2247 checks of crossjump(A,B). In order to prevent redundant
2248 checks of crossjump(B,A), require that A be the block
2249 with the lowest index. */
2250 if (e->src->index > e2->src->index)
2251 continue;
2252
2253 /* If nothing changed since the last attempt, there is nothing
2254 we can do. */
2255 if (!first_pass
2256 && !((e->src->flags & BB_MODIFIED)
2257 || (e2->src->flags & BB_MODIFIED)))
2258 continue;
2259
2260 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2261 direction. */
2262 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2263 {
2264 changed = true;
2265 ix = 0;
2266 break;
2267 }
2268 }
2269 }
2270
2271 if (changed)
2272 crossjumps_occured = true;
2273
2274 return changed;
2275 }
2276
2277 /* Search the successors of BB for common insn sequences. When found,
2278 share code between them by moving it across the basic block
2279 boundary. Return true if any changes made. */
2280
2281 static bool
2282 try_head_merge_bb (basic_block bb)
2283 {
2284 basic_block final_dest_bb = NULL;
2285 int max_match = INT_MAX;
2286 edge e0;
2287 rtx_insn **headptr, **currptr, **nextptr;
2288 bool changed, moveall;
2289 unsigned ix;
2290 rtx_insn *e0_last_head;
2291 rtx cond;
2292 rtx_insn *move_before;
2293 unsigned nedges = EDGE_COUNT (bb->succs);
2294 rtx_insn *jump = BB_END (bb);
2295 regset live, live_union;
2296
2297 /* Nothing to do if there is not at least two outgoing edges. */
2298 if (nedges < 2)
2299 return false;
2300
2301 /* Don't crossjump if this block ends in a computed jump,
2302 unless we are optimizing for size. */
2303 if (optimize_bb_for_size_p (bb)
2304 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2305 && computed_jump_p (BB_END (bb)))
2306 return false;
2307
2308 cond = get_condition (jump, &move_before, true, false);
2309 if (cond == NULL_RTX)
2310 {
2311 #ifdef HAVE_cc0
2312 if (reg_mentioned_p (cc0_rtx, jump))
2313 move_before = prev_nonnote_nondebug_insn (jump);
2314 else
2315 #endif
2316 move_before = jump;
2317 }
2318
2319 for (ix = 0; ix < nedges; ix++)
2320 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
2321 return false;
2322
2323 for (ix = 0; ix < nedges; ix++)
2324 {
2325 edge e = EDGE_SUCC (bb, ix);
2326 basic_block other_bb = e->dest;
2327
2328 if (df_get_bb_dirty (other_bb))
2329 {
2330 block_was_dirty = true;
2331 return false;
2332 }
2333
2334 if (e->flags & EDGE_ABNORMAL)
2335 return false;
2336
2337 /* Normally, all destination blocks must only be reachable from this
2338 block, i.e. they must have one incoming edge.
2339
2340 There is one special case we can handle, that of multiple consecutive
2341 jumps where the first jumps to one of the targets of the second jump.
2342 This happens frequently in switch statements for default labels.
2343 The structure is as follows:
2344 FINAL_DEST_BB
2345 ....
2346 if (cond) jump A;
2347 fall through
2348 BB
2349 jump with targets A, B, C, D...
2350 A
2351 has two incoming edges, from FINAL_DEST_BB and BB
2352
2353 In this case, we can try to move the insns through BB and into
2354 FINAL_DEST_BB. */
2355 if (EDGE_COUNT (other_bb->preds) != 1)
2356 {
2357 edge incoming_edge, incoming_bb_other_edge;
2358 edge_iterator ei;
2359
2360 if (final_dest_bb != NULL
2361 || EDGE_COUNT (other_bb->preds) != 2)
2362 return false;
2363
2364 /* We must be able to move the insns across the whole block. */
2365 move_before = BB_HEAD (bb);
2366 while (!NONDEBUG_INSN_P (move_before))
2367 move_before = NEXT_INSN (move_before);
2368
2369 if (EDGE_COUNT (bb->preds) != 1)
2370 return false;
2371 incoming_edge = EDGE_PRED (bb, 0);
2372 final_dest_bb = incoming_edge->src;
2373 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2374 return false;
2375 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2376 if (incoming_bb_other_edge != incoming_edge)
2377 break;
2378 if (incoming_bb_other_edge->dest != other_bb)
2379 return false;
2380 }
2381 }
2382
2383 e0 = EDGE_SUCC (bb, 0);
2384 e0_last_head = NULL;
2385 changed = false;
2386
2387 for (ix = 1; ix < nedges; ix++)
2388 {
2389 edge e = EDGE_SUCC (bb, ix);
2390 rtx_insn *e0_last, *e_last;
2391 int nmatch;
2392
2393 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2394 &e0_last, &e_last, 0);
2395 if (nmatch == 0)
2396 return false;
2397
2398 if (nmatch < max_match)
2399 {
2400 max_match = nmatch;
2401 e0_last_head = e0_last;
2402 }
2403 }
2404
2405 /* If we matched an entire block, we probably have to avoid moving the
2406 last insn. */
2407 if (max_match > 0
2408 && e0_last_head == BB_END (e0->dest)
2409 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2410 || control_flow_insn_p (e0_last_head)))
2411 {
2412 max_match--;
2413 if (max_match == 0)
2414 return false;
2415 do
2416 e0_last_head = prev_real_insn (e0_last_head);
2417 while (DEBUG_INSN_P (e0_last_head));
2418 }
2419
2420 if (max_match == 0)
2421 return false;
2422
2423 /* We must find a union of the live registers at each of the end points. */
2424 live = BITMAP_ALLOC (NULL);
2425 live_union = BITMAP_ALLOC (NULL);
2426
2427 currptr = XNEWVEC (rtx_insn *, nedges);
2428 headptr = XNEWVEC (rtx_insn *, nedges);
2429 nextptr = XNEWVEC (rtx_insn *, nedges);
2430
2431 for (ix = 0; ix < nedges; ix++)
2432 {
2433 int j;
2434 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2435 rtx_insn *head = BB_HEAD (merge_bb);
2436
2437 while (!NONDEBUG_INSN_P (head))
2438 head = NEXT_INSN (head);
2439 headptr[ix] = head;
2440 currptr[ix] = head;
2441
2442 /* Compute the end point and live information */
2443 for (j = 1; j < max_match; j++)
2444 do
2445 head = NEXT_INSN (head);
2446 while (!NONDEBUG_INSN_P (head));
2447 simulate_backwards_to_point (merge_bb, live, head);
2448 IOR_REG_SET (live_union, live);
2449 }
2450
2451 /* If we're moving across two blocks, verify the validity of the
2452 first move, then adjust the target and let the loop below deal
2453 with the final move. */
2454 if (final_dest_bb != NULL)
2455 {
2456 rtx_insn *move_upto;
2457
2458 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2459 jump, e0->dest, live_union,
2460 NULL, &move_upto);
2461 if (!moveall)
2462 {
2463 if (move_upto == NULL_RTX)
2464 goto out;
2465
2466 while (e0_last_head != move_upto)
2467 {
2468 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2469 live_union);
2470 e0_last_head = PREV_INSN (e0_last_head);
2471 }
2472 }
2473 if (e0_last_head == NULL_RTX)
2474 goto out;
2475
2476 jump = BB_END (final_dest_bb);
2477 cond = get_condition (jump, &move_before, true, false);
2478 if (cond == NULL_RTX)
2479 {
2480 #ifdef HAVE_cc0
2481 if (reg_mentioned_p (cc0_rtx, jump))
2482 move_before = prev_nonnote_nondebug_insn (jump);
2483 else
2484 #endif
2485 move_before = jump;
2486 }
2487 }
2488
2489 do
2490 {
2491 rtx_insn *move_upto;
2492 moveall = can_move_insns_across (currptr[0], e0_last_head,
2493 move_before, jump, e0->dest, live_union,
2494 NULL, &move_upto);
2495 if (!moveall && move_upto == NULL_RTX)
2496 {
2497 if (jump == move_before)
2498 break;
2499
2500 /* Try again, using a different insertion point. */
2501 move_before = jump;
2502
2503 #ifdef HAVE_cc0
2504 /* Don't try moving before a cc0 user, as that may invalidate
2505 the cc0. */
2506 if (reg_mentioned_p (cc0_rtx, jump))
2507 break;
2508 #endif
2509
2510 continue;
2511 }
2512
2513 if (final_dest_bb && !moveall)
2514 /* We haven't checked whether a partial move would be OK for the first
2515 move, so we have to fail this case. */
2516 break;
2517
2518 changed = true;
2519 for (;;)
2520 {
2521 if (currptr[0] == move_upto)
2522 break;
2523 for (ix = 0; ix < nedges; ix++)
2524 {
2525 rtx_insn *curr = currptr[ix];
2526 do
2527 curr = NEXT_INSN (curr);
2528 while (!NONDEBUG_INSN_P (curr));
2529 currptr[ix] = curr;
2530 }
2531 }
2532
2533 /* If we can't currently move all of the identical insns, remember
2534 each insn after the range that we'll merge. */
2535 if (!moveall)
2536 for (ix = 0; ix < nedges; ix++)
2537 {
2538 rtx_insn *curr = currptr[ix];
2539 do
2540 curr = NEXT_INSN (curr);
2541 while (!NONDEBUG_INSN_P (curr));
2542 nextptr[ix] = curr;
2543 }
2544
2545 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2546 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2547 if (final_dest_bb != NULL)
2548 df_set_bb_dirty (final_dest_bb);
2549 df_set_bb_dirty (bb);
2550 for (ix = 1; ix < nedges; ix++)
2551 {
2552 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2553 delete_insn_chain (headptr[ix], currptr[ix], false);
2554 }
2555 if (!moveall)
2556 {
2557 if (jump == move_before)
2558 break;
2559
2560 /* For the unmerged insns, try a different insertion point. */
2561 move_before = jump;
2562
2563 #ifdef HAVE_cc0
2564 /* Don't try moving before a cc0 user, as that may invalidate
2565 the cc0. */
2566 if (reg_mentioned_p (cc0_rtx, jump))
2567 break;
2568 #endif
2569
2570 for (ix = 0; ix < nedges; ix++)
2571 currptr[ix] = headptr[ix] = nextptr[ix];
2572 }
2573 }
2574 while (!moveall);
2575
2576 out:
2577 free (currptr);
2578 free (headptr);
2579 free (nextptr);
2580
2581 crossjumps_occured |= changed;
2582
2583 return changed;
2584 }
2585
2586 /* Return true if BB contains just bb note, or bb note followed
2587 by only DEBUG_INSNs. */
2588
2589 static bool
2590 trivially_empty_bb_p (basic_block bb)
2591 {
2592 rtx_insn *insn = BB_END (bb);
2593
2594 while (1)
2595 {
2596 if (insn == BB_HEAD (bb))
2597 return true;
2598 if (!DEBUG_INSN_P (insn))
2599 return false;
2600 insn = PREV_INSN (insn);
2601 }
2602 }
2603
2604 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2605 instructions etc. Return nonzero if changes were made. */
2606
2607 static bool
2608 try_optimize_cfg (int mode)
2609 {
2610 bool changed_overall = false;
2611 bool changed;
2612 int iterations = 0;
2613 basic_block bb, b, next;
2614
2615 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2616 clear_bb_flags ();
2617
2618 crossjumps_occured = false;
2619
2620 FOR_EACH_BB_FN (bb, cfun)
2621 update_forwarder_flag (bb);
2622
2623 if (! targetm.cannot_modify_jumps_p ())
2624 {
2625 first_pass = true;
2626 /* Attempt to merge blocks as made possible by edge removal. If
2627 a block has only one successor, and the successor has only
2628 one predecessor, they may be combined. */
2629 do
2630 {
2631 block_was_dirty = false;
2632 changed = false;
2633 iterations++;
2634
2635 if (dump_file)
2636 fprintf (dump_file,
2637 "\n\ntry_optimize_cfg iteration %i\n\n",
2638 iterations);
2639
2640 for (b = ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb; b
2641 != EXIT_BLOCK_PTR_FOR_FN (cfun);)
2642 {
2643 basic_block c;
2644 edge s;
2645 bool changed_here = false;
2646
2647 /* Delete trivially dead basic blocks. This is either
2648 blocks with no predecessors, or empty blocks with no
2649 successors. However if the empty block with no
2650 successors is the successor of the ENTRY_BLOCK, it is
2651 kept. This ensures that the ENTRY_BLOCK will have a
2652 successor which is a precondition for many RTL
2653 passes. Empty blocks may result from expanding
2654 __builtin_unreachable (). */
2655 if (EDGE_COUNT (b->preds) == 0
2656 || (EDGE_COUNT (b->succs) == 0
2657 && trivially_empty_bb_p (b)
2658 && single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))->dest
2659 != b))
2660 {
2661 c = b->prev_bb;
2662 if (EDGE_COUNT (b->preds) > 0)
2663 {
2664 edge e;
2665 edge_iterator ei;
2666
2667 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2668 {
2669 if (BB_FOOTER (b)
2670 && BARRIER_P (BB_FOOTER (b)))
2671 FOR_EACH_EDGE (e, ei, b->preds)
2672 if ((e->flags & EDGE_FALLTHRU)
2673 && BB_FOOTER (e->src) == NULL)
2674 {
2675 if (BB_FOOTER (b))
2676 {
2677 BB_FOOTER (e->src) = BB_FOOTER (b);
2678 BB_FOOTER (b) = NULL;
2679 }
2680 else
2681 {
2682 start_sequence ();
2683 BB_FOOTER (e->src) = emit_barrier ();
2684 end_sequence ();
2685 }
2686 }
2687 }
2688 else
2689 {
2690 rtx_insn *last = get_last_bb_insn (b);
2691 if (last && BARRIER_P (last))
2692 FOR_EACH_EDGE (e, ei, b->preds)
2693 if ((e->flags & EDGE_FALLTHRU))
2694 emit_barrier_after (BB_END (e->src));
2695 }
2696 }
2697 delete_basic_block (b);
2698 changed = true;
2699 /* Avoid trying to remove the exit block. */
2700 b = (c == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? c->next_bb : c);
2701 continue;
2702 }
2703
2704 /* Remove code labels no longer used. */
2705 if (single_pred_p (b)
2706 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2707 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2708 && LABEL_P (BB_HEAD (b))
2709 && !LABEL_PRESERVE_P (BB_HEAD (b))
2710 /* If the previous block ends with a branch to this
2711 block, we can't delete the label. Normally this
2712 is a condjump that is yet to be simplified, but
2713 if CASE_DROPS_THRU, this can be a tablejump with
2714 some element going to the same place as the
2715 default (fallthru). */
2716 && (single_pred (b) == ENTRY_BLOCK_PTR_FOR_FN (cfun)
2717 || !JUMP_P (BB_END (single_pred (b)))
2718 || ! label_is_jump_target_p (BB_HEAD (b),
2719 BB_END (single_pred (b)))))
2720 {
2721 delete_insn (BB_HEAD (b));
2722 if (dump_file)
2723 fprintf (dump_file, "Deleted label in block %i.\n",
2724 b->index);
2725 }
2726
2727 /* If we fall through an empty block, we can remove it. */
2728 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2729 && single_pred_p (b)
2730 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2731 && !LABEL_P (BB_HEAD (b))
2732 && FORWARDER_BLOCK_P (b)
2733 /* Note that forwarder_block_p true ensures that
2734 there is a successor for this block. */
2735 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2736 && n_basic_blocks_for_fn (cfun) > NUM_FIXED_BLOCKS + 1)
2737 {
2738 if (dump_file)
2739 fprintf (dump_file,
2740 "Deleting fallthru block %i.\n",
2741 b->index);
2742
2743 c = ((b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun))
2744 ? b->next_bb : b->prev_bb);
2745 redirect_edge_succ_nodup (single_pred_edge (b),
2746 single_succ (b));
2747 delete_basic_block (b);
2748 changed = true;
2749 b = c;
2750 continue;
2751 }
2752
2753 /* Merge B with its single successor, if any. */
2754 if (single_succ_p (b)
2755 && (s = single_succ_edge (b))
2756 && !(s->flags & EDGE_COMPLEX)
2757 && (c = s->dest) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2758 && single_pred_p (c)
2759 && b != c)
2760 {
2761 /* When not in cfg_layout mode use code aware of reordering
2762 INSN. This code possibly creates new basic blocks so it
2763 does not fit merge_blocks interface and is kept here in
2764 hope that it will become useless once more of compiler
2765 is transformed to use cfg_layout mode. */
2766
2767 if ((mode & CLEANUP_CFGLAYOUT)
2768 && can_merge_blocks_p (b, c))
2769 {
2770 merge_blocks (b, c);
2771 update_forwarder_flag (b);
2772 changed_here = true;
2773 }
2774 else if (!(mode & CLEANUP_CFGLAYOUT)
2775 /* If the jump insn has side effects,
2776 we can't kill the edge. */
2777 && (!JUMP_P (BB_END (b))
2778 || (reload_completed
2779 ? simplejump_p (BB_END (b))
2780 : (onlyjump_p (BB_END (b))
2781 && !tablejump_p (BB_END (b),
2782 NULL, NULL))))
2783 && (next = merge_blocks_move (s, b, c, mode)))
2784 {
2785 b = next;
2786 changed_here = true;
2787 }
2788 }
2789
2790 /* Simplify branch over branch. */
2791 if ((mode & CLEANUP_EXPENSIVE)
2792 && !(mode & CLEANUP_CFGLAYOUT)
2793 && try_simplify_condjump (b))
2794 changed_here = true;
2795
2796 /* If B has a single outgoing edge, but uses a
2797 non-trivial jump instruction without side-effects, we
2798 can either delete the jump entirely, or replace it
2799 with a simple unconditional jump. */
2800 if (single_succ_p (b)
2801 && single_succ (b) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2802 && onlyjump_p (BB_END (b))
2803 && !CROSSING_JUMP_P (BB_END (b))
2804 && try_redirect_by_replacing_jump (single_succ_edge (b),
2805 single_succ (b),
2806 (mode & CLEANUP_CFGLAYOUT) != 0))
2807 {
2808 update_forwarder_flag (b);
2809 changed_here = true;
2810 }
2811
2812 /* Simplify branch to branch. */
2813 if (try_forward_edges (mode, b))
2814 {
2815 update_forwarder_flag (b);
2816 changed_here = true;
2817 }
2818
2819 /* Look for shared code between blocks. */
2820 if ((mode & CLEANUP_CROSSJUMP)
2821 && try_crossjump_bb (mode, b))
2822 changed_here = true;
2823
2824 if ((mode & CLEANUP_CROSSJUMP)
2825 /* This can lengthen register lifetimes. Do it only after
2826 reload. */
2827 && reload_completed
2828 && try_head_merge_bb (b))
2829 changed_here = true;
2830
2831 /* Don't get confused by the index shift caused by
2832 deleting blocks. */
2833 if (!changed_here)
2834 b = b->next_bb;
2835 else
2836 changed = true;
2837 }
2838
2839 if ((mode & CLEANUP_CROSSJUMP)
2840 && try_crossjump_bb (mode, EXIT_BLOCK_PTR_FOR_FN (cfun)))
2841 changed = true;
2842
2843 if (block_was_dirty)
2844 {
2845 /* This should only be set by head-merging. */
2846 gcc_assert (mode & CLEANUP_CROSSJUMP);
2847 df_analyze ();
2848 }
2849
2850 if (changed)
2851 {
2852 /* Edge forwarding in particular can cause hot blocks previously
2853 reached by both hot and cold blocks to become dominated only
2854 by cold blocks. This will cause the verification below to fail,
2855 and lead to now cold code in the hot section. This is not easy
2856 to detect and fix during edge forwarding, and in some cases
2857 is only visible after newly unreachable blocks are deleted,
2858 which will be done in fixup_partitions. */
2859 fixup_partitions ();
2860
2861 #ifdef ENABLE_CHECKING
2862 verify_flow_info ();
2863 #endif
2864 }
2865
2866 changed_overall |= changed;
2867 first_pass = false;
2868 }
2869 while (changed);
2870 }
2871
2872 FOR_ALL_BB_FN (b, cfun)
2873 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
2874
2875 return changed_overall;
2876 }
2877 \f
2878 /* Delete all unreachable basic blocks. */
2879
2880 bool
2881 delete_unreachable_blocks (void)
2882 {
2883 bool changed = false;
2884 basic_block b, prev_bb;
2885
2886 find_unreachable_blocks ();
2887
2888 /* When we're in GIMPLE mode and there may be debug insns, we should
2889 delete blocks in reverse dominator order, so as to get a chance
2890 to substitute all released DEFs into debug stmts. If we don't
2891 have dominators information, walking blocks backward gets us a
2892 better chance of retaining most debug information than
2893 otherwise. */
2894 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
2895 && dom_info_available_p (CDI_DOMINATORS))
2896 {
2897 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
2898 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
2899 {
2900 prev_bb = b->prev_bb;
2901
2902 if (!(b->flags & BB_REACHABLE))
2903 {
2904 /* Speed up the removal of blocks that don't dominate
2905 others. Walking backwards, this should be the common
2906 case. */
2907 if (!first_dom_son (CDI_DOMINATORS, b))
2908 delete_basic_block (b);
2909 else
2910 {
2911 vec<basic_block> h
2912 = get_all_dominated_blocks (CDI_DOMINATORS, b);
2913
2914 while (h.length ())
2915 {
2916 b = h.pop ();
2917
2918 prev_bb = b->prev_bb;
2919
2920 gcc_assert (!(b->flags & BB_REACHABLE));
2921
2922 delete_basic_block (b);
2923 }
2924
2925 h.release ();
2926 }
2927
2928 changed = true;
2929 }
2930 }
2931 }
2932 else
2933 {
2934 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
2935 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
2936 {
2937 prev_bb = b->prev_bb;
2938
2939 if (!(b->flags & BB_REACHABLE))
2940 {
2941 delete_basic_block (b);
2942 changed = true;
2943 }
2944 }
2945 }
2946
2947 if (changed)
2948 tidy_fallthru_edges ();
2949 return changed;
2950 }
2951
2952 /* Delete any jump tables never referenced. We can't delete them at the
2953 time of removing tablejump insn as they are referenced by the preceding
2954 insns computing the destination, so we delay deleting and garbagecollect
2955 them once life information is computed. */
2956 void
2957 delete_dead_jumptables (void)
2958 {
2959 basic_block bb;
2960
2961 /* A dead jump table does not belong to any basic block. Scan insns
2962 between two adjacent basic blocks. */
2963 FOR_EACH_BB_FN (bb, cfun)
2964 {
2965 rtx_insn *insn, *next;
2966
2967 for (insn = NEXT_INSN (BB_END (bb));
2968 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
2969 insn = next)
2970 {
2971 next = NEXT_INSN (insn);
2972 if (LABEL_P (insn)
2973 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
2974 && JUMP_TABLE_DATA_P (next))
2975 {
2976 rtx_insn *label = insn, *jump = next;
2977
2978 if (dump_file)
2979 fprintf (dump_file, "Dead jumptable %i removed\n",
2980 INSN_UID (insn));
2981
2982 next = NEXT_INSN (next);
2983 delete_insn (jump);
2984 delete_insn (label);
2985 }
2986 }
2987 }
2988 }
2989
2990 \f
2991 /* Tidy the CFG by deleting unreachable code and whatnot. */
2992
2993 bool
2994 cleanup_cfg (int mode)
2995 {
2996 bool changed = false;
2997
2998 /* Set the cfglayout mode flag here. We could update all the callers
2999 but that is just inconvenient, especially given that we eventually
3000 want to have cfglayout mode as the default. */
3001 if (current_ir_type () == IR_RTL_CFGLAYOUT)
3002 mode |= CLEANUP_CFGLAYOUT;
3003
3004 timevar_push (TV_CLEANUP_CFG);
3005 if (delete_unreachable_blocks ())
3006 {
3007 changed = true;
3008 /* We've possibly created trivially dead code. Cleanup it right
3009 now to introduce more opportunities for try_optimize_cfg. */
3010 if (!(mode & (CLEANUP_NO_INSN_DEL))
3011 && !reload_completed)
3012 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3013 }
3014
3015 compact_blocks ();
3016
3017 /* To tail-merge blocks ending in the same noreturn function (e.g.
3018 a call to abort) we have to insert fake edges to exit. Do this
3019 here once. The fake edges do not interfere with any other CFG
3020 cleanups. */
3021 if (mode & CLEANUP_CROSSJUMP)
3022 add_noreturn_fake_exit_edges ();
3023
3024 if (!dbg_cnt (cfg_cleanup))
3025 return changed;
3026
3027 while (try_optimize_cfg (mode))
3028 {
3029 delete_unreachable_blocks (), changed = true;
3030 if (!(mode & CLEANUP_NO_INSN_DEL))
3031 {
3032 /* Try to remove some trivially dead insns when doing an expensive
3033 cleanup. But delete_trivially_dead_insns doesn't work after
3034 reload (it only handles pseudos) and run_fast_dce is too costly
3035 to run in every iteration.
3036
3037 For effective cross jumping, we really want to run a fast DCE to
3038 clean up any dead conditions, or they get in the way of performing
3039 useful tail merges.
3040
3041 Other transformations in cleanup_cfg are not so sensitive to dead
3042 code, so delete_trivially_dead_insns or even doing nothing at all
3043 is good enough. */
3044 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3045 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3046 break;
3047 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occured)
3048 run_fast_dce ();
3049 }
3050 else
3051 break;
3052 }
3053
3054 if (mode & CLEANUP_CROSSJUMP)
3055 remove_fake_exit_edges ();
3056
3057 /* Don't call delete_dead_jumptables in cfglayout mode, because
3058 that function assumes that jump tables are in the insns stream.
3059 But we also don't _have_ to delete dead jumptables in cfglayout
3060 mode because we shouldn't even be looking at things that are
3061 not in a basic block. Dead jumptables are cleaned up when
3062 going out of cfglayout mode. */
3063 if (!(mode & CLEANUP_CFGLAYOUT))
3064 delete_dead_jumptables ();
3065
3066 /* ??? We probably do this way too often. */
3067 if (current_loops
3068 && (changed
3069 || (mode & CLEANUP_CFG_CHANGED)))
3070 {
3071 timevar_push (TV_REPAIR_LOOPS);
3072 /* The above doesn't preserve dominance info if available. */
3073 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3074 calculate_dominance_info (CDI_DOMINATORS);
3075 fix_loop_structure (NULL);
3076 free_dominance_info (CDI_DOMINATORS);
3077 timevar_pop (TV_REPAIR_LOOPS);
3078 }
3079
3080 timevar_pop (TV_CLEANUP_CFG);
3081
3082 return changed;
3083 }
3084 \f
3085 namespace {
3086
3087 const pass_data pass_data_jump =
3088 {
3089 RTL_PASS, /* type */
3090 "jump", /* name */
3091 OPTGROUP_NONE, /* optinfo_flags */
3092 TV_JUMP, /* tv_id */
3093 0, /* properties_required */
3094 0, /* properties_provided */
3095 0, /* properties_destroyed */
3096 0, /* todo_flags_start */
3097 0, /* todo_flags_finish */
3098 };
3099
3100 class pass_jump : public rtl_opt_pass
3101 {
3102 public:
3103 pass_jump (gcc::context *ctxt)
3104 : rtl_opt_pass (pass_data_jump, ctxt)
3105 {}
3106
3107 /* opt_pass methods: */
3108 virtual unsigned int execute (function *);
3109
3110 }; // class pass_jump
3111
3112 unsigned int
3113 pass_jump::execute (function *)
3114 {
3115 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3116 if (dump_file)
3117 dump_flow_info (dump_file, dump_flags);
3118 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3119 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3120 return 0;
3121 }
3122
3123 } // anon namespace
3124
3125 rtl_opt_pass *
3126 make_pass_jump (gcc::context *ctxt)
3127 {
3128 return new pass_jump (ctxt);
3129 }
3130 \f
3131 namespace {
3132
3133 const pass_data pass_data_jump2 =
3134 {
3135 RTL_PASS, /* type */
3136 "jump2", /* name */
3137 OPTGROUP_NONE, /* optinfo_flags */
3138 TV_JUMP, /* tv_id */
3139 0, /* properties_required */
3140 0, /* properties_provided */
3141 0, /* properties_destroyed */
3142 0, /* todo_flags_start */
3143 0, /* todo_flags_finish */
3144 };
3145
3146 class pass_jump2 : public rtl_opt_pass
3147 {
3148 public:
3149 pass_jump2 (gcc::context *ctxt)
3150 : rtl_opt_pass (pass_data_jump2, ctxt)
3151 {}
3152
3153 /* opt_pass methods: */
3154 virtual unsigned int execute (function *)
3155 {
3156 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3157 return 0;
3158 }
3159
3160 }; // class pass_jump2
3161
3162 } // anon namespace
3163
3164 rtl_opt_pass *
3165 make_pass_jump2 (gcc::context *ctxt)
3166 {
3167 return new pass_jump2 (ctxt);
3168 }