2014-09-10 Michael Meissner <meissner@linux.vnet.ibm.com>
[official-gcc.git] / gcc / cfgcleanup.c
bloba00816887535fc338816506c252e0f55379d7f26
1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987-2014 Free Software Foundation, Inc.
4 This file is part of GCC.
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.
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.
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/>. */
20 /* This file contains optimizer of the control flow. The main entry point is
21 cleanup_cfg. Following optimizations are performed:
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. */
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 "function.h" /* For inline functions in emit-rtl.h they need crtl. */
49 #include "emit-rtl.h"
50 #include "tree-pass.h"
51 #include "cfgloop.h"
52 #include "expr.h"
53 #include "df.h"
54 #include "dce.h"
55 #include "dbgcnt.h"
56 #include "emit-rtl.h"
57 #include "rtl-iter.h"
59 #define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
61 /* Set to true when we are running first pass of try_optimize_cfg loop. */
62 static bool first_pass;
64 /* Set to true if crossjumps occurred in the latest run of try_optimize_cfg. */
65 static bool crossjumps_occured;
67 /* Set to true if we couldn't run an optimization due to stale liveness
68 information; we should run df_analyze to enable more opportunities. */
69 static bool block_was_dirty;
71 static bool try_crossjump_to_edge (int, edge, edge, enum replace_direction);
72 static bool try_crossjump_bb (int, basic_block);
73 static bool outgoing_edges_match (int, basic_block, basic_block);
74 static enum replace_direction old_insns_match_p (int, rtx_insn *, rtx_insn *);
76 static void merge_blocks_move_predecessor_nojumps (basic_block, basic_block);
77 static void merge_blocks_move_successor_nojumps (basic_block, basic_block);
78 static bool try_optimize_cfg (int);
79 static bool try_simplify_condjump (basic_block);
80 static bool try_forward_edges (int, basic_block);
81 static edge thread_jump (edge, basic_block);
82 static bool mark_effect (rtx, bitmap);
83 static void notice_new_block (basic_block);
84 static void update_forwarder_flag (basic_block);
85 static void merge_memattrs (rtx, rtx);
87 /* Set flags for newly created block. */
89 static void
90 notice_new_block (basic_block bb)
92 if (!bb)
93 return;
95 if (forwarder_block_p (bb))
96 bb->flags |= BB_FORWARDER_BLOCK;
99 /* Recompute forwarder flag after block has been modified. */
101 static void
102 update_forwarder_flag (basic_block bb)
104 if (forwarder_block_p (bb))
105 bb->flags |= BB_FORWARDER_BLOCK;
106 else
107 bb->flags &= ~BB_FORWARDER_BLOCK;
110 /* Simplify a conditional jump around an unconditional jump.
111 Return true if something changed. */
113 static bool
114 try_simplify_condjump (basic_block cbranch_block)
116 basic_block jump_block, jump_dest_block, cbranch_dest_block;
117 edge cbranch_jump_edge, cbranch_fallthru_edge;
118 rtx_insn *cbranch_insn;
120 /* Verify that there are exactly two successors. */
121 if (EDGE_COUNT (cbranch_block->succs) != 2)
122 return false;
124 /* Verify that we've got a normal conditional branch at the end
125 of the block. */
126 cbranch_insn = BB_END (cbranch_block);
127 if (!any_condjump_p (cbranch_insn))
128 return false;
130 cbranch_fallthru_edge = FALLTHRU_EDGE (cbranch_block);
131 cbranch_jump_edge = BRANCH_EDGE (cbranch_block);
133 /* The next block must not have multiple predecessors, must not
134 be the last block in the function, and must contain just the
135 unconditional jump. */
136 jump_block = cbranch_fallthru_edge->dest;
137 if (!single_pred_p (jump_block)
138 || jump_block->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
139 || !FORWARDER_BLOCK_P (jump_block))
140 return false;
141 jump_dest_block = single_succ (jump_block);
143 /* If we are partitioning hot/cold basic blocks, we don't want to
144 mess up unconditional or indirect jumps that cross between hot
145 and cold sections.
147 Basic block partitioning may result in some jumps that appear to
148 be optimizable (or blocks that appear to be mergeable), but which really
149 must be left untouched (they are required to make it safely across
150 partition boundaries). See the comments at the top of
151 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
153 if (BB_PARTITION (jump_block) != BB_PARTITION (jump_dest_block)
154 || (cbranch_jump_edge->flags & EDGE_CROSSING))
155 return false;
157 /* The conditional branch must target the block after the
158 unconditional branch. */
159 cbranch_dest_block = cbranch_jump_edge->dest;
161 if (cbranch_dest_block == EXIT_BLOCK_PTR_FOR_FN (cfun)
162 || !can_fallthru (jump_block, cbranch_dest_block))
163 return false;
165 /* Invert the conditional branch. */
166 if (!invert_jump (cbranch_insn, block_label (jump_dest_block), 0))
167 return false;
169 if (dump_file)
170 fprintf (dump_file, "Simplifying condjump %i around jump %i\n",
171 INSN_UID (cbranch_insn), INSN_UID (BB_END (jump_block)));
173 /* Success. Update the CFG to match. Note that after this point
174 the edge variable names appear backwards; the redirection is done
175 this way to preserve edge profile data. */
176 cbranch_jump_edge = redirect_edge_succ_nodup (cbranch_jump_edge,
177 cbranch_dest_block);
178 cbranch_fallthru_edge = redirect_edge_succ_nodup (cbranch_fallthru_edge,
179 jump_dest_block);
180 cbranch_jump_edge->flags |= EDGE_FALLTHRU;
181 cbranch_fallthru_edge->flags &= ~EDGE_FALLTHRU;
182 update_br_prob_note (cbranch_block);
184 /* Delete the block with the unconditional jump, and clean up the mess. */
185 delete_basic_block (jump_block);
186 tidy_fallthru_edge (cbranch_jump_edge);
187 update_forwarder_flag (cbranch_block);
189 return true;
192 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
193 on register. Used by jump threading. */
195 static bool
196 mark_effect (rtx exp, regset nonequal)
198 int regno;
199 rtx dest;
200 switch (GET_CODE (exp))
202 /* In case we do clobber the register, mark it as equal, as we know the
203 value is dead so it don't have to match. */
204 case CLOBBER:
205 if (REG_P (XEXP (exp, 0)))
207 dest = XEXP (exp, 0);
208 regno = REGNO (dest);
209 if (HARD_REGISTER_NUM_P (regno))
210 bitmap_clear_range (nonequal, regno,
211 hard_regno_nregs[regno][GET_MODE (dest)]);
212 else
213 bitmap_clear_bit (nonequal, regno);
215 return false;
217 case SET:
218 if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
219 return false;
220 dest = SET_DEST (exp);
221 if (dest == pc_rtx)
222 return false;
223 if (!REG_P (dest))
224 return true;
225 regno = REGNO (dest);
226 if (HARD_REGISTER_NUM_P (regno))
227 bitmap_set_range (nonequal, regno,
228 hard_regno_nregs[regno][GET_MODE (dest)]);
229 else
230 bitmap_set_bit (nonequal, regno);
231 return false;
233 default:
234 return false;
238 /* Return true if X contains a register in NONEQUAL. */
239 static bool
240 mentions_nonequal_regs (const_rtx x, regset nonequal)
242 subrtx_iterator::array_type array;
243 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
245 const_rtx x = *iter;
246 if (REG_P (x))
248 unsigned int regno = REGNO (x);
249 if (REGNO_REG_SET_P (nonequal, regno))
250 return true;
251 if (regno < FIRST_PSEUDO_REGISTER)
253 int n = hard_regno_nregs[regno][GET_MODE (x)];
254 while (--n > 0)
255 if (REGNO_REG_SET_P (nonequal, regno + n))
256 return true;
260 return false;
263 /* Attempt to prove that the basic block B will have no side effects and
264 always continues in the same edge if reached via E. Return the edge
265 if exist, NULL otherwise. */
267 static edge
268 thread_jump (edge e, basic_block b)
270 rtx set1, set2, cond1, cond2;
271 rtx_insn *insn;
272 enum rtx_code code1, code2, reversed_code2;
273 bool reverse1 = false;
274 unsigned i;
275 regset nonequal;
276 bool failed = false;
277 reg_set_iterator rsi;
279 if (b->flags & BB_NONTHREADABLE_BLOCK)
280 return NULL;
282 /* At the moment, we do handle only conditional jumps, but later we may
283 want to extend this code to tablejumps and others. */
284 if (EDGE_COUNT (e->src->succs) != 2)
285 return NULL;
286 if (EDGE_COUNT (b->succs) != 2)
288 b->flags |= BB_NONTHREADABLE_BLOCK;
289 return NULL;
292 /* Second branch must end with onlyjump, as we will eliminate the jump. */
293 if (!any_condjump_p (BB_END (e->src)))
294 return NULL;
296 if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
298 b->flags |= BB_NONTHREADABLE_BLOCK;
299 return NULL;
302 set1 = pc_set (BB_END (e->src));
303 set2 = pc_set (BB_END (b));
304 if (((e->flags & EDGE_FALLTHRU) != 0)
305 != (XEXP (SET_SRC (set1), 1) == pc_rtx))
306 reverse1 = true;
308 cond1 = XEXP (SET_SRC (set1), 0);
309 cond2 = XEXP (SET_SRC (set2), 0);
310 if (reverse1)
311 code1 = reversed_comparison_code (cond1, BB_END (e->src));
312 else
313 code1 = GET_CODE (cond1);
315 code2 = GET_CODE (cond2);
316 reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
318 if (!comparison_dominates_p (code1, code2)
319 && !comparison_dominates_p (code1, reversed_code2))
320 return NULL;
322 /* Ensure that the comparison operators are equivalent.
323 ??? This is far too pessimistic. We should allow swapped operands,
324 different CCmodes, or for example comparisons for interval, that
325 dominate even when operands are not equivalent. */
326 if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
327 || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
328 return NULL;
330 /* Short circuit cases where block B contains some side effects, as we can't
331 safely bypass it. */
332 for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
333 insn = NEXT_INSN (insn))
334 if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
336 b->flags |= BB_NONTHREADABLE_BLOCK;
337 return NULL;
340 cselib_init (0);
342 /* First process all values computed in the source basic block. */
343 for (insn = NEXT_INSN (BB_HEAD (e->src));
344 insn != NEXT_INSN (BB_END (e->src));
345 insn = NEXT_INSN (insn))
346 if (INSN_P (insn))
347 cselib_process_insn (insn);
349 nonequal = BITMAP_ALLOC (NULL);
350 CLEAR_REG_SET (nonequal);
352 /* Now assume that we've continued by the edge E to B and continue
353 processing as if it were same basic block.
354 Our goal is to prove that whole block is an NOOP. */
356 for (insn = NEXT_INSN (BB_HEAD (b));
357 insn != NEXT_INSN (BB_END (b)) && !failed;
358 insn = NEXT_INSN (insn))
360 if (INSN_P (insn))
362 rtx pat = PATTERN (insn);
364 if (GET_CODE (pat) == PARALLEL)
366 for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
367 failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
369 else
370 failed |= mark_effect (pat, nonequal);
373 cselib_process_insn (insn);
376 /* Later we should clear nonequal of dead registers. So far we don't
377 have life information in cfg_cleanup. */
378 if (failed)
380 b->flags |= BB_NONTHREADABLE_BLOCK;
381 goto failed_exit;
384 /* cond2 must not mention any register that is not equal to the
385 former block. */
386 if (mentions_nonequal_regs (cond2, nonequal))
387 goto failed_exit;
389 EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
390 goto failed_exit;
392 BITMAP_FREE (nonequal);
393 cselib_finish ();
394 if ((comparison_dominates_p (code1, code2) != 0)
395 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
396 return BRANCH_EDGE (b);
397 else
398 return FALLTHRU_EDGE (b);
400 failed_exit:
401 BITMAP_FREE (nonequal);
402 cselib_finish ();
403 return NULL;
406 /* Attempt to forward edges leaving basic block B.
407 Return true if successful. */
409 static bool
410 try_forward_edges (int mode, basic_block b)
412 bool changed = false;
413 edge_iterator ei;
414 edge e, *threaded_edges = NULL;
416 /* If we are partitioning hot/cold basic blocks, we don't want to
417 mess up unconditional or indirect jumps that cross between hot
418 and cold sections.
420 Basic block partitioning may result in some jumps that appear to
421 be optimizable (or blocks that appear to be mergeable), but which really
422 must be left untouched (they are required to make it safely across
423 partition boundaries). See the comments at the top of
424 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
426 if (JUMP_P (BB_END (b)) && CROSSING_JUMP_P (BB_END (b)))
427 return false;
429 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
431 basic_block target, first;
432 location_t goto_locus;
433 int counter;
434 bool threaded = false;
435 int nthreaded_edges = 0;
436 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
438 /* Skip complex edges because we don't know how to update them.
440 Still handle fallthru edges, as we can succeed to forward fallthru
441 edge to the same place as the branch edge of conditional branch
442 and turn conditional branch to an unconditional branch. */
443 if (e->flags & EDGE_COMPLEX)
445 ei_next (&ei);
446 continue;
449 target = first = e->dest;
450 counter = NUM_FIXED_BLOCKS;
451 goto_locus = e->goto_locus;
453 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
454 up jumps that cross between hot/cold sections.
456 Basic block partitioning may result in some jumps that appear
457 to be optimizable (or blocks that appear to be mergeable), but which
458 really must be left untouched (they are required to make it safely
459 across partition boundaries). See the comments at the top of
460 bb-reorder.c:partition_hot_cold_basic_blocks for complete
461 details. */
463 if (first != EXIT_BLOCK_PTR_FOR_FN (cfun)
464 && JUMP_P (BB_END (first))
465 && CROSSING_JUMP_P (BB_END (first)))
466 return changed;
468 while (counter < n_basic_blocks_for_fn (cfun))
470 basic_block new_target = NULL;
471 bool new_target_threaded = false;
472 may_thread |= (target->flags & BB_MODIFIED) != 0;
474 if (FORWARDER_BLOCK_P (target)
475 && !(single_succ_edge (target)->flags & EDGE_CROSSING)
476 && single_succ (target) != EXIT_BLOCK_PTR_FOR_FN (cfun))
478 /* Bypass trivial infinite loops. */
479 new_target = single_succ (target);
480 if (target == new_target)
481 counter = n_basic_blocks_for_fn (cfun);
482 else if (!optimize)
484 /* When not optimizing, ensure that edges or forwarder
485 blocks with different locus are not optimized out. */
486 location_t new_locus = single_succ_edge (target)->goto_locus;
487 location_t locus = goto_locus;
489 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
490 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
491 && new_locus != locus)
492 new_target = NULL;
493 else
495 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
496 locus = new_locus;
498 rtx_insn *last = BB_END (target);
499 if (DEBUG_INSN_P (last))
500 last = prev_nondebug_insn (last);
501 if (last && INSN_P (last))
502 new_locus = INSN_LOCATION (last);
503 else
504 new_locus = UNKNOWN_LOCATION;
506 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
507 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
508 && new_locus != locus)
509 new_target = NULL;
510 else
512 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
513 locus = new_locus;
515 goto_locus = locus;
521 /* Allow to thread only over one edge at time to simplify updating
522 of probabilities. */
523 else if ((mode & CLEANUP_THREADING) && may_thread)
525 edge t = thread_jump (e, target);
526 if (t)
528 if (!threaded_edges)
529 threaded_edges = XNEWVEC (edge,
530 n_basic_blocks_for_fn (cfun));
531 else
533 int i;
535 /* Detect an infinite loop across blocks not
536 including the start block. */
537 for (i = 0; i < nthreaded_edges; ++i)
538 if (threaded_edges[i] == t)
539 break;
540 if (i < nthreaded_edges)
542 counter = n_basic_blocks_for_fn (cfun);
543 break;
547 /* Detect an infinite loop across the start block. */
548 if (t->dest == b)
549 break;
551 gcc_assert (nthreaded_edges
552 < (n_basic_blocks_for_fn (cfun)
553 - NUM_FIXED_BLOCKS));
554 threaded_edges[nthreaded_edges++] = t;
556 new_target = t->dest;
557 new_target_threaded = true;
561 if (!new_target)
562 break;
564 counter++;
565 target = new_target;
566 threaded |= new_target_threaded;
569 if (counter >= n_basic_blocks_for_fn (cfun))
571 if (dump_file)
572 fprintf (dump_file, "Infinite loop in BB %i.\n",
573 target->index);
575 else if (target == first)
576 ; /* We didn't do anything. */
577 else
579 /* Save the values now, as the edge may get removed. */
580 gcov_type edge_count = e->count;
581 int edge_probability = e->probability;
582 int edge_frequency;
583 int n = 0;
585 e->goto_locus = goto_locus;
587 /* Don't force if target is exit block. */
588 if (threaded && target != EXIT_BLOCK_PTR_FOR_FN (cfun))
590 notice_new_block (redirect_edge_and_branch_force (e, target));
591 if (dump_file)
592 fprintf (dump_file, "Conditionals threaded.\n");
594 else if (!redirect_edge_and_branch (e, target))
596 if (dump_file)
597 fprintf (dump_file,
598 "Forwarding edge %i->%i to %i failed.\n",
599 b->index, e->dest->index, target->index);
600 ei_next (&ei);
601 continue;
604 /* We successfully forwarded the edge. Now update profile
605 data: for each edge we traversed in the chain, remove
606 the original edge's execution count. */
607 edge_frequency = apply_probability (b->frequency, edge_probability);
611 edge t;
613 if (!single_succ_p (first))
615 gcc_assert (n < nthreaded_edges);
616 t = threaded_edges [n++];
617 gcc_assert (t->src == first);
618 update_bb_profile_for_threading (first, edge_frequency,
619 edge_count, t);
620 update_br_prob_note (first);
622 else
624 first->count -= edge_count;
625 if (first->count < 0)
626 first->count = 0;
627 first->frequency -= edge_frequency;
628 if (first->frequency < 0)
629 first->frequency = 0;
630 /* It is possible that as the result of
631 threading we've removed edge as it is
632 threaded to the fallthru edge. Avoid
633 getting out of sync. */
634 if (n < nthreaded_edges
635 && first == threaded_edges [n]->src)
636 n++;
637 t = single_succ_edge (first);
640 t->count -= edge_count;
641 if (t->count < 0)
642 t->count = 0;
643 first = t->dest;
645 while (first != target);
647 changed = true;
648 continue;
650 ei_next (&ei);
653 free (threaded_edges);
654 return changed;
658 /* Blocks A and B are to be merged into a single block. A has no incoming
659 fallthru edge, so it can be moved before B without adding or modifying
660 any jumps (aside from the jump from A to B). */
662 static void
663 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
665 rtx_insn *barrier;
667 /* If we are partitioning hot/cold basic blocks, we don't want to
668 mess up unconditional or indirect jumps that cross between hot
669 and cold sections.
671 Basic block partitioning may result in some jumps that appear to
672 be optimizable (or blocks that appear to be mergeable), but which really
673 must be left untouched (they are required to make it safely across
674 partition boundaries). See the comments at the top of
675 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
677 if (BB_PARTITION (a) != BB_PARTITION (b))
678 return;
680 barrier = next_nonnote_insn (BB_END (a));
681 gcc_assert (BARRIER_P (barrier));
682 delete_insn (barrier);
684 /* Scramble the insn chain. */
685 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
686 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
687 df_set_bb_dirty (a);
689 if (dump_file)
690 fprintf (dump_file, "Moved block %d before %d and merged.\n",
691 a->index, b->index);
693 /* Swap the records for the two blocks around. */
695 unlink_block (a);
696 link_block (a, b->prev_bb);
698 /* Now blocks A and B are contiguous. Merge them. */
699 merge_blocks (a, b);
702 /* Blocks A and B are to be merged into a single block. B has no outgoing
703 fallthru edge, so it can be moved after A without adding or modifying
704 any jumps (aside from the jump from A to B). */
706 static void
707 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
709 rtx_insn *barrier, *real_b_end;
710 rtx label;
711 rtx_jump_table_data *table;
713 /* If we are partitioning hot/cold basic blocks, we don't want to
714 mess up unconditional or indirect jumps that cross between hot
715 and cold sections.
717 Basic block partitioning may result in some jumps that appear to
718 be optimizable (or blocks that appear to be mergeable), but which really
719 must be left untouched (they are required to make it safely across
720 partition boundaries). See the comments at the top of
721 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
723 if (BB_PARTITION (a) != BB_PARTITION (b))
724 return;
726 real_b_end = BB_END (b);
728 /* If there is a jump table following block B temporarily add the jump table
729 to block B so that it will also be moved to the correct location. */
730 if (tablejump_p (BB_END (b), &label, &table)
731 && prev_active_insn (label) == BB_END (b))
733 BB_END (b) = table;
736 /* There had better have been a barrier there. Delete it. */
737 barrier = NEXT_INSN (BB_END (b));
738 if (barrier && BARRIER_P (barrier))
739 delete_insn (barrier);
742 /* Scramble the insn chain. */
743 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
745 /* Restore the real end of b. */
746 BB_END (b) = real_b_end;
748 if (dump_file)
749 fprintf (dump_file, "Moved block %d after %d and merged.\n",
750 b->index, a->index);
752 /* Now blocks A and B are contiguous. Merge them. */
753 merge_blocks (a, b);
756 /* Attempt to merge basic blocks that are potentially non-adjacent.
757 Return NULL iff the attempt failed, otherwise return basic block
758 where cleanup_cfg should continue. Because the merging commonly
759 moves basic block away or introduces another optimization
760 possibility, return basic block just before B so cleanup_cfg don't
761 need to iterate.
763 It may be good idea to return basic block before C in the case
764 C has been moved after B and originally appeared earlier in the
765 insn sequence, but we have no information available about the
766 relative ordering of these two. Hopefully it is not too common. */
768 static basic_block
769 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
771 basic_block next;
773 /* If we are partitioning hot/cold basic blocks, we don't want to
774 mess up unconditional or indirect jumps that cross between hot
775 and cold sections.
777 Basic block partitioning may result in some jumps that appear to
778 be optimizable (or blocks that appear to be mergeable), but which really
779 must be left untouched (they are required to make it safely across
780 partition boundaries). See the comments at the top of
781 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
783 if (BB_PARTITION (b) != BB_PARTITION (c))
784 return NULL;
786 /* If B has a fallthru edge to C, no need to move anything. */
787 if (e->flags & EDGE_FALLTHRU)
789 int b_index = b->index, c_index = c->index;
791 /* Protect the loop latches. */
792 if (current_loops && c->loop_father->latch == c)
793 return NULL;
795 merge_blocks (b, c);
796 update_forwarder_flag (b);
798 if (dump_file)
799 fprintf (dump_file, "Merged %d and %d without moving.\n",
800 b_index, c_index);
802 return b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? b : b->prev_bb;
805 /* Otherwise we will need to move code around. Do that only if expensive
806 transformations are allowed. */
807 else if (mode & CLEANUP_EXPENSIVE)
809 edge tmp_edge, b_fallthru_edge;
810 bool c_has_outgoing_fallthru;
811 bool b_has_incoming_fallthru;
813 /* Avoid overactive code motion, as the forwarder blocks should be
814 eliminated by edge redirection instead. One exception might have
815 been if B is a forwarder block and C has no fallthru edge, but
816 that should be cleaned up by bb-reorder instead. */
817 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
818 return NULL;
820 /* We must make sure to not munge nesting of lexical blocks,
821 and loop notes. This is done by squeezing out all the notes
822 and leaving them there to lie. Not ideal, but functional. */
824 tmp_edge = find_fallthru_edge (c->succs);
825 c_has_outgoing_fallthru = (tmp_edge != NULL);
827 tmp_edge = find_fallthru_edge (b->preds);
828 b_has_incoming_fallthru = (tmp_edge != NULL);
829 b_fallthru_edge = tmp_edge;
830 next = b->prev_bb;
831 if (next == c)
832 next = next->prev_bb;
834 /* Otherwise, we're going to try to move C after B. If C does
835 not have an outgoing fallthru, then it can be moved
836 immediately after B without introducing or modifying jumps. */
837 if (! c_has_outgoing_fallthru)
839 merge_blocks_move_successor_nojumps (b, c);
840 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
843 /* If B does not have an incoming fallthru, then it can be moved
844 immediately before C without introducing or modifying jumps.
845 C cannot be the first block, so we do not have to worry about
846 accessing a non-existent block. */
848 if (b_has_incoming_fallthru)
850 basic_block bb;
852 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
853 return NULL;
854 bb = force_nonfallthru (b_fallthru_edge);
855 if (bb)
856 notice_new_block (bb);
859 merge_blocks_move_predecessor_nojumps (b, c);
860 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
863 return NULL;
867 /* Removes the memory attributes of MEM expression
868 if they are not equal. */
870 void
871 merge_memattrs (rtx x, rtx y)
873 int i;
874 int j;
875 enum rtx_code code;
876 const char *fmt;
878 if (x == y)
879 return;
880 if (x == 0 || y == 0)
881 return;
883 code = GET_CODE (x);
885 if (code != GET_CODE (y))
886 return;
888 if (GET_MODE (x) != GET_MODE (y))
889 return;
891 if (code == MEM && !mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
893 if (! MEM_ATTRS (x))
894 MEM_ATTRS (y) = 0;
895 else if (! MEM_ATTRS (y))
896 MEM_ATTRS (x) = 0;
897 else
899 HOST_WIDE_INT mem_size;
901 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
903 set_mem_alias_set (x, 0);
904 set_mem_alias_set (y, 0);
907 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
909 set_mem_expr (x, 0);
910 set_mem_expr (y, 0);
911 clear_mem_offset (x);
912 clear_mem_offset (y);
914 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
915 || (MEM_OFFSET_KNOWN_P (x)
916 && MEM_OFFSET (x) != MEM_OFFSET (y)))
918 clear_mem_offset (x);
919 clear_mem_offset (y);
922 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
924 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
925 set_mem_size (x, mem_size);
926 set_mem_size (y, mem_size);
928 else
930 clear_mem_size (x);
931 clear_mem_size (y);
934 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
935 set_mem_align (y, MEM_ALIGN (x));
938 if (code == MEM)
940 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
942 MEM_READONLY_P (x) = 0;
943 MEM_READONLY_P (y) = 0;
945 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
947 MEM_NOTRAP_P (x) = 0;
948 MEM_NOTRAP_P (y) = 0;
950 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
952 MEM_VOLATILE_P (x) = 1;
953 MEM_VOLATILE_P (y) = 1;
957 fmt = GET_RTX_FORMAT (code);
958 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
960 switch (fmt[i])
962 case 'E':
963 /* Two vectors must have the same length. */
964 if (XVECLEN (x, i) != XVECLEN (y, i))
965 return;
967 for (j = 0; j < XVECLEN (x, i); j++)
968 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
970 break;
972 case 'e':
973 merge_memattrs (XEXP (x, i), XEXP (y, i));
976 return;
980 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
981 different single sets S1 and S2. */
983 static bool
984 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
986 int i;
987 rtx e1, e2;
989 if (p1 == s1 && p2 == s2)
990 return true;
992 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
993 return false;
995 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
996 return false;
998 for (i = 0; i < XVECLEN (p1, 0); i++)
1000 e1 = XVECEXP (p1, 0, i);
1001 e2 = XVECEXP (p2, 0, i);
1002 if (e1 == s1 && e2 == s2)
1003 continue;
1004 if (reload_completed
1005 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
1006 continue;
1008 return false;
1011 return true;
1014 /* Examine register notes on I1 and I2 and return:
1015 - dir_forward if I1 can be replaced by I2, or
1016 - dir_backward if I2 can be replaced by I1, or
1017 - dir_both if both are the case. */
1019 static enum replace_direction
1020 can_replace_by (rtx_insn *i1, rtx_insn *i2)
1022 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1023 bool c1, c2;
1025 /* Check for 2 sets. */
1026 s1 = single_set (i1);
1027 s2 = single_set (i2);
1028 if (s1 == NULL_RTX || s2 == NULL_RTX)
1029 return dir_none;
1031 /* Check that the 2 sets set the same dest. */
1032 d1 = SET_DEST (s1);
1033 d2 = SET_DEST (s2);
1034 if (!(reload_completed
1035 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1036 return dir_none;
1038 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1039 set dest to the same value. */
1040 note1 = find_reg_equal_equiv_note (i1);
1041 note2 = find_reg_equal_equiv_note (i2);
1042 if (!note1 || !note2 || !rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0))
1043 || !CONST_INT_P (XEXP (note1, 0)))
1044 return dir_none;
1046 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1047 return dir_none;
1049 /* Although the 2 sets set dest to the same value, we cannot replace
1050 (set (dest) (const_int))
1052 (set (dest) (reg))
1053 because we don't know if the reg is live and has the same value at the
1054 location of replacement. */
1055 src1 = SET_SRC (s1);
1056 src2 = SET_SRC (s2);
1057 c1 = CONST_INT_P (src1);
1058 c2 = CONST_INT_P (src2);
1059 if (c1 && c2)
1060 return dir_both;
1061 else if (c2)
1062 return dir_forward;
1063 else if (c1)
1064 return dir_backward;
1066 return dir_none;
1069 /* Merges directions A and B. */
1071 static enum replace_direction
1072 merge_dir (enum replace_direction a, enum replace_direction b)
1074 /* Implements the following table:
1075 |bo fw bw no
1076 ---+-----------
1077 bo |bo fw bw no
1078 fw |-- fw no no
1079 bw |-- -- bw no
1080 no |-- -- -- no. */
1082 if (a == b)
1083 return a;
1085 if (a == dir_both)
1086 return b;
1087 if (b == dir_both)
1088 return a;
1090 return dir_none;
1093 /* Examine I1 and I2 and return:
1094 - dir_forward if I1 can be replaced by I2, or
1095 - dir_backward if I2 can be replaced by I1, or
1096 - dir_both if both are the case. */
1098 static enum replace_direction
1099 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx_insn *i1, rtx_insn *i2)
1101 rtx p1, p2;
1103 /* Verify that I1 and I2 are equivalent. */
1104 if (GET_CODE (i1) != GET_CODE (i2))
1105 return dir_none;
1107 /* __builtin_unreachable() may lead to empty blocks (ending with
1108 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1109 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1110 return dir_both;
1112 /* ??? Do not allow cross-jumping between different stack levels. */
1113 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1114 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1115 if (p1 && p2)
1117 p1 = XEXP (p1, 0);
1118 p2 = XEXP (p2, 0);
1119 if (!rtx_equal_p (p1, p2))
1120 return dir_none;
1122 /* ??? Worse, this adjustment had better be constant lest we
1123 have differing incoming stack levels. */
1124 if (!frame_pointer_needed
1125 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1126 return dir_none;
1128 else if (p1 || p2)
1129 return dir_none;
1131 p1 = PATTERN (i1);
1132 p2 = PATTERN (i2);
1134 if (GET_CODE (p1) != GET_CODE (p2))
1135 return dir_none;
1137 /* If this is a CALL_INSN, compare register usage information.
1138 If we don't check this on stack register machines, the two
1139 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1140 numbers of stack registers in the same basic block.
1141 If we don't check this on machines with delay slots, a delay slot may
1142 be filled that clobbers a parameter expected by the subroutine.
1144 ??? We take the simple route for now and assume that if they're
1145 equal, they were constructed identically.
1147 Also check for identical exception regions. */
1149 if (CALL_P (i1))
1151 /* Ensure the same EH region. */
1152 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1153 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1155 if (!n1 && n2)
1156 return dir_none;
1158 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1159 return dir_none;
1161 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1162 CALL_INSN_FUNCTION_USAGE (i2))
1163 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1164 return dir_none;
1166 /* For address sanitizer, never crossjump __asan_report_* builtins,
1167 otherwise errors might be reported on incorrect lines. */
1168 if (flag_sanitize & SANITIZE_ADDRESS)
1170 rtx call = get_call_rtx_from (i1);
1171 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1173 rtx symbol = XEXP (XEXP (call, 0), 0);
1174 if (SYMBOL_REF_DECL (symbol)
1175 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1177 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1178 == BUILT_IN_NORMAL)
1179 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1180 >= BUILT_IN_ASAN_REPORT_LOAD1
1181 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1182 <= BUILT_IN_ASAN_STOREN)
1183 return dir_none;
1189 #ifdef STACK_REGS
1190 /* If cross_jump_death_matters is not 0, the insn's mode
1191 indicates whether or not the insn contains any stack-like
1192 regs. */
1194 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1196 /* If register stack conversion has already been done, then
1197 death notes must also be compared before it is certain that
1198 the two instruction streams match. */
1200 rtx note;
1201 HARD_REG_SET i1_regset, i2_regset;
1203 CLEAR_HARD_REG_SET (i1_regset);
1204 CLEAR_HARD_REG_SET (i2_regset);
1206 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1207 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1208 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1210 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1211 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1212 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1214 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1215 return dir_none;
1217 #endif
1219 if (reload_completed
1220 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1221 return dir_both;
1223 return can_replace_by (i1, i2);
1226 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1227 flow_find_head_matching_sequence, ensure the notes match. */
1229 static void
1230 merge_notes (rtx_insn *i1, rtx_insn *i2)
1232 /* If the merged insns have different REG_EQUAL notes, then
1233 remove them. */
1234 rtx equiv1 = find_reg_equal_equiv_note (i1);
1235 rtx equiv2 = find_reg_equal_equiv_note (i2);
1237 if (equiv1 && !equiv2)
1238 remove_note (i1, equiv1);
1239 else if (!equiv1 && equiv2)
1240 remove_note (i2, equiv2);
1241 else if (equiv1 && equiv2
1242 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1244 remove_note (i1, equiv1);
1245 remove_note (i2, equiv2);
1249 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1250 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1251 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1252 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1253 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1255 static void
1256 walk_to_nondebug_insn (rtx_insn **i1, basic_block *bb1, bool follow_fallthru,
1257 bool *did_fallthru)
1259 edge fallthru;
1261 *did_fallthru = false;
1263 /* Ignore notes. */
1264 while (!NONDEBUG_INSN_P (*i1))
1266 if (*i1 != BB_HEAD (*bb1))
1268 *i1 = PREV_INSN (*i1);
1269 continue;
1272 if (!follow_fallthru)
1273 return;
1275 fallthru = find_fallthru_edge ((*bb1)->preds);
1276 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)
1277 || !single_succ_p (fallthru->src))
1278 return;
1280 *bb1 = fallthru->src;
1281 *i1 = BB_END (*bb1);
1282 *did_fallthru = true;
1286 /* Look through the insns at the end of BB1 and BB2 and find the longest
1287 sequence that are either equivalent, or allow forward or backward
1288 replacement. Store the first insns for that sequence in *F1 and *F2 and
1289 return the sequence length.
1291 DIR_P indicates the allowed replacement direction on function entry, and
1292 the actual replacement direction on function exit. If NULL, only equivalent
1293 sequences are allowed.
1295 To simplify callers of this function, if the blocks match exactly,
1296 store the head of the blocks in *F1 and *F2. */
1299 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx_insn **f1,
1300 rtx_insn **f2, enum replace_direction *dir_p)
1302 rtx_insn *i1, *i2, *last1, *last2, *afterlast1, *afterlast2;
1303 int ninsns = 0;
1304 enum replace_direction dir, last_dir, afterlast_dir;
1305 bool follow_fallthru, did_fallthru;
1307 if (dir_p)
1308 dir = *dir_p;
1309 else
1310 dir = dir_both;
1311 afterlast_dir = dir;
1312 last_dir = afterlast_dir;
1314 /* Skip simple jumps at the end of the blocks. Complex jumps still
1315 need to be compared for equivalence, which we'll do below. */
1317 i1 = BB_END (bb1);
1318 last1 = afterlast1 = last2 = afterlast2 = NULL;
1319 if (onlyjump_p (i1)
1320 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1322 last1 = i1;
1323 i1 = PREV_INSN (i1);
1326 i2 = BB_END (bb2);
1327 if (onlyjump_p (i2)
1328 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1330 last2 = i2;
1331 /* Count everything except for unconditional jump as insn.
1332 Don't count any jumps if dir_p is NULL. */
1333 if (!simplejump_p (i2) && !returnjump_p (i2) && last1 && dir_p)
1334 ninsns++;
1335 i2 = PREV_INSN (i2);
1338 while (true)
1340 /* In the following example, we can replace all jumps to C by jumps to A.
1342 This removes 4 duplicate insns.
1343 [bb A] insn1 [bb C] insn1
1344 insn2 insn2
1345 [bb B] insn3 insn3
1346 insn4 insn4
1347 jump_insn jump_insn
1349 We could also replace all jumps to A by jumps to C, but that leaves B
1350 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1351 step, all jumps to B would be replaced with jumps to the middle of C,
1352 achieving the same result with more effort.
1353 So we allow only the first possibility, which means that we don't allow
1354 fallthru in the block that's being replaced. */
1356 follow_fallthru = dir_p && dir != dir_forward;
1357 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1358 if (did_fallthru)
1359 dir = dir_backward;
1361 follow_fallthru = dir_p && dir != dir_backward;
1362 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1363 if (did_fallthru)
1364 dir = dir_forward;
1366 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1367 break;
1369 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1370 if (dir == dir_none || (!dir_p && dir != dir_both))
1371 break;
1373 merge_memattrs (i1, i2);
1375 /* Don't begin a cross-jump with a NOTE insn. */
1376 if (INSN_P (i1))
1378 merge_notes (i1, i2);
1380 afterlast1 = last1, afterlast2 = last2;
1381 last1 = i1, last2 = i2;
1382 afterlast_dir = last_dir;
1383 last_dir = dir;
1384 if (active_insn_p (i1))
1385 ninsns++;
1388 i1 = PREV_INSN (i1);
1389 i2 = PREV_INSN (i2);
1392 #ifdef HAVE_cc0
1393 /* Don't allow the insn after a compare to be shared by
1394 cross-jumping unless the compare is also shared. */
1395 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && ! sets_cc0_p (last1))
1396 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1397 #endif
1399 /* Include preceding notes and labels in the cross-jump. One,
1400 this may bring us to the head of the blocks as requested above.
1401 Two, it keeps line number notes as matched as may be. */
1402 if (ninsns)
1404 bb1 = BLOCK_FOR_INSN (last1);
1405 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1406 last1 = PREV_INSN (last1);
1408 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1409 last1 = PREV_INSN (last1);
1411 bb2 = BLOCK_FOR_INSN (last2);
1412 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1413 last2 = PREV_INSN (last2);
1415 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1416 last2 = PREV_INSN (last2);
1418 *f1 = last1;
1419 *f2 = last2;
1422 if (dir_p)
1423 *dir_p = last_dir;
1424 return ninsns;
1427 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1428 the head of the two blocks. Do not include jumps at the end.
1429 If STOP_AFTER is nonzero, stop after finding that many matching
1430 instructions. If STOP_AFTER is zero, count all INSN_P insns, if it is
1431 non-zero, only count active insns. */
1434 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx_insn **f1,
1435 rtx_insn **f2, int stop_after)
1437 rtx_insn *i1, *i2, *last1, *last2, *beforelast1, *beforelast2;
1438 int ninsns = 0;
1439 edge e;
1440 edge_iterator ei;
1441 int nehedges1 = 0, nehedges2 = 0;
1443 FOR_EACH_EDGE (e, ei, bb1->succs)
1444 if (e->flags & EDGE_EH)
1445 nehedges1++;
1446 FOR_EACH_EDGE (e, ei, bb2->succs)
1447 if (e->flags & EDGE_EH)
1448 nehedges2++;
1450 i1 = BB_HEAD (bb1);
1451 i2 = BB_HEAD (bb2);
1452 last1 = beforelast1 = last2 = beforelast2 = NULL;
1454 while (true)
1456 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1457 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1459 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1460 break;
1461 i1 = NEXT_INSN (i1);
1464 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1466 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1467 break;
1468 i2 = NEXT_INSN (i2);
1471 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1472 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1473 break;
1475 if (NOTE_P (i1) || NOTE_P (i2)
1476 || JUMP_P (i1) || JUMP_P (i2))
1477 break;
1479 /* A sanity check to make sure we're not merging insns with different
1480 effects on EH. If only one of them ends a basic block, it shouldn't
1481 have an EH edge; if both end a basic block, there should be the same
1482 number of EH edges. */
1483 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1484 && nehedges1 > 0)
1485 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1486 && nehedges2 > 0)
1487 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1488 && nehedges1 != nehedges2))
1489 break;
1491 if (old_insns_match_p (0, i1, i2) != dir_both)
1492 break;
1494 merge_memattrs (i1, i2);
1496 /* Don't begin a cross-jump with a NOTE insn. */
1497 if (INSN_P (i1))
1499 merge_notes (i1, i2);
1501 beforelast1 = last1, beforelast2 = last2;
1502 last1 = i1, last2 = i2;
1503 if (!stop_after || active_insn_p (i1))
1504 ninsns++;
1507 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1508 || (stop_after > 0 && ninsns == stop_after))
1509 break;
1511 i1 = NEXT_INSN (i1);
1512 i2 = NEXT_INSN (i2);
1515 #ifdef HAVE_cc0
1516 /* Don't allow a compare to be shared by cross-jumping unless the insn
1517 after the compare is also shared. */
1518 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && sets_cc0_p (last1))
1519 last1 = beforelast1, last2 = beforelast2, ninsns--;
1520 #endif
1522 if (ninsns)
1524 *f1 = last1;
1525 *f2 = last2;
1528 return ninsns;
1531 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1532 the branch instruction. This means that if we commonize the control
1533 flow before end of the basic block, the semantic remains unchanged.
1535 We may assume that there exists one edge with a common destination. */
1537 static bool
1538 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1540 int nehedges1 = 0, nehedges2 = 0;
1541 edge fallthru1 = 0, fallthru2 = 0;
1542 edge e1, e2;
1543 edge_iterator ei;
1545 /* If we performed shrink-wrapping, edges to the exit block can
1546 only be distinguished for JUMP_INSNs. The two paths may differ in
1547 whether they went through the prologue. Sibcalls are fine, we know
1548 that we either didn't need or inserted an epilogue before them. */
1549 if (crtl->shrink_wrapped
1550 && single_succ_p (bb1)
1551 && single_succ (bb1) == EXIT_BLOCK_PTR_FOR_FN (cfun)
1552 && !JUMP_P (BB_END (bb1))
1553 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1554 return false;
1556 /* If BB1 has only one successor, we may be looking at either an
1557 unconditional jump, or a fake edge to exit. */
1558 if (single_succ_p (bb1)
1559 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1560 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1561 return (single_succ_p (bb2)
1562 && (single_succ_edge (bb2)->flags
1563 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1564 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1566 /* Match conditional jumps - this may get tricky when fallthru and branch
1567 edges are crossed. */
1568 if (EDGE_COUNT (bb1->succs) == 2
1569 && any_condjump_p (BB_END (bb1))
1570 && onlyjump_p (BB_END (bb1)))
1572 edge b1, f1, b2, f2;
1573 bool reverse, match;
1574 rtx set1, set2, cond1, cond2;
1575 enum rtx_code code1, code2;
1577 if (EDGE_COUNT (bb2->succs) != 2
1578 || !any_condjump_p (BB_END (bb2))
1579 || !onlyjump_p (BB_END (bb2)))
1580 return false;
1582 b1 = BRANCH_EDGE (bb1);
1583 b2 = BRANCH_EDGE (bb2);
1584 f1 = FALLTHRU_EDGE (bb1);
1585 f2 = FALLTHRU_EDGE (bb2);
1587 /* Get around possible forwarders on fallthru edges. Other cases
1588 should be optimized out already. */
1589 if (FORWARDER_BLOCK_P (f1->dest))
1590 f1 = single_succ_edge (f1->dest);
1592 if (FORWARDER_BLOCK_P (f2->dest))
1593 f2 = single_succ_edge (f2->dest);
1595 /* To simplify use of this function, return false if there are
1596 unneeded forwarder blocks. These will get eliminated later
1597 during cleanup_cfg. */
1598 if (FORWARDER_BLOCK_P (f1->dest)
1599 || FORWARDER_BLOCK_P (f2->dest)
1600 || FORWARDER_BLOCK_P (b1->dest)
1601 || FORWARDER_BLOCK_P (b2->dest))
1602 return false;
1604 if (f1->dest == f2->dest && b1->dest == b2->dest)
1605 reverse = false;
1606 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1607 reverse = true;
1608 else
1609 return false;
1611 set1 = pc_set (BB_END (bb1));
1612 set2 = pc_set (BB_END (bb2));
1613 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1614 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1615 reverse = !reverse;
1617 cond1 = XEXP (SET_SRC (set1), 0);
1618 cond2 = XEXP (SET_SRC (set2), 0);
1619 code1 = GET_CODE (cond1);
1620 if (reverse)
1621 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1622 else
1623 code2 = GET_CODE (cond2);
1625 if (code2 == UNKNOWN)
1626 return false;
1628 /* Verify codes and operands match. */
1629 match = ((code1 == code2
1630 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1631 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1632 || (code1 == swap_condition (code2)
1633 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1634 XEXP (cond2, 0))
1635 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1636 XEXP (cond2, 1))));
1638 /* If we return true, we will join the blocks. Which means that
1639 we will only have one branch prediction bit to work with. Thus
1640 we require the existing branches to have probabilities that are
1641 roughly similar. */
1642 if (match
1643 && optimize_bb_for_speed_p (bb1)
1644 && optimize_bb_for_speed_p (bb2))
1646 int prob2;
1648 if (b1->dest == b2->dest)
1649 prob2 = b2->probability;
1650 else
1651 /* Do not use f2 probability as f2 may be forwarded. */
1652 prob2 = REG_BR_PROB_BASE - b2->probability;
1654 /* Fail if the difference in probabilities is greater than 50%.
1655 This rules out two well-predicted branches with opposite
1656 outcomes. */
1657 if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
1659 if (dump_file)
1660 fprintf (dump_file,
1661 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1662 bb1->index, bb2->index, b1->probability, prob2);
1664 return false;
1668 if (dump_file && match)
1669 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1670 bb1->index, bb2->index);
1672 return match;
1675 /* Generic case - we are seeing a computed jump, table jump or trapping
1676 instruction. */
1678 /* Check whether there are tablejumps in the end of BB1 and BB2.
1679 Return true if they are identical. */
1681 rtx label1, label2;
1682 rtx_jump_table_data *table1, *table2;
1684 if (tablejump_p (BB_END (bb1), &label1, &table1)
1685 && tablejump_p (BB_END (bb2), &label2, &table2)
1686 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1688 /* The labels should never be the same rtx. If they really are same
1689 the jump tables are same too. So disable crossjumping of blocks BB1
1690 and BB2 because when deleting the common insns in the end of BB1
1691 by delete_basic_block () the jump table would be deleted too. */
1692 /* If LABEL2 is referenced in BB1->END do not do anything
1693 because we would loose information when replacing
1694 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1695 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1697 /* Set IDENTICAL to true when the tables are identical. */
1698 bool identical = false;
1699 rtx p1, p2;
1701 p1 = PATTERN (table1);
1702 p2 = PATTERN (table2);
1703 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1705 identical = true;
1707 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1708 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1709 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1710 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1712 int i;
1714 identical = true;
1715 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1716 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1717 identical = false;
1720 if (identical)
1722 bool match;
1724 /* Temporarily replace references to LABEL1 with LABEL2
1725 in BB1->END so that we could compare the instructions. */
1726 replace_label_in_insn (BB_END (bb1), label1, label2, false);
1728 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1729 == dir_both);
1730 if (dump_file && match)
1731 fprintf (dump_file,
1732 "Tablejumps in bb %i and %i match.\n",
1733 bb1->index, bb2->index);
1735 /* Set the original label in BB1->END because when deleting
1736 a block whose end is a tablejump, the tablejump referenced
1737 from the instruction is deleted too. */
1738 replace_label_in_insn (BB_END (bb1), label2, label1, false);
1740 return match;
1743 return false;
1747 /* Find the last non-debug non-note instruction in each bb, except
1748 stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
1749 handles that case specially. old_insns_match_p does not handle
1750 other types of instruction notes. */
1751 rtx_insn *last1 = BB_END (bb1);
1752 rtx_insn *last2 = BB_END (bb2);
1753 while (!NOTE_INSN_BASIC_BLOCK_P (last1) &&
1754 (DEBUG_INSN_P (last1) || NOTE_P (last1)))
1755 last1 = PREV_INSN (last1);
1756 while (!NOTE_INSN_BASIC_BLOCK_P (last2) &&
1757 (DEBUG_INSN_P (last2) || NOTE_P (last2)))
1758 last2 = PREV_INSN (last2);
1759 gcc_assert (last1 && last2);
1761 /* First ensure that the instructions match. There may be many outgoing
1762 edges so this test is generally cheaper. */
1763 if (old_insns_match_p (mode, last1, last2) != dir_both)
1764 return false;
1766 /* Search the outgoing edges, ensure that the counts do match, find possible
1767 fallthru and exception handling edges since these needs more
1768 validation. */
1769 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1770 return false;
1772 bool nonfakeedges = false;
1773 FOR_EACH_EDGE (e1, ei, bb1->succs)
1775 e2 = EDGE_SUCC (bb2, ei.index);
1777 if ((e1->flags & EDGE_FAKE) == 0)
1778 nonfakeedges = true;
1780 if (e1->flags & EDGE_EH)
1781 nehedges1++;
1783 if (e2->flags & EDGE_EH)
1784 nehedges2++;
1786 if (e1->flags & EDGE_FALLTHRU)
1787 fallthru1 = e1;
1788 if (e2->flags & EDGE_FALLTHRU)
1789 fallthru2 = e2;
1792 /* If number of edges of various types does not match, fail. */
1793 if (nehedges1 != nehedges2
1794 || (fallthru1 != 0) != (fallthru2 != 0))
1795 return false;
1797 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1798 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1799 attempt to optimize, as the two basic blocks might have different
1800 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1801 traps there should be REG_ARG_SIZE notes, they could be missing
1802 for __builtin_unreachable () uses though. */
1803 if (!nonfakeedges
1804 && !ACCUMULATE_OUTGOING_ARGS
1805 && (!INSN_P (last1)
1806 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1807 return false;
1809 /* fallthru edges must be forwarded to the same destination. */
1810 if (fallthru1)
1812 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1813 ? single_succ (fallthru1->dest): fallthru1->dest);
1814 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1815 ? single_succ (fallthru2->dest): fallthru2->dest);
1817 if (d1 != d2)
1818 return false;
1821 /* Ensure the same EH region. */
1823 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1824 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1826 if (!n1 && n2)
1827 return false;
1829 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1830 return false;
1833 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1834 version of sequence abstraction. */
1835 FOR_EACH_EDGE (e1, ei, bb2->succs)
1837 edge e2;
1838 edge_iterator ei;
1839 basic_block d1 = e1->dest;
1841 if (FORWARDER_BLOCK_P (d1))
1842 d1 = EDGE_SUCC (d1, 0)->dest;
1844 FOR_EACH_EDGE (e2, ei, bb1->succs)
1846 basic_block d2 = e2->dest;
1847 if (FORWARDER_BLOCK_P (d2))
1848 d2 = EDGE_SUCC (d2, 0)->dest;
1849 if (d1 == d2)
1850 break;
1853 if (!e2)
1854 return false;
1857 return true;
1860 /* Returns true if BB basic block has a preserve label. */
1862 static bool
1863 block_has_preserve_label (basic_block bb)
1865 return (bb
1866 && block_label (bb)
1867 && LABEL_PRESERVE_P (block_label (bb)));
1870 /* E1 and E2 are edges with the same destination block. Search their
1871 predecessors for common code. If found, redirect control flow from
1872 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1873 or the other way around (dir_backward). DIR specifies the allowed
1874 replacement direction. */
1876 static bool
1877 try_crossjump_to_edge (int mode, edge e1, edge e2,
1878 enum replace_direction dir)
1880 int nmatch;
1881 basic_block src1 = e1->src, src2 = e2->src;
1882 basic_block redirect_to, redirect_from, to_remove;
1883 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1884 rtx_insn *newpos1, *newpos2;
1885 edge s;
1886 edge_iterator ei;
1888 newpos1 = newpos2 = NULL;
1890 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1891 to try this optimization.
1893 Basic block partitioning may result in some jumps that appear to
1894 be optimizable (or blocks that appear to be mergeable), but which really
1895 must be left untouched (they are required to make it safely across
1896 partition boundaries). See the comments at the top of
1897 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1899 if (crtl->has_bb_partition && reload_completed)
1900 return false;
1902 /* Search backward through forwarder blocks. We don't need to worry
1903 about multiple entry or chained forwarders, as they will be optimized
1904 away. We do this to look past the unconditional jump following a
1905 conditional jump that is required due to the current CFG shape. */
1906 if (single_pred_p (src1)
1907 && FORWARDER_BLOCK_P (src1))
1908 e1 = single_pred_edge (src1), src1 = e1->src;
1910 if (single_pred_p (src2)
1911 && FORWARDER_BLOCK_P (src2))
1912 e2 = single_pred_edge (src2), src2 = e2->src;
1914 /* Nothing to do if we reach ENTRY, or a common source block. */
1915 if (src1 == ENTRY_BLOCK_PTR_FOR_FN (cfun) || src2
1916 == ENTRY_BLOCK_PTR_FOR_FN (cfun))
1917 return false;
1918 if (src1 == src2)
1919 return false;
1921 /* Seeing more than 1 forwarder blocks would confuse us later... */
1922 if (FORWARDER_BLOCK_P (e1->dest)
1923 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1924 return false;
1926 if (FORWARDER_BLOCK_P (e2->dest)
1927 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1928 return false;
1930 /* Likewise with dead code (possibly newly created by the other optimizations
1931 of cfg_cleanup). */
1932 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1933 return false;
1935 /* Look for the common insn sequence, part the first ... */
1936 if (!outgoing_edges_match (mode, src1, src2))
1937 return false;
1939 /* ... and part the second. */
1940 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
1942 osrc1 = src1;
1943 osrc2 = src2;
1944 if (newpos1 != NULL_RTX)
1945 src1 = BLOCK_FOR_INSN (newpos1);
1946 if (newpos2 != NULL_RTX)
1947 src2 = BLOCK_FOR_INSN (newpos2);
1949 if (dir == dir_backward)
1951 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
1952 SWAP (basic_block, osrc1, osrc2);
1953 SWAP (basic_block, src1, src2);
1954 SWAP (edge, e1, e2);
1955 SWAP (rtx_insn *, newpos1, newpos2);
1956 #undef SWAP
1959 /* Don't proceed with the crossjump unless we found a sufficient number
1960 of matching instructions or the 'from' block was totally matched
1961 (such that its predecessors will hopefully be redirected and the
1962 block removed). */
1963 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
1964 && (newpos1 != BB_HEAD (src1)))
1965 return false;
1967 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
1968 if (block_has_preserve_label (e1->dest)
1969 && (e1->flags & EDGE_ABNORMAL))
1970 return false;
1972 /* Here we know that the insns in the end of SRC1 which are common with SRC2
1973 will be deleted.
1974 If we have tablejumps in the end of SRC1 and SRC2
1975 they have been already compared for equivalence in outgoing_edges_match ()
1976 so replace the references to TABLE1 by references to TABLE2. */
1978 rtx label1, label2;
1979 rtx_jump_table_data *table1, *table2;
1981 if (tablejump_p (BB_END (osrc1), &label1, &table1)
1982 && tablejump_p (BB_END (osrc2), &label2, &table2)
1983 && label1 != label2)
1985 rtx_insn *insn;
1987 /* Replace references to LABEL1 with LABEL2. */
1988 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
1990 /* Do not replace the label in SRC1->END because when deleting
1991 a block whose end is a tablejump, the tablejump referenced
1992 from the instruction is deleted too. */
1993 if (insn != BB_END (osrc1))
1994 replace_label_in_insn (insn, label1, label2, true);
1999 /* Avoid splitting if possible. We must always split when SRC2 has
2000 EH predecessor edges, or we may end up with basic blocks with both
2001 normal and EH predecessor edges. */
2002 if (newpos2 == BB_HEAD (src2)
2003 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
2004 redirect_to = src2;
2005 else
2007 if (newpos2 == BB_HEAD (src2))
2009 /* Skip possible basic block header. */
2010 if (LABEL_P (newpos2))
2011 newpos2 = NEXT_INSN (newpos2);
2012 while (DEBUG_INSN_P (newpos2))
2013 newpos2 = NEXT_INSN (newpos2);
2014 if (NOTE_P (newpos2))
2015 newpos2 = NEXT_INSN (newpos2);
2016 while (DEBUG_INSN_P (newpos2))
2017 newpos2 = NEXT_INSN (newpos2);
2020 if (dump_file)
2021 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2022 src2->index, nmatch);
2023 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2026 if (dump_file)
2027 fprintf (dump_file,
2028 "Cross jumping from bb %i to bb %i; %i common insns\n",
2029 src1->index, src2->index, nmatch);
2031 /* We may have some registers visible through the block. */
2032 df_set_bb_dirty (redirect_to);
2034 if (osrc2 == src2)
2035 redirect_edges_to = redirect_to;
2036 else
2037 redirect_edges_to = osrc2;
2039 /* Recompute the frequencies and counts of outgoing edges. */
2040 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2042 edge s2;
2043 edge_iterator ei;
2044 basic_block d = s->dest;
2046 if (FORWARDER_BLOCK_P (d))
2047 d = single_succ (d);
2049 FOR_EACH_EDGE (s2, ei, src1->succs)
2051 basic_block d2 = s2->dest;
2052 if (FORWARDER_BLOCK_P (d2))
2053 d2 = single_succ (d2);
2054 if (d == d2)
2055 break;
2058 s->count += s2->count;
2060 /* Take care to update possible forwarder blocks. We verified
2061 that there is no more than one in the chain, so we can't run
2062 into infinite loop. */
2063 if (FORWARDER_BLOCK_P (s->dest))
2065 single_succ_edge (s->dest)->count += s2->count;
2066 s->dest->count += s2->count;
2067 s->dest->frequency += EDGE_FREQUENCY (s);
2070 if (FORWARDER_BLOCK_P (s2->dest))
2072 single_succ_edge (s2->dest)->count -= s2->count;
2073 if (single_succ_edge (s2->dest)->count < 0)
2074 single_succ_edge (s2->dest)->count = 0;
2075 s2->dest->count -= s2->count;
2076 s2->dest->frequency -= EDGE_FREQUENCY (s);
2077 if (s2->dest->frequency < 0)
2078 s2->dest->frequency = 0;
2079 if (s2->dest->count < 0)
2080 s2->dest->count = 0;
2083 if (!redirect_edges_to->frequency && !src1->frequency)
2084 s->probability = (s->probability + s2->probability) / 2;
2085 else
2086 s->probability
2087 = ((s->probability * redirect_edges_to->frequency +
2088 s2->probability * src1->frequency)
2089 / (redirect_edges_to->frequency + src1->frequency));
2092 /* Adjust count and frequency for the block. An earlier jump
2093 threading pass may have left the profile in an inconsistent
2094 state (see update_bb_profile_for_threading) so we must be
2095 prepared for overflows. */
2096 tmp = redirect_to;
2099 tmp->count += src1->count;
2100 tmp->frequency += src1->frequency;
2101 if (tmp->frequency > BB_FREQ_MAX)
2102 tmp->frequency = BB_FREQ_MAX;
2103 if (tmp == redirect_edges_to)
2104 break;
2105 tmp = find_fallthru_edge (tmp->succs)->dest;
2107 while (true);
2108 update_br_prob_note (redirect_edges_to);
2110 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2112 /* Skip possible basic block header. */
2113 if (LABEL_P (newpos1))
2114 newpos1 = NEXT_INSN (newpos1);
2116 while (DEBUG_INSN_P (newpos1))
2117 newpos1 = NEXT_INSN (newpos1);
2119 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2120 newpos1 = NEXT_INSN (newpos1);
2122 while (DEBUG_INSN_P (newpos1))
2123 newpos1 = NEXT_INSN (newpos1);
2125 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2126 to_remove = single_succ (redirect_from);
2128 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2129 delete_basic_block (to_remove);
2131 update_forwarder_flag (redirect_from);
2132 if (redirect_to != src2)
2133 update_forwarder_flag (src2);
2135 return true;
2138 /* Search the predecessors of BB for common insn sequences. When found,
2139 share code between them by redirecting control flow. Return true if
2140 any changes made. */
2142 static bool
2143 try_crossjump_bb (int mode, basic_block bb)
2145 edge e, e2, fallthru;
2146 bool changed;
2147 unsigned max, ix, ix2;
2149 /* Nothing to do if there is not at least two incoming edges. */
2150 if (EDGE_COUNT (bb->preds) < 2)
2151 return false;
2153 /* Don't crossjump if this block ends in a computed jump,
2154 unless we are optimizing for size. */
2155 if (optimize_bb_for_size_p (bb)
2156 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2157 && computed_jump_p (BB_END (bb)))
2158 return false;
2160 /* If we are partitioning hot/cold basic blocks, we don't want to
2161 mess up unconditional or indirect jumps that cross between hot
2162 and cold sections.
2164 Basic block partitioning may result in some jumps that appear to
2165 be optimizable (or blocks that appear to be mergeable), but which really
2166 must be left untouched (they are required to make it safely across
2167 partition boundaries). See the comments at the top of
2168 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2170 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2171 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2172 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2173 return false;
2175 /* It is always cheapest to redirect a block that ends in a branch to
2176 a block that falls through into BB, as that adds no branches to the
2177 program. We'll try that combination first. */
2178 fallthru = NULL;
2179 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2181 if (EDGE_COUNT (bb->preds) > max)
2182 return false;
2184 fallthru = find_fallthru_edge (bb->preds);
2186 changed = false;
2187 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2189 e = EDGE_PRED (bb, ix);
2190 ix++;
2192 /* As noted above, first try with the fallthru predecessor (or, a
2193 fallthru predecessor if we are in cfglayout mode). */
2194 if (fallthru)
2196 /* Don't combine the fallthru edge into anything else.
2197 If there is a match, we'll do it the other way around. */
2198 if (e == fallthru)
2199 continue;
2200 /* If nothing changed since the last attempt, there is nothing
2201 we can do. */
2202 if (!first_pass
2203 && !((e->src->flags & BB_MODIFIED)
2204 || (fallthru->src->flags & BB_MODIFIED)))
2205 continue;
2207 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2209 changed = true;
2210 ix = 0;
2211 continue;
2215 /* Non-obvious work limiting check: Recognize that we're going
2216 to call try_crossjump_bb on every basic block. So if we have
2217 two blocks with lots of outgoing edges (a switch) and they
2218 share lots of common destinations, then we would do the
2219 cross-jump check once for each common destination.
2221 Now, if the blocks actually are cross-jump candidates, then
2222 all of their destinations will be shared. Which means that
2223 we only need check them for cross-jump candidacy once. We
2224 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2225 choosing to do the check from the block for which the edge
2226 in question is the first successor of A. */
2227 if (EDGE_SUCC (e->src, 0) != e)
2228 continue;
2230 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2232 e2 = EDGE_PRED (bb, ix2);
2234 if (e2 == e)
2235 continue;
2237 /* We've already checked the fallthru edge above. */
2238 if (e2 == fallthru)
2239 continue;
2241 /* The "first successor" check above only prevents multiple
2242 checks of crossjump(A,B). In order to prevent redundant
2243 checks of crossjump(B,A), require that A be the block
2244 with the lowest index. */
2245 if (e->src->index > e2->src->index)
2246 continue;
2248 /* If nothing changed since the last attempt, there is nothing
2249 we can do. */
2250 if (!first_pass
2251 && !((e->src->flags & BB_MODIFIED)
2252 || (e2->src->flags & BB_MODIFIED)))
2253 continue;
2255 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2256 direction. */
2257 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2259 changed = true;
2260 ix = 0;
2261 break;
2266 if (changed)
2267 crossjumps_occured = true;
2269 return changed;
2272 /* Search the successors of BB for common insn sequences. When found,
2273 share code between them by moving it across the basic block
2274 boundary. Return true if any changes made. */
2276 static bool
2277 try_head_merge_bb (basic_block bb)
2279 basic_block final_dest_bb = NULL;
2280 int max_match = INT_MAX;
2281 edge e0;
2282 rtx_insn **headptr, **currptr, **nextptr;
2283 bool changed, moveall;
2284 unsigned ix;
2285 rtx_insn *e0_last_head;
2286 rtx cond;
2287 rtx_insn *move_before;
2288 unsigned nedges = EDGE_COUNT (bb->succs);
2289 rtx_insn *jump = BB_END (bb);
2290 regset live, live_union;
2292 /* Nothing to do if there is not at least two outgoing edges. */
2293 if (nedges < 2)
2294 return false;
2296 /* Don't crossjump if this block ends in a computed jump,
2297 unless we are optimizing for size. */
2298 if (optimize_bb_for_size_p (bb)
2299 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2300 && computed_jump_p (BB_END (bb)))
2301 return false;
2303 cond = get_condition (jump, &move_before, true, false);
2304 if (cond == NULL_RTX)
2306 #ifdef HAVE_cc0
2307 if (reg_mentioned_p (cc0_rtx, jump))
2308 move_before = prev_nonnote_nondebug_insn (jump);
2309 else
2310 #endif
2311 move_before = jump;
2314 for (ix = 0; ix < nedges; ix++)
2315 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
2316 return false;
2318 for (ix = 0; ix < nedges; ix++)
2320 edge e = EDGE_SUCC (bb, ix);
2321 basic_block other_bb = e->dest;
2323 if (df_get_bb_dirty (other_bb))
2325 block_was_dirty = true;
2326 return false;
2329 if (e->flags & EDGE_ABNORMAL)
2330 return false;
2332 /* Normally, all destination blocks must only be reachable from this
2333 block, i.e. they must have one incoming edge.
2335 There is one special case we can handle, that of multiple consecutive
2336 jumps where the first jumps to one of the targets of the second jump.
2337 This happens frequently in switch statements for default labels.
2338 The structure is as follows:
2339 FINAL_DEST_BB
2340 ....
2341 if (cond) jump A;
2342 fall through
2344 jump with targets A, B, C, D...
2346 has two incoming edges, from FINAL_DEST_BB and BB
2348 In this case, we can try to move the insns through BB and into
2349 FINAL_DEST_BB. */
2350 if (EDGE_COUNT (other_bb->preds) != 1)
2352 edge incoming_edge, incoming_bb_other_edge;
2353 edge_iterator ei;
2355 if (final_dest_bb != NULL
2356 || EDGE_COUNT (other_bb->preds) != 2)
2357 return false;
2359 /* We must be able to move the insns across the whole block. */
2360 move_before = BB_HEAD (bb);
2361 while (!NONDEBUG_INSN_P (move_before))
2362 move_before = NEXT_INSN (move_before);
2364 if (EDGE_COUNT (bb->preds) != 1)
2365 return false;
2366 incoming_edge = EDGE_PRED (bb, 0);
2367 final_dest_bb = incoming_edge->src;
2368 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2369 return false;
2370 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2371 if (incoming_bb_other_edge != incoming_edge)
2372 break;
2373 if (incoming_bb_other_edge->dest != other_bb)
2374 return false;
2378 e0 = EDGE_SUCC (bb, 0);
2379 e0_last_head = NULL;
2380 changed = false;
2382 for (ix = 1; ix < nedges; ix++)
2384 edge e = EDGE_SUCC (bb, ix);
2385 rtx_insn *e0_last, *e_last;
2386 int nmatch;
2388 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2389 &e0_last, &e_last, 0);
2390 if (nmatch == 0)
2391 return false;
2393 if (nmatch < max_match)
2395 max_match = nmatch;
2396 e0_last_head = e0_last;
2400 /* If we matched an entire block, we probably have to avoid moving the
2401 last insn. */
2402 if (max_match > 0
2403 && e0_last_head == BB_END (e0->dest)
2404 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2405 || control_flow_insn_p (e0_last_head)))
2407 max_match--;
2408 if (max_match == 0)
2409 return false;
2411 e0_last_head = prev_real_insn (e0_last_head);
2412 while (DEBUG_INSN_P (e0_last_head));
2415 if (max_match == 0)
2416 return false;
2418 /* We must find a union of the live registers at each of the end points. */
2419 live = BITMAP_ALLOC (NULL);
2420 live_union = BITMAP_ALLOC (NULL);
2422 currptr = XNEWVEC (rtx_insn *, nedges);
2423 headptr = XNEWVEC (rtx_insn *, nedges);
2424 nextptr = XNEWVEC (rtx_insn *, nedges);
2426 for (ix = 0; ix < nedges; ix++)
2428 int j;
2429 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2430 rtx_insn *head = BB_HEAD (merge_bb);
2432 while (!NONDEBUG_INSN_P (head))
2433 head = NEXT_INSN (head);
2434 headptr[ix] = head;
2435 currptr[ix] = head;
2437 /* Compute the end point and live information */
2438 for (j = 1; j < max_match; j++)
2440 head = NEXT_INSN (head);
2441 while (!NONDEBUG_INSN_P (head));
2442 simulate_backwards_to_point (merge_bb, live, head);
2443 IOR_REG_SET (live_union, live);
2446 /* If we're moving across two blocks, verify the validity of the
2447 first move, then adjust the target and let the loop below deal
2448 with the final move. */
2449 if (final_dest_bb != NULL)
2451 rtx_insn *move_upto;
2453 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2454 jump, e0->dest, live_union,
2455 NULL, &move_upto);
2456 if (!moveall)
2458 if (move_upto == NULL_RTX)
2459 goto out;
2461 while (e0_last_head != move_upto)
2463 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2464 live_union);
2465 e0_last_head = PREV_INSN (e0_last_head);
2468 if (e0_last_head == NULL_RTX)
2469 goto out;
2471 jump = BB_END (final_dest_bb);
2472 cond = get_condition (jump, &move_before, true, false);
2473 if (cond == NULL_RTX)
2475 #ifdef HAVE_cc0
2476 if (reg_mentioned_p (cc0_rtx, jump))
2477 move_before = prev_nonnote_nondebug_insn (jump);
2478 else
2479 #endif
2480 move_before = jump;
2486 rtx_insn *move_upto;
2487 moveall = can_move_insns_across (currptr[0], e0_last_head,
2488 move_before, jump, e0->dest, live_union,
2489 NULL, &move_upto);
2490 if (!moveall && move_upto == NULL_RTX)
2492 if (jump == move_before)
2493 break;
2495 /* Try again, using a different insertion point. */
2496 move_before = jump;
2498 #ifdef HAVE_cc0
2499 /* Don't try moving before a cc0 user, as that may invalidate
2500 the cc0. */
2501 if (reg_mentioned_p (cc0_rtx, jump))
2502 break;
2503 #endif
2505 continue;
2508 if (final_dest_bb && !moveall)
2509 /* We haven't checked whether a partial move would be OK for the first
2510 move, so we have to fail this case. */
2511 break;
2513 changed = true;
2514 for (;;)
2516 if (currptr[0] == move_upto)
2517 break;
2518 for (ix = 0; ix < nedges; ix++)
2520 rtx_insn *curr = currptr[ix];
2522 curr = NEXT_INSN (curr);
2523 while (!NONDEBUG_INSN_P (curr));
2524 currptr[ix] = curr;
2528 /* If we can't currently move all of the identical insns, remember
2529 each insn after the range that we'll merge. */
2530 if (!moveall)
2531 for (ix = 0; ix < nedges; ix++)
2533 rtx_insn *curr = currptr[ix];
2535 curr = NEXT_INSN (curr);
2536 while (!NONDEBUG_INSN_P (curr));
2537 nextptr[ix] = curr;
2540 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2541 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2542 if (final_dest_bb != NULL)
2543 df_set_bb_dirty (final_dest_bb);
2544 df_set_bb_dirty (bb);
2545 for (ix = 1; ix < nedges; ix++)
2547 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2548 delete_insn_chain (headptr[ix], currptr[ix], false);
2550 if (!moveall)
2552 if (jump == move_before)
2553 break;
2555 /* For the unmerged insns, try a different insertion point. */
2556 move_before = jump;
2558 #ifdef HAVE_cc0
2559 /* Don't try moving before a cc0 user, as that may invalidate
2560 the cc0. */
2561 if (reg_mentioned_p (cc0_rtx, jump))
2562 break;
2563 #endif
2565 for (ix = 0; ix < nedges; ix++)
2566 currptr[ix] = headptr[ix] = nextptr[ix];
2569 while (!moveall);
2571 out:
2572 free (currptr);
2573 free (headptr);
2574 free (nextptr);
2576 crossjumps_occured |= changed;
2578 return changed;
2581 /* Return true if BB contains just bb note, or bb note followed
2582 by only DEBUG_INSNs. */
2584 static bool
2585 trivially_empty_bb_p (basic_block bb)
2587 rtx_insn *insn = BB_END (bb);
2589 while (1)
2591 if (insn == BB_HEAD (bb))
2592 return true;
2593 if (!DEBUG_INSN_P (insn))
2594 return false;
2595 insn = PREV_INSN (insn);
2599 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2600 instructions etc. Return nonzero if changes were made. */
2602 static bool
2603 try_optimize_cfg (int mode)
2605 bool changed_overall = false;
2606 bool changed;
2607 int iterations = 0;
2608 basic_block bb, b, next;
2610 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2611 clear_bb_flags ();
2613 crossjumps_occured = false;
2615 FOR_EACH_BB_FN (bb, cfun)
2616 update_forwarder_flag (bb);
2618 if (! targetm.cannot_modify_jumps_p ())
2620 first_pass = true;
2621 /* Attempt to merge blocks as made possible by edge removal. If
2622 a block has only one successor, and the successor has only
2623 one predecessor, they may be combined. */
2626 block_was_dirty = false;
2627 changed = false;
2628 iterations++;
2630 if (dump_file)
2631 fprintf (dump_file,
2632 "\n\ntry_optimize_cfg iteration %i\n\n",
2633 iterations);
2635 for (b = ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb; b
2636 != EXIT_BLOCK_PTR_FOR_FN (cfun);)
2638 basic_block c;
2639 edge s;
2640 bool changed_here = false;
2642 /* Delete trivially dead basic blocks. This is either
2643 blocks with no predecessors, or empty blocks with no
2644 successors. However if the empty block with no
2645 successors is the successor of the ENTRY_BLOCK, it is
2646 kept. This ensures that the ENTRY_BLOCK will have a
2647 successor which is a precondition for many RTL
2648 passes. Empty blocks may result from expanding
2649 __builtin_unreachable (). */
2650 if (EDGE_COUNT (b->preds) == 0
2651 || (EDGE_COUNT (b->succs) == 0
2652 && trivially_empty_bb_p (b)
2653 && single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))->dest
2654 != b))
2656 c = b->prev_bb;
2657 if (EDGE_COUNT (b->preds) > 0)
2659 edge e;
2660 edge_iterator ei;
2662 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2664 if (BB_FOOTER (b)
2665 && BARRIER_P (BB_FOOTER (b)))
2666 FOR_EACH_EDGE (e, ei, b->preds)
2667 if ((e->flags & EDGE_FALLTHRU)
2668 && BB_FOOTER (e->src) == NULL)
2670 if (BB_FOOTER (b))
2672 BB_FOOTER (e->src) = BB_FOOTER (b);
2673 BB_FOOTER (b) = NULL;
2675 else
2677 start_sequence ();
2678 BB_FOOTER (e->src) = emit_barrier ();
2679 end_sequence ();
2683 else
2685 rtx_insn *last = get_last_bb_insn (b);
2686 if (last && BARRIER_P (last))
2687 FOR_EACH_EDGE (e, ei, b->preds)
2688 if ((e->flags & EDGE_FALLTHRU))
2689 emit_barrier_after (BB_END (e->src));
2692 delete_basic_block (b);
2693 changed = true;
2694 /* Avoid trying to remove the exit block. */
2695 b = (c == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? c->next_bb : c);
2696 continue;
2699 /* Remove code labels no longer used. */
2700 if (single_pred_p (b)
2701 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2702 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2703 && LABEL_P (BB_HEAD (b))
2704 /* If the previous block ends with a branch to this
2705 block, we can't delete the label. Normally this
2706 is a condjump that is yet to be simplified, but
2707 if CASE_DROPS_THRU, this can be a tablejump with
2708 some element going to the same place as the
2709 default (fallthru). */
2710 && (single_pred (b) == ENTRY_BLOCK_PTR_FOR_FN (cfun)
2711 || !JUMP_P (BB_END (single_pred (b)))
2712 || ! label_is_jump_target_p (BB_HEAD (b),
2713 BB_END (single_pred (b)))))
2715 delete_insn (BB_HEAD (b));
2716 if (dump_file)
2717 fprintf (dump_file, "Deleted label in block %i.\n",
2718 b->index);
2721 /* If we fall through an empty block, we can remove it. */
2722 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2723 && single_pred_p (b)
2724 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2725 && !LABEL_P (BB_HEAD (b))
2726 && FORWARDER_BLOCK_P (b)
2727 /* Note that forwarder_block_p true ensures that
2728 there is a successor for this block. */
2729 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2730 && n_basic_blocks_for_fn (cfun) > NUM_FIXED_BLOCKS + 1)
2732 if (dump_file)
2733 fprintf (dump_file,
2734 "Deleting fallthru block %i.\n",
2735 b->index);
2737 c = ((b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun))
2738 ? b->next_bb : b->prev_bb);
2739 redirect_edge_succ_nodup (single_pred_edge (b),
2740 single_succ (b));
2741 delete_basic_block (b);
2742 changed = true;
2743 b = c;
2744 continue;
2747 /* Merge B with its single successor, if any. */
2748 if (single_succ_p (b)
2749 && (s = single_succ_edge (b))
2750 && !(s->flags & EDGE_COMPLEX)
2751 && (c = s->dest) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2752 && single_pred_p (c)
2753 && b != c)
2755 /* When not in cfg_layout mode use code aware of reordering
2756 INSN. This code possibly creates new basic blocks so it
2757 does not fit merge_blocks interface and is kept here in
2758 hope that it will become useless once more of compiler
2759 is transformed to use cfg_layout mode. */
2761 if ((mode & CLEANUP_CFGLAYOUT)
2762 && can_merge_blocks_p (b, c))
2764 merge_blocks (b, c);
2765 update_forwarder_flag (b);
2766 changed_here = true;
2768 else if (!(mode & CLEANUP_CFGLAYOUT)
2769 /* If the jump insn has side effects,
2770 we can't kill the edge. */
2771 && (!JUMP_P (BB_END (b))
2772 || (reload_completed
2773 ? simplejump_p (BB_END (b))
2774 : (onlyjump_p (BB_END (b))
2775 && !tablejump_p (BB_END (b),
2776 NULL, NULL))))
2777 && (next = merge_blocks_move (s, b, c, mode)))
2779 b = next;
2780 changed_here = true;
2784 /* Simplify branch over branch. */
2785 if ((mode & CLEANUP_EXPENSIVE)
2786 && !(mode & CLEANUP_CFGLAYOUT)
2787 && try_simplify_condjump (b))
2788 changed_here = true;
2790 /* If B has a single outgoing edge, but uses a
2791 non-trivial jump instruction without side-effects, we
2792 can either delete the jump entirely, or replace it
2793 with a simple unconditional jump. */
2794 if (single_succ_p (b)
2795 && single_succ (b) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2796 && onlyjump_p (BB_END (b))
2797 && !CROSSING_JUMP_P (BB_END (b))
2798 && try_redirect_by_replacing_jump (single_succ_edge (b),
2799 single_succ (b),
2800 (mode & CLEANUP_CFGLAYOUT) != 0))
2802 update_forwarder_flag (b);
2803 changed_here = true;
2806 /* Simplify branch to branch. */
2807 if (try_forward_edges (mode, b))
2809 update_forwarder_flag (b);
2810 changed_here = true;
2813 /* Look for shared code between blocks. */
2814 if ((mode & CLEANUP_CROSSJUMP)
2815 && try_crossjump_bb (mode, b))
2816 changed_here = true;
2818 if ((mode & CLEANUP_CROSSJUMP)
2819 /* This can lengthen register lifetimes. Do it only after
2820 reload. */
2821 && reload_completed
2822 && try_head_merge_bb (b))
2823 changed_here = true;
2825 /* Don't get confused by the index shift caused by
2826 deleting blocks. */
2827 if (!changed_here)
2828 b = b->next_bb;
2829 else
2830 changed = true;
2833 if ((mode & CLEANUP_CROSSJUMP)
2834 && try_crossjump_bb (mode, EXIT_BLOCK_PTR_FOR_FN (cfun)))
2835 changed = true;
2837 if (block_was_dirty)
2839 /* This should only be set by head-merging. */
2840 gcc_assert (mode & CLEANUP_CROSSJUMP);
2841 df_analyze ();
2844 if (changed)
2846 /* Edge forwarding in particular can cause hot blocks previously
2847 reached by both hot and cold blocks to become dominated only
2848 by cold blocks. This will cause the verification below to fail,
2849 and lead to now cold code in the hot section. This is not easy
2850 to detect and fix during edge forwarding, and in some cases
2851 is only visible after newly unreachable blocks are deleted,
2852 which will be done in fixup_partitions. */
2853 fixup_partitions ();
2855 #ifdef ENABLE_CHECKING
2856 verify_flow_info ();
2857 #endif
2860 changed_overall |= changed;
2861 first_pass = false;
2863 while (changed);
2866 FOR_ALL_BB_FN (b, cfun)
2867 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
2869 return changed_overall;
2872 /* Delete all unreachable basic blocks. */
2874 bool
2875 delete_unreachable_blocks (void)
2877 bool changed = false;
2878 basic_block b, prev_bb;
2880 find_unreachable_blocks ();
2882 /* When we're in GIMPLE mode and there may be debug insns, we should
2883 delete blocks in reverse dominator order, so as to get a chance
2884 to substitute all released DEFs into debug stmts. If we don't
2885 have dominators information, walking blocks backward gets us a
2886 better chance of retaining most debug information than
2887 otherwise. */
2888 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
2889 && dom_info_available_p (CDI_DOMINATORS))
2891 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
2892 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
2894 prev_bb = b->prev_bb;
2896 if (!(b->flags & BB_REACHABLE))
2898 /* Speed up the removal of blocks that don't dominate
2899 others. Walking backwards, this should be the common
2900 case. */
2901 if (!first_dom_son (CDI_DOMINATORS, b))
2902 delete_basic_block (b);
2903 else
2905 vec<basic_block> h
2906 = get_all_dominated_blocks (CDI_DOMINATORS, b);
2908 while (h.length ())
2910 b = h.pop ();
2912 prev_bb = b->prev_bb;
2914 gcc_assert (!(b->flags & BB_REACHABLE));
2916 delete_basic_block (b);
2919 h.release ();
2922 changed = true;
2926 else
2928 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
2929 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
2931 prev_bb = b->prev_bb;
2933 if (!(b->flags & BB_REACHABLE))
2935 delete_basic_block (b);
2936 changed = true;
2941 if (changed)
2942 tidy_fallthru_edges ();
2943 return changed;
2946 /* Delete any jump tables never referenced. We can't delete them at the
2947 time of removing tablejump insn as they are referenced by the preceding
2948 insns computing the destination, so we delay deleting and garbagecollect
2949 them once life information is computed. */
2950 void
2951 delete_dead_jumptables (void)
2953 basic_block bb;
2955 /* A dead jump table does not belong to any basic block. Scan insns
2956 between two adjacent basic blocks. */
2957 FOR_EACH_BB_FN (bb, cfun)
2959 rtx_insn *insn, *next;
2961 for (insn = NEXT_INSN (BB_END (bb));
2962 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
2963 insn = next)
2965 next = NEXT_INSN (insn);
2966 if (LABEL_P (insn)
2967 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
2968 && JUMP_TABLE_DATA_P (next))
2970 rtx_insn *label = insn, *jump = next;
2972 if (dump_file)
2973 fprintf (dump_file, "Dead jumptable %i removed\n",
2974 INSN_UID (insn));
2976 next = NEXT_INSN (next);
2977 delete_insn (jump);
2978 delete_insn (label);
2985 /* Tidy the CFG by deleting unreachable code and whatnot. */
2987 bool
2988 cleanup_cfg (int mode)
2990 bool changed = false;
2992 /* Set the cfglayout mode flag here. We could update all the callers
2993 but that is just inconvenient, especially given that we eventually
2994 want to have cfglayout mode as the default. */
2995 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2996 mode |= CLEANUP_CFGLAYOUT;
2998 timevar_push (TV_CLEANUP_CFG);
2999 if (delete_unreachable_blocks ())
3001 changed = true;
3002 /* We've possibly created trivially dead code. Cleanup it right
3003 now to introduce more opportunities for try_optimize_cfg. */
3004 if (!(mode & (CLEANUP_NO_INSN_DEL))
3005 && !reload_completed)
3006 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3009 compact_blocks ();
3011 /* To tail-merge blocks ending in the same noreturn function (e.g.
3012 a call to abort) we have to insert fake edges to exit. Do this
3013 here once. The fake edges do not interfere with any other CFG
3014 cleanups. */
3015 if (mode & CLEANUP_CROSSJUMP)
3016 add_noreturn_fake_exit_edges ();
3018 if (!dbg_cnt (cfg_cleanup))
3019 return changed;
3021 while (try_optimize_cfg (mode))
3023 delete_unreachable_blocks (), changed = true;
3024 if (!(mode & CLEANUP_NO_INSN_DEL))
3026 /* Try to remove some trivially dead insns when doing an expensive
3027 cleanup. But delete_trivially_dead_insns doesn't work after
3028 reload (it only handles pseudos) and run_fast_dce is too costly
3029 to run in every iteration.
3031 For effective cross jumping, we really want to run a fast DCE to
3032 clean up any dead conditions, or they get in the way of performing
3033 useful tail merges.
3035 Other transformations in cleanup_cfg are not so sensitive to dead
3036 code, so delete_trivially_dead_insns or even doing nothing at all
3037 is good enough. */
3038 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3039 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3040 break;
3041 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occured)
3042 run_fast_dce ();
3044 else
3045 break;
3048 if (mode & CLEANUP_CROSSJUMP)
3049 remove_fake_exit_edges ();
3051 /* Don't call delete_dead_jumptables in cfglayout mode, because
3052 that function assumes that jump tables are in the insns stream.
3053 But we also don't _have_ to delete dead jumptables in cfglayout
3054 mode because we shouldn't even be looking at things that are
3055 not in a basic block. Dead jumptables are cleaned up when
3056 going out of cfglayout mode. */
3057 if (!(mode & CLEANUP_CFGLAYOUT))
3058 delete_dead_jumptables ();
3060 /* ??? We probably do this way too often. */
3061 if (current_loops
3062 && (changed
3063 || (mode & CLEANUP_CFG_CHANGED)))
3065 timevar_push (TV_REPAIR_LOOPS);
3066 /* The above doesn't preserve dominance info if available. */
3067 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3068 calculate_dominance_info (CDI_DOMINATORS);
3069 fix_loop_structure (NULL);
3070 free_dominance_info (CDI_DOMINATORS);
3071 timevar_pop (TV_REPAIR_LOOPS);
3074 timevar_pop (TV_CLEANUP_CFG);
3076 return changed;
3079 namespace {
3081 const pass_data pass_data_jump =
3083 RTL_PASS, /* type */
3084 "jump", /* name */
3085 OPTGROUP_NONE, /* optinfo_flags */
3086 TV_JUMP, /* tv_id */
3087 0, /* properties_required */
3088 0, /* properties_provided */
3089 0, /* properties_destroyed */
3090 0, /* todo_flags_start */
3091 0, /* todo_flags_finish */
3094 class pass_jump : public rtl_opt_pass
3096 public:
3097 pass_jump (gcc::context *ctxt)
3098 : rtl_opt_pass (pass_data_jump, ctxt)
3101 /* opt_pass methods: */
3102 virtual unsigned int execute (function *);
3104 }; // class pass_jump
3106 unsigned int
3107 pass_jump::execute (function *)
3109 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3110 if (dump_file)
3111 dump_flow_info (dump_file, dump_flags);
3112 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3113 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3114 return 0;
3117 } // anon namespace
3119 rtl_opt_pass *
3120 make_pass_jump (gcc::context *ctxt)
3122 return new pass_jump (ctxt);
3125 namespace {
3127 const pass_data pass_data_jump2 =
3129 RTL_PASS, /* type */
3130 "jump2", /* name */
3131 OPTGROUP_NONE, /* optinfo_flags */
3132 TV_JUMP, /* tv_id */
3133 0, /* properties_required */
3134 0, /* properties_provided */
3135 0, /* properties_destroyed */
3136 0, /* todo_flags_start */
3137 0, /* todo_flags_finish */
3140 class pass_jump2 : public rtl_opt_pass
3142 public:
3143 pass_jump2 (gcc::context *ctxt)
3144 : rtl_opt_pass (pass_data_jump2, ctxt)
3147 /* opt_pass methods: */
3148 virtual unsigned int execute (function *)
3150 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3151 return 0;
3154 }; // class pass_jump2
3156 } // anon namespace
3158 rtl_opt_pass *
3159 make_pass_jump2 (gcc::context *ctxt)
3161 return new pass_jump2 (ctxt);