libgo: add misc/cgo files
[official-gcc.git] / gcc / cfgcleanup.c
blob1201148fa88d728019e0768424d878365b26d19b
1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987-2017 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 "backend.h"
36 #include "target.h"
37 #include "rtl.h"
38 #include "tree.h"
39 #include "cfghooks.h"
40 #include "df.h"
41 #include "memmodel.h"
42 #include "tm_p.h"
43 #include "insn-config.h"
44 #include "emit-rtl.h"
45 #include "cselib.h"
46 #include "params.h"
47 #include "tree-pass.h"
48 #include "cfgloop.h"
49 #include "cfgrtl.h"
50 #include "cfganal.h"
51 #include "cfgbuild.h"
52 #include "cfgcleanup.h"
53 #include "dce.h"
54 #include "dbgcnt.h"
55 #include "rtl-iter.h"
57 #define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
59 /* Set to true when we are running first pass of try_optimize_cfg loop. */
60 static bool first_pass;
62 /* Set to true if crossjumps occurred in the latest run of try_optimize_cfg. */
63 static bool crossjumps_occurred;
65 /* Set to true if we couldn't run an optimization due to stale liveness
66 information; we should run df_analyze to enable more opportunities. */
67 static bool block_was_dirty;
69 static bool try_crossjump_to_edge (int, edge, edge, enum replace_direction);
70 static bool try_crossjump_bb (int, basic_block);
71 static bool outgoing_edges_match (int, basic_block, basic_block);
72 static enum replace_direction old_insns_match_p (int, rtx_insn *, rtx_insn *);
74 static void merge_blocks_move_predecessor_nojumps (basic_block, basic_block);
75 static void merge_blocks_move_successor_nojumps (basic_block, basic_block);
76 static bool try_optimize_cfg (int);
77 static bool try_simplify_condjump (basic_block);
78 static bool try_forward_edges (int, basic_block);
79 static edge thread_jump (edge, basic_block);
80 static bool mark_effect (rtx, bitmap);
81 static void notice_new_block (basic_block);
82 static void update_forwarder_flag (basic_block);
83 static void merge_memattrs (rtx, rtx);
85 /* Set flags for newly created block. */
87 static void
88 notice_new_block (basic_block bb)
90 if (!bb)
91 return;
93 if (forwarder_block_p (bb))
94 bb->flags |= BB_FORWARDER_BLOCK;
97 /* Recompute forwarder flag after block has been modified. */
99 static void
100 update_forwarder_flag (basic_block bb)
102 if (forwarder_block_p (bb))
103 bb->flags |= BB_FORWARDER_BLOCK;
104 else
105 bb->flags &= ~BB_FORWARDER_BLOCK;
108 /* Simplify a conditional jump around an unconditional jump.
109 Return true if something changed. */
111 static bool
112 try_simplify_condjump (basic_block cbranch_block)
114 basic_block jump_block, jump_dest_block, cbranch_dest_block;
115 edge cbranch_jump_edge, cbranch_fallthru_edge;
116 rtx_insn *cbranch_insn;
118 /* Verify that there are exactly two successors. */
119 if (EDGE_COUNT (cbranch_block->succs) != 2)
120 return false;
122 /* Verify that we've got a normal conditional branch at the end
123 of the block. */
124 cbranch_insn = BB_END (cbranch_block);
125 if (!any_condjump_p (cbranch_insn))
126 return false;
128 cbranch_fallthru_edge = FALLTHRU_EDGE (cbranch_block);
129 cbranch_jump_edge = BRANCH_EDGE (cbranch_block);
131 /* The next block must not have multiple predecessors, must not
132 be the last block in the function, and must contain just the
133 unconditional jump. */
134 jump_block = cbranch_fallthru_edge->dest;
135 if (!single_pred_p (jump_block)
136 || jump_block->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
137 || !FORWARDER_BLOCK_P (jump_block))
138 return false;
139 jump_dest_block = single_succ (jump_block);
141 /* If we are partitioning hot/cold basic blocks, we don't want to
142 mess up unconditional or indirect jumps that cross between hot
143 and cold sections.
145 Basic block partitioning may result in some jumps that appear to
146 be optimizable (or blocks that appear to be mergeable), but which really
147 must be left untouched (they are required to make it safely across
148 partition boundaries). See the comments at the top of
149 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
151 if (BB_PARTITION (jump_block) != BB_PARTITION (jump_dest_block)
152 || (cbranch_jump_edge->flags & EDGE_CROSSING))
153 return false;
155 /* The conditional branch must target the block after the
156 unconditional branch. */
157 cbranch_dest_block = cbranch_jump_edge->dest;
159 if (cbranch_dest_block == EXIT_BLOCK_PTR_FOR_FN (cfun)
160 || jump_dest_block == EXIT_BLOCK_PTR_FOR_FN (cfun)
161 || !can_fallthru (jump_block, cbranch_dest_block))
162 return false;
164 /* Invert the conditional branch. */
165 if (!invert_jump (as_a <rtx_jump_insn *> (cbranch_insn),
166 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 rtx dest;
199 switch (GET_CODE (exp))
201 /* In case we do clobber the register, mark it as equal, as we know the
202 value is dead so it don't have to match. */
203 case CLOBBER:
204 dest = XEXP (exp, 0);
205 if (REG_P (dest))
206 bitmap_clear_range (nonequal, REGNO (dest), REG_NREGS (dest));
207 return false;
209 case SET:
210 if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
211 return false;
212 dest = SET_DEST (exp);
213 if (dest == pc_rtx)
214 return false;
215 if (!REG_P (dest))
216 return true;
217 bitmap_set_range (nonequal, REGNO (dest), REG_NREGS (dest));
218 return false;
220 default:
221 return false;
225 /* Return true if X contains a register in NONEQUAL. */
226 static bool
227 mentions_nonequal_regs (const_rtx x, regset nonequal)
229 subrtx_iterator::array_type array;
230 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
232 const_rtx x = *iter;
233 if (REG_P (x))
235 unsigned int end_regno = END_REGNO (x);
236 for (unsigned int regno = REGNO (x); regno < end_regno; ++regno)
237 if (REGNO_REG_SET_P (nonequal, regno))
238 return true;
241 return false;
244 /* Attempt to prove that the basic block B will have no side effects and
245 always continues in the same edge if reached via E. Return the edge
246 if exist, NULL otherwise. */
248 static edge
249 thread_jump (edge e, basic_block b)
251 rtx set1, set2, cond1, cond2;
252 rtx_insn *insn;
253 enum rtx_code code1, code2, reversed_code2;
254 bool reverse1 = false;
255 unsigned i;
256 regset nonequal;
257 bool failed = false;
258 reg_set_iterator rsi;
260 if (b->flags & BB_NONTHREADABLE_BLOCK)
261 return NULL;
263 /* At the moment, we do handle only conditional jumps, but later we may
264 want to extend this code to tablejumps and others. */
265 if (EDGE_COUNT (e->src->succs) != 2)
266 return NULL;
267 if (EDGE_COUNT (b->succs) != 2)
269 b->flags |= BB_NONTHREADABLE_BLOCK;
270 return NULL;
273 /* Second branch must end with onlyjump, as we will eliminate the jump. */
274 if (!any_condjump_p (BB_END (e->src)))
275 return NULL;
277 if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
279 b->flags |= BB_NONTHREADABLE_BLOCK;
280 return NULL;
283 set1 = pc_set (BB_END (e->src));
284 set2 = pc_set (BB_END (b));
285 if (((e->flags & EDGE_FALLTHRU) != 0)
286 != (XEXP (SET_SRC (set1), 1) == pc_rtx))
287 reverse1 = true;
289 cond1 = XEXP (SET_SRC (set1), 0);
290 cond2 = XEXP (SET_SRC (set2), 0);
291 if (reverse1)
292 code1 = reversed_comparison_code (cond1, BB_END (e->src));
293 else
294 code1 = GET_CODE (cond1);
296 code2 = GET_CODE (cond2);
297 reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
299 if (!comparison_dominates_p (code1, code2)
300 && !comparison_dominates_p (code1, reversed_code2))
301 return NULL;
303 /* Ensure that the comparison operators are equivalent.
304 ??? This is far too pessimistic. We should allow swapped operands,
305 different CCmodes, or for example comparisons for interval, that
306 dominate even when operands are not equivalent. */
307 if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
308 || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
309 return NULL;
311 /* Short circuit cases where block B contains some side effects, as we can't
312 safely bypass it. */
313 for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
314 insn = NEXT_INSN (insn))
315 if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
317 b->flags |= BB_NONTHREADABLE_BLOCK;
318 return NULL;
321 cselib_init (0);
323 /* First process all values computed in the source basic block. */
324 for (insn = NEXT_INSN (BB_HEAD (e->src));
325 insn != NEXT_INSN (BB_END (e->src));
326 insn = NEXT_INSN (insn))
327 if (INSN_P (insn))
328 cselib_process_insn (insn);
330 nonequal = BITMAP_ALLOC (NULL);
331 CLEAR_REG_SET (nonequal);
333 /* Now assume that we've continued by the edge E to B and continue
334 processing as if it were same basic block.
335 Our goal is to prove that whole block is an NOOP. */
337 for (insn = NEXT_INSN (BB_HEAD (b));
338 insn != NEXT_INSN (BB_END (b)) && !failed;
339 insn = NEXT_INSN (insn))
341 if (INSN_P (insn))
343 rtx pat = PATTERN (insn);
345 if (GET_CODE (pat) == PARALLEL)
347 for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
348 failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
350 else
351 failed |= mark_effect (pat, nonequal);
354 cselib_process_insn (insn);
357 /* Later we should clear nonequal of dead registers. So far we don't
358 have life information in cfg_cleanup. */
359 if (failed)
361 b->flags |= BB_NONTHREADABLE_BLOCK;
362 goto failed_exit;
365 /* cond2 must not mention any register that is not equal to the
366 former block. */
367 if (mentions_nonequal_regs (cond2, nonequal))
368 goto failed_exit;
370 EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
371 goto failed_exit;
373 BITMAP_FREE (nonequal);
374 cselib_finish ();
375 if ((comparison_dominates_p (code1, code2) != 0)
376 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
377 return BRANCH_EDGE (b);
378 else
379 return FALLTHRU_EDGE (b);
381 failed_exit:
382 BITMAP_FREE (nonequal);
383 cselib_finish ();
384 return NULL;
387 /* Attempt to forward edges leaving basic block B.
388 Return true if successful. */
390 static bool
391 try_forward_edges (int mode, basic_block b)
393 bool changed = false;
394 edge_iterator ei;
395 edge e, *threaded_edges = NULL;
397 /* If we are partitioning hot/cold basic blocks, we don't want to
398 mess up unconditional or indirect jumps that cross between hot
399 and cold sections.
401 Basic block partitioning may result in some jumps that appear to
402 be optimizable (or blocks that appear to be mergeable), but which really
403 must be left untouched (they are required to make it safely across
404 partition boundaries). See the comments at the top of
405 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
407 if (JUMP_P (BB_END (b)) && CROSSING_JUMP_P (BB_END (b)))
408 return false;
410 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
412 basic_block target, first;
413 location_t goto_locus;
414 int counter;
415 bool threaded = false;
416 int nthreaded_edges = 0;
417 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
419 /* Skip complex edges because we don't know how to update them.
421 Still handle fallthru edges, as we can succeed to forward fallthru
422 edge to the same place as the branch edge of conditional branch
423 and turn conditional branch to an unconditional branch. */
424 if (e->flags & EDGE_COMPLEX)
426 ei_next (&ei);
427 continue;
430 target = first = e->dest;
431 counter = NUM_FIXED_BLOCKS;
432 goto_locus = e->goto_locus;
434 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
435 up jumps that cross between hot/cold sections.
437 Basic block partitioning may result in some jumps that appear
438 to be optimizable (or blocks that appear to be mergeable), but which
439 really must be left untouched (they are required to make it safely
440 across partition boundaries). See the comments at the top of
441 bb-reorder.c:partition_hot_cold_basic_blocks for complete
442 details. */
444 if (first != EXIT_BLOCK_PTR_FOR_FN (cfun)
445 && JUMP_P (BB_END (first))
446 && CROSSING_JUMP_P (BB_END (first)))
447 return changed;
449 while (counter < n_basic_blocks_for_fn (cfun))
451 basic_block new_target = NULL;
452 bool new_target_threaded = false;
453 may_thread |= (target->flags & BB_MODIFIED) != 0;
455 if (FORWARDER_BLOCK_P (target)
456 && !(single_succ_edge (target)->flags & EDGE_CROSSING)
457 && single_succ (target) != EXIT_BLOCK_PTR_FOR_FN (cfun))
459 /* Bypass trivial infinite loops. */
460 new_target = single_succ (target);
461 if (target == new_target)
462 counter = n_basic_blocks_for_fn (cfun);
463 else if (!optimize)
465 /* When not optimizing, ensure that edges or forwarder
466 blocks with different locus are not optimized out. */
467 location_t new_locus = single_succ_edge (target)->goto_locus;
468 location_t locus = goto_locus;
470 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
471 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
472 && new_locus != locus)
473 new_target = NULL;
474 else
476 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
477 locus = new_locus;
479 rtx_insn *last = BB_END (target);
480 if (DEBUG_INSN_P (last))
481 last = prev_nondebug_insn (last);
482 if (last && INSN_P (last))
483 new_locus = INSN_LOCATION (last);
484 else
485 new_locus = UNKNOWN_LOCATION;
487 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
488 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
489 && new_locus != locus)
490 new_target = NULL;
491 else
493 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
494 locus = new_locus;
496 goto_locus = locus;
502 /* Allow to thread only over one edge at time to simplify updating
503 of probabilities. */
504 else if ((mode & CLEANUP_THREADING) && may_thread)
506 edge t = thread_jump (e, target);
507 if (t)
509 if (!threaded_edges)
510 threaded_edges = XNEWVEC (edge,
511 n_basic_blocks_for_fn (cfun));
512 else
514 int i;
516 /* Detect an infinite loop across blocks not
517 including the start block. */
518 for (i = 0; i < nthreaded_edges; ++i)
519 if (threaded_edges[i] == t)
520 break;
521 if (i < nthreaded_edges)
523 counter = n_basic_blocks_for_fn (cfun);
524 break;
528 /* Detect an infinite loop across the start block. */
529 if (t->dest == b)
530 break;
532 gcc_assert (nthreaded_edges
533 < (n_basic_blocks_for_fn (cfun)
534 - NUM_FIXED_BLOCKS));
535 threaded_edges[nthreaded_edges++] = t;
537 new_target = t->dest;
538 new_target_threaded = true;
542 if (!new_target)
543 break;
545 counter++;
546 target = new_target;
547 threaded |= new_target_threaded;
550 if (counter >= n_basic_blocks_for_fn (cfun))
552 if (dump_file)
553 fprintf (dump_file, "Infinite loop in BB %i.\n",
554 target->index);
556 else if (target == first)
557 ; /* We didn't do anything. */
558 else
560 /* Save the values now, as the edge may get removed. */
561 profile_count edge_count = e->count;
562 int edge_probability = e->probability;
563 int edge_frequency;
564 int n = 0;
566 e->goto_locus = goto_locus;
568 /* Don't force if target is exit block. */
569 if (threaded && target != EXIT_BLOCK_PTR_FOR_FN (cfun))
571 notice_new_block (redirect_edge_and_branch_force (e, target));
572 if (dump_file)
573 fprintf (dump_file, "Conditionals threaded.\n");
575 else if (!redirect_edge_and_branch (e, target))
577 if (dump_file)
578 fprintf (dump_file,
579 "Forwarding edge %i->%i to %i failed.\n",
580 b->index, e->dest->index, target->index);
581 ei_next (&ei);
582 continue;
585 /* We successfully forwarded the edge. Now update profile
586 data: for each edge we traversed in the chain, remove
587 the original edge's execution count. */
588 edge_frequency = apply_probability (b->frequency, edge_probability);
592 edge t;
594 if (!single_succ_p (first))
596 gcc_assert (n < nthreaded_edges);
597 t = threaded_edges [n++];
598 gcc_assert (t->src == first);
599 update_bb_profile_for_threading (first, edge_frequency,
600 edge_count, t);
601 update_br_prob_note (first);
603 else
605 first->count -= edge_count;
606 first->frequency -= edge_frequency;
607 if (first->frequency < 0)
608 first->frequency = 0;
609 /* It is possible that as the result of
610 threading we've removed edge as it is
611 threaded to the fallthru edge. Avoid
612 getting out of sync. */
613 if (n < nthreaded_edges
614 && first == threaded_edges [n]->src)
615 n++;
616 t = single_succ_edge (first);
619 t->count -= edge_count;
620 first = t->dest;
622 while (first != target);
624 changed = true;
625 continue;
627 ei_next (&ei);
630 free (threaded_edges);
631 return changed;
635 /* Blocks A and B are to be merged into a single block. A has no incoming
636 fallthru edge, so it can be moved before B without adding or modifying
637 any jumps (aside from the jump from A to B). */
639 static void
640 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
642 rtx_insn *barrier;
644 /* If we are partitioning hot/cold basic blocks, we don't want to
645 mess up unconditional or indirect jumps that cross between hot
646 and cold sections.
648 Basic block partitioning may result in some jumps that appear to
649 be optimizable (or blocks that appear to be mergeable), but which really
650 must be left untouched (they are required to make it safely across
651 partition boundaries). See the comments at the top of
652 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
654 if (BB_PARTITION (a) != BB_PARTITION (b))
655 return;
657 barrier = next_nonnote_insn (BB_END (a));
658 gcc_assert (BARRIER_P (barrier));
659 delete_insn (barrier);
661 /* Scramble the insn chain. */
662 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
663 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
664 df_set_bb_dirty (a);
666 if (dump_file)
667 fprintf (dump_file, "Moved block %d before %d and merged.\n",
668 a->index, b->index);
670 /* Swap the records for the two blocks around. */
672 unlink_block (a);
673 link_block (a, b->prev_bb);
675 /* Now blocks A and B are contiguous. Merge them. */
676 merge_blocks (a, b);
679 /* Blocks A and B are to be merged into a single block. B has no outgoing
680 fallthru edge, so it can be moved after A without adding or modifying
681 any jumps (aside from the jump from A to B). */
683 static void
684 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
686 rtx_insn *barrier, *real_b_end;
687 rtx_insn *label;
688 rtx_jump_table_data *table;
690 /* If we are partitioning hot/cold basic blocks, we don't want to
691 mess up unconditional or indirect jumps that cross between hot
692 and cold sections.
694 Basic block partitioning may result in some jumps that appear to
695 be optimizable (or blocks that appear to be mergeable), but which really
696 must be left untouched (they are required to make it safely across
697 partition boundaries). See the comments at the top of
698 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
700 if (BB_PARTITION (a) != BB_PARTITION (b))
701 return;
703 real_b_end = BB_END (b);
705 /* If there is a jump table following block B temporarily add the jump table
706 to block B so that it will also be moved to the correct location. */
707 if (tablejump_p (BB_END (b), &label, &table)
708 && prev_active_insn (label) == BB_END (b))
710 BB_END (b) = table;
713 /* There had better have been a barrier there. Delete it. */
714 barrier = NEXT_INSN (BB_END (b));
715 if (barrier && BARRIER_P (barrier))
716 delete_insn (barrier);
719 /* Scramble the insn chain. */
720 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
722 /* Restore the real end of b. */
723 BB_END (b) = real_b_end;
725 if (dump_file)
726 fprintf (dump_file, "Moved block %d after %d and merged.\n",
727 b->index, a->index);
729 /* Now blocks A and B are contiguous. Merge them. */
730 merge_blocks (a, b);
733 /* Attempt to merge basic blocks that are potentially non-adjacent.
734 Return NULL iff the attempt failed, otherwise return basic block
735 where cleanup_cfg should continue. Because the merging commonly
736 moves basic block away or introduces another optimization
737 possibility, return basic block just before B so cleanup_cfg don't
738 need to iterate.
740 It may be good idea to return basic block before C in the case
741 C has been moved after B and originally appeared earlier in the
742 insn sequence, but we have no information available about the
743 relative ordering of these two. Hopefully it is not too common. */
745 static basic_block
746 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
748 basic_block next;
750 /* If we are partitioning hot/cold basic blocks, we don't want to
751 mess up unconditional or indirect jumps that cross between hot
752 and cold sections.
754 Basic block partitioning may result in some jumps that appear to
755 be optimizable (or blocks that appear to be mergeable), but which really
756 must be left untouched (they are required to make it safely across
757 partition boundaries). See the comments at the top of
758 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
760 if (BB_PARTITION (b) != BB_PARTITION (c))
761 return NULL;
763 /* If B has a fallthru edge to C, no need to move anything. */
764 if (e->flags & EDGE_FALLTHRU)
766 int b_index = b->index, c_index = c->index;
768 /* Protect the loop latches. */
769 if (current_loops && c->loop_father->latch == c)
770 return NULL;
772 merge_blocks (b, c);
773 update_forwarder_flag (b);
775 if (dump_file)
776 fprintf (dump_file, "Merged %d and %d without moving.\n",
777 b_index, c_index);
779 return b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? b : b->prev_bb;
782 /* Otherwise we will need to move code around. Do that only if expensive
783 transformations are allowed. */
784 else if (mode & CLEANUP_EXPENSIVE)
786 edge tmp_edge, b_fallthru_edge;
787 bool c_has_outgoing_fallthru;
788 bool b_has_incoming_fallthru;
790 /* Avoid overactive code motion, as the forwarder blocks should be
791 eliminated by edge redirection instead. One exception might have
792 been if B is a forwarder block and C has no fallthru edge, but
793 that should be cleaned up by bb-reorder instead. */
794 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
795 return NULL;
797 /* We must make sure to not munge nesting of lexical blocks,
798 and loop notes. This is done by squeezing out all the notes
799 and leaving them there to lie. Not ideal, but functional. */
801 tmp_edge = find_fallthru_edge (c->succs);
802 c_has_outgoing_fallthru = (tmp_edge != NULL);
804 tmp_edge = find_fallthru_edge (b->preds);
805 b_has_incoming_fallthru = (tmp_edge != NULL);
806 b_fallthru_edge = tmp_edge;
807 next = b->prev_bb;
808 if (next == c)
809 next = next->prev_bb;
811 /* Otherwise, we're going to try to move C after B. If C does
812 not have an outgoing fallthru, then it can be moved
813 immediately after B without introducing or modifying jumps. */
814 if (! c_has_outgoing_fallthru)
816 merge_blocks_move_successor_nojumps (b, c);
817 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
820 /* If B does not have an incoming fallthru, then it can be moved
821 immediately before C without introducing or modifying jumps.
822 C cannot be the first block, so we do not have to worry about
823 accessing a non-existent block. */
825 if (b_has_incoming_fallthru)
827 basic_block bb;
829 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
830 return NULL;
831 bb = force_nonfallthru (b_fallthru_edge);
832 if (bb)
833 notice_new_block (bb);
836 merge_blocks_move_predecessor_nojumps (b, c);
837 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
840 return NULL;
844 /* Removes the memory attributes of MEM expression
845 if they are not equal. */
847 static void
848 merge_memattrs (rtx x, rtx y)
850 int i;
851 int j;
852 enum rtx_code code;
853 const char *fmt;
855 if (x == y)
856 return;
857 if (x == 0 || y == 0)
858 return;
860 code = GET_CODE (x);
862 if (code != GET_CODE (y))
863 return;
865 if (GET_MODE (x) != GET_MODE (y))
866 return;
868 if (code == MEM && !mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
870 if (! MEM_ATTRS (x))
871 MEM_ATTRS (y) = 0;
872 else if (! MEM_ATTRS (y))
873 MEM_ATTRS (x) = 0;
874 else
876 HOST_WIDE_INT mem_size;
878 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
880 set_mem_alias_set (x, 0);
881 set_mem_alias_set (y, 0);
884 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
886 set_mem_expr (x, 0);
887 set_mem_expr (y, 0);
888 clear_mem_offset (x);
889 clear_mem_offset (y);
891 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
892 || (MEM_OFFSET_KNOWN_P (x)
893 && MEM_OFFSET (x) != MEM_OFFSET (y)))
895 clear_mem_offset (x);
896 clear_mem_offset (y);
899 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
901 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
902 set_mem_size (x, mem_size);
903 set_mem_size (y, mem_size);
905 else
907 clear_mem_size (x);
908 clear_mem_size (y);
911 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
912 set_mem_align (y, MEM_ALIGN (x));
915 if (code == MEM)
917 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
919 MEM_READONLY_P (x) = 0;
920 MEM_READONLY_P (y) = 0;
922 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
924 MEM_NOTRAP_P (x) = 0;
925 MEM_NOTRAP_P (y) = 0;
927 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
929 MEM_VOLATILE_P (x) = 1;
930 MEM_VOLATILE_P (y) = 1;
934 fmt = GET_RTX_FORMAT (code);
935 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
937 switch (fmt[i])
939 case 'E':
940 /* Two vectors must have the same length. */
941 if (XVECLEN (x, i) != XVECLEN (y, i))
942 return;
944 for (j = 0; j < XVECLEN (x, i); j++)
945 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
947 break;
949 case 'e':
950 merge_memattrs (XEXP (x, i), XEXP (y, i));
953 return;
957 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
958 different single sets S1 and S2. */
960 static bool
961 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
963 int i;
964 rtx e1, e2;
966 if (p1 == s1 && p2 == s2)
967 return true;
969 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
970 return false;
972 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
973 return false;
975 for (i = 0; i < XVECLEN (p1, 0); i++)
977 e1 = XVECEXP (p1, 0, i);
978 e2 = XVECEXP (p2, 0, i);
979 if (e1 == s1 && e2 == s2)
980 continue;
981 if (reload_completed
982 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
983 continue;
985 return false;
988 return true;
992 /* NOTE1 is the REG_EQUAL note, if any, attached to an insn
993 that is a single_set with a SET_SRC of SRC1. Similarly
994 for NOTE2/SRC2.
996 So effectively NOTE1/NOTE2 are an alternate form of
997 SRC1/SRC2 respectively.
999 Return nonzero if SRC1 or NOTE1 has the same constant
1000 integer value as SRC2 or NOTE2. Else return zero. */
1001 static int
1002 values_equal_p (rtx note1, rtx note2, rtx src1, rtx src2)
1004 if (note1
1005 && note2
1006 && CONST_INT_P (XEXP (note1, 0))
1007 && rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0)))
1008 return 1;
1010 if (!note1
1011 && !note2
1012 && CONST_INT_P (src1)
1013 && CONST_INT_P (src2)
1014 && rtx_equal_p (src1, src2))
1015 return 1;
1017 if (note1
1018 && CONST_INT_P (src2)
1019 && rtx_equal_p (XEXP (note1, 0), src2))
1020 return 1;
1022 if (note2
1023 && CONST_INT_P (src1)
1024 && rtx_equal_p (XEXP (note2, 0), src1))
1025 return 1;
1027 return 0;
1030 /* Examine register notes on I1 and I2 and return:
1031 - dir_forward if I1 can be replaced by I2, or
1032 - dir_backward if I2 can be replaced by I1, or
1033 - dir_both if both are the case. */
1035 static enum replace_direction
1036 can_replace_by (rtx_insn *i1, rtx_insn *i2)
1038 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1039 bool c1, c2;
1041 /* Check for 2 sets. */
1042 s1 = single_set (i1);
1043 s2 = single_set (i2);
1044 if (s1 == NULL_RTX || s2 == NULL_RTX)
1045 return dir_none;
1047 /* Check that the 2 sets set the same dest. */
1048 d1 = SET_DEST (s1);
1049 d2 = SET_DEST (s2);
1050 if (!(reload_completed
1051 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1052 return dir_none;
1054 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1055 set dest to the same value. */
1056 note1 = find_reg_equal_equiv_note (i1);
1057 note2 = find_reg_equal_equiv_note (i2);
1059 src1 = SET_SRC (s1);
1060 src2 = SET_SRC (s2);
1062 if (!values_equal_p (note1, note2, src1, src2))
1063 return dir_none;
1065 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1066 return dir_none;
1068 /* Although the 2 sets set dest to the same value, we cannot replace
1069 (set (dest) (const_int))
1071 (set (dest) (reg))
1072 because we don't know if the reg is live and has the same value at the
1073 location of replacement. */
1074 c1 = CONST_INT_P (src1);
1075 c2 = CONST_INT_P (src2);
1076 if (c1 && c2)
1077 return dir_both;
1078 else if (c2)
1079 return dir_forward;
1080 else if (c1)
1081 return dir_backward;
1083 return dir_none;
1086 /* Merges directions A and B. */
1088 static enum replace_direction
1089 merge_dir (enum replace_direction a, enum replace_direction b)
1091 /* Implements the following table:
1092 |bo fw bw no
1093 ---+-----------
1094 bo |bo fw bw no
1095 fw |-- fw no no
1096 bw |-- -- bw no
1097 no |-- -- -- no. */
1099 if (a == b)
1100 return a;
1102 if (a == dir_both)
1103 return b;
1104 if (b == dir_both)
1105 return a;
1107 return dir_none;
1110 /* Array of flags indexed by reg note kind, true if the given
1111 reg note is CFA related. */
1112 static const bool reg_note_cfa_p[] = {
1113 #undef REG_CFA_NOTE
1114 #define DEF_REG_NOTE(NAME) false,
1115 #define REG_CFA_NOTE(NAME) true,
1116 #include "reg-notes.def"
1117 #undef REG_CFA_NOTE
1118 #undef DEF_REG_NOTE
1119 false
1122 /* Return true if I1 and I2 have identical CFA notes (the same order
1123 and equivalent content). */
1125 static bool
1126 insns_have_identical_cfa_notes (rtx_insn *i1, rtx_insn *i2)
1128 rtx n1, n2;
1129 for (n1 = REG_NOTES (i1), n2 = REG_NOTES (i2); ;
1130 n1 = XEXP (n1, 1), n2 = XEXP (n2, 1))
1132 /* Skip over reg notes not related to CFI information. */
1133 while (n1 && !reg_note_cfa_p[REG_NOTE_KIND (n1)])
1134 n1 = XEXP (n1, 1);
1135 while (n2 && !reg_note_cfa_p[REG_NOTE_KIND (n2)])
1136 n2 = XEXP (n2, 1);
1137 if (n1 == NULL_RTX && n2 == NULL_RTX)
1138 return true;
1139 if (n1 == NULL_RTX || n2 == NULL_RTX)
1140 return false;
1141 if (XEXP (n1, 0) == XEXP (n2, 0))
1143 else if (XEXP (n1, 0) == NULL_RTX || XEXP (n2, 0) == NULL_RTX)
1144 return false;
1145 else if (!(reload_completed
1146 ? rtx_renumbered_equal_p (XEXP (n1, 0), XEXP (n2, 0))
1147 : rtx_equal_p (XEXP (n1, 0), XEXP (n2, 0))))
1148 return false;
1152 /* Examine I1 and I2 and return:
1153 - dir_forward if I1 can be replaced by I2, or
1154 - dir_backward if I2 can be replaced by I1, or
1155 - dir_both if both are the case. */
1157 static enum replace_direction
1158 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx_insn *i1, rtx_insn *i2)
1160 rtx p1, p2;
1162 /* Verify that I1 and I2 are equivalent. */
1163 if (GET_CODE (i1) != GET_CODE (i2))
1164 return dir_none;
1166 /* __builtin_unreachable() may lead to empty blocks (ending with
1167 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1168 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1169 return dir_both;
1171 /* ??? Do not allow cross-jumping between different stack levels. */
1172 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1173 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1174 if (p1 && p2)
1176 p1 = XEXP (p1, 0);
1177 p2 = XEXP (p2, 0);
1178 if (!rtx_equal_p (p1, p2))
1179 return dir_none;
1181 /* ??? Worse, this adjustment had better be constant lest we
1182 have differing incoming stack levels. */
1183 if (!frame_pointer_needed
1184 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1185 return dir_none;
1187 else if (p1 || p2)
1188 return dir_none;
1190 /* Do not allow cross-jumping between frame related insns and other
1191 insns. */
1192 if (RTX_FRAME_RELATED_P (i1) != RTX_FRAME_RELATED_P (i2))
1193 return dir_none;
1195 p1 = PATTERN (i1);
1196 p2 = PATTERN (i2);
1198 if (GET_CODE (p1) != GET_CODE (p2))
1199 return dir_none;
1201 /* If this is a CALL_INSN, compare register usage information.
1202 If we don't check this on stack register machines, the two
1203 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1204 numbers of stack registers in the same basic block.
1205 If we don't check this on machines with delay slots, a delay slot may
1206 be filled that clobbers a parameter expected by the subroutine.
1208 ??? We take the simple route for now and assume that if they're
1209 equal, they were constructed identically.
1211 Also check for identical exception regions. */
1213 if (CALL_P (i1))
1215 /* Ensure the same EH region. */
1216 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1217 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1219 if (!n1 && n2)
1220 return dir_none;
1222 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1223 return dir_none;
1225 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1226 CALL_INSN_FUNCTION_USAGE (i2))
1227 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1228 return dir_none;
1230 /* For address sanitizer, never crossjump __asan_report_* builtins,
1231 otherwise errors might be reported on incorrect lines. */
1232 if (flag_sanitize & SANITIZE_ADDRESS)
1234 rtx call = get_call_rtx_from (i1);
1235 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1237 rtx symbol = XEXP (XEXP (call, 0), 0);
1238 if (SYMBOL_REF_DECL (symbol)
1239 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1241 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1242 == BUILT_IN_NORMAL)
1243 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1244 >= BUILT_IN_ASAN_REPORT_LOAD1
1245 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1246 <= BUILT_IN_ASAN_STOREN)
1247 return dir_none;
1253 /* If both i1 and i2 are frame related, verify all the CFA notes
1254 in the same order and with the same content. */
1255 if (RTX_FRAME_RELATED_P (i1) && !insns_have_identical_cfa_notes (i1, i2))
1256 return dir_none;
1258 #ifdef STACK_REGS
1259 /* If cross_jump_death_matters is not 0, the insn's mode
1260 indicates whether or not the insn contains any stack-like
1261 regs. */
1263 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1265 /* If register stack conversion has already been done, then
1266 death notes must also be compared before it is certain that
1267 the two instruction streams match. */
1269 rtx note;
1270 HARD_REG_SET i1_regset, i2_regset;
1272 CLEAR_HARD_REG_SET (i1_regset);
1273 CLEAR_HARD_REG_SET (i2_regset);
1275 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1276 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1277 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1279 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1280 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1281 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1283 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1284 return dir_none;
1286 #endif
1288 if (reload_completed
1289 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1290 return dir_both;
1292 return can_replace_by (i1, i2);
1295 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1296 flow_find_head_matching_sequence, ensure the notes match. */
1298 static void
1299 merge_notes (rtx_insn *i1, rtx_insn *i2)
1301 /* If the merged insns have different REG_EQUAL notes, then
1302 remove them. */
1303 rtx equiv1 = find_reg_equal_equiv_note (i1);
1304 rtx equiv2 = find_reg_equal_equiv_note (i2);
1306 if (equiv1 && !equiv2)
1307 remove_note (i1, equiv1);
1308 else if (!equiv1 && equiv2)
1309 remove_note (i2, equiv2);
1310 else if (equiv1 && equiv2
1311 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1313 remove_note (i1, equiv1);
1314 remove_note (i2, equiv2);
1318 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1319 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1320 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1321 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1322 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1324 static void
1325 walk_to_nondebug_insn (rtx_insn **i1, basic_block *bb1, bool follow_fallthru,
1326 bool *did_fallthru)
1328 edge fallthru;
1330 *did_fallthru = false;
1332 /* Ignore notes. */
1333 while (!NONDEBUG_INSN_P (*i1))
1335 if (*i1 != BB_HEAD (*bb1))
1337 *i1 = PREV_INSN (*i1);
1338 continue;
1341 if (!follow_fallthru)
1342 return;
1344 fallthru = find_fallthru_edge ((*bb1)->preds);
1345 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)
1346 || !single_succ_p (fallthru->src))
1347 return;
1349 *bb1 = fallthru->src;
1350 *i1 = BB_END (*bb1);
1351 *did_fallthru = true;
1355 /* Look through the insns at the end of BB1 and BB2 and find the longest
1356 sequence that are either equivalent, or allow forward or backward
1357 replacement. Store the first insns for that sequence in *F1 and *F2 and
1358 return the sequence length.
1360 DIR_P indicates the allowed replacement direction on function entry, and
1361 the actual replacement direction on function exit. If NULL, only equivalent
1362 sequences are allowed.
1364 To simplify callers of this function, if the blocks match exactly,
1365 store the head of the blocks in *F1 and *F2. */
1368 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx_insn **f1,
1369 rtx_insn **f2, enum replace_direction *dir_p)
1371 rtx_insn *i1, *i2, *last1, *last2, *afterlast1, *afterlast2;
1372 int ninsns = 0;
1373 enum replace_direction dir, last_dir, afterlast_dir;
1374 bool follow_fallthru, did_fallthru;
1376 if (dir_p)
1377 dir = *dir_p;
1378 else
1379 dir = dir_both;
1380 afterlast_dir = dir;
1381 last_dir = afterlast_dir;
1383 /* Skip simple jumps at the end of the blocks. Complex jumps still
1384 need to be compared for equivalence, which we'll do below. */
1386 i1 = BB_END (bb1);
1387 last1 = afterlast1 = last2 = afterlast2 = NULL;
1388 if (onlyjump_p (i1)
1389 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1391 last1 = i1;
1392 i1 = PREV_INSN (i1);
1395 i2 = BB_END (bb2);
1396 if (onlyjump_p (i2)
1397 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1399 last2 = i2;
1400 /* Count everything except for unconditional jump as insn.
1401 Don't count any jumps if dir_p is NULL. */
1402 if (!simplejump_p (i2) && !returnjump_p (i2) && last1 && dir_p)
1403 ninsns++;
1404 i2 = PREV_INSN (i2);
1407 while (true)
1409 /* In the following example, we can replace all jumps to C by jumps to A.
1411 This removes 4 duplicate insns.
1412 [bb A] insn1 [bb C] insn1
1413 insn2 insn2
1414 [bb B] insn3 insn3
1415 insn4 insn4
1416 jump_insn jump_insn
1418 We could also replace all jumps to A by jumps to C, but that leaves B
1419 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1420 step, all jumps to B would be replaced with jumps to the middle of C,
1421 achieving the same result with more effort.
1422 So we allow only the first possibility, which means that we don't allow
1423 fallthru in the block that's being replaced. */
1425 follow_fallthru = dir_p && dir != dir_forward;
1426 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1427 if (did_fallthru)
1428 dir = dir_backward;
1430 follow_fallthru = dir_p && dir != dir_backward;
1431 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1432 if (did_fallthru)
1433 dir = dir_forward;
1435 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1436 break;
1438 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1439 if (dir == dir_none || (!dir_p && dir != dir_both))
1440 break;
1442 merge_memattrs (i1, i2);
1444 /* Don't begin a cross-jump with a NOTE insn. */
1445 if (INSN_P (i1))
1447 merge_notes (i1, i2);
1449 afterlast1 = last1, afterlast2 = last2;
1450 last1 = i1, last2 = i2;
1451 afterlast_dir = last_dir;
1452 last_dir = dir;
1453 if (active_insn_p (i1))
1454 ninsns++;
1457 i1 = PREV_INSN (i1);
1458 i2 = PREV_INSN (i2);
1461 /* Don't allow the insn after a compare to be shared by
1462 cross-jumping unless the compare is also shared. */
1463 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1464 && ! sets_cc0_p (last1))
1465 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1467 /* Include preceding notes and labels in the cross-jump. One,
1468 this may bring us to the head of the blocks as requested above.
1469 Two, it keeps line number notes as matched as may be. */
1470 if (ninsns)
1472 bb1 = BLOCK_FOR_INSN (last1);
1473 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1474 last1 = PREV_INSN (last1);
1476 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1477 last1 = PREV_INSN (last1);
1479 bb2 = BLOCK_FOR_INSN (last2);
1480 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1481 last2 = PREV_INSN (last2);
1483 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1484 last2 = PREV_INSN (last2);
1486 *f1 = last1;
1487 *f2 = last2;
1490 if (dir_p)
1491 *dir_p = last_dir;
1492 return ninsns;
1495 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1496 the head of the two blocks. Do not include jumps at the end.
1497 If STOP_AFTER is nonzero, stop after finding that many matching
1498 instructions. If STOP_AFTER is zero, count all INSN_P insns, if it is
1499 non-zero, only count active insns. */
1502 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx_insn **f1,
1503 rtx_insn **f2, int stop_after)
1505 rtx_insn *i1, *i2, *last1, *last2, *beforelast1, *beforelast2;
1506 int ninsns = 0;
1507 edge e;
1508 edge_iterator ei;
1509 int nehedges1 = 0, nehedges2 = 0;
1511 FOR_EACH_EDGE (e, ei, bb1->succs)
1512 if (e->flags & EDGE_EH)
1513 nehedges1++;
1514 FOR_EACH_EDGE (e, ei, bb2->succs)
1515 if (e->flags & EDGE_EH)
1516 nehedges2++;
1518 i1 = BB_HEAD (bb1);
1519 i2 = BB_HEAD (bb2);
1520 last1 = beforelast1 = last2 = beforelast2 = NULL;
1522 while (true)
1524 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1525 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1527 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1528 break;
1529 i1 = NEXT_INSN (i1);
1532 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1534 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1535 break;
1536 i2 = NEXT_INSN (i2);
1539 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1540 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1541 break;
1543 if (NOTE_P (i1) || NOTE_P (i2)
1544 || JUMP_P (i1) || JUMP_P (i2))
1545 break;
1547 /* A sanity check to make sure we're not merging insns with different
1548 effects on EH. If only one of them ends a basic block, it shouldn't
1549 have an EH edge; if both end a basic block, there should be the same
1550 number of EH edges. */
1551 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1552 && nehedges1 > 0)
1553 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1554 && nehedges2 > 0)
1555 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1556 && nehedges1 != nehedges2))
1557 break;
1559 if (old_insns_match_p (0, i1, i2) != dir_both)
1560 break;
1562 merge_memattrs (i1, i2);
1564 /* Don't begin a cross-jump with a NOTE insn. */
1565 if (INSN_P (i1))
1567 merge_notes (i1, i2);
1569 beforelast1 = last1, beforelast2 = last2;
1570 last1 = i1, last2 = i2;
1571 if (!stop_after || active_insn_p (i1))
1572 ninsns++;
1575 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1576 || (stop_after > 0 && ninsns == stop_after))
1577 break;
1579 i1 = NEXT_INSN (i1);
1580 i2 = NEXT_INSN (i2);
1583 /* Don't allow a compare to be shared by cross-jumping unless the insn
1584 after the compare is also shared. */
1585 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1586 && sets_cc0_p (last1))
1587 last1 = beforelast1, last2 = beforelast2, ninsns--;
1589 if (ninsns)
1591 *f1 = last1;
1592 *f2 = last2;
1595 return ninsns;
1598 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1599 the branch instruction. This means that if we commonize the control
1600 flow before end of the basic block, the semantic remains unchanged.
1602 We may assume that there exists one edge with a common destination. */
1604 static bool
1605 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1607 int nehedges1 = 0, nehedges2 = 0;
1608 edge fallthru1 = 0, fallthru2 = 0;
1609 edge e1, e2;
1610 edge_iterator ei;
1612 /* If we performed shrink-wrapping, edges to the exit block can
1613 only be distinguished for JUMP_INSNs. The two paths may differ in
1614 whether they went through the prologue. Sibcalls are fine, we know
1615 that we either didn't need or inserted an epilogue before them. */
1616 if (crtl->shrink_wrapped
1617 && single_succ_p (bb1)
1618 && single_succ (bb1) == EXIT_BLOCK_PTR_FOR_FN (cfun)
1619 && !JUMP_P (BB_END (bb1))
1620 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1621 return false;
1623 /* If BB1 has only one successor, we may be looking at either an
1624 unconditional jump, or a fake edge to exit. */
1625 if (single_succ_p (bb1)
1626 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1627 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1628 return (single_succ_p (bb2)
1629 && (single_succ_edge (bb2)->flags
1630 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1631 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1633 /* Match conditional jumps - this may get tricky when fallthru and branch
1634 edges are crossed. */
1635 if (EDGE_COUNT (bb1->succs) == 2
1636 && any_condjump_p (BB_END (bb1))
1637 && onlyjump_p (BB_END (bb1)))
1639 edge b1, f1, b2, f2;
1640 bool reverse, match;
1641 rtx set1, set2, cond1, cond2;
1642 enum rtx_code code1, code2;
1644 if (EDGE_COUNT (bb2->succs) != 2
1645 || !any_condjump_p (BB_END (bb2))
1646 || !onlyjump_p (BB_END (bb2)))
1647 return false;
1649 b1 = BRANCH_EDGE (bb1);
1650 b2 = BRANCH_EDGE (bb2);
1651 f1 = FALLTHRU_EDGE (bb1);
1652 f2 = FALLTHRU_EDGE (bb2);
1654 /* Get around possible forwarders on fallthru edges. Other cases
1655 should be optimized out already. */
1656 if (FORWARDER_BLOCK_P (f1->dest))
1657 f1 = single_succ_edge (f1->dest);
1659 if (FORWARDER_BLOCK_P (f2->dest))
1660 f2 = single_succ_edge (f2->dest);
1662 /* To simplify use of this function, return false if there are
1663 unneeded forwarder blocks. These will get eliminated later
1664 during cleanup_cfg. */
1665 if (FORWARDER_BLOCK_P (f1->dest)
1666 || FORWARDER_BLOCK_P (f2->dest)
1667 || FORWARDER_BLOCK_P (b1->dest)
1668 || FORWARDER_BLOCK_P (b2->dest))
1669 return false;
1671 if (f1->dest == f2->dest && b1->dest == b2->dest)
1672 reverse = false;
1673 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1674 reverse = true;
1675 else
1676 return false;
1678 set1 = pc_set (BB_END (bb1));
1679 set2 = pc_set (BB_END (bb2));
1680 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1681 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1682 reverse = !reverse;
1684 cond1 = XEXP (SET_SRC (set1), 0);
1685 cond2 = XEXP (SET_SRC (set2), 0);
1686 code1 = GET_CODE (cond1);
1687 if (reverse)
1688 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1689 else
1690 code2 = GET_CODE (cond2);
1692 if (code2 == UNKNOWN)
1693 return false;
1695 /* Verify codes and operands match. */
1696 match = ((code1 == code2
1697 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1698 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1699 || (code1 == swap_condition (code2)
1700 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1701 XEXP (cond2, 0))
1702 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1703 XEXP (cond2, 1))));
1705 /* If we return true, we will join the blocks. Which means that
1706 we will only have one branch prediction bit to work with. Thus
1707 we require the existing branches to have probabilities that are
1708 roughly similar. */
1709 if (match
1710 && optimize_bb_for_speed_p (bb1)
1711 && optimize_bb_for_speed_p (bb2))
1713 int prob2;
1715 if (b1->dest == b2->dest)
1716 prob2 = b2->probability;
1717 else
1718 /* Do not use f2 probability as f2 may be forwarded. */
1719 prob2 = REG_BR_PROB_BASE - b2->probability;
1721 /* Fail if the difference in probabilities is greater than 50%.
1722 This rules out two well-predicted branches with opposite
1723 outcomes. */
1724 if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
1726 if (dump_file)
1727 fprintf (dump_file,
1728 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1729 bb1->index, bb2->index, b1->probability, prob2);
1731 return false;
1735 if (dump_file && match)
1736 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1737 bb1->index, bb2->index);
1739 return match;
1742 /* Generic case - we are seeing a computed jump, table jump or trapping
1743 instruction. */
1745 /* Check whether there are tablejumps in the end of BB1 and BB2.
1746 Return true if they are identical. */
1748 rtx_insn *label1, *label2;
1749 rtx_jump_table_data *table1, *table2;
1751 if (tablejump_p (BB_END (bb1), &label1, &table1)
1752 && tablejump_p (BB_END (bb2), &label2, &table2)
1753 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1755 /* The labels should never be the same rtx. If they really are same
1756 the jump tables are same too. So disable crossjumping of blocks BB1
1757 and BB2 because when deleting the common insns in the end of BB1
1758 by delete_basic_block () the jump table would be deleted too. */
1759 /* If LABEL2 is referenced in BB1->END do not do anything
1760 because we would loose information when replacing
1761 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1762 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1764 /* Set IDENTICAL to true when the tables are identical. */
1765 bool identical = false;
1766 rtx p1, p2;
1768 p1 = PATTERN (table1);
1769 p2 = PATTERN (table2);
1770 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1772 identical = true;
1774 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1775 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1776 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1777 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1779 int i;
1781 identical = true;
1782 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1783 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1784 identical = false;
1787 if (identical)
1789 bool match;
1791 /* Temporarily replace references to LABEL1 with LABEL2
1792 in BB1->END so that we could compare the instructions. */
1793 replace_label_in_insn (BB_END (bb1), label1, label2, false);
1795 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1796 == dir_both);
1797 if (dump_file && match)
1798 fprintf (dump_file,
1799 "Tablejumps in bb %i and %i match.\n",
1800 bb1->index, bb2->index);
1802 /* Set the original label in BB1->END because when deleting
1803 a block whose end is a tablejump, the tablejump referenced
1804 from the instruction is deleted too. */
1805 replace_label_in_insn (BB_END (bb1), label2, label1, false);
1807 return match;
1810 return false;
1814 /* Find the last non-debug non-note instruction in each bb, except
1815 stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
1816 handles that case specially. old_insns_match_p does not handle
1817 other types of instruction notes. */
1818 rtx_insn *last1 = BB_END (bb1);
1819 rtx_insn *last2 = BB_END (bb2);
1820 while (!NOTE_INSN_BASIC_BLOCK_P (last1) &&
1821 (DEBUG_INSN_P (last1) || NOTE_P (last1)))
1822 last1 = PREV_INSN (last1);
1823 while (!NOTE_INSN_BASIC_BLOCK_P (last2) &&
1824 (DEBUG_INSN_P (last2) || NOTE_P (last2)))
1825 last2 = PREV_INSN (last2);
1826 gcc_assert (last1 && last2);
1828 /* First ensure that the instructions match. There may be many outgoing
1829 edges so this test is generally cheaper. */
1830 if (old_insns_match_p (mode, last1, last2) != dir_both)
1831 return false;
1833 /* Search the outgoing edges, ensure that the counts do match, find possible
1834 fallthru and exception handling edges since these needs more
1835 validation. */
1836 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1837 return false;
1839 bool nonfakeedges = false;
1840 FOR_EACH_EDGE (e1, ei, bb1->succs)
1842 e2 = EDGE_SUCC (bb2, ei.index);
1844 if ((e1->flags & EDGE_FAKE) == 0)
1845 nonfakeedges = true;
1847 if (e1->flags & EDGE_EH)
1848 nehedges1++;
1850 if (e2->flags & EDGE_EH)
1851 nehedges2++;
1853 if (e1->flags & EDGE_FALLTHRU)
1854 fallthru1 = e1;
1855 if (e2->flags & EDGE_FALLTHRU)
1856 fallthru2 = e2;
1859 /* If number of edges of various types does not match, fail. */
1860 if (nehedges1 != nehedges2
1861 || (fallthru1 != 0) != (fallthru2 != 0))
1862 return false;
1864 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1865 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1866 attempt to optimize, as the two basic blocks might have different
1867 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1868 traps there should be REG_ARG_SIZE notes, they could be missing
1869 for __builtin_unreachable () uses though. */
1870 if (!nonfakeedges
1871 && !ACCUMULATE_OUTGOING_ARGS
1872 && (!INSN_P (last1)
1873 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1874 return false;
1876 /* fallthru edges must be forwarded to the same destination. */
1877 if (fallthru1)
1879 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1880 ? single_succ (fallthru1->dest): fallthru1->dest);
1881 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1882 ? single_succ (fallthru2->dest): fallthru2->dest);
1884 if (d1 != d2)
1885 return false;
1888 /* Ensure the same EH region. */
1890 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1891 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1893 if (!n1 && n2)
1894 return false;
1896 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1897 return false;
1900 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1901 version of sequence abstraction. */
1902 FOR_EACH_EDGE (e1, ei, bb2->succs)
1904 edge e2;
1905 edge_iterator ei;
1906 basic_block d1 = e1->dest;
1908 if (FORWARDER_BLOCK_P (d1))
1909 d1 = EDGE_SUCC (d1, 0)->dest;
1911 FOR_EACH_EDGE (e2, ei, bb1->succs)
1913 basic_block d2 = e2->dest;
1914 if (FORWARDER_BLOCK_P (d2))
1915 d2 = EDGE_SUCC (d2, 0)->dest;
1916 if (d1 == d2)
1917 break;
1920 if (!e2)
1921 return false;
1924 return true;
1927 /* Returns true if BB basic block has a preserve label. */
1929 static bool
1930 block_has_preserve_label (basic_block bb)
1932 return (bb
1933 && block_label (bb)
1934 && LABEL_PRESERVE_P (block_label (bb)));
1937 /* E1 and E2 are edges with the same destination block. Search their
1938 predecessors for common code. If found, redirect control flow from
1939 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1940 or the other way around (dir_backward). DIR specifies the allowed
1941 replacement direction. */
1943 static bool
1944 try_crossjump_to_edge (int mode, edge e1, edge e2,
1945 enum replace_direction dir)
1947 int nmatch;
1948 basic_block src1 = e1->src, src2 = e2->src;
1949 basic_block redirect_to, redirect_from, to_remove;
1950 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1951 rtx_insn *newpos1, *newpos2;
1952 edge s;
1953 edge_iterator ei;
1955 newpos1 = newpos2 = NULL;
1957 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1958 to try this optimization.
1960 Basic block partitioning may result in some jumps that appear to
1961 be optimizable (or blocks that appear to be mergeable), but which really
1962 must be left untouched (they are required to make it safely across
1963 partition boundaries). See the comments at the top of
1964 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1966 if (crtl->has_bb_partition && reload_completed)
1967 return false;
1969 /* Search backward through forwarder blocks. We don't need to worry
1970 about multiple entry or chained forwarders, as they will be optimized
1971 away. We do this to look past the unconditional jump following a
1972 conditional jump that is required due to the current CFG shape. */
1973 if (single_pred_p (src1)
1974 && FORWARDER_BLOCK_P (src1))
1975 e1 = single_pred_edge (src1), src1 = e1->src;
1977 if (single_pred_p (src2)
1978 && FORWARDER_BLOCK_P (src2))
1979 e2 = single_pred_edge (src2), src2 = e2->src;
1981 /* Nothing to do if we reach ENTRY, or a common source block. */
1982 if (src1 == ENTRY_BLOCK_PTR_FOR_FN (cfun) || src2
1983 == ENTRY_BLOCK_PTR_FOR_FN (cfun))
1984 return false;
1985 if (src1 == src2)
1986 return false;
1988 /* Seeing more than 1 forwarder blocks would confuse us later... */
1989 if (FORWARDER_BLOCK_P (e1->dest)
1990 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1991 return false;
1993 if (FORWARDER_BLOCK_P (e2->dest)
1994 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1995 return false;
1997 /* Likewise with dead code (possibly newly created by the other optimizations
1998 of cfg_cleanup). */
1999 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
2000 return false;
2002 /* Look for the common insn sequence, part the first ... */
2003 if (!outgoing_edges_match (mode, src1, src2))
2004 return false;
2006 /* ... and part the second. */
2007 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
2009 osrc1 = src1;
2010 osrc2 = src2;
2011 if (newpos1 != NULL_RTX)
2012 src1 = BLOCK_FOR_INSN (newpos1);
2013 if (newpos2 != NULL_RTX)
2014 src2 = BLOCK_FOR_INSN (newpos2);
2016 /* Check that SRC1 and SRC2 have preds again. They may have changed
2017 above due to the call to flow_find_cross_jump. */
2018 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
2019 return false;
2021 if (dir == dir_backward)
2023 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
2024 SWAP (basic_block, osrc1, osrc2);
2025 SWAP (basic_block, src1, src2);
2026 SWAP (edge, e1, e2);
2027 SWAP (rtx_insn *, newpos1, newpos2);
2028 #undef SWAP
2031 /* Don't proceed with the crossjump unless we found a sufficient number
2032 of matching instructions or the 'from' block was totally matched
2033 (such that its predecessors will hopefully be redirected and the
2034 block removed). */
2035 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
2036 && (newpos1 != BB_HEAD (src1)))
2037 return false;
2039 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
2040 if (block_has_preserve_label (e1->dest)
2041 && (e1->flags & EDGE_ABNORMAL))
2042 return false;
2044 /* Here we know that the insns in the end of SRC1 which are common with SRC2
2045 will be deleted.
2046 If we have tablejumps in the end of SRC1 and SRC2
2047 they have been already compared for equivalence in outgoing_edges_match ()
2048 so replace the references to TABLE1 by references to TABLE2. */
2050 rtx_insn *label1, *label2;
2051 rtx_jump_table_data *table1, *table2;
2053 if (tablejump_p (BB_END (osrc1), &label1, &table1)
2054 && tablejump_p (BB_END (osrc2), &label2, &table2)
2055 && label1 != label2)
2057 rtx_insn *insn;
2059 /* Replace references to LABEL1 with LABEL2. */
2060 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2062 /* Do not replace the label in SRC1->END because when deleting
2063 a block whose end is a tablejump, the tablejump referenced
2064 from the instruction is deleted too. */
2065 if (insn != BB_END (osrc1))
2066 replace_label_in_insn (insn, label1, label2, true);
2071 /* Avoid splitting if possible. We must always split when SRC2 has
2072 EH predecessor edges, or we may end up with basic blocks with both
2073 normal and EH predecessor edges. */
2074 if (newpos2 == BB_HEAD (src2)
2075 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
2076 redirect_to = src2;
2077 else
2079 if (newpos2 == BB_HEAD (src2))
2081 /* Skip possible basic block header. */
2082 if (LABEL_P (newpos2))
2083 newpos2 = NEXT_INSN (newpos2);
2084 while (DEBUG_INSN_P (newpos2))
2085 newpos2 = NEXT_INSN (newpos2);
2086 if (NOTE_P (newpos2))
2087 newpos2 = NEXT_INSN (newpos2);
2088 while (DEBUG_INSN_P (newpos2))
2089 newpos2 = NEXT_INSN (newpos2);
2092 if (dump_file)
2093 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2094 src2->index, nmatch);
2095 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2098 if (dump_file)
2099 fprintf (dump_file,
2100 "Cross jumping from bb %i to bb %i; %i common insns\n",
2101 src1->index, src2->index, nmatch);
2103 /* We may have some registers visible through the block. */
2104 df_set_bb_dirty (redirect_to);
2106 if (osrc2 == src2)
2107 redirect_edges_to = redirect_to;
2108 else
2109 redirect_edges_to = osrc2;
2111 /* Recompute the frequencies and counts of outgoing edges. */
2112 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2114 edge s2;
2115 edge_iterator ei;
2116 basic_block d = s->dest;
2118 if (FORWARDER_BLOCK_P (d))
2119 d = single_succ (d);
2121 FOR_EACH_EDGE (s2, ei, src1->succs)
2123 basic_block d2 = s2->dest;
2124 if (FORWARDER_BLOCK_P (d2))
2125 d2 = single_succ (d2);
2126 if (d == d2)
2127 break;
2130 s->count += s2->count;
2132 /* Take care to update possible forwarder blocks. We verified
2133 that there is no more than one in the chain, so we can't run
2134 into infinite loop. */
2135 if (FORWARDER_BLOCK_P (s->dest))
2137 single_succ_edge (s->dest)->count += s2->count;
2138 s->dest->count += s2->count;
2139 s->dest->frequency += EDGE_FREQUENCY (s);
2142 if (FORWARDER_BLOCK_P (s2->dest))
2144 single_succ_edge (s2->dest)->count -= s2->count;
2145 s2->dest->count -= s2->count;
2146 s2->dest->frequency -= EDGE_FREQUENCY (s);
2147 if (s2->dest->frequency < 0)
2148 s2->dest->frequency = 0;
2151 if (!redirect_edges_to->frequency && !src1->frequency)
2152 s->probability = (s->probability + s2->probability) / 2;
2153 else
2154 s->probability
2155 = ((s->probability * redirect_edges_to->frequency +
2156 s2->probability * src1->frequency)
2157 / (redirect_edges_to->frequency + src1->frequency));
2160 /* Adjust count and frequency for the block. An earlier jump
2161 threading pass may have left the profile in an inconsistent
2162 state (see update_bb_profile_for_threading) so we must be
2163 prepared for overflows. */
2164 tmp = redirect_to;
2167 tmp->count += src1->count;
2168 tmp->frequency += src1->frequency;
2169 if (tmp->frequency > BB_FREQ_MAX)
2170 tmp->frequency = BB_FREQ_MAX;
2171 if (tmp == redirect_edges_to)
2172 break;
2173 tmp = find_fallthru_edge (tmp->succs)->dest;
2175 while (true);
2176 update_br_prob_note (redirect_edges_to);
2178 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2180 /* Skip possible basic block header. */
2181 if (LABEL_P (newpos1))
2182 newpos1 = NEXT_INSN (newpos1);
2184 while (DEBUG_INSN_P (newpos1))
2185 newpos1 = NEXT_INSN (newpos1);
2187 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2188 newpos1 = NEXT_INSN (newpos1);
2190 while (DEBUG_INSN_P (newpos1))
2191 newpos1 = NEXT_INSN (newpos1);
2193 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2194 to_remove = single_succ (redirect_from);
2196 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2197 delete_basic_block (to_remove);
2199 update_forwarder_flag (redirect_from);
2200 if (redirect_to != src2)
2201 update_forwarder_flag (src2);
2203 return true;
2206 /* Search the predecessors of BB for common insn sequences. When found,
2207 share code between them by redirecting control flow. Return true if
2208 any changes made. */
2210 static bool
2211 try_crossjump_bb (int mode, basic_block bb)
2213 edge e, e2, fallthru;
2214 bool changed;
2215 unsigned max, ix, ix2;
2217 /* Nothing to do if there is not at least two incoming edges. */
2218 if (EDGE_COUNT (bb->preds) < 2)
2219 return false;
2221 /* Don't crossjump if this block ends in a computed jump,
2222 unless we are optimizing for size. */
2223 if (optimize_bb_for_size_p (bb)
2224 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2225 && computed_jump_p (BB_END (bb)))
2226 return false;
2228 /* If we are partitioning hot/cold basic blocks, we don't want to
2229 mess up unconditional or indirect jumps that cross between hot
2230 and cold sections.
2232 Basic block partitioning may result in some jumps that appear to
2233 be optimizable (or blocks that appear to be mergeable), but which really
2234 must be left untouched (they are required to make it safely across
2235 partition boundaries). See the comments at the top of
2236 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2238 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2239 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2240 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2241 return false;
2243 /* It is always cheapest to redirect a block that ends in a branch to
2244 a block that falls through into BB, as that adds no branches to the
2245 program. We'll try that combination first. */
2246 fallthru = NULL;
2247 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2249 if (EDGE_COUNT (bb->preds) > max)
2250 return false;
2252 fallthru = find_fallthru_edge (bb->preds);
2254 changed = false;
2255 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2257 e = EDGE_PRED (bb, ix);
2258 ix++;
2260 /* As noted above, first try with the fallthru predecessor (or, a
2261 fallthru predecessor if we are in cfglayout mode). */
2262 if (fallthru)
2264 /* Don't combine the fallthru edge into anything else.
2265 If there is a match, we'll do it the other way around. */
2266 if (e == fallthru)
2267 continue;
2268 /* If nothing changed since the last attempt, there is nothing
2269 we can do. */
2270 if (!first_pass
2271 && !((e->src->flags & BB_MODIFIED)
2272 || (fallthru->src->flags & BB_MODIFIED)))
2273 continue;
2275 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2277 changed = true;
2278 ix = 0;
2279 continue;
2283 /* Non-obvious work limiting check: Recognize that we're going
2284 to call try_crossjump_bb on every basic block. So if we have
2285 two blocks with lots of outgoing edges (a switch) and they
2286 share lots of common destinations, then we would do the
2287 cross-jump check once for each common destination.
2289 Now, if the blocks actually are cross-jump candidates, then
2290 all of their destinations will be shared. Which means that
2291 we only need check them for cross-jump candidacy once. We
2292 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2293 choosing to do the check from the block for which the edge
2294 in question is the first successor of A. */
2295 if (EDGE_SUCC (e->src, 0) != e)
2296 continue;
2298 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2300 e2 = EDGE_PRED (bb, ix2);
2302 if (e2 == e)
2303 continue;
2305 /* We've already checked the fallthru edge above. */
2306 if (e2 == fallthru)
2307 continue;
2309 /* The "first successor" check above only prevents multiple
2310 checks of crossjump(A,B). In order to prevent redundant
2311 checks of crossjump(B,A), require that A be the block
2312 with the lowest index. */
2313 if (e->src->index > e2->src->index)
2314 continue;
2316 /* If nothing changed since the last attempt, there is nothing
2317 we can do. */
2318 if (!first_pass
2319 && !((e->src->flags & BB_MODIFIED)
2320 || (e2->src->flags & BB_MODIFIED)))
2321 continue;
2323 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2324 direction. */
2325 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2327 changed = true;
2328 ix = 0;
2329 break;
2334 if (changed)
2335 crossjumps_occurred = true;
2337 return changed;
2340 /* Search the successors of BB for common insn sequences. When found,
2341 share code between them by moving it across the basic block
2342 boundary. Return true if any changes made. */
2344 static bool
2345 try_head_merge_bb (basic_block bb)
2347 basic_block final_dest_bb = NULL;
2348 int max_match = INT_MAX;
2349 edge e0;
2350 rtx_insn **headptr, **currptr, **nextptr;
2351 bool changed, moveall;
2352 unsigned ix;
2353 rtx_insn *e0_last_head;
2354 rtx cond;
2355 rtx_insn *move_before;
2356 unsigned nedges = EDGE_COUNT (bb->succs);
2357 rtx_insn *jump = BB_END (bb);
2358 regset live, live_union;
2360 /* Nothing to do if there is not at least two outgoing edges. */
2361 if (nedges < 2)
2362 return false;
2364 /* Don't crossjump if this block ends in a computed jump,
2365 unless we are optimizing for size. */
2366 if (optimize_bb_for_size_p (bb)
2367 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2368 && computed_jump_p (BB_END (bb)))
2369 return false;
2371 cond = get_condition (jump, &move_before, true, false);
2372 if (cond == NULL_RTX)
2374 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2375 move_before = prev_nonnote_nondebug_insn (jump);
2376 else
2377 move_before = jump;
2380 for (ix = 0; ix < nedges; ix++)
2381 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
2382 return false;
2384 for (ix = 0; ix < nedges; ix++)
2386 edge e = EDGE_SUCC (bb, ix);
2387 basic_block other_bb = e->dest;
2389 if (df_get_bb_dirty (other_bb))
2391 block_was_dirty = true;
2392 return false;
2395 if (e->flags & EDGE_ABNORMAL)
2396 return false;
2398 /* Normally, all destination blocks must only be reachable from this
2399 block, i.e. they must have one incoming edge.
2401 There is one special case we can handle, that of multiple consecutive
2402 jumps where the first jumps to one of the targets of the second jump.
2403 This happens frequently in switch statements for default labels.
2404 The structure is as follows:
2405 FINAL_DEST_BB
2406 ....
2407 if (cond) jump A;
2408 fall through
2410 jump with targets A, B, C, D...
2412 has two incoming edges, from FINAL_DEST_BB and BB
2414 In this case, we can try to move the insns through BB and into
2415 FINAL_DEST_BB. */
2416 if (EDGE_COUNT (other_bb->preds) != 1)
2418 edge incoming_edge, incoming_bb_other_edge;
2419 edge_iterator ei;
2421 if (final_dest_bb != NULL
2422 || EDGE_COUNT (other_bb->preds) != 2)
2423 return false;
2425 /* We must be able to move the insns across the whole block. */
2426 move_before = BB_HEAD (bb);
2427 while (!NONDEBUG_INSN_P (move_before))
2428 move_before = NEXT_INSN (move_before);
2430 if (EDGE_COUNT (bb->preds) != 1)
2431 return false;
2432 incoming_edge = EDGE_PRED (bb, 0);
2433 final_dest_bb = incoming_edge->src;
2434 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2435 return false;
2436 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2437 if (incoming_bb_other_edge != incoming_edge)
2438 break;
2439 if (incoming_bb_other_edge->dest != other_bb)
2440 return false;
2444 e0 = EDGE_SUCC (bb, 0);
2445 e0_last_head = NULL;
2446 changed = false;
2448 for (ix = 1; ix < nedges; ix++)
2450 edge e = EDGE_SUCC (bb, ix);
2451 rtx_insn *e0_last, *e_last;
2452 int nmatch;
2454 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2455 &e0_last, &e_last, 0);
2456 if (nmatch == 0)
2457 return false;
2459 if (nmatch < max_match)
2461 max_match = nmatch;
2462 e0_last_head = e0_last;
2466 /* If we matched an entire block, we probably have to avoid moving the
2467 last insn. */
2468 if (max_match > 0
2469 && e0_last_head == BB_END (e0->dest)
2470 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2471 || control_flow_insn_p (e0_last_head)))
2473 max_match--;
2474 if (max_match == 0)
2475 return false;
2477 e0_last_head = prev_real_insn (e0_last_head);
2478 while (DEBUG_INSN_P (e0_last_head));
2481 if (max_match == 0)
2482 return false;
2484 /* We must find a union of the live registers at each of the end points. */
2485 live = BITMAP_ALLOC (NULL);
2486 live_union = BITMAP_ALLOC (NULL);
2488 currptr = XNEWVEC (rtx_insn *, nedges);
2489 headptr = XNEWVEC (rtx_insn *, nedges);
2490 nextptr = XNEWVEC (rtx_insn *, nedges);
2492 for (ix = 0; ix < nedges; ix++)
2494 int j;
2495 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2496 rtx_insn *head = BB_HEAD (merge_bb);
2498 while (!NONDEBUG_INSN_P (head))
2499 head = NEXT_INSN (head);
2500 headptr[ix] = head;
2501 currptr[ix] = head;
2503 /* Compute the end point and live information */
2504 for (j = 1; j < max_match; j++)
2506 head = NEXT_INSN (head);
2507 while (!NONDEBUG_INSN_P (head));
2508 simulate_backwards_to_point (merge_bb, live, head);
2509 IOR_REG_SET (live_union, live);
2512 /* If we're moving across two blocks, verify the validity of the
2513 first move, then adjust the target and let the loop below deal
2514 with the final move. */
2515 if (final_dest_bb != NULL)
2517 rtx_insn *move_upto;
2519 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2520 jump, e0->dest, live_union,
2521 NULL, &move_upto);
2522 if (!moveall)
2524 if (move_upto == NULL_RTX)
2525 goto out;
2527 while (e0_last_head != move_upto)
2529 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2530 live_union);
2531 e0_last_head = PREV_INSN (e0_last_head);
2534 if (e0_last_head == NULL_RTX)
2535 goto out;
2537 jump = BB_END (final_dest_bb);
2538 cond = get_condition (jump, &move_before, true, false);
2539 if (cond == NULL_RTX)
2541 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2542 move_before = prev_nonnote_nondebug_insn (jump);
2543 else
2544 move_before = jump;
2550 rtx_insn *move_upto;
2551 moveall = can_move_insns_across (currptr[0], e0_last_head,
2552 move_before, jump, e0->dest, live_union,
2553 NULL, &move_upto);
2554 if (!moveall && move_upto == NULL_RTX)
2556 if (jump == move_before)
2557 break;
2559 /* Try again, using a different insertion point. */
2560 move_before = jump;
2562 /* Don't try moving before a cc0 user, as that may invalidate
2563 the cc0. */
2564 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2565 break;
2567 continue;
2570 if (final_dest_bb && !moveall)
2571 /* We haven't checked whether a partial move would be OK for the first
2572 move, so we have to fail this case. */
2573 break;
2575 changed = true;
2576 for (;;)
2578 if (currptr[0] == move_upto)
2579 break;
2580 for (ix = 0; ix < nedges; ix++)
2582 rtx_insn *curr = currptr[ix];
2584 curr = NEXT_INSN (curr);
2585 while (!NONDEBUG_INSN_P (curr));
2586 currptr[ix] = curr;
2590 /* If we can't currently move all of the identical insns, remember
2591 each insn after the range that we'll merge. */
2592 if (!moveall)
2593 for (ix = 0; ix < nedges; ix++)
2595 rtx_insn *curr = currptr[ix];
2597 curr = NEXT_INSN (curr);
2598 while (!NONDEBUG_INSN_P (curr));
2599 nextptr[ix] = curr;
2602 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2603 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2604 if (final_dest_bb != NULL)
2605 df_set_bb_dirty (final_dest_bb);
2606 df_set_bb_dirty (bb);
2607 for (ix = 1; ix < nedges; ix++)
2609 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2610 delete_insn_chain (headptr[ix], currptr[ix], false);
2612 if (!moveall)
2614 if (jump == move_before)
2615 break;
2617 /* For the unmerged insns, try a different insertion point. */
2618 move_before = jump;
2620 /* Don't try moving before a cc0 user, as that may invalidate
2621 the cc0. */
2622 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2623 break;
2625 for (ix = 0; ix < nedges; ix++)
2626 currptr[ix] = headptr[ix] = nextptr[ix];
2629 while (!moveall);
2631 out:
2632 free (currptr);
2633 free (headptr);
2634 free (nextptr);
2636 crossjumps_occurred |= changed;
2638 return changed;
2641 /* Return true if BB contains just bb note, or bb note followed
2642 by only DEBUG_INSNs. */
2644 static bool
2645 trivially_empty_bb_p (basic_block bb)
2647 rtx_insn *insn = BB_END (bb);
2649 while (1)
2651 if (insn == BB_HEAD (bb))
2652 return true;
2653 if (!DEBUG_INSN_P (insn))
2654 return false;
2655 insn = PREV_INSN (insn);
2659 /* Return true if BB contains just a return and possibly a USE of the
2660 return value. Fill in *RET and *USE with the return and use insns
2661 if any found, otherwise NULL. All CLOBBERs are ignored. */
2663 static bool
2664 bb_is_just_return (basic_block bb, rtx_insn **ret, rtx_insn **use)
2666 *ret = *use = NULL;
2667 rtx_insn *insn;
2669 if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
2670 return false;
2672 FOR_BB_INSNS (bb, insn)
2673 if (NONDEBUG_INSN_P (insn))
2675 rtx pat = PATTERN (insn);
2677 if (!*ret && ANY_RETURN_P (pat))
2678 *ret = insn;
2679 else if (!*ret && !*use && GET_CODE (pat) == USE
2680 && REG_P (XEXP (pat, 0))
2681 && REG_FUNCTION_VALUE_P (XEXP (pat, 0)))
2682 *use = insn;
2683 else if (GET_CODE (pat) != CLOBBER)
2684 return false;
2687 return !!*ret;
2690 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2691 instructions etc. Return nonzero if changes were made. */
2693 static bool
2694 try_optimize_cfg (int mode)
2696 bool changed_overall = false;
2697 bool changed;
2698 int iterations = 0;
2699 basic_block bb, b, next;
2701 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2702 clear_bb_flags ();
2704 crossjumps_occurred = false;
2706 FOR_EACH_BB_FN (bb, cfun)
2707 update_forwarder_flag (bb);
2709 if (! targetm.cannot_modify_jumps_p ())
2711 first_pass = true;
2712 /* Attempt to merge blocks as made possible by edge removal. If
2713 a block has only one successor, and the successor has only
2714 one predecessor, they may be combined. */
2717 block_was_dirty = false;
2718 changed = false;
2719 iterations++;
2721 if (dump_file)
2722 fprintf (dump_file,
2723 "\n\ntry_optimize_cfg iteration %i\n\n",
2724 iterations);
2726 for (b = ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb; b
2727 != EXIT_BLOCK_PTR_FOR_FN (cfun);)
2729 basic_block c;
2730 edge s;
2731 bool changed_here = false;
2733 /* Delete trivially dead basic blocks. This is either
2734 blocks with no predecessors, or empty blocks with no
2735 successors. However if the empty block with no
2736 successors is the successor of the ENTRY_BLOCK, it is
2737 kept. This ensures that the ENTRY_BLOCK will have a
2738 successor which is a precondition for many RTL
2739 passes. Empty blocks may result from expanding
2740 __builtin_unreachable (). */
2741 if (EDGE_COUNT (b->preds) == 0
2742 || (EDGE_COUNT (b->succs) == 0
2743 && trivially_empty_bb_p (b)
2744 && single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))->dest
2745 != b))
2747 c = b->prev_bb;
2748 if (EDGE_COUNT (b->preds) > 0)
2750 edge e;
2751 edge_iterator ei;
2753 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2755 if (BB_FOOTER (b)
2756 && BARRIER_P (BB_FOOTER (b)))
2757 FOR_EACH_EDGE (e, ei, b->preds)
2758 if ((e->flags & EDGE_FALLTHRU)
2759 && BB_FOOTER (e->src) == NULL)
2761 if (BB_FOOTER (b))
2763 BB_FOOTER (e->src) = BB_FOOTER (b);
2764 BB_FOOTER (b) = NULL;
2766 else
2768 start_sequence ();
2769 BB_FOOTER (e->src) = emit_barrier ();
2770 end_sequence ();
2774 else
2776 rtx_insn *last = get_last_bb_insn (b);
2777 if (last && BARRIER_P (last))
2778 FOR_EACH_EDGE (e, ei, b->preds)
2779 if ((e->flags & EDGE_FALLTHRU))
2780 emit_barrier_after (BB_END (e->src));
2783 delete_basic_block (b);
2784 changed = true;
2785 /* Avoid trying to remove the exit block. */
2786 b = (c == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? c->next_bb : c);
2787 continue;
2790 /* Remove code labels no longer used. */
2791 if (single_pred_p (b)
2792 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2793 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2794 && LABEL_P (BB_HEAD (b))
2795 && !LABEL_PRESERVE_P (BB_HEAD (b))
2796 /* If the previous block ends with a branch to this
2797 block, we can't delete the label. Normally this
2798 is a condjump that is yet to be simplified, but
2799 if CASE_DROPS_THRU, this can be a tablejump with
2800 some element going to the same place as the
2801 default (fallthru). */
2802 && (single_pred (b) == ENTRY_BLOCK_PTR_FOR_FN (cfun)
2803 || !JUMP_P (BB_END (single_pred (b)))
2804 || ! label_is_jump_target_p (BB_HEAD (b),
2805 BB_END (single_pred (b)))))
2807 delete_insn (BB_HEAD (b));
2808 if (dump_file)
2809 fprintf (dump_file, "Deleted label in block %i.\n",
2810 b->index);
2813 /* If we fall through an empty block, we can remove it. */
2814 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2815 && single_pred_p (b)
2816 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2817 && !LABEL_P (BB_HEAD (b))
2818 && FORWARDER_BLOCK_P (b)
2819 /* Note that forwarder_block_p true ensures that
2820 there is a successor for this block. */
2821 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2822 && n_basic_blocks_for_fn (cfun) > NUM_FIXED_BLOCKS + 1)
2824 if (dump_file)
2825 fprintf (dump_file,
2826 "Deleting fallthru block %i.\n",
2827 b->index);
2829 c = ((b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun))
2830 ? b->next_bb : b->prev_bb);
2831 redirect_edge_succ_nodup (single_pred_edge (b),
2832 single_succ (b));
2833 delete_basic_block (b);
2834 changed = true;
2835 b = c;
2836 continue;
2839 /* Merge B with its single successor, if any. */
2840 if (single_succ_p (b)
2841 && (s = single_succ_edge (b))
2842 && !(s->flags & EDGE_COMPLEX)
2843 && (c = s->dest) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2844 && single_pred_p (c)
2845 && b != c)
2847 /* When not in cfg_layout mode use code aware of reordering
2848 INSN. This code possibly creates new basic blocks so it
2849 does not fit merge_blocks interface and is kept here in
2850 hope that it will become useless once more of compiler
2851 is transformed to use cfg_layout mode. */
2853 if ((mode & CLEANUP_CFGLAYOUT)
2854 && can_merge_blocks_p (b, c))
2856 merge_blocks (b, c);
2857 update_forwarder_flag (b);
2858 changed_here = true;
2860 else if (!(mode & CLEANUP_CFGLAYOUT)
2861 /* If the jump insn has side effects,
2862 we can't kill the edge. */
2863 && (!JUMP_P (BB_END (b))
2864 || (reload_completed
2865 ? simplejump_p (BB_END (b))
2866 : (onlyjump_p (BB_END (b))
2867 && !tablejump_p (BB_END (b),
2868 NULL, NULL))))
2869 && (next = merge_blocks_move (s, b, c, mode)))
2871 b = next;
2872 changed_here = true;
2876 /* Try to change a branch to a return to just that return. */
2877 rtx_insn *ret, *use;
2878 if (single_succ_p (b)
2879 && onlyjump_p (BB_END (b))
2880 && bb_is_just_return (single_succ (b), &ret, &use))
2882 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2883 PATTERN (ret), 0))
2885 if (use)
2886 emit_insn_before (copy_insn (PATTERN (use)),
2887 BB_END (b));
2888 if (dump_file)
2889 fprintf (dump_file, "Changed jump %d->%d to return.\n",
2890 b->index, single_succ (b)->index);
2891 redirect_edge_succ (single_succ_edge (b),
2892 EXIT_BLOCK_PTR_FOR_FN (cfun));
2893 single_succ_edge (b)->flags &= ~EDGE_CROSSING;
2894 changed_here = true;
2898 /* Try to change a conditional branch to a return to the
2899 respective conditional return. */
2900 if (EDGE_COUNT (b->succs) == 2
2901 && any_condjump_p (BB_END (b))
2902 && bb_is_just_return (BRANCH_EDGE (b)->dest, &ret, &use))
2904 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2905 PATTERN (ret), 0))
2907 if (use)
2908 emit_insn_before (copy_insn (PATTERN (use)),
2909 BB_END (b));
2910 if (dump_file)
2911 fprintf (dump_file, "Changed conditional jump %d->%d "
2912 "to conditional return.\n",
2913 b->index, BRANCH_EDGE (b)->dest->index);
2914 redirect_edge_succ (BRANCH_EDGE (b),
2915 EXIT_BLOCK_PTR_FOR_FN (cfun));
2916 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2917 changed_here = true;
2921 /* Try to flip a conditional branch that falls through to
2922 a return so that it becomes a conditional return and a
2923 new jump to the original branch target. */
2924 if (EDGE_COUNT (b->succs) == 2
2925 && BRANCH_EDGE (b)->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
2926 && any_condjump_p (BB_END (b))
2927 && bb_is_just_return (FALLTHRU_EDGE (b)->dest, &ret, &use))
2929 if (invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2930 JUMP_LABEL (BB_END (b)), 0))
2932 basic_block new_ft = BRANCH_EDGE (b)->dest;
2933 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2934 PATTERN (ret), 0))
2936 if (use)
2937 emit_insn_before (copy_insn (PATTERN (use)),
2938 BB_END (b));
2939 if (dump_file)
2940 fprintf (dump_file, "Changed conditional jump "
2941 "%d->%d to conditional return, adding "
2942 "fall-through jump.\n",
2943 b->index, BRANCH_EDGE (b)->dest->index);
2944 redirect_edge_succ (BRANCH_EDGE (b),
2945 EXIT_BLOCK_PTR_FOR_FN (cfun));
2946 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2947 std::swap (BRANCH_EDGE (b)->probability,
2948 FALLTHRU_EDGE (b)->probability);
2949 update_br_prob_note (b);
2950 basic_block jb = force_nonfallthru (FALLTHRU_EDGE (b));
2951 notice_new_block (jb);
2952 if (!redirect_jump (as_a <rtx_jump_insn *> (BB_END (jb)),
2953 block_label (new_ft), 0))
2954 gcc_unreachable ();
2955 redirect_edge_succ (single_succ_edge (jb), new_ft);
2956 changed_here = true;
2958 else
2960 /* Invert the jump back to what it was. This should
2961 never fail. */
2962 if (!invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2963 JUMP_LABEL (BB_END (b)), 0))
2964 gcc_unreachable ();
2969 /* Simplify branch over branch. */
2970 if ((mode & CLEANUP_EXPENSIVE)
2971 && !(mode & CLEANUP_CFGLAYOUT)
2972 && try_simplify_condjump (b))
2973 changed_here = true;
2975 /* If B has a single outgoing edge, but uses a
2976 non-trivial jump instruction without side-effects, we
2977 can either delete the jump entirely, or replace it
2978 with a simple unconditional jump. */
2979 if (single_succ_p (b)
2980 && single_succ (b) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2981 && onlyjump_p (BB_END (b))
2982 && !CROSSING_JUMP_P (BB_END (b))
2983 && try_redirect_by_replacing_jump (single_succ_edge (b),
2984 single_succ (b),
2985 (mode & CLEANUP_CFGLAYOUT) != 0))
2987 update_forwarder_flag (b);
2988 changed_here = true;
2991 /* Simplify branch to branch. */
2992 if (try_forward_edges (mode, b))
2994 update_forwarder_flag (b);
2995 changed_here = true;
2998 /* Look for shared code between blocks. */
2999 if ((mode & CLEANUP_CROSSJUMP)
3000 && try_crossjump_bb (mode, b))
3001 changed_here = true;
3003 if ((mode & CLEANUP_CROSSJUMP)
3004 /* This can lengthen register lifetimes. Do it only after
3005 reload. */
3006 && reload_completed
3007 && try_head_merge_bb (b))
3008 changed_here = true;
3010 /* Don't get confused by the index shift caused by
3011 deleting blocks. */
3012 if (!changed_here)
3013 b = b->next_bb;
3014 else
3015 changed = true;
3018 if ((mode & CLEANUP_CROSSJUMP)
3019 && try_crossjump_bb (mode, EXIT_BLOCK_PTR_FOR_FN (cfun)))
3020 changed = true;
3022 if (block_was_dirty)
3024 /* This should only be set by head-merging. */
3025 gcc_assert (mode & CLEANUP_CROSSJUMP);
3026 df_analyze ();
3029 if (changed)
3031 /* Edge forwarding in particular can cause hot blocks previously
3032 reached by both hot and cold blocks to become dominated only
3033 by cold blocks. This will cause the verification below to fail,
3034 and lead to now cold code in the hot section. This is not easy
3035 to detect and fix during edge forwarding, and in some cases
3036 is only visible after newly unreachable blocks are deleted,
3037 which will be done in fixup_partitions. */
3038 fixup_partitions ();
3039 checking_verify_flow_info ();
3042 changed_overall |= changed;
3043 first_pass = false;
3045 while (changed);
3048 FOR_ALL_BB_FN (b, cfun)
3049 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
3051 return changed_overall;
3054 /* Delete all unreachable basic blocks. */
3056 bool
3057 delete_unreachable_blocks (void)
3059 bool changed = false;
3060 basic_block b, prev_bb;
3062 find_unreachable_blocks ();
3064 /* When we're in GIMPLE mode and there may be debug insns, we should
3065 delete blocks in reverse dominator order, so as to get a chance
3066 to substitute all released DEFs into debug stmts. If we don't
3067 have dominators information, walking blocks backward gets us a
3068 better chance of retaining most debug information than
3069 otherwise. */
3070 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
3071 && dom_info_available_p (CDI_DOMINATORS))
3073 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3074 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3076 prev_bb = b->prev_bb;
3078 if (!(b->flags & BB_REACHABLE))
3080 /* Speed up the removal of blocks that don't dominate
3081 others. Walking backwards, this should be the common
3082 case. */
3083 if (!first_dom_son (CDI_DOMINATORS, b))
3084 delete_basic_block (b);
3085 else
3087 vec<basic_block> h
3088 = get_all_dominated_blocks (CDI_DOMINATORS, b);
3090 while (h.length ())
3092 b = h.pop ();
3094 prev_bb = b->prev_bb;
3096 gcc_assert (!(b->flags & BB_REACHABLE));
3098 delete_basic_block (b);
3101 h.release ();
3104 changed = true;
3108 else
3110 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3111 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3113 prev_bb = b->prev_bb;
3115 if (!(b->flags & BB_REACHABLE))
3117 delete_basic_block (b);
3118 changed = true;
3123 if (changed)
3124 tidy_fallthru_edges ();
3125 return changed;
3128 /* Delete any jump tables never referenced. We can't delete them at the
3129 time of removing tablejump insn as they are referenced by the preceding
3130 insns computing the destination, so we delay deleting and garbagecollect
3131 them once life information is computed. */
3132 void
3133 delete_dead_jumptables (void)
3135 basic_block bb;
3137 /* A dead jump table does not belong to any basic block. Scan insns
3138 between two adjacent basic blocks. */
3139 FOR_EACH_BB_FN (bb, cfun)
3141 rtx_insn *insn, *next;
3143 for (insn = NEXT_INSN (BB_END (bb));
3144 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
3145 insn = next)
3147 next = NEXT_INSN (insn);
3148 if (LABEL_P (insn)
3149 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
3150 && JUMP_TABLE_DATA_P (next))
3152 rtx_insn *label = insn, *jump = next;
3154 if (dump_file)
3155 fprintf (dump_file, "Dead jumptable %i removed\n",
3156 INSN_UID (insn));
3158 next = NEXT_INSN (next);
3159 delete_insn (jump);
3160 delete_insn (label);
3167 /* Tidy the CFG by deleting unreachable code and whatnot. */
3169 bool
3170 cleanup_cfg (int mode)
3172 bool changed = false;
3174 /* Set the cfglayout mode flag here. We could update all the callers
3175 but that is just inconvenient, especially given that we eventually
3176 want to have cfglayout mode as the default. */
3177 if (current_ir_type () == IR_RTL_CFGLAYOUT)
3178 mode |= CLEANUP_CFGLAYOUT;
3180 timevar_push (TV_CLEANUP_CFG);
3181 if (delete_unreachable_blocks ())
3183 changed = true;
3184 /* We've possibly created trivially dead code. Cleanup it right
3185 now to introduce more opportunities for try_optimize_cfg. */
3186 if (!(mode & (CLEANUP_NO_INSN_DEL))
3187 && !reload_completed)
3188 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3191 compact_blocks ();
3193 /* To tail-merge blocks ending in the same noreturn function (e.g.
3194 a call to abort) we have to insert fake edges to exit. Do this
3195 here once. The fake edges do not interfere with any other CFG
3196 cleanups. */
3197 if (mode & CLEANUP_CROSSJUMP)
3198 add_noreturn_fake_exit_edges ();
3200 if (!dbg_cnt (cfg_cleanup))
3201 return changed;
3203 while (try_optimize_cfg (mode))
3205 delete_unreachable_blocks (), changed = true;
3206 if (!(mode & CLEANUP_NO_INSN_DEL))
3208 /* Try to remove some trivially dead insns when doing an expensive
3209 cleanup. But delete_trivially_dead_insns doesn't work after
3210 reload (it only handles pseudos) and run_fast_dce is too costly
3211 to run in every iteration.
3213 For effective cross jumping, we really want to run a fast DCE to
3214 clean up any dead conditions, or they get in the way of performing
3215 useful tail merges.
3217 Other transformations in cleanup_cfg are not so sensitive to dead
3218 code, so delete_trivially_dead_insns or even doing nothing at all
3219 is good enough. */
3220 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3221 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3222 break;
3223 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occurred)
3224 run_fast_dce ();
3226 else
3227 break;
3230 if (mode & CLEANUP_CROSSJUMP)
3231 remove_fake_exit_edges ();
3233 /* Don't call delete_dead_jumptables in cfglayout mode, because
3234 that function assumes that jump tables are in the insns stream.
3235 But we also don't _have_ to delete dead jumptables in cfglayout
3236 mode because we shouldn't even be looking at things that are
3237 not in a basic block. Dead jumptables are cleaned up when
3238 going out of cfglayout mode. */
3239 if (!(mode & CLEANUP_CFGLAYOUT))
3240 delete_dead_jumptables ();
3242 /* ??? We probably do this way too often. */
3243 if (current_loops
3244 && (changed
3245 || (mode & CLEANUP_CFG_CHANGED)))
3247 timevar_push (TV_REPAIR_LOOPS);
3248 /* The above doesn't preserve dominance info if available. */
3249 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3250 calculate_dominance_info (CDI_DOMINATORS);
3251 fix_loop_structure (NULL);
3252 free_dominance_info (CDI_DOMINATORS);
3253 timevar_pop (TV_REPAIR_LOOPS);
3256 timevar_pop (TV_CLEANUP_CFG);
3258 return changed;
3261 namespace {
3263 const pass_data pass_data_jump =
3265 RTL_PASS, /* type */
3266 "jump", /* name */
3267 OPTGROUP_NONE, /* optinfo_flags */
3268 TV_JUMP, /* tv_id */
3269 0, /* properties_required */
3270 0, /* properties_provided */
3271 0, /* properties_destroyed */
3272 0, /* todo_flags_start */
3273 0, /* todo_flags_finish */
3276 class pass_jump : public rtl_opt_pass
3278 public:
3279 pass_jump (gcc::context *ctxt)
3280 : rtl_opt_pass (pass_data_jump, ctxt)
3283 /* opt_pass methods: */
3284 virtual unsigned int execute (function *);
3286 }; // class pass_jump
3288 unsigned int
3289 pass_jump::execute (function *)
3291 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3292 if (dump_file)
3293 dump_flow_info (dump_file, dump_flags);
3294 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3295 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3296 return 0;
3299 } // anon namespace
3301 rtl_opt_pass *
3302 make_pass_jump (gcc::context *ctxt)
3304 return new pass_jump (ctxt);
3307 namespace {
3309 const pass_data pass_data_jump2 =
3311 RTL_PASS, /* type */
3312 "jump2", /* name */
3313 OPTGROUP_NONE, /* optinfo_flags */
3314 TV_JUMP, /* tv_id */
3315 0, /* properties_required */
3316 0, /* properties_provided */
3317 0, /* properties_destroyed */
3318 0, /* todo_flags_start */
3319 0, /* todo_flags_finish */
3322 class pass_jump2 : public rtl_opt_pass
3324 public:
3325 pass_jump2 (gcc::context *ctxt)
3326 : rtl_opt_pass (pass_data_jump2, ctxt)
3329 /* opt_pass methods: */
3330 virtual unsigned int execute (function *)
3332 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3333 return 0;
3336 }; // class pass_jump2
3338 } // anon namespace
3340 rtl_opt_pass *
3341 make_pass_jump2 (gcc::context *ctxt)
3343 return new pass_jump2 (ctxt);