* cp-tree.h (struct deferred_access_check): Add location.
[official-gcc.git] / gcc / cfgcleanup.c
blob331d87efa8bff234a9460b498b2e0c2401cfd024
1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987, 1988, 1992, 1993, 1994, 1995, 1996, 1997, 1998,
3 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2010, 2011
4 Free Software Foundation, Inc.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 /* This file contains optimizer of the control flow. The main entry point is
23 cleanup_cfg. Following optimizations are performed:
25 - Unreachable blocks removal
26 - Edge forwarding (edge to the forwarder block is forwarded to its
27 successor. Simplification of the branch instruction is performed by
28 underlying infrastructure so branch can be converted to simplejump or
29 eliminated).
30 - Cross jumping (tail merging)
31 - Conditional jump-around-simplejump simplification
32 - Basic block merging. */
34 #include "config.h"
35 #include "system.h"
36 #include "coretypes.h"
37 #include "tm.h"
38 #include "rtl.h"
39 #include "hard-reg-set.h"
40 #include "regs.h"
41 #include "insn-config.h"
42 #include "flags.h"
43 #include "recog.h"
44 #include "diagnostic-core.h"
45 #include "cselib.h"
46 #include "params.h"
47 #include "tm_p.h"
48 #include "target.h"
49 #include "function.h" /* For inline functions in emit-rtl.h they need crtl. */
50 #include "emit-rtl.h"
51 #include "tree-pass.h"
52 #include "cfgloop.h"
53 #include "expr.h"
54 #include "df.h"
55 #include "dce.h"
56 #include "dbgcnt.h"
58 #define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
60 /* Set to true when we are running first pass of try_optimize_cfg loop. */
61 static bool first_pass;
63 /* Set to true if crossjumps occurred in the latest run of try_optimize_cfg. */
64 static bool crossjumps_occured;
66 /* Set to true if we couldn't run an optimization due to stale liveness
67 information; we should run df_analyze to enable more opportunities. */
68 static bool block_was_dirty;
70 static bool try_crossjump_to_edge (int, edge, edge, enum replace_direction);
71 static bool try_crossjump_bb (int, basic_block);
72 static bool outgoing_edges_match (int, basic_block, basic_block);
73 static enum replace_direction old_insns_match_p (int, rtx, rtx);
75 static void merge_blocks_move_predecessor_nojumps (basic_block, basic_block);
76 static void merge_blocks_move_successor_nojumps (basic_block, basic_block);
77 static bool try_optimize_cfg (int);
78 static bool try_simplify_condjump (basic_block);
79 static bool try_forward_edges (int, basic_block);
80 static edge thread_jump (edge, basic_block);
81 static bool mark_effect (rtx, bitmap);
82 static void notice_new_block (basic_block);
83 static void update_forwarder_flag (basic_block);
84 static int mentions_nonequal_regs (rtx *, void *);
85 static void merge_memattrs (rtx, rtx);
87 /* Set flags for newly created block. */
89 static void
90 notice_new_block (basic_block bb)
92 if (!bb)
93 return;
95 if (forwarder_block_p (bb))
96 bb->flags |= BB_FORWARDER_BLOCK;
99 /* Recompute forwarder flag after block has been modified. */
101 static void
102 update_forwarder_flag (basic_block bb)
104 if (forwarder_block_p (bb))
105 bb->flags |= BB_FORWARDER_BLOCK;
106 else
107 bb->flags &= ~BB_FORWARDER_BLOCK;
110 /* Simplify a conditional jump around an unconditional jump.
111 Return true if something changed. */
113 static bool
114 try_simplify_condjump (basic_block cbranch_block)
116 basic_block jump_block, jump_dest_block, cbranch_dest_block;
117 edge cbranch_jump_edge, cbranch_fallthru_edge;
118 rtx cbranch_insn;
120 /* Verify that there are exactly two successors. */
121 if (EDGE_COUNT (cbranch_block->succs) != 2)
122 return false;
124 /* Verify that we've got a normal conditional branch at the end
125 of the block. */
126 cbranch_insn = BB_END (cbranch_block);
127 if (!any_condjump_p (cbranch_insn))
128 return false;
130 cbranch_fallthru_edge = FALLTHRU_EDGE (cbranch_block);
131 cbranch_jump_edge = BRANCH_EDGE (cbranch_block);
133 /* The next block must not have multiple predecessors, must not
134 be the last block in the function, and must contain just the
135 unconditional jump. */
136 jump_block = cbranch_fallthru_edge->dest;
137 if (!single_pred_p (jump_block)
138 || jump_block->next_bb == EXIT_BLOCK_PTR
139 || !FORWARDER_BLOCK_P (jump_block))
140 return false;
141 jump_dest_block = single_succ (jump_block);
143 /* If we are partitioning hot/cold basic blocks, we don't want to
144 mess up unconditional or indirect jumps that cross between hot
145 and cold sections.
147 Basic block partitioning may result in some jumps that appear to
148 be optimizable (or blocks that appear to be mergeable), but which really
149 must be left untouched (they are required to make it safely across
150 partition boundaries). See the comments at the top of
151 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
153 if (BB_PARTITION (jump_block) != BB_PARTITION (jump_dest_block)
154 || (cbranch_jump_edge->flags & EDGE_CROSSING))
155 return false;
157 /* The conditional branch must target the block after the
158 unconditional branch. */
159 cbranch_dest_block = cbranch_jump_edge->dest;
161 if (cbranch_dest_block == EXIT_BLOCK_PTR
162 || !can_fallthru (jump_block, cbranch_dest_block))
163 return false;
165 /* Invert the conditional branch. */
166 if (!invert_jump (cbranch_insn, block_label (jump_dest_block), 0))
167 return false;
169 if (dump_file)
170 fprintf (dump_file, "Simplifying condjump %i around jump %i\n",
171 INSN_UID (cbranch_insn), INSN_UID (BB_END (jump_block)));
173 /* Success. Update the CFG to match. Note that after this point
174 the edge variable names appear backwards; the redirection is done
175 this way to preserve edge profile data. */
176 cbranch_jump_edge = redirect_edge_succ_nodup (cbranch_jump_edge,
177 cbranch_dest_block);
178 cbranch_fallthru_edge = redirect_edge_succ_nodup (cbranch_fallthru_edge,
179 jump_dest_block);
180 cbranch_jump_edge->flags |= EDGE_FALLTHRU;
181 cbranch_fallthru_edge->flags &= ~EDGE_FALLTHRU;
182 update_br_prob_note (cbranch_block);
184 /* Delete the block with the unconditional jump, and clean up the mess. */
185 delete_basic_block (jump_block);
186 tidy_fallthru_edge (cbranch_jump_edge);
187 update_forwarder_flag (cbranch_block);
189 return true;
192 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
193 on register. Used by jump threading. */
195 static bool
196 mark_effect (rtx exp, regset nonequal)
198 int regno;
199 rtx dest;
200 switch (GET_CODE (exp))
202 /* In case we do clobber the register, mark it as equal, as we know the
203 value is dead so it don't have to match. */
204 case CLOBBER:
205 if (REG_P (XEXP (exp, 0)))
207 dest = XEXP (exp, 0);
208 regno = REGNO (dest);
209 if (HARD_REGISTER_NUM_P (regno))
210 bitmap_clear_range (nonequal, regno,
211 hard_regno_nregs[regno][GET_MODE (dest)]);
212 else
213 bitmap_clear_bit (nonequal, regno);
215 return false;
217 case SET:
218 if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
219 return false;
220 dest = SET_DEST (exp);
221 if (dest == pc_rtx)
222 return false;
223 if (!REG_P (dest))
224 return true;
225 regno = REGNO (dest);
226 if (HARD_REGISTER_NUM_P (regno))
227 bitmap_set_range (nonequal, regno,
228 hard_regno_nregs[regno][GET_MODE (dest)]);
229 else
230 bitmap_set_bit (nonequal, regno);
231 return false;
233 default:
234 return false;
238 /* Return nonzero if X is a register set in regset DATA.
239 Called via for_each_rtx. */
240 static int
241 mentions_nonequal_regs (rtx *x, void *data)
243 regset nonequal = (regset) data;
244 if (REG_P (*x))
246 int regno;
248 regno = REGNO (*x);
249 if (REGNO_REG_SET_P (nonequal, regno))
250 return 1;
251 if (regno < FIRST_PSEUDO_REGISTER)
253 int n = hard_regno_nregs[regno][GET_MODE (*x)];
254 while (--n > 0)
255 if (REGNO_REG_SET_P (nonequal, regno + n))
256 return 1;
259 return 0;
261 /* Attempt to prove that the basic block B will have no side effects and
262 always continues in the same edge if reached via E. Return the edge
263 if exist, NULL otherwise. */
265 static edge
266 thread_jump (edge e, basic_block b)
268 rtx set1, set2, cond1, cond2, insn;
269 enum rtx_code code1, code2, reversed_code2;
270 bool reverse1 = false;
271 unsigned i;
272 regset nonequal;
273 bool failed = false;
274 reg_set_iterator rsi;
276 if (b->flags & BB_NONTHREADABLE_BLOCK)
277 return NULL;
279 /* At the moment, we do handle only conditional jumps, but later we may
280 want to extend this code to tablejumps and others. */
281 if (EDGE_COUNT (e->src->succs) != 2)
282 return NULL;
283 if (EDGE_COUNT (b->succs) != 2)
285 b->flags |= BB_NONTHREADABLE_BLOCK;
286 return NULL;
289 /* Second branch must end with onlyjump, as we will eliminate the jump. */
290 if (!any_condjump_p (BB_END (e->src)))
291 return NULL;
293 if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
295 b->flags |= BB_NONTHREADABLE_BLOCK;
296 return NULL;
299 set1 = pc_set (BB_END (e->src));
300 set2 = pc_set (BB_END (b));
301 if (((e->flags & EDGE_FALLTHRU) != 0)
302 != (XEXP (SET_SRC (set1), 1) == pc_rtx))
303 reverse1 = true;
305 cond1 = XEXP (SET_SRC (set1), 0);
306 cond2 = XEXP (SET_SRC (set2), 0);
307 if (reverse1)
308 code1 = reversed_comparison_code (cond1, BB_END (e->src));
309 else
310 code1 = GET_CODE (cond1);
312 code2 = GET_CODE (cond2);
313 reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
315 if (!comparison_dominates_p (code1, code2)
316 && !comparison_dominates_p (code1, reversed_code2))
317 return NULL;
319 /* Ensure that the comparison operators are equivalent.
320 ??? This is far too pessimistic. We should allow swapped operands,
321 different CCmodes, or for example comparisons for interval, that
322 dominate even when operands are not equivalent. */
323 if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
324 || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
325 return NULL;
327 /* Short circuit cases where block B contains some side effects, as we can't
328 safely bypass it. */
329 for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
330 insn = NEXT_INSN (insn))
331 if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
333 b->flags |= BB_NONTHREADABLE_BLOCK;
334 return NULL;
337 cselib_init (0);
339 /* First process all values computed in the source basic block. */
340 for (insn = NEXT_INSN (BB_HEAD (e->src));
341 insn != NEXT_INSN (BB_END (e->src));
342 insn = NEXT_INSN (insn))
343 if (INSN_P (insn))
344 cselib_process_insn (insn);
346 nonequal = BITMAP_ALLOC (NULL);
347 CLEAR_REG_SET (nonequal);
349 /* Now assume that we've continued by the edge E to B and continue
350 processing as if it were same basic block.
351 Our goal is to prove that whole block is an NOOP. */
353 for (insn = NEXT_INSN (BB_HEAD (b));
354 insn != NEXT_INSN (BB_END (b)) && !failed;
355 insn = NEXT_INSN (insn))
357 if (INSN_P (insn))
359 rtx pat = PATTERN (insn);
361 if (GET_CODE (pat) == PARALLEL)
363 for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
364 failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
366 else
367 failed |= mark_effect (pat, nonequal);
370 cselib_process_insn (insn);
373 /* Later we should clear nonequal of dead registers. So far we don't
374 have life information in cfg_cleanup. */
375 if (failed)
377 b->flags |= BB_NONTHREADABLE_BLOCK;
378 goto failed_exit;
381 /* cond2 must not mention any register that is not equal to the
382 former block. */
383 if (for_each_rtx (&cond2, mentions_nonequal_regs, nonequal))
384 goto failed_exit;
386 EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
387 goto failed_exit;
389 BITMAP_FREE (nonequal);
390 cselib_finish ();
391 if ((comparison_dominates_p (code1, code2) != 0)
392 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
393 return BRANCH_EDGE (b);
394 else
395 return FALLTHRU_EDGE (b);
397 failed_exit:
398 BITMAP_FREE (nonequal);
399 cselib_finish ();
400 return NULL;
403 /* Attempt to forward edges leaving basic block B.
404 Return true if successful. */
406 static bool
407 try_forward_edges (int mode, basic_block b)
409 bool changed = false;
410 edge_iterator ei;
411 edge e, *threaded_edges = NULL;
413 /* If we are partitioning hot/cold basic blocks, we don't want to
414 mess up unconditional or indirect jumps that cross between hot
415 and cold sections.
417 Basic block partitioning may result in some jumps that appear to
418 be optimizable (or blocks that appear to be mergeable), but which really
419 must be left untouched (they are required to make it safely across
420 partition boundaries). See the comments at the top of
421 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
423 if (find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX))
424 return false;
426 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
428 basic_block target, first;
429 int counter, goto_locus;
430 bool threaded = false;
431 int nthreaded_edges = 0;
432 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
434 /* Skip complex edges because we don't know how to update them.
436 Still handle fallthru edges, as we can succeed to forward fallthru
437 edge to the same place as the branch edge of conditional branch
438 and turn conditional branch to an unconditional branch. */
439 if (e->flags & EDGE_COMPLEX)
441 ei_next (&ei);
442 continue;
445 target = first = e->dest;
446 counter = NUM_FIXED_BLOCKS;
447 goto_locus = e->goto_locus;
449 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
450 up jumps that cross between hot/cold sections.
452 Basic block partitioning may result in some jumps that appear
453 to be optimizable (or blocks that appear to be mergeable), but which
454 really must be left untouched (they are required to make it safely
455 across partition boundaries). See the comments at the top of
456 bb-reorder.c:partition_hot_cold_basic_blocks for complete
457 details. */
459 if (first != EXIT_BLOCK_PTR
460 && find_reg_note (BB_END (first), REG_CROSSING_JUMP, NULL_RTX))
461 return false;
463 while (counter < n_basic_blocks)
465 basic_block new_target = NULL;
466 bool new_target_threaded = false;
467 may_thread |= (target->flags & BB_MODIFIED) != 0;
469 if (FORWARDER_BLOCK_P (target)
470 && !(single_succ_edge (target)->flags & EDGE_CROSSING)
471 && single_succ (target) != EXIT_BLOCK_PTR)
473 /* Bypass trivial infinite loops. */
474 new_target = single_succ (target);
475 if (target == new_target)
476 counter = n_basic_blocks;
477 else if (!optimize)
479 /* When not optimizing, ensure that edges or forwarder
480 blocks with different locus are not optimized out. */
481 int new_locus = single_succ_edge (target)->goto_locus;
482 int locus = goto_locus;
484 if (new_locus && locus && !locator_eq (new_locus, locus))
485 new_target = NULL;
486 else
488 rtx last;
490 if (new_locus)
491 locus = new_locus;
493 last = BB_END (target);
494 if (DEBUG_INSN_P (last))
495 last = prev_nondebug_insn (last);
497 new_locus = last && INSN_P (last)
498 ? INSN_LOCATOR (last) : 0;
500 if (new_locus && locus && !locator_eq (new_locus, locus))
501 new_target = NULL;
502 else
504 if (new_locus)
505 locus = new_locus;
507 goto_locus = locus;
513 /* Allow to thread only over one edge at time to simplify updating
514 of probabilities. */
515 else if ((mode & CLEANUP_THREADING) && may_thread)
517 edge t = thread_jump (e, target);
518 if (t)
520 if (!threaded_edges)
521 threaded_edges = XNEWVEC (edge, n_basic_blocks);
522 else
524 int i;
526 /* Detect an infinite loop across blocks not
527 including the start block. */
528 for (i = 0; i < nthreaded_edges; ++i)
529 if (threaded_edges[i] == t)
530 break;
531 if (i < nthreaded_edges)
533 counter = n_basic_blocks;
534 break;
538 /* Detect an infinite loop across the start block. */
539 if (t->dest == b)
540 break;
542 gcc_assert (nthreaded_edges < n_basic_blocks - NUM_FIXED_BLOCKS);
543 threaded_edges[nthreaded_edges++] = t;
545 new_target = t->dest;
546 new_target_threaded = true;
550 if (!new_target)
551 break;
553 counter++;
554 target = new_target;
555 threaded |= new_target_threaded;
558 if (counter >= n_basic_blocks)
560 if (dump_file)
561 fprintf (dump_file, "Infinite loop in BB %i.\n",
562 target->index);
564 else if (target == first)
565 ; /* We didn't do anything. */
566 else
568 /* Save the values now, as the edge may get removed. */
569 gcov_type edge_count = e->count;
570 int edge_probability = e->probability;
571 int edge_frequency;
572 int n = 0;
574 e->goto_locus = goto_locus;
576 /* Don't force if target is exit block. */
577 if (threaded && target != EXIT_BLOCK_PTR)
579 notice_new_block (redirect_edge_and_branch_force (e, target));
580 if (dump_file)
581 fprintf (dump_file, "Conditionals threaded.\n");
583 else if (!redirect_edge_and_branch (e, target))
585 if (dump_file)
586 fprintf (dump_file,
587 "Forwarding edge %i->%i to %i failed.\n",
588 b->index, e->dest->index, target->index);
589 ei_next (&ei);
590 continue;
593 /* We successfully forwarded the edge. Now update profile
594 data: for each edge we traversed in the chain, remove
595 the original edge's execution count. */
596 edge_frequency = ((edge_probability * b->frequency
597 + REG_BR_PROB_BASE / 2)
598 / REG_BR_PROB_BASE);
602 edge t;
604 if (!single_succ_p (first))
606 gcc_assert (n < nthreaded_edges);
607 t = threaded_edges [n++];
608 gcc_assert (t->src == first);
609 update_bb_profile_for_threading (first, edge_frequency,
610 edge_count, t);
611 update_br_prob_note (first);
613 else
615 first->count -= edge_count;
616 if (first->count < 0)
617 first->count = 0;
618 first->frequency -= edge_frequency;
619 if (first->frequency < 0)
620 first->frequency = 0;
621 /* It is possible that as the result of
622 threading we've removed edge as it is
623 threaded to the fallthru edge. Avoid
624 getting out of sync. */
625 if (n < nthreaded_edges
626 && first == threaded_edges [n]->src)
627 n++;
628 t = single_succ_edge (first);
631 t->count -= edge_count;
632 if (t->count < 0)
633 t->count = 0;
634 first = t->dest;
636 while (first != target);
638 changed = true;
639 continue;
641 ei_next (&ei);
644 free (threaded_edges);
645 return changed;
649 /* Blocks A and B are to be merged into a single block. A has no incoming
650 fallthru edge, so it can be moved before B without adding or modifying
651 any jumps (aside from the jump from A to B). */
653 static void
654 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
656 rtx barrier;
658 /* If we are partitioning hot/cold basic blocks, we don't want to
659 mess up unconditional or indirect jumps that cross between hot
660 and cold sections.
662 Basic block partitioning may result in some jumps that appear to
663 be optimizable (or blocks that appear to be mergeable), but which really
664 must be left untouched (they are required to make it safely across
665 partition boundaries). See the comments at the top of
666 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
668 if (BB_PARTITION (a) != BB_PARTITION (b))
669 return;
671 barrier = next_nonnote_insn (BB_END (a));
672 gcc_assert (BARRIER_P (barrier));
673 delete_insn (barrier);
675 /* Scramble the insn chain. */
676 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
677 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
678 df_set_bb_dirty (a);
680 if (dump_file)
681 fprintf (dump_file, "Moved block %d before %d and merged.\n",
682 a->index, b->index);
684 /* Swap the records for the two blocks around. */
686 unlink_block (a);
687 link_block (a, b->prev_bb);
689 /* Now blocks A and B are contiguous. Merge them. */
690 merge_blocks (a, b);
693 /* Blocks A and B are to be merged into a single block. B has no outgoing
694 fallthru edge, so it can be moved after A without adding or modifying
695 any jumps (aside from the jump from A to B). */
697 static void
698 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
700 rtx barrier, real_b_end;
701 rtx label, table;
703 /* If we are partitioning hot/cold basic blocks, we don't want to
704 mess up unconditional or indirect jumps that cross between hot
705 and cold sections.
707 Basic block partitioning may result in some jumps that appear to
708 be optimizable (or blocks that appear to be mergeable), but which really
709 must be left untouched (they are required to make it safely across
710 partition boundaries). See the comments at the top of
711 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
713 if (BB_PARTITION (a) != BB_PARTITION (b))
714 return;
716 real_b_end = BB_END (b);
718 /* If there is a jump table following block B temporarily add the jump table
719 to block B so that it will also be moved to the correct location. */
720 if (tablejump_p (BB_END (b), &label, &table)
721 && prev_active_insn (label) == BB_END (b))
723 BB_END (b) = table;
726 /* There had better have been a barrier there. Delete it. */
727 barrier = NEXT_INSN (BB_END (b));
728 if (barrier && BARRIER_P (barrier))
729 delete_insn (barrier);
732 /* Scramble the insn chain. */
733 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
735 /* Restore the real end of b. */
736 BB_END (b) = real_b_end;
738 if (dump_file)
739 fprintf (dump_file, "Moved block %d after %d and merged.\n",
740 b->index, a->index);
742 /* Now blocks A and B are contiguous. Merge them. */
743 merge_blocks (a, b);
746 /* Attempt to merge basic blocks that are potentially non-adjacent.
747 Return NULL iff the attempt failed, otherwise return basic block
748 where cleanup_cfg should continue. Because the merging commonly
749 moves basic block away or introduces another optimization
750 possibility, return basic block just before B so cleanup_cfg don't
751 need to iterate.
753 It may be good idea to return basic block before C in the case
754 C has been moved after B and originally appeared earlier in the
755 insn sequence, but we have no information available about the
756 relative ordering of these two. Hopefully it is not too common. */
758 static basic_block
759 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
761 basic_block next;
763 /* If we are partitioning hot/cold basic blocks, we don't want to
764 mess up unconditional or indirect jumps that cross between hot
765 and cold sections.
767 Basic block partitioning may result in some jumps that appear to
768 be optimizable (or blocks that appear to be mergeable), but which really
769 must be left untouched (they are required to make it safely across
770 partition boundaries). See the comments at the top of
771 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
773 if (BB_PARTITION (b) != BB_PARTITION (c))
774 return NULL;
776 /* If B has a fallthru edge to C, no need to move anything. */
777 if (e->flags & EDGE_FALLTHRU)
779 int b_index = b->index, c_index = c->index;
781 /* Protect the loop latches. */
782 if (current_loops && c->loop_father->latch == c)
783 return NULL;
785 merge_blocks (b, c);
786 update_forwarder_flag (b);
788 if (dump_file)
789 fprintf (dump_file, "Merged %d and %d without moving.\n",
790 b_index, c_index);
792 return b->prev_bb == ENTRY_BLOCK_PTR ? b : b->prev_bb;
795 /* Otherwise we will need to move code around. Do that only if expensive
796 transformations are allowed. */
797 else if (mode & CLEANUP_EXPENSIVE)
799 edge tmp_edge, b_fallthru_edge;
800 bool c_has_outgoing_fallthru;
801 bool b_has_incoming_fallthru;
803 /* Avoid overactive code motion, as the forwarder blocks should be
804 eliminated by edge redirection instead. One exception might have
805 been if B is a forwarder block and C has no fallthru edge, but
806 that should be cleaned up by bb-reorder instead. */
807 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
808 return NULL;
810 /* We must make sure to not munge nesting of lexical blocks,
811 and loop notes. This is done by squeezing out all the notes
812 and leaving them there to lie. Not ideal, but functional. */
814 tmp_edge = find_fallthru_edge (c->succs);
815 c_has_outgoing_fallthru = (tmp_edge != NULL);
817 tmp_edge = find_fallthru_edge (b->preds);
818 b_has_incoming_fallthru = (tmp_edge != NULL);
819 b_fallthru_edge = tmp_edge;
820 next = b->prev_bb;
821 if (next == c)
822 next = next->prev_bb;
824 /* Otherwise, we're going to try to move C after B. If C does
825 not have an outgoing fallthru, then it can be moved
826 immediately after B without introducing or modifying jumps. */
827 if (! c_has_outgoing_fallthru)
829 merge_blocks_move_successor_nojumps (b, c);
830 return next == ENTRY_BLOCK_PTR ? next->next_bb : next;
833 /* If B does not have an incoming fallthru, then it can be moved
834 immediately before C without introducing or modifying jumps.
835 C cannot be the first block, so we do not have to worry about
836 accessing a non-existent block. */
838 if (b_has_incoming_fallthru)
840 basic_block bb;
842 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR)
843 return NULL;
844 bb = force_nonfallthru (b_fallthru_edge);
845 if (bb)
846 notice_new_block (bb);
849 merge_blocks_move_predecessor_nojumps (b, c);
850 return next == ENTRY_BLOCK_PTR ? next->next_bb : next;
853 return NULL;
857 /* Removes the memory attributes of MEM expression
858 if they are not equal. */
860 void
861 merge_memattrs (rtx x, rtx y)
863 int i;
864 int j;
865 enum rtx_code code;
866 const char *fmt;
868 if (x == y)
869 return;
870 if (x == 0 || y == 0)
871 return;
873 code = GET_CODE (x);
875 if (code != GET_CODE (y))
876 return;
878 if (GET_MODE (x) != GET_MODE (y))
879 return;
881 if (code == MEM && MEM_ATTRS (x) != MEM_ATTRS (y))
883 if (! MEM_ATTRS (x))
884 MEM_ATTRS (y) = 0;
885 else if (! MEM_ATTRS (y))
886 MEM_ATTRS (x) = 0;
887 else
889 HOST_WIDE_INT mem_size;
891 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
893 set_mem_alias_set (x, 0);
894 set_mem_alias_set (y, 0);
897 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
899 set_mem_expr (x, 0);
900 set_mem_expr (y, 0);
901 clear_mem_offset (x);
902 clear_mem_offset (y);
904 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
905 || (MEM_OFFSET_KNOWN_P (x)
906 && MEM_OFFSET (x) != MEM_OFFSET (y)))
908 clear_mem_offset (x);
909 clear_mem_offset (y);
912 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
914 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
915 set_mem_size (x, mem_size);
916 set_mem_size (y, mem_size);
918 else
920 clear_mem_size (x);
921 clear_mem_size (y);
924 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
925 set_mem_align (y, MEM_ALIGN (x));
929 fmt = GET_RTX_FORMAT (code);
930 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
932 switch (fmt[i])
934 case 'E':
935 /* Two vectors must have the same length. */
936 if (XVECLEN (x, i) != XVECLEN (y, i))
937 return;
939 for (j = 0; j < XVECLEN (x, i); j++)
940 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
942 break;
944 case 'e':
945 merge_memattrs (XEXP (x, i), XEXP (y, i));
948 return;
952 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
953 different single sets S1 and S2. */
955 static bool
956 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
958 int i;
959 rtx e1, e2;
961 if (p1 == s1 && p2 == s2)
962 return true;
964 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
965 return false;
967 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
968 return false;
970 for (i = 0; i < XVECLEN (p1, 0); i++)
972 e1 = XVECEXP (p1, 0, i);
973 e2 = XVECEXP (p2, 0, i);
974 if (e1 == s1 && e2 == s2)
975 continue;
976 if (reload_completed
977 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
978 continue;
980 return false;
983 return true;
986 /* Examine register notes on I1 and I2 and return:
987 - dir_forward if I1 can be replaced by I2, or
988 - dir_backward if I2 can be replaced by I1, or
989 - dir_both if both are the case. */
991 static enum replace_direction
992 can_replace_by (rtx i1, rtx i2)
994 rtx s1, s2, d1, d2, src1, src2, note1, note2;
995 bool c1, c2;
997 /* Check for 2 sets. */
998 s1 = single_set (i1);
999 s2 = single_set (i2);
1000 if (s1 == NULL_RTX || s2 == NULL_RTX)
1001 return dir_none;
1003 /* Check that the 2 sets set the same dest. */
1004 d1 = SET_DEST (s1);
1005 d2 = SET_DEST (s2);
1006 if (!(reload_completed
1007 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1008 return dir_none;
1010 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1011 set dest to the same value. */
1012 note1 = find_reg_equal_equiv_note (i1);
1013 note2 = find_reg_equal_equiv_note (i2);
1014 if (!note1 || !note2 || !rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0))
1015 || !CONST_INT_P (XEXP (note1, 0)))
1016 return dir_none;
1018 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1019 return dir_none;
1021 /* Although the 2 sets set dest to the same value, we cannot replace
1022 (set (dest) (const_int))
1024 (set (dest) (reg))
1025 because we don't know if the reg is live and has the same value at the
1026 location of replacement. */
1027 src1 = SET_SRC (s1);
1028 src2 = SET_SRC (s2);
1029 c1 = CONST_INT_P (src1);
1030 c2 = CONST_INT_P (src2);
1031 if (c1 && c2)
1032 return dir_both;
1033 else if (c2)
1034 return dir_forward;
1035 else if (c1)
1036 return dir_backward;
1038 return dir_none;
1041 /* Merges directions A and B. */
1043 static enum replace_direction
1044 merge_dir (enum replace_direction a, enum replace_direction b)
1046 /* Implements the following table:
1047 |bo fw bw no
1048 ---+-----------
1049 bo |bo fw bw no
1050 fw |-- fw no no
1051 bw |-- -- bw no
1052 no |-- -- -- no. */
1054 if (a == b)
1055 return a;
1057 if (a == dir_both)
1058 return b;
1059 if (b == dir_both)
1060 return a;
1062 return dir_none;
1065 /* Examine I1 and I2 and return:
1066 - dir_forward if I1 can be replaced by I2, or
1067 - dir_backward if I2 can be replaced by I1, or
1068 - dir_both if both are the case. */
1070 static enum replace_direction
1071 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx i1, rtx i2)
1073 rtx p1, p2;
1075 /* Verify that I1 and I2 are equivalent. */
1076 if (GET_CODE (i1) != GET_CODE (i2))
1077 return dir_none;
1079 /* __builtin_unreachable() may lead to empty blocks (ending with
1080 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1081 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1082 return dir_both;
1084 /* ??? Do not allow cross-jumping between different stack levels. */
1085 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1086 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1087 if (p1 && p2)
1089 p1 = XEXP (p1, 0);
1090 p2 = XEXP (p2, 0);
1091 if (!rtx_equal_p (p1, p2))
1092 return dir_none;
1094 /* ??? Worse, this adjustment had better be constant lest we
1095 have differing incoming stack levels. */
1096 if (!frame_pointer_needed
1097 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1098 return dir_none;
1100 else if (p1 || p2)
1101 return dir_none;
1103 p1 = PATTERN (i1);
1104 p2 = PATTERN (i2);
1106 if (GET_CODE (p1) != GET_CODE (p2))
1107 return dir_none;
1109 /* If this is a CALL_INSN, compare register usage information.
1110 If we don't check this on stack register machines, the two
1111 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1112 numbers of stack registers in the same basic block.
1113 If we don't check this on machines with delay slots, a delay slot may
1114 be filled that clobbers a parameter expected by the subroutine.
1116 ??? We take the simple route for now and assume that if they're
1117 equal, they were constructed identically.
1119 Also check for identical exception regions. */
1121 if (CALL_P (i1))
1123 /* Ensure the same EH region. */
1124 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1125 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1127 if (!n1 && n2)
1128 return dir_none;
1130 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1131 return dir_none;
1133 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1134 CALL_INSN_FUNCTION_USAGE (i2))
1135 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1136 return dir_none;
1139 #ifdef STACK_REGS
1140 /* If cross_jump_death_matters is not 0, the insn's mode
1141 indicates whether or not the insn contains any stack-like
1142 regs. */
1144 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1146 /* If register stack conversion has already been done, then
1147 death notes must also be compared before it is certain that
1148 the two instruction streams match. */
1150 rtx note;
1151 HARD_REG_SET i1_regset, i2_regset;
1153 CLEAR_HARD_REG_SET (i1_regset);
1154 CLEAR_HARD_REG_SET (i2_regset);
1156 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1157 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1158 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1160 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1161 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1162 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1164 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1165 return dir_none;
1167 #endif
1169 if (reload_completed
1170 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1171 return dir_both;
1173 return can_replace_by (i1, i2);
1176 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1177 flow_find_head_matching_sequence, ensure the notes match. */
1179 static void
1180 merge_notes (rtx i1, rtx i2)
1182 /* If the merged insns have different REG_EQUAL notes, then
1183 remove them. */
1184 rtx equiv1 = find_reg_equal_equiv_note (i1);
1185 rtx equiv2 = find_reg_equal_equiv_note (i2);
1187 if (equiv1 && !equiv2)
1188 remove_note (i1, equiv1);
1189 else if (!equiv1 && equiv2)
1190 remove_note (i2, equiv2);
1191 else if (equiv1 && equiv2
1192 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1194 remove_note (i1, equiv1);
1195 remove_note (i2, equiv2);
1199 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1200 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1201 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1202 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1203 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1205 static void
1206 walk_to_nondebug_insn (rtx *i1, basic_block *bb1, bool follow_fallthru,
1207 bool *did_fallthru)
1209 edge fallthru;
1211 *did_fallthru = false;
1213 /* Ignore notes. */
1214 while (!NONDEBUG_INSN_P (*i1))
1216 if (*i1 != BB_HEAD (*bb1))
1218 *i1 = PREV_INSN (*i1);
1219 continue;
1222 if (!follow_fallthru)
1223 return;
1225 fallthru = find_fallthru_edge ((*bb1)->preds);
1226 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FUNCTION (cfun)
1227 || !single_succ_p (fallthru->src))
1228 return;
1230 *bb1 = fallthru->src;
1231 *i1 = BB_END (*bb1);
1232 *did_fallthru = true;
1236 /* Look through the insns at the end of BB1 and BB2 and find the longest
1237 sequence that are either equivalent, or allow forward or backward
1238 replacement. Store the first insns for that sequence in *F1 and *F2 and
1239 return the sequence length.
1241 DIR_P indicates the allowed replacement direction on function entry, and
1242 the actual replacement direction on function exit. If NULL, only equivalent
1243 sequences are allowed.
1245 To simplify callers of this function, if the blocks match exactly,
1246 store the head of the blocks in *F1 and *F2. */
1249 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx *f1, rtx *f2,
1250 enum replace_direction *dir_p)
1252 rtx i1, i2, last1, last2, afterlast1, afterlast2;
1253 int ninsns = 0;
1254 rtx p1;
1255 enum replace_direction dir, last_dir, afterlast_dir;
1256 bool follow_fallthru, did_fallthru;
1258 if (dir_p)
1259 dir = *dir_p;
1260 else
1261 dir = dir_both;
1262 afterlast_dir = dir;
1263 last_dir = afterlast_dir;
1265 /* Skip simple jumps at the end of the blocks. Complex jumps still
1266 need to be compared for equivalence, which we'll do below. */
1268 i1 = BB_END (bb1);
1269 last1 = afterlast1 = last2 = afterlast2 = NULL_RTX;
1270 if (onlyjump_p (i1)
1271 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1273 last1 = i1;
1274 i1 = PREV_INSN (i1);
1277 i2 = BB_END (bb2);
1278 if (onlyjump_p (i2)
1279 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1281 last2 = i2;
1282 /* Count everything except for unconditional jump as insn. */
1283 if (!simplejump_p (i2) && !returnjump_p (i2) && last1)
1284 ninsns++;
1285 i2 = PREV_INSN (i2);
1288 while (true)
1290 /* In the following example, we can replace all jumps to C by jumps to A.
1292 This removes 4 duplicate insns.
1293 [bb A] insn1 [bb C] insn1
1294 insn2 insn2
1295 [bb B] insn3 insn3
1296 insn4 insn4
1297 jump_insn jump_insn
1299 We could also replace all jumps to A by jumps to C, but that leaves B
1300 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1301 step, all jumps to B would be replaced with jumps to the middle of C,
1302 achieving the same result with more effort.
1303 So we allow only the first possibility, which means that we don't allow
1304 fallthru in the block that's being replaced. */
1306 follow_fallthru = dir_p && dir != dir_forward;
1307 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1308 if (did_fallthru)
1309 dir = dir_backward;
1311 follow_fallthru = dir_p && dir != dir_backward;
1312 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1313 if (did_fallthru)
1314 dir = dir_forward;
1316 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1317 break;
1319 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1320 if (dir == dir_none || (!dir_p && dir != dir_both))
1321 break;
1323 merge_memattrs (i1, i2);
1325 /* Don't begin a cross-jump with a NOTE insn. */
1326 if (INSN_P (i1))
1328 merge_notes (i1, i2);
1330 afterlast1 = last1, afterlast2 = last2;
1331 last1 = i1, last2 = i2;
1332 afterlast_dir = last_dir;
1333 last_dir = dir;
1334 p1 = PATTERN (i1);
1335 if (!(GET_CODE (p1) == USE || GET_CODE (p1) == CLOBBER))
1336 ninsns++;
1339 i1 = PREV_INSN (i1);
1340 i2 = PREV_INSN (i2);
1343 #ifdef HAVE_cc0
1344 /* Don't allow the insn after a compare to be shared by
1345 cross-jumping unless the compare is also shared. */
1346 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && ! sets_cc0_p (last1))
1347 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1348 #endif
1350 /* Include preceding notes and labels in the cross-jump. One,
1351 this may bring us to the head of the blocks as requested above.
1352 Two, it keeps line number notes as matched as may be. */
1353 if (ninsns)
1355 bb1 = BLOCK_FOR_INSN (last1);
1356 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1357 last1 = PREV_INSN (last1);
1359 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1360 last1 = PREV_INSN (last1);
1362 bb2 = BLOCK_FOR_INSN (last2);
1363 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1364 last2 = PREV_INSN (last2);
1366 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1367 last2 = PREV_INSN (last2);
1369 *f1 = last1;
1370 *f2 = last2;
1373 if (dir_p)
1374 *dir_p = last_dir;
1375 return ninsns;
1378 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1379 the head of the two blocks. Do not include jumps at the end.
1380 If STOP_AFTER is nonzero, stop after finding that many matching
1381 instructions. */
1384 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx *f1,
1385 rtx *f2, int stop_after)
1387 rtx i1, i2, last1, last2, beforelast1, beforelast2;
1388 int ninsns = 0;
1389 edge e;
1390 edge_iterator ei;
1391 int nehedges1 = 0, nehedges2 = 0;
1393 FOR_EACH_EDGE (e, ei, bb1->succs)
1394 if (e->flags & EDGE_EH)
1395 nehedges1++;
1396 FOR_EACH_EDGE (e, ei, bb2->succs)
1397 if (e->flags & EDGE_EH)
1398 nehedges2++;
1400 i1 = BB_HEAD (bb1);
1401 i2 = BB_HEAD (bb2);
1402 last1 = beforelast1 = last2 = beforelast2 = NULL_RTX;
1404 while (true)
1406 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1407 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1409 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1410 break;
1411 i1 = NEXT_INSN (i1);
1414 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1416 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1417 break;
1418 i2 = NEXT_INSN (i2);
1421 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1422 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1423 break;
1425 if (NOTE_P (i1) || NOTE_P (i2)
1426 || JUMP_P (i1) || JUMP_P (i2))
1427 break;
1429 /* A sanity check to make sure we're not merging insns with different
1430 effects on EH. If only one of them ends a basic block, it shouldn't
1431 have an EH edge; if both end a basic block, there should be the same
1432 number of EH edges. */
1433 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1434 && nehedges1 > 0)
1435 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1436 && nehedges2 > 0)
1437 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1438 && nehedges1 != nehedges2))
1439 break;
1441 if (old_insns_match_p (0, i1, i2) != dir_both)
1442 break;
1444 merge_memattrs (i1, i2);
1446 /* Don't begin a cross-jump with a NOTE insn. */
1447 if (INSN_P (i1))
1449 merge_notes (i1, i2);
1451 beforelast1 = last1, beforelast2 = last2;
1452 last1 = i1, last2 = i2;
1453 ninsns++;
1456 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1457 || (stop_after > 0 && ninsns == stop_after))
1458 break;
1460 i1 = NEXT_INSN (i1);
1461 i2 = NEXT_INSN (i2);
1464 #ifdef HAVE_cc0
1465 /* Don't allow a compare to be shared by cross-jumping unless the insn
1466 after the compare is also shared. */
1467 if (ninsns && reg_mentioned_p (cc0_rtx, last1) && sets_cc0_p (last1))
1468 last1 = beforelast1, last2 = beforelast2, ninsns--;
1469 #endif
1471 if (ninsns)
1473 *f1 = last1;
1474 *f2 = last2;
1477 return ninsns;
1480 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1481 the branch instruction. This means that if we commonize the control
1482 flow before end of the basic block, the semantic remains unchanged.
1484 We may assume that there exists one edge with a common destination. */
1486 static bool
1487 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1489 int nehedges1 = 0, nehedges2 = 0;
1490 edge fallthru1 = 0, fallthru2 = 0;
1491 edge e1, e2;
1492 edge_iterator ei;
1494 /* If we performed shrink-wrapping, edges to the EXIT_BLOCK_PTR can
1495 only be distinguished for JUMP_INSNs. The two paths may differ in
1496 whether they went through the prologue. Sibcalls are fine, we know
1497 that we either didn't need or inserted an epilogue before them. */
1498 if (crtl->shrink_wrapped
1499 && single_succ_p (bb1) && single_succ (bb1) == EXIT_BLOCK_PTR
1500 && !JUMP_P (BB_END (bb1))
1501 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1502 return false;
1504 /* If BB1 has only one successor, we may be looking at either an
1505 unconditional jump, or a fake edge to exit. */
1506 if (single_succ_p (bb1)
1507 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1508 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1509 return (single_succ_p (bb2)
1510 && (single_succ_edge (bb2)->flags
1511 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1512 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1514 /* Match conditional jumps - this may get tricky when fallthru and branch
1515 edges are crossed. */
1516 if (EDGE_COUNT (bb1->succs) == 2
1517 && any_condjump_p (BB_END (bb1))
1518 && onlyjump_p (BB_END (bb1)))
1520 edge b1, f1, b2, f2;
1521 bool reverse, match;
1522 rtx set1, set2, cond1, cond2;
1523 enum rtx_code code1, code2;
1525 if (EDGE_COUNT (bb2->succs) != 2
1526 || !any_condjump_p (BB_END (bb2))
1527 || !onlyjump_p (BB_END (bb2)))
1528 return false;
1530 b1 = BRANCH_EDGE (bb1);
1531 b2 = BRANCH_EDGE (bb2);
1532 f1 = FALLTHRU_EDGE (bb1);
1533 f2 = FALLTHRU_EDGE (bb2);
1535 /* Get around possible forwarders on fallthru edges. Other cases
1536 should be optimized out already. */
1537 if (FORWARDER_BLOCK_P (f1->dest))
1538 f1 = single_succ_edge (f1->dest);
1540 if (FORWARDER_BLOCK_P (f2->dest))
1541 f2 = single_succ_edge (f2->dest);
1543 /* To simplify use of this function, return false if there are
1544 unneeded forwarder blocks. These will get eliminated later
1545 during cleanup_cfg. */
1546 if (FORWARDER_BLOCK_P (f1->dest)
1547 || FORWARDER_BLOCK_P (f2->dest)
1548 || FORWARDER_BLOCK_P (b1->dest)
1549 || FORWARDER_BLOCK_P (b2->dest))
1550 return false;
1552 if (f1->dest == f2->dest && b1->dest == b2->dest)
1553 reverse = false;
1554 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1555 reverse = true;
1556 else
1557 return false;
1559 set1 = pc_set (BB_END (bb1));
1560 set2 = pc_set (BB_END (bb2));
1561 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1562 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1563 reverse = !reverse;
1565 cond1 = XEXP (SET_SRC (set1), 0);
1566 cond2 = XEXP (SET_SRC (set2), 0);
1567 code1 = GET_CODE (cond1);
1568 if (reverse)
1569 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1570 else
1571 code2 = GET_CODE (cond2);
1573 if (code2 == UNKNOWN)
1574 return false;
1576 /* Verify codes and operands match. */
1577 match = ((code1 == code2
1578 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1579 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1580 || (code1 == swap_condition (code2)
1581 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1582 XEXP (cond2, 0))
1583 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1584 XEXP (cond2, 1))));
1586 /* If we return true, we will join the blocks. Which means that
1587 we will only have one branch prediction bit to work with. Thus
1588 we require the existing branches to have probabilities that are
1589 roughly similar. */
1590 if (match
1591 && optimize_bb_for_speed_p (bb1)
1592 && optimize_bb_for_speed_p (bb2))
1594 int prob2;
1596 if (b1->dest == b2->dest)
1597 prob2 = b2->probability;
1598 else
1599 /* Do not use f2 probability as f2 may be forwarded. */
1600 prob2 = REG_BR_PROB_BASE - b2->probability;
1602 /* Fail if the difference in probabilities is greater than 50%.
1603 This rules out two well-predicted branches with opposite
1604 outcomes. */
1605 if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
1607 if (dump_file)
1608 fprintf (dump_file,
1609 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1610 bb1->index, bb2->index, b1->probability, prob2);
1612 return false;
1616 if (dump_file && match)
1617 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1618 bb1->index, bb2->index);
1620 return match;
1623 /* Generic case - we are seeing a computed jump, table jump or trapping
1624 instruction. */
1626 /* Check whether there are tablejumps in the end of BB1 and BB2.
1627 Return true if they are identical. */
1629 rtx label1, label2;
1630 rtx table1, table2;
1632 if (tablejump_p (BB_END (bb1), &label1, &table1)
1633 && tablejump_p (BB_END (bb2), &label2, &table2)
1634 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1636 /* The labels should never be the same rtx. If they really are same
1637 the jump tables are same too. So disable crossjumping of blocks BB1
1638 and BB2 because when deleting the common insns in the end of BB1
1639 by delete_basic_block () the jump table would be deleted too. */
1640 /* If LABEL2 is referenced in BB1->END do not do anything
1641 because we would loose information when replacing
1642 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1643 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1645 /* Set IDENTICAL to true when the tables are identical. */
1646 bool identical = false;
1647 rtx p1, p2;
1649 p1 = PATTERN (table1);
1650 p2 = PATTERN (table2);
1651 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1653 identical = true;
1655 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1656 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1657 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1658 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1660 int i;
1662 identical = true;
1663 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1664 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1665 identical = false;
1668 if (identical)
1670 replace_label_data rr;
1671 bool match;
1673 /* Temporarily replace references to LABEL1 with LABEL2
1674 in BB1->END so that we could compare the instructions. */
1675 rr.r1 = label1;
1676 rr.r2 = label2;
1677 rr.update_label_nuses = false;
1678 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1680 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1681 == dir_both);
1682 if (dump_file && match)
1683 fprintf (dump_file,
1684 "Tablejumps in bb %i and %i match.\n",
1685 bb1->index, bb2->index);
1687 /* Set the original label in BB1->END because when deleting
1688 a block whose end is a tablejump, the tablejump referenced
1689 from the instruction is deleted too. */
1690 rr.r1 = label2;
1691 rr.r2 = label1;
1692 for_each_rtx (&BB_END (bb1), replace_label, &rr);
1694 return match;
1697 return false;
1701 /* First ensure that the instructions match. There may be many outgoing
1702 edges so this test is generally cheaper. */
1703 if (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2)) != dir_both)
1704 return false;
1706 /* Search the outgoing edges, ensure that the counts do match, find possible
1707 fallthru and exception handling edges since these needs more
1708 validation. */
1709 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1710 return false;
1712 FOR_EACH_EDGE (e1, ei, bb1->succs)
1714 e2 = EDGE_SUCC (bb2, ei.index);
1716 if (e1->flags & EDGE_EH)
1717 nehedges1++;
1719 if (e2->flags & EDGE_EH)
1720 nehedges2++;
1722 if (e1->flags & EDGE_FALLTHRU)
1723 fallthru1 = e1;
1724 if (e2->flags & EDGE_FALLTHRU)
1725 fallthru2 = e2;
1728 /* If number of edges of various types does not match, fail. */
1729 if (nehedges1 != nehedges2
1730 || (fallthru1 != 0) != (fallthru2 != 0))
1731 return false;
1733 /* fallthru edges must be forwarded to the same destination. */
1734 if (fallthru1)
1736 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1737 ? single_succ (fallthru1->dest): fallthru1->dest);
1738 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1739 ? single_succ (fallthru2->dest): fallthru2->dest);
1741 if (d1 != d2)
1742 return false;
1745 /* Ensure the same EH region. */
1747 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1748 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1750 if (!n1 && n2)
1751 return false;
1753 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1754 return false;
1757 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1758 version of sequence abstraction. */
1759 FOR_EACH_EDGE (e1, ei, bb2->succs)
1761 edge e2;
1762 edge_iterator ei;
1763 basic_block d1 = e1->dest;
1765 if (FORWARDER_BLOCK_P (d1))
1766 d1 = EDGE_SUCC (d1, 0)->dest;
1768 FOR_EACH_EDGE (e2, ei, bb1->succs)
1770 basic_block d2 = e2->dest;
1771 if (FORWARDER_BLOCK_P (d2))
1772 d2 = EDGE_SUCC (d2, 0)->dest;
1773 if (d1 == d2)
1774 break;
1777 if (!e2)
1778 return false;
1781 return true;
1784 /* Returns true if BB basic block has a preserve label. */
1786 static bool
1787 block_has_preserve_label (basic_block bb)
1789 return (bb
1790 && block_label (bb)
1791 && LABEL_PRESERVE_P (block_label (bb)));
1794 /* E1 and E2 are edges with the same destination block. Search their
1795 predecessors for common code. If found, redirect control flow from
1796 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1797 or the other way around (dir_backward). DIR specifies the allowed
1798 replacement direction. */
1800 static bool
1801 try_crossjump_to_edge (int mode, edge e1, edge e2,
1802 enum replace_direction dir)
1804 int nmatch;
1805 basic_block src1 = e1->src, src2 = e2->src;
1806 basic_block redirect_to, redirect_from, to_remove;
1807 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1808 rtx newpos1, newpos2;
1809 edge s;
1810 edge_iterator ei;
1812 newpos1 = newpos2 = NULL_RTX;
1814 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1815 to try this optimization.
1817 Basic block partitioning may result in some jumps that appear to
1818 be optimizable (or blocks that appear to be mergeable), but which really
1819 must be left untouched (they are required to make it safely across
1820 partition boundaries). See the comments at the top of
1821 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1823 if (flag_reorder_blocks_and_partition && reload_completed)
1824 return false;
1826 /* Search backward through forwarder blocks. We don't need to worry
1827 about multiple entry or chained forwarders, as they will be optimized
1828 away. We do this to look past the unconditional jump following a
1829 conditional jump that is required due to the current CFG shape. */
1830 if (single_pred_p (src1)
1831 && FORWARDER_BLOCK_P (src1))
1832 e1 = single_pred_edge (src1), src1 = e1->src;
1834 if (single_pred_p (src2)
1835 && FORWARDER_BLOCK_P (src2))
1836 e2 = single_pred_edge (src2), src2 = e2->src;
1838 /* Nothing to do if we reach ENTRY, or a common source block. */
1839 if (src1 == ENTRY_BLOCK_PTR || src2 == ENTRY_BLOCK_PTR)
1840 return false;
1841 if (src1 == src2)
1842 return false;
1844 /* Seeing more than 1 forwarder blocks would confuse us later... */
1845 if (FORWARDER_BLOCK_P (e1->dest)
1846 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1847 return false;
1849 if (FORWARDER_BLOCK_P (e2->dest)
1850 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1851 return false;
1853 /* Likewise with dead code (possibly newly created by the other optimizations
1854 of cfg_cleanup). */
1855 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1856 return false;
1858 /* Look for the common insn sequence, part the first ... */
1859 if (!outgoing_edges_match (mode, src1, src2))
1860 return false;
1862 /* ... and part the second. */
1863 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
1865 osrc1 = src1;
1866 osrc2 = src2;
1867 if (newpos1 != NULL_RTX)
1868 src1 = BLOCK_FOR_INSN (newpos1);
1869 if (newpos2 != NULL_RTX)
1870 src2 = BLOCK_FOR_INSN (newpos2);
1872 if (dir == dir_backward)
1874 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
1875 SWAP (basic_block, osrc1, osrc2);
1876 SWAP (basic_block, src1, src2);
1877 SWAP (edge, e1, e2);
1878 SWAP (rtx, newpos1, newpos2);
1879 #undef SWAP
1882 /* Don't proceed with the crossjump unless we found a sufficient number
1883 of matching instructions or the 'from' block was totally matched
1884 (such that its predecessors will hopefully be redirected and the
1885 block removed). */
1886 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
1887 && (newpos1 != BB_HEAD (src1)))
1888 return false;
1890 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
1891 if (block_has_preserve_label (e1->dest)
1892 && (e1->flags & EDGE_ABNORMAL))
1893 return false;
1895 /* Here we know that the insns in the end of SRC1 which are common with SRC2
1896 will be deleted.
1897 If we have tablejumps in the end of SRC1 and SRC2
1898 they have been already compared for equivalence in outgoing_edges_match ()
1899 so replace the references to TABLE1 by references to TABLE2. */
1901 rtx label1, label2;
1902 rtx table1, table2;
1904 if (tablejump_p (BB_END (osrc1), &label1, &table1)
1905 && tablejump_p (BB_END (osrc2), &label2, &table2)
1906 && label1 != label2)
1908 replace_label_data rr;
1909 rtx insn;
1911 /* Replace references to LABEL1 with LABEL2. */
1912 rr.r1 = label1;
1913 rr.r2 = label2;
1914 rr.update_label_nuses = true;
1915 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
1917 /* Do not replace the label in SRC1->END because when deleting
1918 a block whose end is a tablejump, the tablejump referenced
1919 from the instruction is deleted too. */
1920 if (insn != BB_END (osrc1))
1921 for_each_rtx (&insn, replace_label, &rr);
1926 /* Avoid splitting if possible. We must always split when SRC2 has
1927 EH predecessor edges, or we may end up with basic blocks with both
1928 normal and EH predecessor edges. */
1929 if (newpos2 == BB_HEAD (src2)
1930 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
1931 redirect_to = src2;
1932 else
1934 if (newpos2 == BB_HEAD (src2))
1936 /* Skip possible basic block header. */
1937 if (LABEL_P (newpos2))
1938 newpos2 = NEXT_INSN (newpos2);
1939 while (DEBUG_INSN_P (newpos2))
1940 newpos2 = NEXT_INSN (newpos2);
1941 if (NOTE_P (newpos2))
1942 newpos2 = NEXT_INSN (newpos2);
1943 while (DEBUG_INSN_P (newpos2))
1944 newpos2 = NEXT_INSN (newpos2);
1947 if (dump_file)
1948 fprintf (dump_file, "Splitting bb %i before %i insns\n",
1949 src2->index, nmatch);
1950 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
1953 if (dump_file)
1954 fprintf (dump_file,
1955 "Cross jumping from bb %i to bb %i; %i common insns\n",
1956 src1->index, src2->index, nmatch);
1958 /* We may have some registers visible through the block. */
1959 df_set_bb_dirty (redirect_to);
1961 if (osrc2 == src2)
1962 redirect_edges_to = redirect_to;
1963 else
1964 redirect_edges_to = osrc2;
1966 /* Recompute the frequencies and counts of outgoing edges. */
1967 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
1969 edge s2;
1970 edge_iterator ei;
1971 basic_block d = s->dest;
1973 if (FORWARDER_BLOCK_P (d))
1974 d = single_succ (d);
1976 FOR_EACH_EDGE (s2, ei, src1->succs)
1978 basic_block d2 = s2->dest;
1979 if (FORWARDER_BLOCK_P (d2))
1980 d2 = single_succ (d2);
1981 if (d == d2)
1982 break;
1985 s->count += s2->count;
1987 /* Take care to update possible forwarder blocks. We verified
1988 that there is no more than one in the chain, so we can't run
1989 into infinite loop. */
1990 if (FORWARDER_BLOCK_P (s->dest))
1992 single_succ_edge (s->dest)->count += s2->count;
1993 s->dest->count += s2->count;
1994 s->dest->frequency += EDGE_FREQUENCY (s);
1997 if (FORWARDER_BLOCK_P (s2->dest))
1999 single_succ_edge (s2->dest)->count -= s2->count;
2000 if (single_succ_edge (s2->dest)->count < 0)
2001 single_succ_edge (s2->dest)->count = 0;
2002 s2->dest->count -= s2->count;
2003 s2->dest->frequency -= EDGE_FREQUENCY (s);
2004 if (s2->dest->frequency < 0)
2005 s2->dest->frequency = 0;
2006 if (s2->dest->count < 0)
2007 s2->dest->count = 0;
2010 if (!redirect_edges_to->frequency && !src1->frequency)
2011 s->probability = (s->probability + s2->probability) / 2;
2012 else
2013 s->probability
2014 = ((s->probability * redirect_edges_to->frequency +
2015 s2->probability * src1->frequency)
2016 / (redirect_edges_to->frequency + src1->frequency));
2019 /* Adjust count and frequency for the block. An earlier jump
2020 threading pass may have left the profile in an inconsistent
2021 state (see update_bb_profile_for_threading) so we must be
2022 prepared for overflows. */
2023 tmp = redirect_to;
2026 tmp->count += src1->count;
2027 tmp->frequency += src1->frequency;
2028 if (tmp->frequency > BB_FREQ_MAX)
2029 tmp->frequency = BB_FREQ_MAX;
2030 if (tmp == redirect_edges_to)
2031 break;
2032 tmp = find_fallthru_edge (tmp->succs)->dest;
2034 while (true);
2035 update_br_prob_note (redirect_edges_to);
2037 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2039 /* Skip possible basic block header. */
2040 if (LABEL_P (newpos1))
2041 newpos1 = NEXT_INSN (newpos1);
2043 while (DEBUG_INSN_P (newpos1))
2044 newpos1 = NEXT_INSN (newpos1);
2046 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2047 newpos1 = NEXT_INSN (newpos1);
2049 while (DEBUG_INSN_P (newpos1))
2050 newpos1 = NEXT_INSN (newpos1);
2052 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2053 to_remove = single_succ (redirect_from);
2055 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2056 delete_basic_block (to_remove);
2058 update_forwarder_flag (redirect_from);
2059 if (redirect_to != src2)
2060 update_forwarder_flag (src2);
2062 return true;
2065 /* Search the predecessors of BB for common insn sequences. When found,
2066 share code between them by redirecting control flow. Return true if
2067 any changes made. */
2069 static bool
2070 try_crossjump_bb (int mode, basic_block bb)
2072 edge e, e2, fallthru;
2073 bool changed;
2074 unsigned max, ix, ix2;
2076 /* Nothing to do if there is not at least two incoming edges. */
2077 if (EDGE_COUNT (bb->preds) < 2)
2078 return false;
2080 /* Don't crossjump if this block ends in a computed jump,
2081 unless we are optimizing for size. */
2082 if (optimize_bb_for_size_p (bb)
2083 && bb != EXIT_BLOCK_PTR
2084 && computed_jump_p (BB_END (bb)))
2085 return false;
2087 /* If we are partitioning hot/cold basic blocks, we don't want to
2088 mess up unconditional or indirect jumps that cross between hot
2089 and cold sections.
2091 Basic block partitioning may result in some jumps that appear to
2092 be optimizable (or blocks that appear to be mergeable), but which really
2093 must be left untouched (they are required to make it safely across
2094 partition boundaries). See the comments at the top of
2095 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2097 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2098 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2099 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2100 return false;
2102 /* It is always cheapest to redirect a block that ends in a branch to
2103 a block that falls through into BB, as that adds no branches to the
2104 program. We'll try that combination first. */
2105 fallthru = NULL;
2106 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2108 if (EDGE_COUNT (bb->preds) > max)
2109 return false;
2111 fallthru = find_fallthru_edge (bb->preds);
2113 changed = false;
2114 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2116 e = EDGE_PRED (bb, ix);
2117 ix++;
2119 /* As noted above, first try with the fallthru predecessor (or, a
2120 fallthru predecessor if we are in cfglayout mode). */
2121 if (fallthru)
2123 /* Don't combine the fallthru edge into anything else.
2124 If there is a match, we'll do it the other way around. */
2125 if (e == fallthru)
2126 continue;
2127 /* If nothing changed since the last attempt, there is nothing
2128 we can do. */
2129 if (!first_pass
2130 && !((e->src->flags & BB_MODIFIED)
2131 || (fallthru->src->flags & BB_MODIFIED)))
2132 continue;
2134 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2136 changed = true;
2137 ix = 0;
2138 continue;
2142 /* Non-obvious work limiting check: Recognize that we're going
2143 to call try_crossjump_bb on every basic block. So if we have
2144 two blocks with lots of outgoing edges (a switch) and they
2145 share lots of common destinations, then we would do the
2146 cross-jump check once for each common destination.
2148 Now, if the blocks actually are cross-jump candidates, then
2149 all of their destinations will be shared. Which means that
2150 we only need check them for cross-jump candidacy once. We
2151 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2152 choosing to do the check from the block for which the edge
2153 in question is the first successor of A. */
2154 if (EDGE_SUCC (e->src, 0) != e)
2155 continue;
2157 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2159 e2 = EDGE_PRED (bb, ix2);
2161 if (e2 == e)
2162 continue;
2164 /* We've already checked the fallthru edge above. */
2165 if (e2 == fallthru)
2166 continue;
2168 /* The "first successor" check above only prevents multiple
2169 checks of crossjump(A,B). In order to prevent redundant
2170 checks of crossjump(B,A), require that A be the block
2171 with the lowest index. */
2172 if (e->src->index > e2->src->index)
2173 continue;
2175 /* If nothing changed since the last attempt, there is nothing
2176 we can do. */
2177 if (!first_pass
2178 && !((e->src->flags & BB_MODIFIED)
2179 || (e2->src->flags & BB_MODIFIED)))
2180 continue;
2182 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2183 direction. */
2184 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2186 changed = true;
2187 ix = 0;
2188 break;
2193 if (changed)
2194 crossjumps_occured = true;
2196 return changed;
2199 /* Search the successors of BB for common insn sequences. When found,
2200 share code between them by moving it across the basic block
2201 boundary. Return true if any changes made. */
2203 static bool
2204 try_head_merge_bb (basic_block bb)
2206 basic_block final_dest_bb = NULL;
2207 int max_match = INT_MAX;
2208 edge e0;
2209 rtx *headptr, *currptr, *nextptr;
2210 bool changed, moveall;
2211 unsigned ix;
2212 rtx e0_last_head, cond, move_before;
2213 unsigned nedges = EDGE_COUNT (bb->succs);
2214 rtx jump = BB_END (bb);
2215 regset live, live_union;
2217 /* Nothing to do if there is not at least two outgoing edges. */
2218 if (nedges < 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
2225 && computed_jump_p (BB_END (bb)))
2226 return false;
2228 cond = get_condition (jump, &move_before, true, false);
2229 if (cond == NULL_RTX)
2231 #ifdef HAVE_cc0
2232 if (reg_mentioned_p (cc0_rtx, jump))
2233 move_before = prev_nonnote_nondebug_insn (jump);
2234 else
2235 #endif
2236 move_before = jump;
2239 for (ix = 0; ix < nedges; ix++)
2240 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR)
2241 return false;
2243 for (ix = 0; ix < nedges; ix++)
2245 edge e = EDGE_SUCC (bb, ix);
2246 basic_block other_bb = e->dest;
2248 if (df_get_bb_dirty (other_bb))
2250 block_was_dirty = true;
2251 return false;
2254 if (e->flags & EDGE_ABNORMAL)
2255 return false;
2257 /* Normally, all destination blocks must only be reachable from this
2258 block, i.e. they must have one incoming edge.
2260 There is one special case we can handle, that of multiple consecutive
2261 jumps where the first jumps to one of the targets of the second jump.
2262 This happens frequently in switch statements for default labels.
2263 The structure is as follows:
2264 FINAL_DEST_BB
2265 ....
2266 if (cond) jump A;
2267 fall through
2269 jump with targets A, B, C, D...
2271 has two incoming edges, from FINAL_DEST_BB and BB
2273 In this case, we can try to move the insns through BB and into
2274 FINAL_DEST_BB. */
2275 if (EDGE_COUNT (other_bb->preds) != 1)
2277 edge incoming_edge, incoming_bb_other_edge;
2278 edge_iterator ei;
2280 if (final_dest_bb != NULL
2281 || EDGE_COUNT (other_bb->preds) != 2)
2282 return false;
2284 /* We must be able to move the insns across the whole block. */
2285 move_before = BB_HEAD (bb);
2286 while (!NONDEBUG_INSN_P (move_before))
2287 move_before = NEXT_INSN (move_before);
2289 if (EDGE_COUNT (bb->preds) != 1)
2290 return false;
2291 incoming_edge = EDGE_PRED (bb, 0);
2292 final_dest_bb = incoming_edge->src;
2293 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2294 return false;
2295 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2296 if (incoming_bb_other_edge != incoming_edge)
2297 break;
2298 if (incoming_bb_other_edge->dest != other_bb)
2299 return false;
2303 e0 = EDGE_SUCC (bb, 0);
2304 e0_last_head = NULL_RTX;
2305 changed = false;
2307 for (ix = 1; ix < nedges; ix++)
2309 edge e = EDGE_SUCC (bb, ix);
2310 rtx e0_last, e_last;
2311 int nmatch;
2313 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2314 &e0_last, &e_last, 0);
2315 if (nmatch == 0)
2316 return false;
2318 if (nmatch < max_match)
2320 max_match = nmatch;
2321 e0_last_head = e0_last;
2325 /* If we matched an entire block, we probably have to avoid moving the
2326 last insn. */
2327 if (max_match > 0
2328 && e0_last_head == BB_END (e0->dest)
2329 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2330 || control_flow_insn_p (e0_last_head)))
2332 max_match--;
2333 if (max_match == 0)
2334 return false;
2336 e0_last_head = prev_real_insn (e0_last_head);
2337 while (DEBUG_INSN_P (e0_last_head));
2340 if (max_match == 0)
2341 return false;
2343 /* We must find a union of the live registers at each of the end points. */
2344 live = BITMAP_ALLOC (NULL);
2345 live_union = BITMAP_ALLOC (NULL);
2347 currptr = XNEWVEC (rtx, nedges);
2348 headptr = XNEWVEC (rtx, nedges);
2349 nextptr = XNEWVEC (rtx, nedges);
2351 for (ix = 0; ix < nedges; ix++)
2353 int j;
2354 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2355 rtx head = BB_HEAD (merge_bb);
2357 while (!NONDEBUG_INSN_P (head))
2358 head = NEXT_INSN (head);
2359 headptr[ix] = head;
2360 currptr[ix] = head;
2362 /* Compute the end point and live information */
2363 for (j = 1; j < max_match; j++)
2365 head = NEXT_INSN (head);
2366 while (!NONDEBUG_INSN_P (head));
2367 simulate_backwards_to_point (merge_bb, live, head);
2368 IOR_REG_SET (live_union, live);
2371 /* If we're moving across two blocks, verify the validity of the
2372 first move, then adjust the target and let the loop below deal
2373 with the final move. */
2374 if (final_dest_bb != NULL)
2376 rtx move_upto;
2378 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2379 jump, e0->dest, live_union,
2380 NULL, &move_upto);
2381 if (!moveall)
2383 if (move_upto == NULL_RTX)
2384 goto out;
2386 while (e0_last_head != move_upto)
2388 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2389 live_union);
2390 e0_last_head = PREV_INSN (e0_last_head);
2393 if (e0_last_head == NULL_RTX)
2394 goto out;
2396 jump = BB_END (final_dest_bb);
2397 cond = get_condition (jump, &move_before, true, false);
2398 if (cond == NULL_RTX)
2400 #ifdef HAVE_cc0
2401 if (reg_mentioned_p (cc0_rtx, jump))
2402 move_before = prev_nonnote_nondebug_insn (jump);
2403 else
2404 #endif
2405 move_before = jump;
2411 rtx move_upto;
2412 moveall = can_move_insns_across (currptr[0], e0_last_head,
2413 move_before, jump, e0->dest, live_union,
2414 NULL, &move_upto);
2415 if (!moveall && move_upto == NULL_RTX)
2417 if (jump == move_before)
2418 break;
2420 /* Try again, using a different insertion point. */
2421 move_before = jump;
2423 #ifdef HAVE_cc0
2424 /* Don't try moving before a cc0 user, as that may invalidate
2425 the cc0. */
2426 if (reg_mentioned_p (cc0_rtx, jump))
2427 break;
2428 #endif
2430 continue;
2433 if (final_dest_bb && !moveall)
2434 /* We haven't checked whether a partial move would be OK for the first
2435 move, so we have to fail this case. */
2436 break;
2438 changed = true;
2439 for (;;)
2441 if (currptr[0] == move_upto)
2442 break;
2443 for (ix = 0; ix < nedges; ix++)
2445 rtx curr = currptr[ix];
2447 curr = NEXT_INSN (curr);
2448 while (!NONDEBUG_INSN_P (curr));
2449 currptr[ix] = curr;
2453 /* If we can't currently move all of the identical insns, remember
2454 each insn after the range that we'll merge. */
2455 if (!moveall)
2456 for (ix = 0; ix < nedges; ix++)
2458 rtx curr = currptr[ix];
2460 curr = NEXT_INSN (curr);
2461 while (!NONDEBUG_INSN_P (curr));
2462 nextptr[ix] = curr;
2465 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2466 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2467 if (final_dest_bb != NULL)
2468 df_set_bb_dirty (final_dest_bb);
2469 df_set_bb_dirty (bb);
2470 for (ix = 1; ix < nedges; ix++)
2472 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2473 delete_insn_chain (headptr[ix], currptr[ix], false);
2475 if (!moveall)
2477 if (jump == move_before)
2478 break;
2480 /* For the unmerged insns, try a different insertion point. */
2481 move_before = jump;
2483 #ifdef HAVE_cc0
2484 /* Don't try moving before a cc0 user, as that may invalidate
2485 the cc0. */
2486 if (reg_mentioned_p (cc0_rtx, jump))
2487 break;
2488 #endif
2490 for (ix = 0; ix < nedges; ix++)
2491 currptr[ix] = headptr[ix] = nextptr[ix];
2494 while (!moveall);
2496 out:
2497 free (currptr);
2498 free (headptr);
2499 free (nextptr);
2501 crossjumps_occured |= changed;
2503 return changed;
2506 /* Return true if BB contains just bb note, or bb note followed
2507 by only DEBUG_INSNs. */
2509 static bool
2510 trivially_empty_bb_p (basic_block bb)
2512 rtx insn = BB_END (bb);
2514 while (1)
2516 if (insn == BB_HEAD (bb))
2517 return true;
2518 if (!DEBUG_INSN_P (insn))
2519 return false;
2520 insn = PREV_INSN (insn);
2524 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2525 instructions etc. Return nonzero if changes were made. */
2527 static bool
2528 try_optimize_cfg (int mode)
2530 bool changed_overall = false;
2531 bool changed;
2532 int iterations = 0;
2533 basic_block bb, b, next;
2535 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2536 clear_bb_flags ();
2538 crossjumps_occured = false;
2540 FOR_EACH_BB (bb)
2541 update_forwarder_flag (bb);
2543 if (! targetm.cannot_modify_jumps_p ())
2545 first_pass = true;
2546 /* Attempt to merge blocks as made possible by edge removal. If
2547 a block has only one successor, and the successor has only
2548 one predecessor, they may be combined. */
2551 block_was_dirty = false;
2552 changed = false;
2553 iterations++;
2555 if (dump_file)
2556 fprintf (dump_file,
2557 "\n\ntry_optimize_cfg iteration %i\n\n",
2558 iterations);
2560 for (b = ENTRY_BLOCK_PTR->next_bb; b != EXIT_BLOCK_PTR;)
2562 basic_block c;
2563 edge s;
2564 bool changed_here = false;
2566 /* Delete trivially dead basic blocks. This is either
2567 blocks with no predecessors, or empty blocks with no
2568 successors. However if the empty block with no
2569 successors is the successor of the ENTRY_BLOCK, it is
2570 kept. This ensures that the ENTRY_BLOCK will have a
2571 successor which is a precondition for many RTL
2572 passes. Empty blocks may result from expanding
2573 __builtin_unreachable (). */
2574 if (EDGE_COUNT (b->preds) == 0
2575 || (EDGE_COUNT (b->succs) == 0
2576 && trivially_empty_bb_p (b)
2577 && single_succ_edge (ENTRY_BLOCK_PTR)->dest != b))
2579 c = b->prev_bb;
2580 if (EDGE_COUNT (b->preds) > 0)
2582 edge e;
2583 edge_iterator ei;
2585 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2587 if (BB_FOOTER (b)
2588 && BARRIER_P (BB_FOOTER (b)))
2589 FOR_EACH_EDGE (e, ei, b->preds)
2590 if ((e->flags & EDGE_FALLTHRU)
2591 && BB_FOOTER (e->src) == NULL)
2593 if (BB_FOOTER (b))
2595 BB_FOOTER (e->src) = BB_FOOTER (b);
2596 BB_FOOTER (b) = NULL;
2598 else
2600 start_sequence ();
2601 BB_FOOTER (e->src) = emit_barrier ();
2602 end_sequence ();
2606 else
2608 rtx last = get_last_bb_insn (b);
2609 if (last && BARRIER_P (last))
2610 FOR_EACH_EDGE (e, ei, b->preds)
2611 if ((e->flags & EDGE_FALLTHRU))
2612 emit_barrier_after (BB_END (e->src));
2615 delete_basic_block (b);
2616 changed = true;
2617 /* Avoid trying to remove ENTRY_BLOCK_PTR. */
2618 b = (c == ENTRY_BLOCK_PTR ? c->next_bb : c);
2619 continue;
2622 /* Remove code labels no longer used. */
2623 if (single_pred_p (b)
2624 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2625 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2626 && LABEL_P (BB_HEAD (b))
2627 /* If the previous block ends with a branch to this
2628 block, we can't delete the label. Normally this
2629 is a condjump that is yet to be simplified, but
2630 if CASE_DROPS_THRU, this can be a tablejump with
2631 some element going to the same place as the
2632 default (fallthru). */
2633 && (single_pred (b) == ENTRY_BLOCK_PTR
2634 || !JUMP_P (BB_END (single_pred (b)))
2635 || ! label_is_jump_target_p (BB_HEAD (b),
2636 BB_END (single_pred (b)))))
2638 delete_insn (BB_HEAD (b));
2639 if (dump_file)
2640 fprintf (dump_file, "Deleted label in block %i.\n",
2641 b->index);
2644 /* If we fall through an empty block, we can remove it. */
2645 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2646 && single_pred_p (b)
2647 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2648 && !LABEL_P (BB_HEAD (b))
2649 && FORWARDER_BLOCK_P (b)
2650 /* Note that forwarder_block_p true ensures that
2651 there is a successor for this block. */
2652 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2653 && n_basic_blocks > NUM_FIXED_BLOCKS + 1)
2655 if (dump_file)
2656 fprintf (dump_file,
2657 "Deleting fallthru block %i.\n",
2658 b->index);
2660 c = b->prev_bb == ENTRY_BLOCK_PTR ? b->next_bb : b->prev_bb;
2661 redirect_edge_succ_nodup (single_pred_edge (b),
2662 single_succ (b));
2663 delete_basic_block (b);
2664 changed = true;
2665 b = c;
2666 continue;
2669 /* Merge B with its single successor, if any. */
2670 if (single_succ_p (b)
2671 && (s = single_succ_edge (b))
2672 && !(s->flags & EDGE_COMPLEX)
2673 && (c = s->dest) != EXIT_BLOCK_PTR
2674 && single_pred_p (c)
2675 && b != c)
2677 /* When not in cfg_layout mode use code aware of reordering
2678 INSN. This code possibly creates new basic blocks so it
2679 does not fit merge_blocks interface and is kept here in
2680 hope that it will become useless once more of compiler
2681 is transformed to use cfg_layout mode. */
2683 if ((mode & CLEANUP_CFGLAYOUT)
2684 && can_merge_blocks_p (b, c))
2686 merge_blocks (b, c);
2687 update_forwarder_flag (b);
2688 changed_here = true;
2690 else if (!(mode & CLEANUP_CFGLAYOUT)
2691 /* If the jump insn has side effects,
2692 we can't kill the edge. */
2693 && (!JUMP_P (BB_END (b))
2694 || (reload_completed
2695 ? simplejump_p (BB_END (b))
2696 : (onlyjump_p (BB_END (b))
2697 && !tablejump_p (BB_END (b),
2698 NULL, NULL))))
2699 && (next = merge_blocks_move (s, b, c, mode)))
2701 b = next;
2702 changed_here = true;
2706 /* Simplify branch over branch. */
2707 if ((mode & CLEANUP_EXPENSIVE)
2708 && !(mode & CLEANUP_CFGLAYOUT)
2709 && try_simplify_condjump (b))
2710 changed_here = true;
2712 /* If B has a single outgoing edge, but uses a
2713 non-trivial jump instruction without side-effects, we
2714 can either delete the jump entirely, or replace it
2715 with a simple unconditional jump. */
2716 if (single_succ_p (b)
2717 && single_succ (b) != EXIT_BLOCK_PTR
2718 && onlyjump_p (BB_END (b))
2719 && !find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX)
2720 && try_redirect_by_replacing_jump (single_succ_edge (b),
2721 single_succ (b),
2722 (mode & CLEANUP_CFGLAYOUT) != 0))
2724 update_forwarder_flag (b);
2725 changed_here = true;
2728 /* Simplify branch to branch. */
2729 if (try_forward_edges (mode, b))
2731 update_forwarder_flag (b);
2732 changed_here = true;
2735 /* Look for shared code between blocks. */
2736 if ((mode & CLEANUP_CROSSJUMP)
2737 && try_crossjump_bb (mode, b))
2738 changed_here = true;
2740 if ((mode & CLEANUP_CROSSJUMP)
2741 /* This can lengthen register lifetimes. Do it only after
2742 reload. */
2743 && reload_completed
2744 && try_head_merge_bb (b))
2745 changed_here = true;
2747 /* Don't get confused by the index shift caused by
2748 deleting blocks. */
2749 if (!changed_here)
2750 b = b->next_bb;
2751 else
2752 changed = true;
2755 if ((mode & CLEANUP_CROSSJUMP)
2756 && try_crossjump_bb (mode, EXIT_BLOCK_PTR))
2757 changed = true;
2759 if (block_was_dirty)
2761 /* This should only be set by head-merging. */
2762 gcc_assert (mode & CLEANUP_CROSSJUMP);
2763 df_analyze ();
2766 #ifdef ENABLE_CHECKING
2767 if (changed)
2768 verify_flow_info ();
2769 #endif
2771 changed_overall |= changed;
2772 first_pass = false;
2774 while (changed);
2777 FOR_ALL_BB (b)
2778 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
2780 return changed_overall;
2783 /* Delete all unreachable basic blocks. */
2785 bool
2786 delete_unreachable_blocks (void)
2788 bool changed = false;
2789 basic_block b, prev_bb;
2791 find_unreachable_blocks ();
2793 /* When we're in GIMPLE mode and there may be debug insns, we should
2794 delete blocks in reverse dominator order, so as to get a chance
2795 to substitute all released DEFs into debug stmts. If we don't
2796 have dominators information, walking blocks backward gets us a
2797 better chance of retaining most debug information than
2798 otherwise. */
2799 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
2800 && dom_info_available_p (CDI_DOMINATORS))
2802 for (b = EXIT_BLOCK_PTR->prev_bb; b != ENTRY_BLOCK_PTR; b = prev_bb)
2804 prev_bb = b->prev_bb;
2806 if (!(b->flags & BB_REACHABLE))
2808 /* Speed up the removal of blocks that don't dominate
2809 others. Walking backwards, this should be the common
2810 case. */
2811 if (!first_dom_son (CDI_DOMINATORS, b))
2812 delete_basic_block (b);
2813 else
2815 VEC (basic_block, heap) *h
2816 = get_all_dominated_blocks (CDI_DOMINATORS, b);
2818 while (VEC_length (basic_block, h))
2820 b = VEC_pop (basic_block, h);
2822 prev_bb = b->prev_bb;
2824 gcc_assert (!(b->flags & BB_REACHABLE));
2826 delete_basic_block (b);
2829 VEC_free (basic_block, heap, h);
2832 changed = true;
2836 else
2838 for (b = EXIT_BLOCK_PTR->prev_bb; b != ENTRY_BLOCK_PTR; b = prev_bb)
2840 prev_bb = b->prev_bb;
2842 if (!(b->flags & BB_REACHABLE))
2844 delete_basic_block (b);
2845 changed = true;
2850 if (changed)
2851 tidy_fallthru_edges ();
2852 return changed;
2855 /* Delete any jump tables never referenced. We can't delete them at the
2856 time of removing tablejump insn as they are referenced by the preceding
2857 insns computing the destination, so we delay deleting and garbagecollect
2858 them once life information is computed. */
2859 void
2860 delete_dead_jumptables (void)
2862 basic_block bb;
2864 /* A dead jump table does not belong to any basic block. Scan insns
2865 between two adjacent basic blocks. */
2866 FOR_EACH_BB (bb)
2868 rtx insn, next;
2870 for (insn = NEXT_INSN (BB_END (bb));
2871 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
2872 insn = next)
2874 next = NEXT_INSN (insn);
2875 if (LABEL_P (insn)
2876 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
2877 && JUMP_TABLE_DATA_P (next))
2879 rtx label = insn, jump = next;
2881 if (dump_file)
2882 fprintf (dump_file, "Dead jumptable %i removed\n",
2883 INSN_UID (insn));
2885 next = NEXT_INSN (next);
2886 delete_insn (jump);
2887 delete_insn (label);
2894 /* Tidy the CFG by deleting unreachable code and whatnot. */
2896 bool
2897 cleanup_cfg (int mode)
2899 bool changed = false;
2901 /* Set the cfglayout mode flag here. We could update all the callers
2902 but that is just inconvenient, especially given that we eventually
2903 want to have cfglayout mode as the default. */
2904 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2905 mode |= CLEANUP_CFGLAYOUT;
2907 timevar_push (TV_CLEANUP_CFG);
2908 if (delete_unreachable_blocks ())
2910 changed = true;
2911 /* We've possibly created trivially dead code. Cleanup it right
2912 now to introduce more opportunities for try_optimize_cfg. */
2913 if (!(mode & (CLEANUP_NO_INSN_DEL))
2914 && !reload_completed)
2915 delete_trivially_dead_insns (get_insns (), max_reg_num ());
2918 compact_blocks ();
2920 /* To tail-merge blocks ending in the same noreturn function (e.g.
2921 a call to abort) we have to insert fake edges to exit. Do this
2922 here once. The fake edges do not interfere with any other CFG
2923 cleanups. */
2924 if (mode & CLEANUP_CROSSJUMP)
2925 add_noreturn_fake_exit_edges ();
2927 if (!dbg_cnt (cfg_cleanup))
2928 return changed;
2930 while (try_optimize_cfg (mode))
2932 delete_unreachable_blocks (), changed = true;
2933 if (!(mode & CLEANUP_NO_INSN_DEL))
2935 /* Try to remove some trivially dead insns when doing an expensive
2936 cleanup. But delete_trivially_dead_insns doesn't work after
2937 reload (it only handles pseudos) and run_fast_dce is too costly
2938 to run in every iteration.
2940 For effective cross jumping, we really want to run a fast DCE to
2941 clean up any dead conditions, or they get in the way of performing
2942 useful tail merges.
2944 Other transformations in cleanup_cfg are not so sensitive to dead
2945 code, so delete_trivially_dead_insns or even doing nothing at all
2946 is good enough. */
2947 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
2948 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
2949 break;
2950 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occured)
2951 run_fast_dce ();
2953 else
2954 break;
2957 if (mode & CLEANUP_CROSSJUMP)
2958 remove_fake_exit_edges ();
2960 /* Don't call delete_dead_jumptables in cfglayout mode, because
2961 that function assumes that jump tables are in the insns stream.
2962 But we also don't _have_ to delete dead jumptables in cfglayout
2963 mode because we shouldn't even be looking at things that are
2964 not in a basic block. Dead jumptables are cleaned up when
2965 going out of cfglayout mode. */
2966 if (!(mode & CLEANUP_CFGLAYOUT))
2967 delete_dead_jumptables ();
2969 /* ??? We probably do this way too often. */
2970 if (current_loops
2971 && (changed
2972 || (mode & CLEANUP_CFG_CHANGED)))
2974 bitmap changed_bbs;
2975 timevar_push (TV_REPAIR_LOOPS);
2976 /* The above doesn't preserve dominance info if available. */
2977 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
2978 calculate_dominance_info (CDI_DOMINATORS);
2979 changed_bbs = BITMAP_ALLOC (NULL);
2980 fix_loop_structure (changed_bbs);
2981 BITMAP_FREE (changed_bbs);
2982 free_dominance_info (CDI_DOMINATORS);
2983 timevar_pop (TV_REPAIR_LOOPS);
2986 timevar_pop (TV_CLEANUP_CFG);
2988 return changed;
2991 static unsigned int
2992 execute_jump (void)
2994 delete_trivially_dead_insns (get_insns (), max_reg_num ());
2995 if (dump_file)
2996 dump_flow_info (dump_file, dump_flags);
2997 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
2998 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
2999 return 0;
3002 struct rtl_opt_pass pass_jump =
3005 RTL_PASS,
3006 "jump", /* name */
3007 NULL, /* gate */
3008 execute_jump, /* execute */
3009 NULL, /* sub */
3010 NULL, /* next */
3011 0, /* static_pass_number */
3012 TV_JUMP, /* tv_id */
3013 0, /* properties_required */
3014 0, /* properties_provided */
3015 0, /* properties_destroyed */
3016 TODO_ggc_collect, /* todo_flags_start */
3017 TODO_verify_rtl_sharing, /* todo_flags_finish */
3021 static unsigned int
3022 execute_jump2 (void)
3024 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3025 return 0;
3028 struct rtl_opt_pass pass_jump2 =
3031 RTL_PASS,
3032 "jump2", /* name */
3033 NULL, /* gate */
3034 execute_jump2, /* execute */
3035 NULL, /* sub */
3036 NULL, /* next */
3037 0, /* static_pass_number */
3038 TV_JUMP, /* tv_id */
3039 0, /* properties_required */
3040 0, /* properties_provided */
3041 0, /* properties_destroyed */
3042 TODO_ggc_collect, /* todo_flags_start */
3043 TODO_verify_rtl_sharing, /* todo_flags_finish */