Fix warnings building linux-atomic.c and fptr.c on hppa64-linux
[official-gcc.git] / gcc / cfgcleanup.c
blob82fc505ff502d3eabd2492fa76d59e228a9f80f5
1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987-2021 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 "tree-pass.h"
47 #include "cfgloop.h"
48 #include "cfgrtl.h"
49 #include "cfganal.h"
50 #include "cfgbuild.h"
51 #include "cfgcleanup.h"
52 #include "dce.h"
53 #include "dbgcnt.h"
54 #include "rtl-iter.h"
55 #include "regs.h"
56 #include "function-abi.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_occurred;
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_insn *, rtx_insn *);
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 void merge_memattrs (rtx, rtx);
86 /* Set flags for newly created block. */
88 static void
89 notice_new_block (basic_block bb)
91 if (!bb)
92 return;
94 if (forwarder_block_p (bb))
95 bb->flags |= BB_FORWARDER_BLOCK;
98 /* Recompute forwarder flag after block has been modified. */
100 static void
101 update_forwarder_flag (basic_block bb)
103 if (forwarder_block_p (bb))
104 bb->flags |= BB_FORWARDER_BLOCK;
105 else
106 bb->flags &= ~BB_FORWARDER_BLOCK;
109 /* Simplify a conditional jump around an unconditional jump.
110 Return true if something changed. */
112 static bool
113 try_simplify_condjump (basic_block cbranch_block)
115 basic_block jump_block, jump_dest_block, cbranch_dest_block;
116 edge cbranch_jump_edge, cbranch_fallthru_edge;
117 rtx_insn *cbranch_insn;
119 /* Verify that there are exactly two successors. */
120 if (EDGE_COUNT (cbranch_block->succs) != 2)
121 return false;
123 /* Verify that we've got a normal conditional branch at the end
124 of the block. */
125 cbranch_insn = BB_END (cbranch_block);
126 if (!any_condjump_p (cbranch_insn))
127 return false;
129 cbranch_fallthru_edge = FALLTHRU_EDGE (cbranch_block);
130 cbranch_jump_edge = BRANCH_EDGE (cbranch_block);
132 /* The next block must not have multiple predecessors, must not
133 be the last block in the function, and must contain just the
134 unconditional jump. */
135 jump_block = cbranch_fallthru_edge->dest;
136 if (!single_pred_p (jump_block)
137 || jump_block->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
138 || !FORWARDER_BLOCK_P (jump_block))
139 return false;
140 jump_dest_block = single_succ (jump_block);
142 /* If we are partitioning hot/cold basic blocks, we don't want to
143 mess up unconditional or indirect jumps that cross between hot
144 and cold sections.
146 Basic block partitioning may result in some jumps that appear to
147 be optimizable (or blocks that appear to be mergeable), but which really
148 must be left untouched (they are required to make it safely across
149 partition boundaries). See the comments at the top of
150 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
152 if (BB_PARTITION (jump_block) != BB_PARTITION (jump_dest_block)
153 || (cbranch_jump_edge->flags & EDGE_CROSSING))
154 return false;
156 /* The conditional branch must target the block after the
157 unconditional branch. */
158 cbranch_dest_block = cbranch_jump_edge->dest;
160 if (cbranch_dest_block == EXIT_BLOCK_PTR_FOR_FN (cfun)
161 || jump_dest_block == EXIT_BLOCK_PTR_FOR_FN (cfun)
162 || !can_fallthru (jump_block, cbranch_dest_block))
163 return false;
165 /* Invert the conditional branch. */
166 if (!invert_jump (as_a <rtx_jump_insn *> (cbranch_insn),
167 block_label (jump_dest_block), 0))
168 return false;
170 if (dump_file)
171 fprintf (dump_file, "Simplifying condjump %i around jump %i\n",
172 INSN_UID (cbranch_insn), INSN_UID (BB_END (jump_block)));
174 /* Success. Update the CFG to match. Note that after this point
175 the edge variable names appear backwards; the redirection is done
176 this way to preserve edge profile data. */
177 cbranch_jump_edge = redirect_edge_succ_nodup (cbranch_jump_edge,
178 cbranch_dest_block);
179 cbranch_fallthru_edge = redirect_edge_succ_nodup (cbranch_fallthru_edge,
180 jump_dest_block);
181 cbranch_jump_edge->flags |= EDGE_FALLTHRU;
182 cbranch_fallthru_edge->flags &= ~EDGE_FALLTHRU;
183 update_br_prob_note (cbranch_block);
185 /* Delete the block with the unconditional jump, and clean up the mess. */
186 delete_basic_block (jump_block);
187 tidy_fallthru_edge (cbranch_jump_edge);
188 update_forwarder_flag (cbranch_block);
190 return true;
193 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
194 on register. Used by jump threading. */
196 static bool
197 mark_effect (rtx exp, regset nonequal)
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 dest = XEXP (exp, 0);
206 if (REG_P (dest))
207 bitmap_clear_range (nonequal, REGNO (dest), REG_NREGS (dest));
208 return false;
210 case SET:
211 if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
212 return false;
213 dest = SET_DEST (exp);
214 if (dest == pc_rtx)
215 return false;
216 if (!REG_P (dest))
217 return true;
218 bitmap_set_range (nonequal, REGNO (dest), REG_NREGS (dest));
219 return false;
221 default:
222 return false;
226 /* Return true if X contains a register in NONEQUAL. */
227 static bool
228 mentions_nonequal_regs (const_rtx x, regset nonequal)
230 subrtx_iterator::array_type array;
231 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
233 const_rtx x = *iter;
234 if (REG_P (x))
236 unsigned int end_regno = END_REGNO (x);
237 for (unsigned int regno = REGNO (x); regno < end_regno; ++regno)
238 if (REGNO_REG_SET_P (nonequal, regno))
239 return true;
242 return false;
245 /* Attempt to prove that the basic block B will have no side effects and
246 always continues in the same edge if reached via E. Return the edge
247 if exist, NULL otherwise. */
249 static edge
250 thread_jump (edge e, basic_block b)
252 rtx set1, set2, cond1, cond2;
253 rtx_insn *insn;
254 enum rtx_code code1, code2, reversed_code2;
255 bool reverse1 = false;
256 unsigned i;
257 regset nonequal;
258 bool failed = false;
259 reg_set_iterator rsi;
261 /* Jump threading may cause fixup_partitions to introduce new crossing edges,
262 which is not allowed after reload. */
263 gcc_checking_assert (!reload_completed || !crtl->has_bb_partition);
265 if (b->flags & BB_NONTHREADABLE_BLOCK)
266 return NULL;
268 /* At the moment, we do handle only conditional jumps, but later we may
269 want to extend this code to tablejumps and others. */
270 if (EDGE_COUNT (e->src->succs) != 2)
271 return NULL;
272 if (EDGE_COUNT (b->succs) != 2)
274 b->flags |= BB_NONTHREADABLE_BLOCK;
275 return NULL;
278 /* Second branch must end with onlyjump, as we will eliminate the jump. */
279 if (!any_condjump_p (BB_END (e->src)))
280 return NULL;
282 if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
284 b->flags |= BB_NONTHREADABLE_BLOCK;
285 return NULL;
288 set1 = pc_set (BB_END (e->src));
289 set2 = pc_set (BB_END (b));
290 if (((e->flags & EDGE_FALLTHRU) != 0)
291 != (XEXP (SET_SRC (set1), 1) == pc_rtx))
292 reverse1 = true;
294 cond1 = XEXP (SET_SRC (set1), 0);
295 cond2 = XEXP (SET_SRC (set2), 0);
296 if (reverse1)
297 code1 = reversed_comparison_code (cond1, BB_END (e->src));
298 else
299 code1 = GET_CODE (cond1);
301 code2 = GET_CODE (cond2);
302 reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
304 if (!comparison_dominates_p (code1, code2)
305 && !comparison_dominates_p (code1, reversed_code2))
306 return NULL;
308 /* Ensure that the comparison operators are equivalent.
309 ??? This is far too pessimistic. We should allow swapped operands,
310 different CCmodes, or for example comparisons for interval, that
311 dominate even when operands are not equivalent. */
312 if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
313 || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
314 return NULL;
316 /* Punt if BB_END (e->src) is doloop-like conditional jump that modifies
317 the registers used in cond1. */
318 if (modified_in_p (cond1, BB_END (e->src)))
319 return NULL;
321 /* Short circuit cases where block B contains some side effects, as we can't
322 safely bypass it. */
323 for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
324 insn = NEXT_INSN (insn))
325 if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
327 b->flags |= BB_NONTHREADABLE_BLOCK;
328 return NULL;
331 cselib_init (0);
333 /* First process all values computed in the source basic block. */
334 for (insn = NEXT_INSN (BB_HEAD (e->src));
335 insn != NEXT_INSN (BB_END (e->src));
336 insn = NEXT_INSN (insn))
337 if (INSN_P (insn))
338 cselib_process_insn (insn);
340 nonequal = BITMAP_ALLOC (NULL);
341 CLEAR_REG_SET (nonequal);
343 /* Now assume that we've continued by the edge E to B and continue
344 processing as if it were same basic block.
345 Our goal is to prove that whole block is an NOOP. */
347 for (insn = NEXT_INSN (BB_HEAD (b));
348 insn != NEXT_INSN (BB_END (b)) && !failed;
349 insn = NEXT_INSN (insn))
351 /* cond2 must not mention any register that is not equal to the
352 former block. Check this before processing that instruction,
353 as BB_END (b) could contain also clobbers. */
354 if (insn == BB_END (b)
355 && mentions_nonequal_regs (cond2, nonequal))
356 goto failed_exit;
358 if (INSN_P (insn))
360 rtx pat = PATTERN (insn);
362 if (GET_CODE (pat) == PARALLEL)
364 for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
365 failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
367 else
368 failed |= mark_effect (pat, nonequal);
371 cselib_process_insn (insn);
374 /* Later we should clear nonequal of dead registers. So far we don't
375 have life information in cfg_cleanup. */
376 if (failed)
378 b->flags |= BB_NONTHREADABLE_BLOCK;
379 goto failed_exit;
382 EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
383 goto failed_exit;
385 BITMAP_FREE (nonequal);
386 cselib_finish ();
387 if ((comparison_dominates_p (code1, code2) != 0)
388 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
389 return BRANCH_EDGE (b);
390 else
391 return FALLTHRU_EDGE (b);
393 failed_exit:
394 BITMAP_FREE (nonequal);
395 cselib_finish ();
396 return NULL;
399 /* Attempt to forward edges leaving basic block B.
400 Return true if successful. */
402 static bool
403 try_forward_edges (int mode, basic_block b)
405 bool changed = false;
406 edge_iterator ei;
407 edge e, *threaded_edges = NULL;
409 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
411 basic_block target, first;
412 location_t goto_locus;
413 int counter;
414 bool threaded = false;
415 int nthreaded_edges = 0;
416 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
417 bool new_target_threaded = false;
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 while (counter < n_basic_blocks_for_fn (cfun))
436 basic_block new_target = NULL;
437 may_thread |= (target->flags & BB_MODIFIED) != 0;
439 if (FORWARDER_BLOCK_P (target)
440 && single_succ (target) != EXIT_BLOCK_PTR_FOR_FN (cfun))
442 /* Bypass trivial infinite loops. */
443 new_target = single_succ (target);
444 if (target == new_target)
445 counter = n_basic_blocks_for_fn (cfun);
446 else if (!optimize)
448 /* When not optimizing, ensure that edges or forwarder
449 blocks with different locus are not optimized out. */
450 location_t new_locus = single_succ_edge (target)->goto_locus;
451 location_t locus = goto_locus;
453 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
454 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
455 && new_locus != locus)
456 new_target = NULL;
457 else
459 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
460 locus = new_locus;
462 rtx_insn *last = BB_END (target);
463 if (DEBUG_INSN_P (last))
464 last = prev_nondebug_insn (last);
465 if (last && INSN_P (last))
466 new_locus = INSN_LOCATION (last);
467 else
468 new_locus = UNKNOWN_LOCATION;
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 goto_locus = locus;
485 /* Allow to thread only over one edge at time to simplify updating
486 of probabilities. */
487 else if ((mode & CLEANUP_THREADING) && may_thread)
489 edge t = thread_jump (e, target);
490 if (t)
492 if (!threaded_edges)
493 threaded_edges = XNEWVEC (edge,
494 n_basic_blocks_for_fn (cfun));
495 else
497 int i;
499 /* Detect an infinite loop across blocks not
500 including the start block. */
501 for (i = 0; i < nthreaded_edges; ++i)
502 if (threaded_edges[i] == t)
503 break;
504 if (i < nthreaded_edges)
506 counter = n_basic_blocks_for_fn (cfun);
507 break;
511 /* Detect an infinite loop across the start block. */
512 if (t->dest == b)
513 break;
515 gcc_assert (nthreaded_edges
516 < (n_basic_blocks_for_fn (cfun)
517 - NUM_FIXED_BLOCKS));
518 threaded_edges[nthreaded_edges++] = t;
520 new_target = t->dest;
521 new_target_threaded = true;
525 if (!new_target)
526 break;
528 counter++;
529 /* Do not turn non-crossing jump to crossing. Depending on target
530 it may require different instruction pattern. */
531 if ((e->flags & EDGE_CROSSING)
532 || BB_PARTITION (first) == BB_PARTITION (new_target))
534 target = new_target;
535 threaded |= new_target_threaded;
539 if (counter >= n_basic_blocks_for_fn (cfun))
541 if (dump_file)
542 fprintf (dump_file, "Infinite loop in BB %i.\n",
543 target->index);
545 else if (target == first)
546 ; /* We didn't do anything. */
547 else
549 /* Save the values now, as the edge may get removed. */
550 profile_count edge_count = e->count ();
551 int n = 0;
553 e->goto_locus = goto_locus;
555 /* Don't force if target is exit block. */
556 if (threaded && target != EXIT_BLOCK_PTR_FOR_FN (cfun))
558 notice_new_block (redirect_edge_and_branch_force (e, target));
559 if (dump_file)
560 fprintf (dump_file, "Conditionals threaded.\n");
562 else if (!redirect_edge_and_branch (e, target))
564 if (dump_file)
565 fprintf (dump_file,
566 "Forwarding edge %i->%i to %i failed.\n",
567 b->index, e->dest->index, target->index);
568 ei_next (&ei);
569 continue;
572 /* We successfully forwarded the edge. Now update profile
573 data: for each edge we traversed in the chain, remove
574 the original edge's execution count. */
577 edge t;
579 if (!single_succ_p (first))
581 gcc_assert (n < nthreaded_edges);
582 t = threaded_edges [n++];
583 gcc_assert (t->src == first);
584 update_bb_profile_for_threading (first, edge_count, t);
585 update_br_prob_note (first);
587 else
589 first->count -= edge_count;
590 /* It is possible that as the result of
591 threading we've removed edge as it is
592 threaded to the fallthru edge. Avoid
593 getting out of sync. */
594 if (n < nthreaded_edges
595 && first == threaded_edges [n]->src)
596 n++;
597 t = single_succ_edge (first);
600 first = t->dest;
602 while (first != target);
604 changed = true;
605 continue;
607 ei_next (&ei);
610 free (threaded_edges);
611 return changed;
615 /* Blocks A and B are to be merged into a single block. A has no incoming
616 fallthru edge, so it can be moved before B without adding or modifying
617 any jumps (aside from the jump from A to B). */
619 static void
620 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
622 rtx_insn *barrier;
624 /* If we are partitioning hot/cold basic blocks, we don't want to
625 mess up unconditional or indirect jumps that cross between hot
626 and cold sections.
628 Basic block partitioning may result in some jumps that appear to
629 be optimizable (or blocks that appear to be mergeable), but which really
630 must be left untouched (they are required to make it safely across
631 partition boundaries). See the comments at the top of
632 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
634 if (BB_PARTITION (a) != BB_PARTITION (b))
635 return;
637 barrier = next_nonnote_insn (BB_END (a));
638 gcc_assert (BARRIER_P (barrier));
639 delete_insn (barrier);
641 /* Scramble the insn chain. */
642 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
643 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
644 df_set_bb_dirty (a);
646 if (dump_file)
647 fprintf (dump_file, "Moved block %d before %d and merged.\n",
648 a->index, b->index);
650 /* Swap the records for the two blocks around. */
652 unlink_block (a);
653 link_block (a, b->prev_bb);
655 /* Now blocks A and B are contiguous. Merge them. */
656 merge_blocks (a, b);
659 /* Blocks A and B are to be merged into a single block. B has no outgoing
660 fallthru edge, so it can be moved after A without adding or modifying
661 any jumps (aside from the jump from A to B). */
663 static void
664 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
666 rtx_insn *barrier, *real_b_end;
667 rtx_insn *label;
668 rtx_jump_table_data *table;
670 /* If we are partitioning hot/cold basic blocks, we don't want to
671 mess up unconditional or indirect jumps that cross between hot
672 and cold sections.
674 Basic block partitioning may result in some jumps that appear to
675 be optimizable (or blocks that appear to be mergeable), but which really
676 must be left untouched (they are required to make it safely across
677 partition boundaries). See the comments at the top of
678 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
680 if (BB_PARTITION (a) != BB_PARTITION (b))
681 return;
683 real_b_end = BB_END (b);
685 /* If there is a jump table following block B temporarily add the jump table
686 to block B so that it will also be moved to the correct location. */
687 if (tablejump_p (BB_END (b), &label, &table)
688 && prev_active_insn (label) == BB_END (b))
690 BB_END (b) = table;
693 /* There had better have been a barrier there. Delete it. */
694 barrier = NEXT_INSN (BB_END (b));
695 if (barrier && BARRIER_P (barrier))
696 delete_insn (barrier);
699 /* Scramble the insn chain. */
700 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
702 /* Restore the real end of b. */
703 BB_END (b) = real_b_end;
705 if (dump_file)
706 fprintf (dump_file, "Moved block %d after %d and merged.\n",
707 b->index, a->index);
709 /* Now blocks A and B are contiguous. Merge them. */
710 merge_blocks (a, b);
713 /* Attempt to merge basic blocks that are potentially non-adjacent.
714 Return NULL iff the attempt failed, otherwise return basic block
715 where cleanup_cfg should continue. Because the merging commonly
716 moves basic block away or introduces another optimization
717 possibility, return basic block just before B so cleanup_cfg don't
718 need to iterate.
720 It may be good idea to return basic block before C in the case
721 C has been moved after B and originally appeared earlier in the
722 insn sequence, but we have no information available about the
723 relative ordering of these two. Hopefully it is not too common. */
725 static basic_block
726 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
728 basic_block next;
730 /* If we are partitioning hot/cold basic blocks, we don't want to
731 mess up unconditional or indirect jumps that cross between hot
732 and cold sections.
734 Basic block partitioning may result in some jumps that appear to
735 be optimizable (or blocks that appear to be mergeable), but which really
736 must be left untouched (they are required to make it safely across
737 partition boundaries). See the comments at the top of
738 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
740 if (BB_PARTITION (b) != BB_PARTITION (c))
741 return NULL;
743 /* If B has a fallthru edge to C, no need to move anything. */
744 if (e->flags & EDGE_FALLTHRU)
746 int b_index = b->index, c_index = c->index;
748 /* Protect the loop latches. */
749 if (current_loops && c->loop_father->latch == c)
750 return NULL;
752 merge_blocks (b, c);
753 update_forwarder_flag (b);
755 if (dump_file)
756 fprintf (dump_file, "Merged %d and %d without moving.\n",
757 b_index, c_index);
759 return b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? b : b->prev_bb;
762 /* Otherwise we will need to move code around. Do that only if expensive
763 transformations are allowed. */
764 else if (mode & CLEANUP_EXPENSIVE)
766 edge tmp_edge, b_fallthru_edge;
767 bool c_has_outgoing_fallthru;
768 bool b_has_incoming_fallthru;
770 /* Avoid overactive code motion, as the forwarder blocks should be
771 eliminated by edge redirection instead. One exception might have
772 been if B is a forwarder block and C has no fallthru edge, but
773 that should be cleaned up by bb-reorder instead. */
774 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
775 return NULL;
777 /* We must make sure to not munge nesting of lexical blocks,
778 and loop notes. This is done by squeezing out all the notes
779 and leaving them there to lie. Not ideal, but functional. */
781 tmp_edge = find_fallthru_edge (c->succs);
782 c_has_outgoing_fallthru = (tmp_edge != NULL);
784 tmp_edge = find_fallthru_edge (b->preds);
785 b_has_incoming_fallthru = (tmp_edge != NULL);
786 b_fallthru_edge = tmp_edge;
787 next = b->prev_bb;
788 if (next == c)
789 next = next->prev_bb;
791 /* Otherwise, we're going to try to move C after B. If C does
792 not have an outgoing fallthru, then it can be moved
793 immediately after B without introducing or modifying jumps. */
794 if (! c_has_outgoing_fallthru)
796 merge_blocks_move_successor_nojumps (b, c);
797 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
800 /* If B does not have an incoming fallthru, then it can be moved
801 immediately before C without introducing or modifying jumps.
802 C cannot be the first block, so we do not have to worry about
803 accessing a non-existent block. */
805 if (b_has_incoming_fallthru)
807 basic_block bb;
809 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
810 return NULL;
811 bb = force_nonfallthru (b_fallthru_edge);
812 if (bb)
813 notice_new_block (bb);
816 merge_blocks_move_predecessor_nojumps (b, c);
817 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
820 return NULL;
824 /* Removes the memory attributes of MEM expression
825 if they are not equal. */
827 static void
828 merge_memattrs (rtx x, rtx y)
830 int i;
831 int j;
832 enum rtx_code code;
833 const char *fmt;
835 if (x == y)
836 return;
837 if (x == 0 || y == 0)
838 return;
840 code = GET_CODE (x);
842 if (code != GET_CODE (y))
843 return;
845 if (GET_MODE (x) != GET_MODE (y))
846 return;
848 if (code == MEM && !mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
850 if (! MEM_ATTRS (x))
851 MEM_ATTRS (y) = 0;
852 else if (! MEM_ATTRS (y))
853 MEM_ATTRS (x) = 0;
854 else
856 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
858 set_mem_alias_set (x, 0);
859 set_mem_alias_set (y, 0);
862 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
864 set_mem_expr (x, 0);
865 set_mem_expr (y, 0);
866 clear_mem_offset (x);
867 clear_mem_offset (y);
869 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
870 || (MEM_OFFSET_KNOWN_P (x)
871 && maybe_ne (MEM_OFFSET (x), MEM_OFFSET (y))))
873 clear_mem_offset (x);
874 clear_mem_offset (y);
877 if (!MEM_SIZE_KNOWN_P (x))
878 clear_mem_size (y);
879 else if (!MEM_SIZE_KNOWN_P (y))
880 clear_mem_size (x);
881 else if (known_le (MEM_SIZE (x), MEM_SIZE (y)))
882 set_mem_size (x, MEM_SIZE (y));
883 else if (known_le (MEM_SIZE (y), MEM_SIZE (x)))
884 set_mem_size (y, MEM_SIZE (x));
885 else
887 /* The sizes aren't ordered, so we can't merge them. */
888 clear_mem_size (x);
889 clear_mem_size (y);
892 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
893 set_mem_align (y, MEM_ALIGN (x));
896 if (code == MEM)
898 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
900 MEM_READONLY_P (x) = 0;
901 MEM_READONLY_P (y) = 0;
903 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
905 MEM_NOTRAP_P (x) = 0;
906 MEM_NOTRAP_P (y) = 0;
908 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
910 MEM_VOLATILE_P (x) = 1;
911 MEM_VOLATILE_P (y) = 1;
915 fmt = GET_RTX_FORMAT (code);
916 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
918 switch (fmt[i])
920 case 'E':
921 /* Two vectors must have the same length. */
922 if (XVECLEN (x, i) != XVECLEN (y, i))
923 return;
925 for (j = 0; j < XVECLEN (x, i); j++)
926 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
928 break;
930 case 'e':
931 merge_memattrs (XEXP (x, i), XEXP (y, i));
934 return;
938 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
939 different single sets S1 and S2. */
941 static bool
942 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
944 int i;
945 rtx e1, e2;
947 if (p1 == s1 && p2 == s2)
948 return true;
950 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
951 return false;
953 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
954 return false;
956 for (i = 0; i < XVECLEN (p1, 0); i++)
958 e1 = XVECEXP (p1, 0, i);
959 e2 = XVECEXP (p2, 0, i);
960 if (e1 == s1 && e2 == s2)
961 continue;
962 if (reload_completed
963 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
964 continue;
966 return false;
969 return true;
973 /* NOTE1 is the REG_EQUAL note, if any, attached to an insn
974 that is a single_set with a SET_SRC of SRC1. Similarly
975 for NOTE2/SRC2.
977 So effectively NOTE1/NOTE2 are an alternate form of
978 SRC1/SRC2 respectively.
980 Return nonzero if SRC1 or NOTE1 has the same constant
981 integer value as SRC2 or NOTE2. Else return zero. */
982 static int
983 values_equal_p (rtx note1, rtx note2, rtx src1, rtx src2)
985 if (note1
986 && note2
987 && CONST_INT_P (XEXP (note1, 0))
988 && rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0)))
989 return 1;
991 if (!note1
992 && !note2
993 && CONST_INT_P (src1)
994 && CONST_INT_P (src2)
995 && rtx_equal_p (src1, src2))
996 return 1;
998 if (note1
999 && CONST_INT_P (src2)
1000 && rtx_equal_p (XEXP (note1, 0), src2))
1001 return 1;
1003 if (note2
1004 && CONST_INT_P (src1)
1005 && rtx_equal_p (XEXP (note2, 0), src1))
1006 return 1;
1008 return 0;
1011 /* Examine register notes on I1 and I2 and return:
1012 - dir_forward if I1 can be replaced by I2, or
1013 - dir_backward if I2 can be replaced by I1, or
1014 - dir_both if both are the case. */
1016 static enum replace_direction
1017 can_replace_by (rtx_insn *i1, rtx_insn *i2)
1019 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1020 bool c1, c2;
1022 /* Check for 2 sets. */
1023 s1 = single_set (i1);
1024 s2 = single_set (i2);
1025 if (s1 == NULL_RTX || s2 == NULL_RTX)
1026 return dir_none;
1028 /* Check that the 2 sets set the same dest. */
1029 d1 = SET_DEST (s1);
1030 d2 = SET_DEST (s2);
1031 if (!(reload_completed
1032 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1033 return dir_none;
1035 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1036 set dest to the same value. */
1037 note1 = find_reg_equal_equiv_note (i1);
1038 note2 = find_reg_equal_equiv_note (i2);
1040 src1 = SET_SRC (s1);
1041 src2 = SET_SRC (s2);
1043 if (!values_equal_p (note1, note2, src1, src2))
1044 return dir_none;
1046 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1047 return dir_none;
1049 /* Although the 2 sets set dest to the same value, we cannot replace
1050 (set (dest) (const_int))
1052 (set (dest) (reg))
1053 because we don't know if the reg is live and has the same value at the
1054 location of replacement. */
1055 c1 = CONST_INT_P (src1);
1056 c2 = CONST_INT_P (src2);
1057 if (c1 && c2)
1058 return dir_both;
1059 else if (c2)
1060 return dir_forward;
1061 else if (c1)
1062 return dir_backward;
1064 return dir_none;
1067 /* Merges directions A and B. */
1069 static enum replace_direction
1070 merge_dir (enum replace_direction a, enum replace_direction b)
1072 /* Implements the following table:
1073 |bo fw bw no
1074 ---+-----------
1075 bo |bo fw bw no
1076 fw |-- fw no no
1077 bw |-- -- bw no
1078 no |-- -- -- no. */
1080 if (a == b)
1081 return a;
1083 if (a == dir_both)
1084 return b;
1085 if (b == dir_both)
1086 return a;
1088 return dir_none;
1091 /* Array of flags indexed by reg note kind, true if the given
1092 reg note is CFA related. */
1093 static const bool reg_note_cfa_p[] = {
1094 #undef REG_CFA_NOTE
1095 #define DEF_REG_NOTE(NAME) false,
1096 #define REG_CFA_NOTE(NAME) true,
1097 #include "reg-notes.def"
1098 #undef REG_CFA_NOTE
1099 #undef DEF_REG_NOTE
1100 false
1103 /* Return true if I1 and I2 have identical CFA notes (the same order
1104 and equivalent content). */
1106 static bool
1107 insns_have_identical_cfa_notes (rtx_insn *i1, rtx_insn *i2)
1109 rtx n1, n2;
1110 for (n1 = REG_NOTES (i1), n2 = REG_NOTES (i2); ;
1111 n1 = XEXP (n1, 1), n2 = XEXP (n2, 1))
1113 /* Skip over reg notes not related to CFI information. */
1114 while (n1 && !reg_note_cfa_p[REG_NOTE_KIND (n1)])
1115 n1 = XEXP (n1, 1);
1116 while (n2 && !reg_note_cfa_p[REG_NOTE_KIND (n2)])
1117 n2 = XEXP (n2, 1);
1118 if (n1 == NULL_RTX && n2 == NULL_RTX)
1119 return true;
1120 if (n1 == NULL_RTX || n2 == NULL_RTX)
1121 return false;
1122 if (XEXP (n1, 0) == XEXP (n2, 0))
1124 else if (XEXP (n1, 0) == NULL_RTX || XEXP (n2, 0) == NULL_RTX)
1125 return false;
1126 else if (!(reload_completed
1127 ? rtx_renumbered_equal_p (XEXP (n1, 0), XEXP (n2, 0))
1128 : rtx_equal_p (XEXP (n1, 0), XEXP (n2, 0))))
1129 return false;
1133 /* Examine I1 and I2 and return:
1134 - dir_forward if I1 can be replaced by I2, or
1135 - dir_backward if I2 can be replaced by I1, or
1136 - dir_both if both are the case. */
1138 static enum replace_direction
1139 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx_insn *i1, rtx_insn *i2)
1141 rtx p1, p2;
1143 /* Verify that I1 and I2 are equivalent. */
1144 if (GET_CODE (i1) != GET_CODE (i2))
1145 return dir_none;
1147 /* __builtin_unreachable() may lead to empty blocks (ending with
1148 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1149 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1150 return dir_both;
1152 /* ??? Do not allow cross-jumping between different stack levels. */
1153 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1154 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1155 if (p1 && p2)
1157 p1 = XEXP (p1, 0);
1158 p2 = XEXP (p2, 0);
1159 if (!rtx_equal_p (p1, p2))
1160 return dir_none;
1162 /* ??? Worse, this adjustment had better be constant lest we
1163 have differing incoming stack levels. */
1164 if (!frame_pointer_needed
1165 && known_eq (find_args_size_adjust (i1), HOST_WIDE_INT_MIN))
1166 return dir_none;
1168 else if (p1 || p2)
1169 return dir_none;
1171 /* Do not allow cross-jumping between frame related insns and other
1172 insns. */
1173 if (RTX_FRAME_RELATED_P (i1) != RTX_FRAME_RELATED_P (i2))
1174 return dir_none;
1176 p1 = PATTERN (i1);
1177 p2 = PATTERN (i2);
1179 if (GET_CODE (p1) != GET_CODE (p2))
1180 return dir_none;
1182 /* If this is a CALL_INSN, compare register usage information.
1183 If we don't check this on stack register machines, the two
1184 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1185 numbers of stack registers in the same basic block.
1186 If we don't check this on machines with delay slots, a delay slot may
1187 be filled that clobbers a parameter expected by the subroutine.
1189 ??? We take the simple route for now and assume that if they're
1190 equal, they were constructed identically.
1192 Also check for identical exception regions. */
1194 if (CALL_P (i1))
1196 /* Ensure the same EH region. */
1197 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1198 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1200 if (!n1 && n2)
1201 return dir_none;
1203 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1204 return dir_none;
1206 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1207 CALL_INSN_FUNCTION_USAGE (i2))
1208 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1209 return dir_none;
1211 /* For address sanitizer, never crossjump __asan_report_* builtins,
1212 otherwise errors might be reported on incorrect lines. */
1213 if (flag_sanitize & SANITIZE_ADDRESS)
1215 rtx call = get_call_rtx_from (i1);
1216 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1218 rtx symbol = XEXP (XEXP (call, 0), 0);
1219 if (SYMBOL_REF_DECL (symbol)
1220 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1222 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1223 == BUILT_IN_NORMAL)
1224 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1225 >= BUILT_IN_ASAN_REPORT_LOAD1
1226 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1227 <= BUILT_IN_ASAN_STOREN)
1228 return dir_none;
1233 if (insn_callee_abi (i1) != insn_callee_abi (i2))
1234 return dir_none;
1237 /* If both i1 and i2 are frame related, verify all the CFA notes
1238 in the same order and with the same content. */
1239 if (RTX_FRAME_RELATED_P (i1) && !insns_have_identical_cfa_notes (i1, i2))
1240 return dir_none;
1242 #ifdef STACK_REGS
1243 /* If cross_jump_death_matters is not 0, the insn's mode
1244 indicates whether or not the insn contains any stack-like
1245 regs. */
1247 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1249 /* If register stack conversion has already been done, then
1250 death notes must also be compared before it is certain that
1251 the two instruction streams match. */
1253 rtx note;
1254 HARD_REG_SET i1_regset, i2_regset;
1256 CLEAR_HARD_REG_SET (i1_regset);
1257 CLEAR_HARD_REG_SET (i2_regset);
1259 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1260 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1261 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1263 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1264 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1265 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1267 if (i1_regset != i2_regset)
1268 return dir_none;
1270 #endif
1272 if (reload_completed
1273 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1274 return dir_both;
1276 return can_replace_by (i1, i2);
1279 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1280 flow_find_head_matching_sequence, ensure the notes match. */
1282 static void
1283 merge_notes (rtx_insn *i1, rtx_insn *i2)
1285 /* If the merged insns have different REG_EQUAL notes, then
1286 remove them. */
1287 rtx equiv1 = find_reg_equal_equiv_note (i1);
1288 rtx equiv2 = find_reg_equal_equiv_note (i2);
1290 if (equiv1 && !equiv2)
1291 remove_note (i1, equiv1);
1292 else if (!equiv1 && equiv2)
1293 remove_note (i2, equiv2);
1294 else if (equiv1 && equiv2
1295 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1297 remove_note (i1, equiv1);
1298 remove_note (i2, equiv2);
1302 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1303 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1304 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1305 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1306 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1308 static void
1309 walk_to_nondebug_insn (rtx_insn **i1, basic_block *bb1, bool follow_fallthru,
1310 bool *did_fallthru)
1312 edge fallthru;
1314 *did_fallthru = false;
1316 /* Ignore notes. */
1317 while (!NONDEBUG_INSN_P (*i1))
1319 if (*i1 != BB_HEAD (*bb1))
1321 *i1 = PREV_INSN (*i1);
1322 continue;
1325 if (!follow_fallthru)
1326 return;
1328 fallthru = find_fallthru_edge ((*bb1)->preds);
1329 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)
1330 || !single_succ_p (fallthru->src))
1331 return;
1333 *bb1 = fallthru->src;
1334 *i1 = BB_END (*bb1);
1335 *did_fallthru = true;
1339 /* Look through the insns at the end of BB1 and BB2 and find the longest
1340 sequence that are either equivalent, or allow forward or backward
1341 replacement. Store the first insns for that sequence in *F1 and *F2 and
1342 return the sequence length.
1344 DIR_P indicates the allowed replacement direction on function entry, and
1345 the actual replacement direction on function exit. If NULL, only equivalent
1346 sequences are allowed.
1348 To simplify callers of this function, if the blocks match exactly,
1349 store the head of the blocks in *F1 and *F2. */
1352 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx_insn **f1,
1353 rtx_insn **f2, enum replace_direction *dir_p)
1355 rtx_insn *i1, *i2, *last1, *last2, *afterlast1, *afterlast2;
1356 int ninsns = 0;
1357 enum replace_direction dir, last_dir, afterlast_dir;
1358 bool follow_fallthru, did_fallthru;
1360 if (dir_p)
1361 dir = *dir_p;
1362 else
1363 dir = dir_both;
1364 afterlast_dir = dir;
1365 last_dir = afterlast_dir;
1367 /* Skip simple jumps at the end of the blocks. Complex jumps still
1368 need to be compared for equivalence, which we'll do below. */
1370 i1 = BB_END (bb1);
1371 last1 = afterlast1 = last2 = afterlast2 = NULL;
1372 if (onlyjump_p (i1)
1373 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1375 last1 = i1;
1376 i1 = PREV_INSN (i1);
1379 i2 = BB_END (bb2);
1380 if (onlyjump_p (i2)
1381 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1383 last2 = i2;
1384 /* Count everything except for unconditional jump as insn.
1385 Don't count any jumps if dir_p is NULL. */
1386 if (!simplejump_p (i2) && !returnjump_p (i2) && last1 && dir_p)
1387 ninsns++;
1388 i2 = PREV_INSN (i2);
1391 while (true)
1393 /* In the following example, we can replace all jumps to C by jumps to A.
1395 This removes 4 duplicate insns.
1396 [bb A] insn1 [bb C] insn1
1397 insn2 insn2
1398 [bb B] insn3 insn3
1399 insn4 insn4
1400 jump_insn jump_insn
1402 We could also replace all jumps to A by jumps to C, but that leaves B
1403 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1404 step, all jumps to B would be replaced with jumps to the middle of C,
1405 achieving the same result with more effort.
1406 So we allow only the first possibility, which means that we don't allow
1407 fallthru in the block that's being replaced. */
1409 follow_fallthru = dir_p && dir != dir_forward;
1410 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1411 if (did_fallthru)
1412 dir = dir_backward;
1414 follow_fallthru = dir_p && dir != dir_backward;
1415 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1416 if (did_fallthru)
1417 dir = dir_forward;
1419 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1420 break;
1422 /* Do not turn corssing edge to non-crossing or vice versa after
1423 reload. */
1424 if (BB_PARTITION (BLOCK_FOR_INSN (i1))
1425 != BB_PARTITION (BLOCK_FOR_INSN (i2))
1426 && reload_completed)
1427 break;
1429 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1430 if (dir == dir_none || (!dir_p && dir != dir_both))
1431 break;
1433 merge_memattrs (i1, i2);
1435 /* Don't begin a cross-jump with a NOTE insn. */
1436 if (INSN_P (i1))
1438 merge_notes (i1, i2);
1440 afterlast1 = last1, afterlast2 = last2;
1441 last1 = i1, last2 = i2;
1442 afterlast_dir = last_dir;
1443 last_dir = dir;
1444 if (active_insn_p (i1))
1445 ninsns++;
1448 i1 = PREV_INSN (i1);
1449 i2 = PREV_INSN (i2);
1452 /* Include preceding notes and labels in the cross-jump. One,
1453 this may bring us to the head of the blocks as requested above.
1454 Two, it keeps line number notes as matched as may be. */
1455 if (ninsns)
1457 bb1 = BLOCK_FOR_INSN (last1);
1458 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1459 last1 = PREV_INSN (last1);
1461 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1462 last1 = PREV_INSN (last1);
1464 bb2 = BLOCK_FOR_INSN (last2);
1465 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1466 last2 = PREV_INSN (last2);
1468 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1469 last2 = PREV_INSN (last2);
1471 *f1 = last1;
1472 *f2 = last2;
1475 if (dir_p)
1476 *dir_p = last_dir;
1477 return ninsns;
1480 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1481 the head of the two blocks. Do not include jumps at the end.
1482 If STOP_AFTER is nonzero, stop after finding that many matching
1483 instructions. If STOP_AFTER is zero, count all INSN_P insns, if it is
1484 non-zero, only count active insns. */
1487 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx_insn **f1,
1488 rtx_insn **f2, int stop_after)
1490 rtx_insn *i1, *i2, *last1, *last2, *beforelast1, *beforelast2;
1491 int ninsns = 0;
1492 edge e;
1493 edge_iterator ei;
1494 int nehedges1 = 0, nehedges2 = 0;
1496 FOR_EACH_EDGE (e, ei, bb1->succs)
1497 if (e->flags & EDGE_EH)
1498 nehedges1++;
1499 FOR_EACH_EDGE (e, ei, bb2->succs)
1500 if (e->flags & EDGE_EH)
1501 nehedges2++;
1503 i1 = BB_HEAD (bb1);
1504 i2 = BB_HEAD (bb2);
1505 last1 = beforelast1 = last2 = beforelast2 = NULL;
1507 while (true)
1509 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1510 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1512 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1513 break;
1514 i1 = NEXT_INSN (i1);
1517 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1519 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1520 break;
1521 i2 = NEXT_INSN (i2);
1524 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1525 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1526 break;
1528 if (NOTE_P (i1) || NOTE_P (i2)
1529 || JUMP_P (i1) || JUMP_P (i2))
1530 break;
1532 /* A sanity check to make sure we're not merging insns with different
1533 effects on EH. If only one of them ends a basic block, it shouldn't
1534 have an EH edge; if both end a basic block, there should be the same
1535 number of EH edges. */
1536 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1537 && nehedges1 > 0)
1538 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1539 && nehedges2 > 0)
1540 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1541 && nehedges1 != nehedges2))
1542 break;
1544 if (old_insns_match_p (0, i1, i2) != dir_both)
1545 break;
1547 merge_memattrs (i1, i2);
1549 /* Don't begin a cross-jump with a NOTE insn. */
1550 if (INSN_P (i1))
1552 merge_notes (i1, i2);
1554 beforelast1 = last1, beforelast2 = last2;
1555 last1 = i1, last2 = i2;
1556 if (!stop_after || active_insn_p (i1))
1557 ninsns++;
1560 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1561 || (stop_after > 0 && ninsns == stop_after))
1562 break;
1564 i1 = NEXT_INSN (i1);
1565 i2 = NEXT_INSN (i2);
1568 if (ninsns)
1570 *f1 = last1;
1571 *f2 = last2;
1574 return ninsns;
1577 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1578 the branch instruction. This means that if we commonize the control
1579 flow before end of the basic block, the semantic remains unchanged.
1581 We may assume that there exists one edge with a common destination. */
1583 static bool
1584 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1586 int nehedges1 = 0, nehedges2 = 0;
1587 edge fallthru1 = 0, fallthru2 = 0;
1588 edge e1, e2;
1589 edge_iterator ei;
1591 /* If we performed shrink-wrapping, edges to the exit block can
1592 only be distinguished for JUMP_INSNs. The two paths may differ in
1593 whether they went through the prologue. Sibcalls are fine, we know
1594 that we either didn't need or inserted an epilogue before them. */
1595 if (crtl->shrink_wrapped
1596 && single_succ_p (bb1)
1597 && single_succ (bb1) == EXIT_BLOCK_PTR_FOR_FN (cfun)
1598 && (!JUMP_P (BB_END (bb1))
1599 /* Punt if the only successor is a fake edge to exit, the jump
1600 must be some weird one. */
1601 || (single_succ_edge (bb1)->flags & EDGE_FAKE) != 0)
1602 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1603 return false;
1605 /* If BB1 has only one successor, we may be looking at either an
1606 unconditional jump, or a fake edge to exit. */
1607 if (single_succ_p (bb1)
1608 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1609 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1610 return (single_succ_p (bb2)
1611 && (single_succ_edge (bb2)->flags
1612 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1613 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1615 /* Match conditional jumps - this may get tricky when fallthru and branch
1616 edges are crossed. */
1617 if (EDGE_COUNT (bb1->succs) == 2
1618 && any_condjump_p (BB_END (bb1))
1619 && onlyjump_p (BB_END (bb1)))
1621 edge b1, f1, b2, f2;
1622 bool reverse, match;
1623 rtx set1, set2, cond1, cond2;
1624 enum rtx_code code1, code2;
1626 if (EDGE_COUNT (bb2->succs) != 2
1627 || !any_condjump_p (BB_END (bb2))
1628 || !onlyjump_p (BB_END (bb2)))
1629 return false;
1631 b1 = BRANCH_EDGE (bb1);
1632 b2 = BRANCH_EDGE (bb2);
1633 f1 = FALLTHRU_EDGE (bb1);
1634 f2 = FALLTHRU_EDGE (bb2);
1636 /* Get around possible forwarders on fallthru edges. Other cases
1637 should be optimized out already. */
1638 if (FORWARDER_BLOCK_P (f1->dest))
1639 f1 = single_succ_edge (f1->dest);
1641 if (FORWARDER_BLOCK_P (f2->dest))
1642 f2 = single_succ_edge (f2->dest);
1644 /* To simplify use of this function, return false if there are
1645 unneeded forwarder blocks. These will get eliminated later
1646 during cleanup_cfg. */
1647 if (FORWARDER_BLOCK_P (f1->dest)
1648 || FORWARDER_BLOCK_P (f2->dest)
1649 || FORWARDER_BLOCK_P (b1->dest)
1650 || FORWARDER_BLOCK_P (b2->dest))
1651 return false;
1653 if (f1->dest == f2->dest && b1->dest == b2->dest)
1654 reverse = false;
1655 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1656 reverse = true;
1657 else
1658 return false;
1660 set1 = pc_set (BB_END (bb1));
1661 set2 = pc_set (BB_END (bb2));
1662 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1663 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1664 reverse = !reverse;
1666 cond1 = XEXP (SET_SRC (set1), 0);
1667 cond2 = XEXP (SET_SRC (set2), 0);
1668 code1 = GET_CODE (cond1);
1669 if (reverse)
1670 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1671 else
1672 code2 = GET_CODE (cond2);
1674 if (code2 == UNKNOWN)
1675 return false;
1677 /* Verify codes and operands match. */
1678 match = ((code1 == code2
1679 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1680 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1681 || (code1 == swap_condition (code2)
1682 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1683 XEXP (cond2, 0))
1684 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1685 XEXP (cond2, 1))));
1687 /* If we return true, we will join the blocks. Which means that
1688 we will only have one branch prediction bit to work with. Thus
1689 we require the existing branches to have probabilities that are
1690 roughly similar. */
1691 if (match
1692 && optimize_bb_for_speed_p (bb1)
1693 && optimize_bb_for_speed_p (bb2))
1695 profile_probability prob2;
1697 if (b1->dest == b2->dest)
1698 prob2 = b2->probability;
1699 else
1700 /* Do not use f2 probability as f2 may be forwarded. */
1701 prob2 = b2->probability.invert ();
1703 /* Fail if the difference in probabilities is greater than 50%.
1704 This rules out two well-predicted branches with opposite
1705 outcomes. */
1706 if (b1->probability.differs_lot_from_p (prob2))
1708 if (dump_file)
1710 fprintf (dump_file,
1711 "Outcomes of branch in bb %i and %i differ too"
1712 " much (", bb1->index, bb2->index);
1713 b1->probability.dump (dump_file);
1714 prob2.dump (dump_file);
1715 fprintf (dump_file, ")\n");
1717 return false;
1721 if (dump_file && match)
1722 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1723 bb1->index, bb2->index);
1725 return match;
1728 /* Generic case - we are seeing a computed jump, table jump or trapping
1729 instruction. */
1731 /* Check whether there are tablejumps in the end of BB1 and BB2.
1732 Return true if they are identical. */
1734 rtx_insn *label1, *label2;
1735 rtx_jump_table_data *table1, *table2;
1737 if (tablejump_p (BB_END (bb1), &label1, &table1)
1738 && tablejump_p (BB_END (bb2), &label2, &table2)
1739 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1741 /* The labels should never be the same rtx. If they really are same
1742 the jump tables are same too. So disable crossjumping of blocks BB1
1743 and BB2 because when deleting the common insns in the end of BB1
1744 by delete_basic_block () the jump table would be deleted too. */
1745 /* If LABEL2 is referenced in BB1->END do not do anything
1746 because we would loose information when replacing
1747 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1748 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1750 /* Set IDENTICAL to true when the tables are identical. */
1751 bool identical = false;
1752 rtx p1, p2;
1754 p1 = PATTERN (table1);
1755 p2 = PATTERN (table2);
1756 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1758 identical = true;
1760 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1761 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1762 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1763 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1765 int i;
1767 identical = true;
1768 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1769 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1770 identical = false;
1773 if (identical)
1775 bool match;
1777 /* Temporarily replace references to LABEL1 with LABEL2
1778 in BB1->END so that we could compare the instructions. */
1779 replace_label_in_insn (BB_END (bb1), label1, label2, false);
1781 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1782 == dir_both);
1783 if (dump_file && match)
1784 fprintf (dump_file,
1785 "Tablejumps in bb %i and %i match.\n",
1786 bb1->index, bb2->index);
1788 /* Set the original label in BB1->END because when deleting
1789 a block whose end is a tablejump, the tablejump referenced
1790 from the instruction is deleted too. */
1791 replace_label_in_insn (BB_END (bb1), label2, label1, false);
1793 return match;
1796 return false;
1800 /* Find the last non-debug non-note instruction in each bb, except
1801 stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
1802 handles that case specially. old_insns_match_p does not handle
1803 other types of instruction notes. */
1804 rtx_insn *last1 = BB_END (bb1);
1805 rtx_insn *last2 = BB_END (bb2);
1806 while (!NOTE_INSN_BASIC_BLOCK_P (last1) &&
1807 (DEBUG_INSN_P (last1) || NOTE_P (last1)))
1808 last1 = PREV_INSN (last1);
1809 while (!NOTE_INSN_BASIC_BLOCK_P (last2) &&
1810 (DEBUG_INSN_P (last2) || NOTE_P (last2)))
1811 last2 = PREV_INSN (last2);
1812 gcc_assert (last1 && last2);
1814 /* First ensure that the instructions match. There may be many outgoing
1815 edges so this test is generally cheaper. */
1816 if (old_insns_match_p (mode, last1, last2) != dir_both)
1817 return false;
1819 /* Search the outgoing edges, ensure that the counts do match, find possible
1820 fallthru and exception handling edges since these needs more
1821 validation. */
1822 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1823 return false;
1825 bool nonfakeedges = false;
1826 FOR_EACH_EDGE (e1, ei, bb1->succs)
1828 e2 = EDGE_SUCC (bb2, ei.index);
1830 if ((e1->flags & EDGE_FAKE) == 0)
1831 nonfakeedges = true;
1833 if (e1->flags & EDGE_EH)
1834 nehedges1++;
1836 if (e2->flags & EDGE_EH)
1837 nehedges2++;
1839 if (e1->flags & EDGE_FALLTHRU)
1840 fallthru1 = e1;
1841 if (e2->flags & EDGE_FALLTHRU)
1842 fallthru2 = e2;
1845 /* If number of edges of various types does not match, fail. */
1846 if (nehedges1 != nehedges2
1847 || (fallthru1 != 0) != (fallthru2 != 0))
1848 return false;
1850 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1851 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1852 attempt to optimize, as the two basic blocks might have different
1853 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1854 traps there should be REG_ARG_SIZE notes, they could be missing
1855 for __builtin_unreachable () uses though. */
1856 if (!nonfakeedges
1857 && !ACCUMULATE_OUTGOING_ARGS
1858 && (!INSN_P (last1)
1859 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1860 return false;
1862 /* fallthru edges must be forwarded to the same destination. */
1863 if (fallthru1)
1865 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1866 ? single_succ (fallthru1->dest): fallthru1->dest);
1867 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1868 ? single_succ (fallthru2->dest): fallthru2->dest);
1870 if (d1 != d2)
1871 return false;
1874 /* Ensure the same EH region. */
1876 rtx n1 = find_reg_note (last1, REG_EH_REGION, 0);
1877 rtx n2 = find_reg_note (last2, REG_EH_REGION, 0);
1879 if (!n1 && n2)
1880 return false;
1882 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1883 return false;
1886 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1887 version of sequence abstraction. */
1888 FOR_EACH_EDGE (e1, ei, bb2->succs)
1890 edge e2;
1891 edge_iterator ei;
1892 basic_block d1 = e1->dest;
1894 if (FORWARDER_BLOCK_P (d1))
1895 d1 = EDGE_SUCC (d1, 0)->dest;
1897 FOR_EACH_EDGE (e2, ei, bb1->succs)
1899 basic_block d2 = e2->dest;
1900 if (FORWARDER_BLOCK_P (d2))
1901 d2 = EDGE_SUCC (d2, 0)->dest;
1902 if (d1 == d2)
1903 break;
1906 if (!e2)
1907 return false;
1910 return true;
1913 /* Returns true if BB basic block has a preserve label. */
1915 static bool
1916 block_has_preserve_label (basic_block bb)
1918 return (bb
1919 && block_label (bb)
1920 && LABEL_PRESERVE_P (block_label (bb)));
1923 /* E1 and E2 are edges with the same destination block. Search their
1924 predecessors for common code. If found, redirect control flow from
1925 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1926 or the other way around (dir_backward). DIR specifies the allowed
1927 replacement direction. */
1929 static bool
1930 try_crossjump_to_edge (int mode, edge e1, edge e2,
1931 enum replace_direction dir)
1933 int nmatch;
1934 basic_block src1 = e1->src, src2 = e2->src;
1935 basic_block redirect_to, redirect_from, to_remove;
1936 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1937 rtx_insn *newpos1, *newpos2;
1938 edge s;
1939 edge_iterator ei;
1941 newpos1 = newpos2 = NULL;
1943 /* Search backward through forwarder blocks. We don't need to worry
1944 about multiple entry or chained forwarders, as they will be optimized
1945 away. We do this to look past the unconditional jump following a
1946 conditional jump that is required due to the current CFG shape. */
1947 if (single_pred_p (src1)
1948 && FORWARDER_BLOCK_P (src1))
1949 e1 = single_pred_edge (src1), src1 = e1->src;
1951 if (single_pred_p (src2)
1952 && FORWARDER_BLOCK_P (src2))
1953 e2 = single_pred_edge (src2), src2 = e2->src;
1955 /* Nothing to do if we reach ENTRY, or a common source block. */
1956 if (src1 == ENTRY_BLOCK_PTR_FOR_FN (cfun) || src2
1957 == ENTRY_BLOCK_PTR_FOR_FN (cfun))
1958 return false;
1959 if (src1 == src2)
1960 return false;
1962 /* Seeing more than 1 forwarder blocks would confuse us later... */
1963 if (FORWARDER_BLOCK_P (e1->dest)
1964 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1965 return false;
1967 if (FORWARDER_BLOCK_P (e2->dest)
1968 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1969 return false;
1971 /* Likewise with dead code (possibly newly created by the other optimizations
1972 of cfg_cleanup). */
1973 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1974 return false;
1976 /* Do not turn corssing edge to non-crossing or vice versa after reload. */
1977 if (BB_PARTITION (src1) != BB_PARTITION (src2)
1978 && reload_completed)
1979 return false;
1981 /* Look for the common insn sequence, part the first ... */
1982 if (!outgoing_edges_match (mode, src1, src2))
1983 return false;
1985 /* ... and part the second. */
1986 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
1988 osrc1 = src1;
1989 osrc2 = src2;
1990 if (newpos1 != NULL_RTX)
1991 src1 = BLOCK_FOR_INSN (newpos1);
1992 if (newpos2 != NULL_RTX)
1993 src2 = BLOCK_FOR_INSN (newpos2);
1995 /* Check that SRC1 and SRC2 have preds again. They may have changed
1996 above due to the call to flow_find_cross_jump. */
1997 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1998 return false;
2000 if (dir == dir_backward)
2002 std::swap (osrc1, osrc2);
2003 std::swap (src1, src2);
2004 std::swap (e1, e2);
2005 std::swap (newpos1, newpos2);
2008 /* Don't proceed with the crossjump unless we found a sufficient number
2009 of matching instructions or the 'from' block was totally matched
2010 (such that its predecessors will hopefully be redirected and the
2011 block removed). */
2012 if ((nmatch < param_min_crossjump_insns)
2013 && (newpos1 != BB_HEAD (src1)))
2014 return false;
2016 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
2017 if (block_has_preserve_label (e1->dest)
2018 && (e1->flags & EDGE_ABNORMAL))
2019 return false;
2021 /* Here we know that the insns in the end of SRC1 which are common with SRC2
2022 will be deleted.
2023 If we have tablejumps in the end of SRC1 and SRC2
2024 they have been already compared for equivalence in outgoing_edges_match ()
2025 so replace the references to TABLE1 by references to TABLE2. */
2027 rtx_insn *label1, *label2;
2028 rtx_jump_table_data *table1, *table2;
2030 if (tablejump_p (BB_END (osrc1), &label1, &table1)
2031 && tablejump_p (BB_END (osrc2), &label2, &table2)
2032 && label1 != label2)
2034 rtx_insn *insn;
2036 /* Replace references to LABEL1 with LABEL2. */
2037 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2039 /* Do not replace the label in SRC1->END because when deleting
2040 a block whose end is a tablejump, the tablejump referenced
2041 from the instruction is deleted too. */
2042 if (insn != BB_END (osrc1))
2043 replace_label_in_insn (insn, label1, label2, true);
2048 /* Avoid splitting if possible. We must always split when SRC2 has
2049 EH predecessor edges, or we may end up with basic blocks with both
2050 normal and EH predecessor edges. */
2051 if (newpos2 == BB_HEAD (src2)
2052 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
2053 redirect_to = src2;
2054 else
2056 if (newpos2 == BB_HEAD (src2))
2058 /* Skip possible basic block header. */
2059 if (LABEL_P (newpos2))
2060 newpos2 = NEXT_INSN (newpos2);
2061 while (DEBUG_INSN_P (newpos2))
2062 newpos2 = NEXT_INSN (newpos2);
2063 if (NOTE_P (newpos2))
2064 newpos2 = NEXT_INSN (newpos2);
2065 while (DEBUG_INSN_P (newpos2))
2066 newpos2 = NEXT_INSN (newpos2);
2069 if (dump_file)
2070 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2071 src2->index, nmatch);
2072 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2075 if (dump_file)
2076 fprintf (dump_file,
2077 "Cross jumping from bb %i to bb %i; %i common insns\n",
2078 src1->index, src2->index, nmatch);
2080 /* We may have some registers visible through the block. */
2081 df_set_bb_dirty (redirect_to);
2083 if (osrc2 == src2)
2084 redirect_edges_to = redirect_to;
2085 else
2086 redirect_edges_to = osrc2;
2088 /* Recompute the counts of destinations of outgoing edges. */
2089 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2091 edge s2;
2092 edge_iterator ei;
2093 basic_block d = s->dest;
2095 if (FORWARDER_BLOCK_P (d))
2096 d = single_succ (d);
2098 FOR_EACH_EDGE (s2, ei, src1->succs)
2100 basic_block d2 = s2->dest;
2101 if (FORWARDER_BLOCK_P (d2))
2102 d2 = single_succ (d2);
2103 if (d == d2)
2104 break;
2107 /* Take care to update possible forwarder blocks. We verified
2108 that there is no more than one in the chain, so we can't run
2109 into infinite loop. */
2110 if (FORWARDER_BLOCK_P (s->dest))
2111 s->dest->count += s->count ();
2113 if (FORWARDER_BLOCK_P (s2->dest))
2114 s2->dest->count -= s->count ();
2116 s->probability = s->probability.combine_with_count
2117 (redirect_edges_to->count,
2118 s2->probability, src1->count);
2121 /* Adjust count for the block. An earlier jump
2122 threading pass may have left the profile in an inconsistent
2123 state (see update_bb_profile_for_threading) so we must be
2124 prepared for overflows. */
2125 tmp = redirect_to;
2128 tmp->count += src1->count;
2129 if (tmp == redirect_edges_to)
2130 break;
2131 tmp = find_fallthru_edge (tmp->succs)->dest;
2133 while (true);
2134 update_br_prob_note (redirect_edges_to);
2136 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2138 /* Skip possible basic block header. */
2139 if (LABEL_P (newpos1))
2140 newpos1 = NEXT_INSN (newpos1);
2142 while (DEBUG_INSN_P (newpos1))
2143 newpos1 = NEXT_INSN (newpos1);
2145 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2146 newpos1 = NEXT_INSN (newpos1);
2148 /* Skip also prologue and function markers. */
2149 while (DEBUG_INSN_P (newpos1)
2150 || (NOTE_P (newpos1)
2151 && (NOTE_KIND (newpos1) == NOTE_INSN_PROLOGUE_END
2152 || NOTE_KIND (newpos1) == NOTE_INSN_FUNCTION_BEG)))
2153 newpos1 = NEXT_INSN (newpos1);
2155 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2156 to_remove = single_succ (redirect_from);
2158 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2159 delete_basic_block (to_remove);
2161 update_forwarder_flag (redirect_from);
2162 if (redirect_to != src2)
2163 update_forwarder_flag (src2);
2165 return true;
2168 /* Search the predecessors of BB for common insn sequences. When found,
2169 share code between them by redirecting control flow. Return true if
2170 any changes made. */
2172 static bool
2173 try_crossjump_bb (int mode, basic_block bb)
2175 edge e, e2, fallthru;
2176 bool changed;
2177 unsigned max, ix, ix2;
2179 /* Nothing to do if there is not at least two incoming edges. */
2180 if (EDGE_COUNT (bb->preds) < 2)
2181 return false;
2183 /* Don't crossjump if this block ends in a computed jump,
2184 unless we are optimizing for size. */
2185 if (optimize_bb_for_size_p (bb)
2186 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2187 && computed_jump_p (BB_END (bb)))
2188 return false;
2190 /* If we are partitioning hot/cold basic blocks, we don't want to
2191 mess up unconditional or indirect jumps that cross between hot
2192 and cold sections.
2194 Basic block partitioning may result in some jumps that appear to
2195 be optimizable (or blocks that appear to be mergeable), but which really
2196 must be left untouched (they are required to make it safely across
2197 partition boundaries). See the comments at the top of
2198 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2200 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2201 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2202 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2203 return false;
2205 /* It is always cheapest to redirect a block that ends in a branch to
2206 a block that falls through into BB, as that adds no branches to the
2207 program. We'll try that combination first. */
2208 fallthru = NULL;
2209 max = param_max_crossjump_edges;
2211 if (EDGE_COUNT (bb->preds) > max)
2212 return false;
2214 fallthru = find_fallthru_edge (bb->preds);
2216 changed = false;
2217 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2219 e = EDGE_PRED (bb, ix);
2220 ix++;
2222 /* As noted above, first try with the fallthru predecessor (or, a
2223 fallthru predecessor if we are in cfglayout mode). */
2224 if (fallthru)
2226 /* Don't combine the fallthru edge into anything else.
2227 If there is a match, we'll do it the other way around. */
2228 if (e == fallthru)
2229 continue;
2230 /* If nothing changed since the last attempt, there is nothing
2231 we can do. */
2232 if (!first_pass
2233 && !((e->src->flags & BB_MODIFIED)
2234 || (fallthru->src->flags & BB_MODIFIED)))
2235 continue;
2237 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2239 changed = true;
2240 ix = 0;
2241 continue;
2245 /* Non-obvious work limiting check: Recognize that we're going
2246 to call try_crossjump_bb on every basic block. So if we have
2247 two blocks with lots of outgoing edges (a switch) and they
2248 share lots of common destinations, then we would do the
2249 cross-jump check once for each common destination.
2251 Now, if the blocks actually are cross-jump candidates, then
2252 all of their destinations will be shared. Which means that
2253 we only need check them for cross-jump candidacy once. We
2254 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2255 choosing to do the check from the block for which the edge
2256 in question is the first successor of A. */
2257 if (EDGE_SUCC (e->src, 0) != e)
2258 continue;
2260 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2262 e2 = EDGE_PRED (bb, ix2);
2264 if (e2 == e)
2265 continue;
2267 /* We've already checked the fallthru edge above. */
2268 if (e2 == fallthru)
2269 continue;
2271 /* The "first successor" check above only prevents multiple
2272 checks of crossjump(A,B). In order to prevent redundant
2273 checks of crossjump(B,A), require that A be the block
2274 with the lowest index. */
2275 if (e->src->index > e2->src->index)
2276 continue;
2278 /* If nothing changed since the last attempt, there is nothing
2279 we can do. */
2280 if (!first_pass
2281 && !((e->src->flags & BB_MODIFIED)
2282 || (e2->src->flags & BB_MODIFIED)))
2283 continue;
2285 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2286 direction. */
2287 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2289 changed = true;
2290 ix = 0;
2291 break;
2296 if (changed)
2297 crossjumps_occurred = true;
2299 return changed;
2302 /* Search the successors of BB for common insn sequences. When found,
2303 share code between them by moving it across the basic block
2304 boundary. Return true if any changes made. */
2306 static bool
2307 try_head_merge_bb (basic_block bb)
2309 basic_block final_dest_bb = NULL;
2310 int max_match = INT_MAX;
2311 edge e0;
2312 rtx_insn **headptr, **currptr, **nextptr;
2313 bool changed, moveall;
2314 unsigned ix;
2315 rtx_insn *e0_last_head;
2316 rtx cond;
2317 rtx_insn *move_before;
2318 unsigned nedges = EDGE_COUNT (bb->succs);
2319 rtx_insn *jump = BB_END (bb);
2320 regset live, live_union;
2322 /* Nothing to do if there is not at least two outgoing edges. */
2323 if (nedges < 2)
2324 return false;
2326 /* Don't crossjump if this block ends in a computed jump,
2327 unless we are optimizing for size. */
2328 if (optimize_bb_for_size_p (bb)
2329 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2330 && computed_jump_p (BB_END (bb)))
2331 return false;
2333 cond = get_condition (jump, &move_before, true, false);
2334 if (cond == NULL_RTX)
2335 move_before = jump;
2337 for (ix = 0; ix < nedges; ix++)
2338 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
2339 return false;
2341 for (ix = 0; ix < nedges; ix++)
2343 edge e = EDGE_SUCC (bb, ix);
2344 basic_block other_bb = e->dest;
2346 if (df_get_bb_dirty (other_bb))
2348 block_was_dirty = true;
2349 return false;
2352 if (e->flags & EDGE_ABNORMAL)
2353 return false;
2355 /* Normally, all destination blocks must only be reachable from this
2356 block, i.e. they must have one incoming edge.
2358 There is one special case we can handle, that of multiple consecutive
2359 jumps where the first jumps to one of the targets of the second jump.
2360 This happens frequently in switch statements for default labels.
2361 The structure is as follows:
2362 FINAL_DEST_BB
2363 ....
2364 if (cond) jump A;
2365 fall through
2367 jump with targets A, B, C, D...
2369 has two incoming edges, from FINAL_DEST_BB and BB
2371 In this case, we can try to move the insns through BB and into
2372 FINAL_DEST_BB. */
2373 if (EDGE_COUNT (other_bb->preds) != 1)
2375 edge incoming_edge, incoming_bb_other_edge;
2376 edge_iterator ei;
2378 if (final_dest_bb != NULL
2379 || EDGE_COUNT (other_bb->preds) != 2)
2380 return false;
2382 /* We must be able to move the insns across the whole block. */
2383 move_before = BB_HEAD (bb);
2384 while (!NONDEBUG_INSN_P (move_before))
2385 move_before = NEXT_INSN (move_before);
2387 if (EDGE_COUNT (bb->preds) != 1)
2388 return false;
2389 incoming_edge = EDGE_PRED (bb, 0);
2390 final_dest_bb = incoming_edge->src;
2391 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2392 return false;
2393 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2394 if (incoming_bb_other_edge != incoming_edge)
2395 break;
2396 if (incoming_bb_other_edge->dest != other_bb)
2397 return false;
2401 e0 = EDGE_SUCC (bb, 0);
2402 e0_last_head = NULL;
2403 changed = false;
2405 for (ix = 1; ix < nedges; ix++)
2407 edge e = EDGE_SUCC (bb, ix);
2408 rtx_insn *e0_last, *e_last;
2409 int nmatch;
2411 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2412 &e0_last, &e_last, 0);
2413 if (nmatch == 0)
2414 return false;
2416 if (nmatch < max_match)
2418 max_match = nmatch;
2419 e0_last_head = e0_last;
2423 /* If we matched an entire block, we probably have to avoid moving the
2424 last insn. */
2425 if (max_match > 0
2426 && e0_last_head == BB_END (e0->dest)
2427 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2428 || control_flow_insn_p (e0_last_head)))
2430 max_match--;
2431 if (max_match == 0)
2432 return false;
2433 e0_last_head = prev_real_nondebug_insn (e0_last_head);
2436 if (max_match == 0)
2437 return false;
2439 /* We must find a union of the live registers at each of the end points. */
2440 live = BITMAP_ALLOC (NULL);
2441 live_union = BITMAP_ALLOC (NULL);
2443 currptr = XNEWVEC (rtx_insn *, nedges);
2444 headptr = XNEWVEC (rtx_insn *, nedges);
2445 nextptr = XNEWVEC (rtx_insn *, nedges);
2447 for (ix = 0; ix < nedges; ix++)
2449 int j;
2450 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2451 rtx_insn *head = BB_HEAD (merge_bb);
2453 while (!NONDEBUG_INSN_P (head))
2454 head = NEXT_INSN (head);
2455 headptr[ix] = head;
2456 currptr[ix] = head;
2458 /* Compute the end point and live information */
2459 for (j = 1; j < max_match; j++)
2461 head = NEXT_INSN (head);
2462 while (!NONDEBUG_INSN_P (head));
2463 simulate_backwards_to_point (merge_bb, live, head);
2464 IOR_REG_SET (live_union, live);
2467 /* If we're moving across two blocks, verify the validity of the
2468 first move, then adjust the target and let the loop below deal
2469 with the final move. */
2470 if (final_dest_bb != NULL)
2472 rtx_insn *move_upto;
2474 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2475 jump, e0->dest, live_union,
2476 NULL, &move_upto);
2477 if (!moveall)
2479 if (move_upto == NULL_RTX)
2480 goto out;
2482 while (e0_last_head != move_upto)
2484 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2485 live_union);
2486 e0_last_head = PREV_INSN (e0_last_head);
2489 if (e0_last_head == NULL_RTX)
2490 goto out;
2492 jump = BB_END (final_dest_bb);
2493 cond = get_condition (jump, &move_before, true, false);
2494 if (cond == NULL_RTX)
2495 move_before = jump;
2500 rtx_insn *move_upto;
2501 moveall = can_move_insns_across (currptr[0], e0_last_head,
2502 move_before, jump, e0->dest, live_union,
2503 NULL, &move_upto);
2504 if (!moveall && move_upto == NULL_RTX)
2506 if (jump == move_before)
2507 break;
2509 /* Try again, using a different insertion point. */
2510 move_before = jump;
2512 continue;
2515 if (final_dest_bb && !moveall)
2516 /* We haven't checked whether a partial move would be OK for the first
2517 move, so we have to fail this case. */
2518 break;
2520 changed = true;
2521 for (;;)
2523 if (currptr[0] == move_upto)
2524 break;
2525 for (ix = 0; ix < nedges; ix++)
2527 rtx_insn *curr = currptr[ix];
2529 curr = NEXT_INSN (curr);
2530 while (!NONDEBUG_INSN_P (curr));
2531 currptr[ix] = curr;
2535 /* If we can't currently move all of the identical insns, remember
2536 each insn after the range that we'll merge. */
2537 if (!moveall)
2538 for (ix = 0; ix < nedges; ix++)
2540 rtx_insn *curr = currptr[ix];
2542 curr = NEXT_INSN (curr);
2543 while (!NONDEBUG_INSN_P (curr));
2544 nextptr[ix] = curr;
2547 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2548 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2549 if (final_dest_bb != NULL)
2550 df_set_bb_dirty (final_dest_bb);
2551 df_set_bb_dirty (bb);
2552 for (ix = 1; ix < nedges; ix++)
2554 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2555 delete_insn_chain (headptr[ix], currptr[ix], false);
2557 if (!moveall)
2559 if (jump == move_before)
2560 break;
2562 /* For the unmerged insns, try a different insertion point. */
2563 move_before = jump;
2565 for (ix = 0; ix < nedges; ix++)
2566 currptr[ix] = headptr[ix] = nextptr[ix];
2569 while (!moveall);
2571 out:
2572 free (currptr);
2573 free (headptr);
2574 free (nextptr);
2576 crossjumps_occurred |= changed;
2578 return changed;
2581 /* Return true if BB contains just bb note, or bb note followed
2582 by only DEBUG_INSNs. */
2584 static bool
2585 trivially_empty_bb_p (basic_block bb)
2587 rtx_insn *insn = BB_END (bb);
2589 while (1)
2591 if (insn == BB_HEAD (bb))
2592 return true;
2593 if (!DEBUG_INSN_P (insn))
2594 return false;
2595 insn = PREV_INSN (insn);
2599 /* Return true if BB contains just a return and possibly a USE of the
2600 return value. Fill in *RET and *USE with the return and use insns
2601 if any found, otherwise NULL. All CLOBBERs are ignored. */
2603 static bool
2604 bb_is_just_return (basic_block bb, rtx_insn **ret, rtx_insn **use)
2606 *ret = *use = NULL;
2607 rtx_insn *insn;
2609 if (bb == EXIT_BLOCK_PTR_FOR_FN (cfun))
2610 return false;
2612 FOR_BB_INSNS (bb, insn)
2613 if (NONDEBUG_INSN_P (insn))
2615 rtx pat = PATTERN (insn);
2617 if (!*ret && ANY_RETURN_P (pat))
2618 *ret = insn;
2619 else if (!*ret && !*use && GET_CODE (pat) == USE
2620 && REG_P (XEXP (pat, 0))
2621 && REG_FUNCTION_VALUE_P (XEXP (pat, 0)))
2622 *use = insn;
2623 else if (GET_CODE (pat) != CLOBBER)
2624 return false;
2627 return !!*ret;
2630 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2631 instructions etc. Return nonzero if changes were made. */
2633 static bool
2634 try_optimize_cfg (int mode)
2636 bool changed_overall = false;
2637 bool changed;
2638 int iterations = 0;
2639 basic_block bb, b, next;
2641 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2642 clear_bb_flags ();
2644 crossjumps_occurred = false;
2646 FOR_EACH_BB_FN (bb, cfun)
2647 update_forwarder_flag (bb);
2649 if (! targetm.cannot_modify_jumps_p ())
2651 first_pass = true;
2652 /* Attempt to merge blocks as made possible by edge removal. If
2653 a block has only one successor, and the successor has only
2654 one predecessor, they may be combined. */
2657 block_was_dirty = false;
2658 changed = false;
2659 iterations++;
2661 if (dump_file)
2662 fprintf (dump_file,
2663 "\n\ntry_optimize_cfg iteration %i\n\n",
2664 iterations);
2666 for (b = ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb; b
2667 != EXIT_BLOCK_PTR_FOR_FN (cfun);)
2669 basic_block c;
2670 edge s;
2671 bool changed_here = false;
2673 /* Delete trivially dead basic blocks. This is either
2674 blocks with no predecessors, or empty blocks with no
2675 successors. However if the empty block with no
2676 successors is the successor of the ENTRY_BLOCK, it is
2677 kept. This ensures that the ENTRY_BLOCK will have a
2678 successor which is a precondition for many RTL
2679 passes. Empty blocks may result from expanding
2680 __builtin_unreachable (). */
2681 if (EDGE_COUNT (b->preds) == 0
2682 || (EDGE_COUNT (b->succs) == 0
2683 && trivially_empty_bb_p (b)
2684 && single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))->dest
2685 != b))
2687 c = b->prev_bb;
2688 if (EDGE_COUNT (b->preds) > 0)
2690 edge e;
2691 edge_iterator ei;
2693 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2695 rtx_insn *insn;
2696 for (insn = BB_FOOTER (b);
2697 insn; insn = NEXT_INSN (insn))
2698 if (BARRIER_P (insn))
2699 break;
2700 if (insn)
2701 FOR_EACH_EDGE (e, ei, b->preds)
2702 if ((e->flags & EDGE_FALLTHRU))
2704 if (BB_FOOTER (b)
2705 && BB_FOOTER (e->src) == NULL)
2707 BB_FOOTER (e->src) = BB_FOOTER (b);
2708 BB_FOOTER (b) = NULL;
2710 else
2711 emit_barrier_after_bb (e->src);
2714 else
2716 rtx_insn *last = get_last_bb_insn (b);
2717 if (last && BARRIER_P (last))
2718 FOR_EACH_EDGE (e, ei, b->preds)
2719 if ((e->flags & EDGE_FALLTHRU))
2720 emit_barrier_after (BB_END (e->src));
2723 delete_basic_block (b);
2724 changed = true;
2725 /* Avoid trying to remove the exit block. */
2726 b = (c == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? c->next_bb : c);
2727 continue;
2730 /* Remove code labels no longer used. */
2731 if (single_pred_p (b)
2732 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2733 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2734 && LABEL_P (BB_HEAD (b))
2735 && !LABEL_PRESERVE_P (BB_HEAD (b))
2736 /* If the previous block ends with a branch to this
2737 block, we can't delete the label. Normally this
2738 is a condjump that is yet to be simplified, but
2739 if CASE_DROPS_THRU, this can be a tablejump with
2740 some element going to the same place as the
2741 default (fallthru). */
2742 && (single_pred (b) == ENTRY_BLOCK_PTR_FOR_FN (cfun)
2743 || !JUMP_P (BB_END (single_pred (b)))
2744 || ! label_is_jump_target_p (BB_HEAD (b),
2745 BB_END (single_pred (b)))))
2747 delete_insn (BB_HEAD (b));
2748 if (dump_file)
2749 fprintf (dump_file, "Deleted label in block %i.\n",
2750 b->index);
2753 /* If we fall through an empty block, we can remove it. */
2754 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2755 && single_pred_p (b)
2756 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2757 && !LABEL_P (BB_HEAD (b))
2758 && FORWARDER_BLOCK_P (b)
2759 /* Note that forwarder_block_p true ensures that
2760 there is a successor for this block. */
2761 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2762 && n_basic_blocks_for_fn (cfun) > NUM_FIXED_BLOCKS + 1)
2764 if (dump_file)
2765 fprintf (dump_file,
2766 "Deleting fallthru block %i.\n",
2767 b->index);
2769 c = ((b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun))
2770 ? b->next_bb : b->prev_bb);
2771 redirect_edge_succ_nodup (single_pred_edge (b),
2772 single_succ (b));
2773 delete_basic_block (b);
2774 changed = true;
2775 b = c;
2776 continue;
2779 /* Merge B with its single successor, if any. */
2780 if (single_succ_p (b)
2781 && (s = single_succ_edge (b))
2782 && !(s->flags & EDGE_COMPLEX)
2783 && (c = s->dest) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2784 && single_pred_p (c)
2785 && b != c)
2787 /* When not in cfg_layout mode use code aware of reordering
2788 INSN. This code possibly creates new basic blocks so it
2789 does not fit merge_blocks interface and is kept here in
2790 hope that it will become useless once more of compiler
2791 is transformed to use cfg_layout mode. */
2793 if ((mode & CLEANUP_CFGLAYOUT)
2794 && can_merge_blocks_p (b, c))
2796 merge_blocks (b, c);
2797 update_forwarder_flag (b);
2798 changed_here = true;
2800 else if (!(mode & CLEANUP_CFGLAYOUT)
2801 /* If the jump insn has side effects,
2802 we can't kill the edge. */
2803 && (!JUMP_P (BB_END (b))
2804 || (reload_completed
2805 ? simplejump_p (BB_END (b))
2806 : (onlyjump_p (BB_END (b))
2807 && !tablejump_p (BB_END (b),
2808 NULL, NULL))))
2809 && (next = merge_blocks_move (s, b, c, mode)))
2811 b = next;
2812 changed_here = true;
2816 /* Try to change a branch to a return to just that return. */
2817 rtx_insn *ret, *use;
2818 if (single_succ_p (b)
2819 && onlyjump_p (BB_END (b))
2820 && bb_is_just_return (single_succ (b), &ret, &use))
2822 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2823 PATTERN (ret), 0))
2825 if (use)
2826 emit_insn_before (copy_insn (PATTERN (use)),
2827 BB_END (b));
2828 if (dump_file)
2829 fprintf (dump_file, "Changed jump %d->%d to return.\n",
2830 b->index, single_succ (b)->index);
2831 redirect_edge_succ (single_succ_edge (b),
2832 EXIT_BLOCK_PTR_FOR_FN (cfun));
2833 single_succ_edge (b)->flags &= ~EDGE_CROSSING;
2834 changed_here = true;
2838 /* Try to change a conditional branch to a return to the
2839 respective conditional return. */
2840 if (EDGE_COUNT (b->succs) == 2
2841 && any_condjump_p (BB_END (b))
2842 && bb_is_just_return (BRANCH_EDGE (b)->dest, &ret, &use))
2844 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2845 PATTERN (ret), 0))
2847 if (use)
2848 emit_insn_before (copy_insn (PATTERN (use)),
2849 BB_END (b));
2850 if (dump_file)
2851 fprintf (dump_file, "Changed conditional jump %d->%d "
2852 "to conditional return.\n",
2853 b->index, BRANCH_EDGE (b)->dest->index);
2854 redirect_edge_succ (BRANCH_EDGE (b),
2855 EXIT_BLOCK_PTR_FOR_FN (cfun));
2856 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2857 changed_here = true;
2861 /* Try to flip a conditional branch that falls through to
2862 a return so that it becomes a conditional return and a
2863 new jump to the original branch target. */
2864 if (EDGE_COUNT (b->succs) == 2
2865 && BRANCH_EDGE (b)->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
2866 && any_condjump_p (BB_END (b))
2867 && bb_is_just_return (FALLTHRU_EDGE (b)->dest, &ret, &use))
2869 if (invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2870 JUMP_LABEL (BB_END (b)), 0))
2872 basic_block new_ft = BRANCH_EDGE (b)->dest;
2873 if (redirect_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2874 PATTERN (ret), 0))
2876 if (use)
2877 emit_insn_before (copy_insn (PATTERN (use)),
2878 BB_END (b));
2879 if (dump_file)
2880 fprintf (dump_file, "Changed conditional jump "
2881 "%d->%d to conditional return, adding "
2882 "fall-through jump.\n",
2883 b->index, BRANCH_EDGE (b)->dest->index);
2884 redirect_edge_succ (BRANCH_EDGE (b),
2885 EXIT_BLOCK_PTR_FOR_FN (cfun));
2886 BRANCH_EDGE (b)->flags &= ~EDGE_CROSSING;
2887 std::swap (BRANCH_EDGE (b)->probability,
2888 FALLTHRU_EDGE (b)->probability);
2889 update_br_prob_note (b);
2890 basic_block jb = force_nonfallthru (FALLTHRU_EDGE (b));
2891 notice_new_block (jb);
2892 if (!redirect_jump (as_a <rtx_jump_insn *> (BB_END (jb)),
2893 block_label (new_ft), 0))
2894 gcc_unreachable ();
2895 redirect_edge_succ (single_succ_edge (jb), new_ft);
2896 changed_here = true;
2898 else
2900 /* Invert the jump back to what it was. This should
2901 never fail. */
2902 if (!invert_jump (as_a <rtx_jump_insn *> (BB_END (b)),
2903 JUMP_LABEL (BB_END (b)), 0))
2904 gcc_unreachable ();
2909 /* Simplify branch over branch. */
2910 if ((mode & CLEANUP_EXPENSIVE)
2911 && !(mode & CLEANUP_CFGLAYOUT)
2912 && try_simplify_condjump (b))
2913 changed_here = true;
2915 /* If B has a single outgoing edge, but uses a
2916 non-trivial jump instruction without side-effects, we
2917 can either delete the jump entirely, or replace it
2918 with a simple unconditional jump. */
2919 if (single_succ_p (b)
2920 && single_succ (b) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2921 && onlyjump_p (BB_END (b))
2922 && !CROSSING_JUMP_P (BB_END (b))
2923 && try_redirect_by_replacing_jump (single_succ_edge (b),
2924 single_succ (b),
2925 (mode & CLEANUP_CFGLAYOUT) != 0))
2927 update_forwarder_flag (b);
2928 changed_here = true;
2931 /* Simplify branch to branch. */
2932 if (try_forward_edges (mode, b))
2934 update_forwarder_flag (b);
2935 changed_here = true;
2938 /* Look for shared code between blocks. */
2939 if ((mode & CLEANUP_CROSSJUMP)
2940 && try_crossjump_bb (mode, b))
2941 changed_here = true;
2943 if ((mode & CLEANUP_CROSSJUMP)
2944 /* This can lengthen register lifetimes. Do it only after
2945 reload. */
2946 && reload_completed
2947 && try_head_merge_bb (b))
2948 changed_here = true;
2950 /* Don't get confused by the index shift caused by
2951 deleting blocks. */
2952 if (!changed_here)
2953 b = b->next_bb;
2954 else
2955 changed = true;
2958 if ((mode & CLEANUP_CROSSJUMP)
2959 && try_crossjump_bb (mode, EXIT_BLOCK_PTR_FOR_FN (cfun)))
2960 changed = true;
2962 if (block_was_dirty)
2964 /* This should only be set by head-merging. */
2965 gcc_assert (mode & CLEANUP_CROSSJUMP);
2966 df_analyze ();
2969 if (changed)
2971 /* Edge forwarding in particular can cause hot blocks previously
2972 reached by both hot and cold blocks to become dominated only
2973 by cold blocks. This will cause the verification below to fail,
2974 and lead to now cold code in the hot section. This is not easy
2975 to detect and fix during edge forwarding, and in some cases
2976 is only visible after newly unreachable blocks are deleted,
2977 which will be done in fixup_partitions. */
2978 if ((mode & CLEANUP_NO_PARTITIONING) == 0)
2980 fixup_partitions ();
2981 checking_verify_flow_info ();
2985 changed_overall |= changed;
2986 first_pass = false;
2988 while (changed);
2991 FOR_ALL_BB_FN (b, cfun)
2992 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
2994 return changed_overall;
2997 /* Delete all unreachable basic blocks. */
2999 bool
3000 delete_unreachable_blocks (void)
3002 bool changed = false;
3003 basic_block b, prev_bb;
3005 find_unreachable_blocks ();
3007 /* When we're in GIMPLE mode and there may be debug bind insns, we
3008 should delete blocks in reverse dominator order, so as to get a
3009 chance to substitute all released DEFs into debug bind stmts. If
3010 we don't have dominators information, walking blocks backward
3011 gets us a better chance of retaining most debug information than
3012 otherwise. */
3013 if (MAY_HAVE_DEBUG_BIND_INSNS && current_ir_type () == IR_GIMPLE
3014 && dom_info_available_p (CDI_DOMINATORS))
3016 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3017 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3019 prev_bb = b->prev_bb;
3021 if (!(b->flags & BB_REACHABLE))
3023 /* Speed up the removal of blocks that don't dominate
3024 others. Walking backwards, this should be the common
3025 case. */
3026 if (!first_dom_son (CDI_DOMINATORS, b))
3027 delete_basic_block (b);
3028 else
3030 auto_vec<basic_block> h
3031 = get_all_dominated_blocks (CDI_DOMINATORS, b);
3033 while (h.length ())
3035 b = h.pop ();
3037 prev_bb = b->prev_bb;
3039 gcc_assert (!(b->flags & BB_REACHABLE));
3041 delete_basic_block (b);
3045 changed = true;
3049 else
3051 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
3052 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
3054 prev_bb = b->prev_bb;
3056 if (!(b->flags & BB_REACHABLE))
3058 delete_basic_block (b);
3059 changed = true;
3064 if (changed)
3065 tidy_fallthru_edges ();
3066 return changed;
3069 /* Delete any jump tables never referenced. We can't delete them at the
3070 time of removing tablejump insn as they are referenced by the preceding
3071 insns computing the destination, so we delay deleting and garbagecollect
3072 them once life information is computed. */
3073 void
3074 delete_dead_jumptables (void)
3076 basic_block bb;
3078 /* A dead jump table does not belong to any basic block. Scan insns
3079 between two adjacent basic blocks. */
3080 FOR_EACH_BB_FN (bb, cfun)
3082 rtx_insn *insn, *next;
3084 for (insn = NEXT_INSN (BB_END (bb));
3085 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
3086 insn = next)
3088 next = NEXT_INSN (insn);
3089 if (LABEL_P (insn)
3090 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
3091 && JUMP_TABLE_DATA_P (next))
3093 rtx_insn *label = insn, *jump = next;
3095 if (dump_file)
3096 fprintf (dump_file, "Dead jumptable %i removed\n",
3097 INSN_UID (insn));
3099 next = NEXT_INSN (next);
3100 delete_insn (jump);
3101 delete_insn (label);
3108 /* Tidy the CFG by deleting unreachable code and whatnot. */
3110 bool
3111 cleanup_cfg (int mode)
3113 bool changed = false;
3115 /* Set the cfglayout mode flag here. We could update all the callers
3116 but that is just inconvenient, especially given that we eventually
3117 want to have cfglayout mode as the default. */
3118 if (current_ir_type () == IR_RTL_CFGLAYOUT)
3119 mode |= CLEANUP_CFGLAYOUT;
3121 timevar_push (TV_CLEANUP_CFG);
3122 if (delete_unreachable_blocks ())
3124 changed = true;
3125 /* We've possibly created trivially dead code. Cleanup it right
3126 now to introduce more opportunities for try_optimize_cfg. */
3127 if (!(mode & (CLEANUP_NO_INSN_DEL))
3128 && !reload_completed)
3129 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3132 compact_blocks ();
3134 /* To tail-merge blocks ending in the same noreturn function (e.g.
3135 a call to abort) we have to insert fake edges to exit. Do this
3136 here once. The fake edges do not interfere with any other CFG
3137 cleanups. */
3138 if (mode & CLEANUP_CROSSJUMP)
3139 add_noreturn_fake_exit_edges ();
3141 if (!dbg_cnt (cfg_cleanup))
3142 return changed;
3144 while (try_optimize_cfg (mode))
3146 delete_unreachable_blocks (), changed = true;
3147 if (!(mode & CLEANUP_NO_INSN_DEL))
3149 /* Try to remove some trivially dead insns when doing an expensive
3150 cleanup. But delete_trivially_dead_insns doesn't work after
3151 reload (it only handles pseudos) and run_fast_dce is too costly
3152 to run in every iteration.
3154 For effective cross jumping, we really want to run a fast DCE to
3155 clean up any dead conditions, or they get in the way of performing
3156 useful tail merges.
3158 Other transformations in cleanup_cfg are not so sensitive to dead
3159 code, so delete_trivially_dead_insns or even doing nothing at all
3160 is good enough. */
3161 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3162 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3163 break;
3164 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occurred)
3166 run_fast_dce ();
3167 mode &= ~CLEANUP_FORCE_FAST_DCE;
3170 else
3171 break;
3174 if (mode & CLEANUP_CROSSJUMP)
3175 remove_fake_exit_edges ();
3177 if (mode & CLEANUP_FORCE_FAST_DCE)
3178 run_fast_dce ();
3180 /* Don't call delete_dead_jumptables in cfglayout mode, because
3181 that function assumes that jump tables are in the insns stream.
3182 But we also don't _have_ to delete dead jumptables in cfglayout
3183 mode because we shouldn't even be looking at things that are
3184 not in a basic block. Dead jumptables are cleaned up when
3185 going out of cfglayout mode. */
3186 if (!(mode & CLEANUP_CFGLAYOUT))
3187 delete_dead_jumptables ();
3189 /* ??? We probably do this way too often. */
3190 if (current_loops
3191 && (changed
3192 || (mode & CLEANUP_CFG_CHANGED)))
3194 timevar_push (TV_REPAIR_LOOPS);
3195 /* The above doesn't preserve dominance info if available. */
3196 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3197 calculate_dominance_info (CDI_DOMINATORS);
3198 fix_loop_structure (NULL);
3199 free_dominance_info (CDI_DOMINATORS);
3200 timevar_pop (TV_REPAIR_LOOPS);
3203 timevar_pop (TV_CLEANUP_CFG);
3205 return changed;
3208 namespace {
3210 const pass_data pass_data_jump =
3212 RTL_PASS, /* type */
3213 "jump", /* name */
3214 OPTGROUP_NONE, /* optinfo_flags */
3215 TV_JUMP, /* tv_id */
3216 0, /* properties_required */
3217 0, /* properties_provided */
3218 0, /* properties_destroyed */
3219 0, /* todo_flags_start */
3220 0, /* todo_flags_finish */
3223 class pass_jump : public rtl_opt_pass
3225 public:
3226 pass_jump (gcc::context *ctxt)
3227 : rtl_opt_pass (pass_data_jump, ctxt)
3230 /* opt_pass methods: */
3231 virtual unsigned int execute (function *);
3233 }; // class pass_jump
3235 unsigned int
3236 pass_jump::execute (function *)
3238 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3239 if (dump_file)
3240 dump_flow_info (dump_file, dump_flags);
3241 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3242 | (flag_thread_jumps && flag_expensive_optimizations
3243 ? CLEANUP_THREADING : 0));
3244 return 0;
3247 } // anon namespace
3249 rtl_opt_pass *
3250 make_pass_jump (gcc::context *ctxt)
3252 return new pass_jump (ctxt);
3255 namespace {
3257 const pass_data pass_data_jump_after_combine =
3259 RTL_PASS, /* type */
3260 "jump_after_combine", /* name */
3261 OPTGROUP_NONE, /* optinfo_flags */
3262 TV_JUMP, /* tv_id */
3263 0, /* properties_required */
3264 0, /* properties_provided */
3265 0, /* properties_destroyed */
3266 0, /* todo_flags_start */
3267 0, /* todo_flags_finish */
3270 class pass_jump_after_combine : public rtl_opt_pass
3272 public:
3273 pass_jump_after_combine (gcc::context *ctxt)
3274 : rtl_opt_pass (pass_data_jump_after_combine, ctxt)
3277 /* opt_pass methods: */
3278 virtual bool gate (function *)
3280 return flag_thread_jumps && flag_expensive_optimizations;
3282 virtual unsigned int execute (function *);
3284 }; // class pass_jump_after_combine
3286 unsigned int
3287 pass_jump_after_combine::execute (function *)
3289 /* Jump threading does not keep dominators up-to-date. */
3290 free_dominance_info (CDI_DOMINATORS);
3291 cleanup_cfg (CLEANUP_THREADING);
3292 return 0;
3295 } // anon namespace
3297 rtl_opt_pass *
3298 make_pass_jump_after_combine (gcc::context *ctxt)
3300 return new pass_jump_after_combine (ctxt);
3303 namespace {
3305 const pass_data pass_data_jump2 =
3307 RTL_PASS, /* type */
3308 "jump2", /* name */
3309 OPTGROUP_NONE, /* optinfo_flags */
3310 TV_JUMP, /* tv_id */
3311 0, /* properties_required */
3312 0, /* properties_provided */
3313 0, /* properties_destroyed */
3314 0, /* todo_flags_start */
3315 0, /* todo_flags_finish */
3318 class pass_jump2 : public rtl_opt_pass
3320 public:
3321 pass_jump2 (gcc::context *ctxt)
3322 : rtl_opt_pass (pass_data_jump2, ctxt)
3325 /* opt_pass methods: */
3326 virtual unsigned int execute (function *)
3328 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3329 return 0;
3332 }; // class pass_jump2
3334 } // anon namespace
3336 rtl_opt_pass *
3337 make_pass_jump2 (gcc::context *ctxt)
3339 return new pass_jump2 (ctxt);