gcc/
[official-gcc.git] / gcc / cfgcleanup.c
blob477b6da2e81621db4a4ac7e9ec8b397378ecdaea
1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987-2015 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 /* This file contains optimizer of the control flow. The main entry point is
21 cleanup_cfg. Following optimizations are performed:
23 - Unreachable blocks removal
24 - Edge forwarding (edge to the forwarder block is forwarded to its
25 successor. Simplification of the branch instruction is performed by
26 underlying infrastructure so branch can be converted to simplejump or
27 eliminated).
28 - Cross jumping (tail merging)
29 - Conditional jump-around-simplejump simplification
30 - Basic block merging. */
32 #include "config.h"
33 #include "system.h"
34 #include "coretypes.h"
35 #include "tm.h"
36 #include "rtl.h"
37 #include "hash-set.h"
38 #include "machmode.h"
39 #include "vec.h"
40 #include "double-int.h"
41 #include "input.h"
42 #include "alias.h"
43 #include "symtab.h"
44 #include "wide-int.h"
45 #include "inchash.h"
46 #include "tree.h"
47 #include "hard-reg-set.h"
48 #include "regs.h"
49 #include "insn-config.h"
50 #include "flags.h"
51 #include "recog.h"
52 #include "diagnostic-core.h"
53 #include "cselib.h"
54 #include "params.h"
55 #include "tm_p.h"
56 #include "target.h"
57 #include "hashtab.h"
58 #include "function.h" /* For inline functions in emit-rtl.h they need crtl. */
59 #include "emit-rtl.h"
60 #include "tree-pass.h"
61 #include "cfgloop.h"
62 #include "function.h"
63 #include "statistics.h"
64 #include "real.h"
65 #include "fixed-value.h"
66 #include "expmed.h"
67 #include "dojump.h"
68 #include "explow.h"
69 #include "calls.h"
70 #include "varasm.h"
71 #include "stmt.h"
72 #include "expr.h"
73 #include "dominance.h"
74 #include "cfg.h"
75 #include "cfgrtl.h"
76 #include "cfganal.h"
77 #include "cfgbuild.h"
78 #include "cfgcleanup.h"
79 #include "predict.h"
80 #include "basic-block.h"
81 #include "df.h"
82 #include "dce.h"
83 #include "dbgcnt.h"
84 #include "rtl-iter.h"
86 #define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
88 /* Set to true when we are running first pass of try_optimize_cfg loop. */
89 static bool first_pass;
91 /* Set to true if crossjumps occurred in the latest run of try_optimize_cfg. */
92 static bool crossjumps_occured;
94 /* Set to true if we couldn't run an optimization due to stale liveness
95 information; we should run df_analyze to enable more opportunities. */
96 static bool block_was_dirty;
98 static bool try_crossjump_to_edge (int, edge, edge, enum replace_direction);
99 static bool try_crossjump_bb (int, basic_block);
100 static bool outgoing_edges_match (int, basic_block, basic_block);
101 static enum replace_direction old_insns_match_p (int, rtx_insn *, rtx_insn *);
103 static void merge_blocks_move_predecessor_nojumps (basic_block, basic_block);
104 static void merge_blocks_move_successor_nojumps (basic_block, basic_block);
105 static bool try_optimize_cfg (int);
106 static bool try_simplify_condjump (basic_block);
107 static bool try_forward_edges (int, basic_block);
108 static edge thread_jump (edge, basic_block);
109 static bool mark_effect (rtx, bitmap);
110 static void notice_new_block (basic_block);
111 static void update_forwarder_flag (basic_block);
112 static void merge_memattrs (rtx, rtx);
114 /* Set flags for newly created block. */
116 static void
117 notice_new_block (basic_block bb)
119 if (!bb)
120 return;
122 if (forwarder_block_p (bb))
123 bb->flags |= BB_FORWARDER_BLOCK;
126 /* Recompute forwarder flag after block has been modified. */
128 static void
129 update_forwarder_flag (basic_block bb)
131 if (forwarder_block_p (bb))
132 bb->flags |= BB_FORWARDER_BLOCK;
133 else
134 bb->flags &= ~BB_FORWARDER_BLOCK;
137 /* Simplify a conditional jump around an unconditional jump.
138 Return true if something changed. */
140 static bool
141 try_simplify_condjump (basic_block cbranch_block)
143 basic_block jump_block, jump_dest_block, cbranch_dest_block;
144 edge cbranch_jump_edge, cbranch_fallthru_edge;
145 rtx_insn *cbranch_insn;
147 /* Verify that there are exactly two successors. */
148 if (EDGE_COUNT (cbranch_block->succs) != 2)
149 return false;
151 /* Verify that we've got a normal conditional branch at the end
152 of the block. */
153 cbranch_insn = BB_END (cbranch_block);
154 if (!any_condjump_p (cbranch_insn))
155 return false;
157 cbranch_fallthru_edge = FALLTHRU_EDGE (cbranch_block);
158 cbranch_jump_edge = BRANCH_EDGE (cbranch_block);
160 /* The next block must not have multiple predecessors, must not
161 be the last block in the function, and must contain just the
162 unconditional jump. */
163 jump_block = cbranch_fallthru_edge->dest;
164 if (!single_pred_p (jump_block)
165 || jump_block->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
166 || !FORWARDER_BLOCK_P (jump_block))
167 return false;
168 jump_dest_block = single_succ (jump_block);
170 /* If we are partitioning hot/cold basic blocks, we don't want to
171 mess up unconditional or indirect jumps that cross between hot
172 and cold sections.
174 Basic block partitioning may result in some jumps that appear to
175 be optimizable (or blocks that appear to be mergeable), but which really
176 must be left untouched (they are required to make it safely across
177 partition boundaries). See the comments at the top of
178 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
180 if (BB_PARTITION (jump_block) != BB_PARTITION (jump_dest_block)
181 || (cbranch_jump_edge->flags & EDGE_CROSSING))
182 return false;
184 /* The conditional branch must target the block after the
185 unconditional branch. */
186 cbranch_dest_block = cbranch_jump_edge->dest;
188 if (cbranch_dest_block == EXIT_BLOCK_PTR_FOR_FN (cfun)
189 || !can_fallthru (jump_block, cbranch_dest_block))
190 return false;
192 /* Invert the conditional branch. */
193 if (!invert_jump (cbranch_insn, block_label (jump_dest_block), 0))
194 return false;
196 if (dump_file)
197 fprintf (dump_file, "Simplifying condjump %i around jump %i\n",
198 INSN_UID (cbranch_insn), INSN_UID (BB_END (jump_block)));
200 /* Success. Update the CFG to match. Note that after this point
201 the edge variable names appear backwards; the redirection is done
202 this way to preserve edge profile data. */
203 cbranch_jump_edge = redirect_edge_succ_nodup (cbranch_jump_edge,
204 cbranch_dest_block);
205 cbranch_fallthru_edge = redirect_edge_succ_nodup (cbranch_fallthru_edge,
206 jump_dest_block);
207 cbranch_jump_edge->flags |= EDGE_FALLTHRU;
208 cbranch_fallthru_edge->flags &= ~EDGE_FALLTHRU;
209 update_br_prob_note (cbranch_block);
211 /* Delete the block with the unconditional jump, and clean up the mess. */
212 delete_basic_block (jump_block);
213 tidy_fallthru_edge (cbranch_jump_edge);
214 update_forwarder_flag (cbranch_block);
216 return true;
219 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
220 on register. Used by jump threading. */
222 static bool
223 mark_effect (rtx exp, regset nonequal)
225 int regno;
226 rtx dest;
227 switch (GET_CODE (exp))
229 /* In case we do clobber the register, mark it as equal, as we know the
230 value is dead so it don't have to match. */
231 case CLOBBER:
232 if (REG_P (XEXP (exp, 0)))
234 dest = XEXP (exp, 0);
235 regno = REGNO (dest);
236 if (HARD_REGISTER_NUM_P (regno))
237 bitmap_clear_range (nonequal, regno,
238 hard_regno_nregs[regno][GET_MODE (dest)]);
239 else
240 bitmap_clear_bit (nonequal, regno);
242 return false;
244 case SET:
245 if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
246 return false;
247 dest = SET_DEST (exp);
248 if (dest == pc_rtx)
249 return false;
250 if (!REG_P (dest))
251 return true;
252 regno = REGNO (dest);
253 if (HARD_REGISTER_NUM_P (regno))
254 bitmap_set_range (nonequal, regno,
255 hard_regno_nregs[regno][GET_MODE (dest)]);
256 else
257 bitmap_set_bit (nonequal, regno);
258 return false;
260 default:
261 return false;
265 /* Return true if X contains a register in NONEQUAL. */
266 static bool
267 mentions_nonequal_regs (const_rtx x, regset nonequal)
269 subrtx_iterator::array_type array;
270 FOR_EACH_SUBRTX (iter, array, x, NONCONST)
272 const_rtx x = *iter;
273 if (REG_P (x))
275 unsigned int regno = REGNO (x);
276 if (REGNO_REG_SET_P (nonequal, regno))
277 return true;
278 if (regno < FIRST_PSEUDO_REGISTER)
280 int n = hard_regno_nregs[regno][GET_MODE (x)];
281 while (--n > 0)
282 if (REGNO_REG_SET_P (nonequal, regno + n))
283 return true;
287 return false;
290 /* Attempt to prove that the basic block B will have no side effects and
291 always continues in the same edge if reached via E. Return the edge
292 if exist, NULL otherwise. */
294 static edge
295 thread_jump (edge e, basic_block b)
297 rtx set1, set2, cond1, cond2;
298 rtx_insn *insn;
299 enum rtx_code code1, code2, reversed_code2;
300 bool reverse1 = false;
301 unsigned i;
302 regset nonequal;
303 bool failed = false;
304 reg_set_iterator rsi;
306 if (b->flags & BB_NONTHREADABLE_BLOCK)
307 return NULL;
309 /* At the moment, we do handle only conditional jumps, but later we may
310 want to extend this code to tablejumps and others. */
311 if (EDGE_COUNT (e->src->succs) != 2)
312 return NULL;
313 if (EDGE_COUNT (b->succs) != 2)
315 b->flags |= BB_NONTHREADABLE_BLOCK;
316 return NULL;
319 /* Second branch must end with onlyjump, as we will eliminate the jump. */
320 if (!any_condjump_p (BB_END (e->src)))
321 return NULL;
323 if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
325 b->flags |= BB_NONTHREADABLE_BLOCK;
326 return NULL;
329 set1 = pc_set (BB_END (e->src));
330 set2 = pc_set (BB_END (b));
331 if (((e->flags & EDGE_FALLTHRU) != 0)
332 != (XEXP (SET_SRC (set1), 1) == pc_rtx))
333 reverse1 = true;
335 cond1 = XEXP (SET_SRC (set1), 0);
336 cond2 = XEXP (SET_SRC (set2), 0);
337 if (reverse1)
338 code1 = reversed_comparison_code (cond1, BB_END (e->src));
339 else
340 code1 = GET_CODE (cond1);
342 code2 = GET_CODE (cond2);
343 reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
345 if (!comparison_dominates_p (code1, code2)
346 && !comparison_dominates_p (code1, reversed_code2))
347 return NULL;
349 /* Ensure that the comparison operators are equivalent.
350 ??? This is far too pessimistic. We should allow swapped operands,
351 different CCmodes, or for example comparisons for interval, that
352 dominate even when operands are not equivalent. */
353 if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
354 || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
355 return NULL;
357 /* Short circuit cases where block B contains some side effects, as we can't
358 safely bypass it. */
359 for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
360 insn = NEXT_INSN (insn))
361 if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
363 b->flags |= BB_NONTHREADABLE_BLOCK;
364 return NULL;
367 cselib_init (0);
369 /* First process all values computed in the source basic block. */
370 for (insn = NEXT_INSN (BB_HEAD (e->src));
371 insn != NEXT_INSN (BB_END (e->src));
372 insn = NEXT_INSN (insn))
373 if (INSN_P (insn))
374 cselib_process_insn (insn);
376 nonequal = BITMAP_ALLOC (NULL);
377 CLEAR_REG_SET (nonequal);
379 /* Now assume that we've continued by the edge E to B and continue
380 processing as if it were same basic block.
381 Our goal is to prove that whole block is an NOOP. */
383 for (insn = NEXT_INSN (BB_HEAD (b));
384 insn != NEXT_INSN (BB_END (b)) && !failed;
385 insn = NEXT_INSN (insn))
387 if (INSN_P (insn))
389 rtx pat = PATTERN (insn);
391 if (GET_CODE (pat) == PARALLEL)
393 for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
394 failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
396 else
397 failed |= mark_effect (pat, nonequal);
400 cselib_process_insn (insn);
403 /* Later we should clear nonequal of dead registers. So far we don't
404 have life information in cfg_cleanup. */
405 if (failed)
407 b->flags |= BB_NONTHREADABLE_BLOCK;
408 goto failed_exit;
411 /* cond2 must not mention any register that is not equal to the
412 former block. */
413 if (mentions_nonequal_regs (cond2, nonequal))
414 goto failed_exit;
416 EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
417 goto failed_exit;
419 BITMAP_FREE (nonequal);
420 cselib_finish ();
421 if ((comparison_dominates_p (code1, code2) != 0)
422 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
423 return BRANCH_EDGE (b);
424 else
425 return FALLTHRU_EDGE (b);
427 failed_exit:
428 BITMAP_FREE (nonequal);
429 cselib_finish ();
430 return NULL;
433 /* Attempt to forward edges leaving basic block B.
434 Return true if successful. */
436 static bool
437 try_forward_edges (int mode, basic_block b)
439 bool changed = false;
440 edge_iterator ei;
441 edge e, *threaded_edges = NULL;
443 /* If we are partitioning hot/cold basic blocks, we don't want to
444 mess up unconditional or indirect jumps that cross between hot
445 and cold sections.
447 Basic block partitioning may result in some jumps that appear to
448 be optimizable (or blocks that appear to be mergeable), but which really
449 must be left untouched (they are required to make it safely across
450 partition boundaries). See the comments at the top of
451 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
453 if (JUMP_P (BB_END (b)) && CROSSING_JUMP_P (BB_END (b)))
454 return false;
456 for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
458 basic_block target, first;
459 location_t goto_locus;
460 int counter;
461 bool threaded = false;
462 int nthreaded_edges = 0;
463 bool may_thread = first_pass || (b->flags & BB_MODIFIED) != 0;
465 /* Skip complex edges because we don't know how to update them.
467 Still handle fallthru edges, as we can succeed to forward fallthru
468 edge to the same place as the branch edge of conditional branch
469 and turn conditional branch to an unconditional branch. */
470 if (e->flags & EDGE_COMPLEX)
472 ei_next (&ei);
473 continue;
476 target = first = e->dest;
477 counter = NUM_FIXED_BLOCKS;
478 goto_locus = e->goto_locus;
480 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
481 up jumps that cross between hot/cold sections.
483 Basic block partitioning may result in some jumps that appear
484 to be optimizable (or blocks that appear to be mergeable), but which
485 really must be left untouched (they are required to make it safely
486 across partition boundaries). See the comments at the top of
487 bb-reorder.c:partition_hot_cold_basic_blocks for complete
488 details. */
490 if (first != EXIT_BLOCK_PTR_FOR_FN (cfun)
491 && JUMP_P (BB_END (first))
492 && CROSSING_JUMP_P (BB_END (first)))
493 return changed;
495 while (counter < n_basic_blocks_for_fn (cfun))
497 basic_block new_target = NULL;
498 bool new_target_threaded = false;
499 may_thread |= (target->flags & BB_MODIFIED) != 0;
501 if (FORWARDER_BLOCK_P (target)
502 && !(single_succ_edge (target)->flags & EDGE_CROSSING)
503 && single_succ (target) != EXIT_BLOCK_PTR_FOR_FN (cfun))
505 /* Bypass trivial infinite loops. */
506 new_target = single_succ (target);
507 if (target == new_target)
508 counter = n_basic_blocks_for_fn (cfun);
509 else if (!optimize)
511 /* When not optimizing, ensure that edges or forwarder
512 blocks with different locus are not optimized out. */
513 location_t new_locus = single_succ_edge (target)->goto_locus;
514 location_t locus = goto_locus;
516 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
517 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
518 && new_locus != locus)
519 new_target = NULL;
520 else
522 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
523 locus = new_locus;
525 rtx_insn *last = BB_END (target);
526 if (DEBUG_INSN_P (last))
527 last = prev_nondebug_insn (last);
528 if (last && INSN_P (last))
529 new_locus = INSN_LOCATION (last);
530 else
531 new_locus = UNKNOWN_LOCATION;
533 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION
534 && LOCATION_LOCUS (locus) != UNKNOWN_LOCATION
535 && new_locus != locus)
536 new_target = NULL;
537 else
539 if (LOCATION_LOCUS (new_locus) != UNKNOWN_LOCATION)
540 locus = new_locus;
542 goto_locus = locus;
548 /* Allow to thread only over one edge at time to simplify updating
549 of probabilities. */
550 else if ((mode & CLEANUP_THREADING) && may_thread)
552 edge t = thread_jump (e, target);
553 if (t)
555 if (!threaded_edges)
556 threaded_edges = XNEWVEC (edge,
557 n_basic_blocks_for_fn (cfun));
558 else
560 int i;
562 /* Detect an infinite loop across blocks not
563 including the start block. */
564 for (i = 0; i < nthreaded_edges; ++i)
565 if (threaded_edges[i] == t)
566 break;
567 if (i < nthreaded_edges)
569 counter = n_basic_blocks_for_fn (cfun);
570 break;
574 /* Detect an infinite loop across the start block. */
575 if (t->dest == b)
576 break;
578 gcc_assert (nthreaded_edges
579 < (n_basic_blocks_for_fn (cfun)
580 - NUM_FIXED_BLOCKS));
581 threaded_edges[nthreaded_edges++] = t;
583 new_target = t->dest;
584 new_target_threaded = true;
588 if (!new_target)
589 break;
591 counter++;
592 target = new_target;
593 threaded |= new_target_threaded;
596 if (counter >= n_basic_blocks_for_fn (cfun))
598 if (dump_file)
599 fprintf (dump_file, "Infinite loop in BB %i.\n",
600 target->index);
602 else if (target == first)
603 ; /* We didn't do anything. */
604 else
606 /* Save the values now, as the edge may get removed. */
607 gcov_type edge_count = e->count;
608 int edge_probability = e->probability;
609 int edge_frequency;
610 int n = 0;
612 e->goto_locus = goto_locus;
614 /* Don't force if target is exit block. */
615 if (threaded && target != EXIT_BLOCK_PTR_FOR_FN (cfun))
617 notice_new_block (redirect_edge_and_branch_force (e, target));
618 if (dump_file)
619 fprintf (dump_file, "Conditionals threaded.\n");
621 else if (!redirect_edge_and_branch (e, target))
623 if (dump_file)
624 fprintf (dump_file,
625 "Forwarding edge %i->%i to %i failed.\n",
626 b->index, e->dest->index, target->index);
627 ei_next (&ei);
628 continue;
631 /* We successfully forwarded the edge. Now update profile
632 data: for each edge we traversed in the chain, remove
633 the original edge's execution count. */
634 edge_frequency = apply_probability (b->frequency, edge_probability);
638 edge t;
640 if (!single_succ_p (first))
642 gcc_assert (n < nthreaded_edges);
643 t = threaded_edges [n++];
644 gcc_assert (t->src == first);
645 update_bb_profile_for_threading (first, edge_frequency,
646 edge_count, t);
647 update_br_prob_note (first);
649 else
651 first->count -= edge_count;
652 if (first->count < 0)
653 first->count = 0;
654 first->frequency -= edge_frequency;
655 if (first->frequency < 0)
656 first->frequency = 0;
657 /* It is possible that as the result of
658 threading we've removed edge as it is
659 threaded to the fallthru edge. Avoid
660 getting out of sync. */
661 if (n < nthreaded_edges
662 && first == threaded_edges [n]->src)
663 n++;
664 t = single_succ_edge (first);
667 t->count -= edge_count;
668 if (t->count < 0)
669 t->count = 0;
670 first = t->dest;
672 while (first != target);
674 changed = true;
675 continue;
677 ei_next (&ei);
680 free (threaded_edges);
681 return changed;
685 /* Blocks A and B are to be merged into a single block. A has no incoming
686 fallthru edge, so it can be moved before B without adding or modifying
687 any jumps (aside from the jump from A to B). */
689 static void
690 merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
692 rtx_insn *barrier;
694 /* If we are partitioning hot/cold basic blocks, we don't want to
695 mess up unconditional or indirect jumps that cross between hot
696 and cold sections.
698 Basic block partitioning may result in some jumps that appear to
699 be optimizable (or blocks that appear to be mergeable), but which really
700 must be left untouched (they are required to make it safely across
701 partition boundaries). See the comments at the top of
702 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
704 if (BB_PARTITION (a) != BB_PARTITION (b))
705 return;
707 barrier = next_nonnote_insn (BB_END (a));
708 gcc_assert (BARRIER_P (barrier));
709 delete_insn (barrier);
711 /* Scramble the insn chain. */
712 if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
713 reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
714 df_set_bb_dirty (a);
716 if (dump_file)
717 fprintf (dump_file, "Moved block %d before %d and merged.\n",
718 a->index, b->index);
720 /* Swap the records for the two blocks around. */
722 unlink_block (a);
723 link_block (a, b->prev_bb);
725 /* Now blocks A and B are contiguous. Merge them. */
726 merge_blocks (a, b);
729 /* Blocks A and B are to be merged into a single block. B has no outgoing
730 fallthru edge, so it can be moved after A without adding or modifying
731 any jumps (aside from the jump from A to B). */
733 static void
734 merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
736 rtx_insn *barrier, *real_b_end;
737 rtx label;
738 rtx_jump_table_data *table;
740 /* If we are partitioning hot/cold basic blocks, we don't want to
741 mess up unconditional or indirect jumps that cross between hot
742 and cold sections.
744 Basic block partitioning may result in some jumps that appear to
745 be optimizable (or blocks that appear to be mergeable), but which really
746 must be left untouched (they are required to make it safely across
747 partition boundaries). See the comments at the top of
748 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
750 if (BB_PARTITION (a) != BB_PARTITION (b))
751 return;
753 real_b_end = BB_END (b);
755 /* If there is a jump table following block B temporarily add the jump table
756 to block B so that it will also be moved to the correct location. */
757 if (tablejump_p (BB_END (b), &label, &table)
758 && prev_active_insn (label) == BB_END (b))
760 BB_END (b) = table;
763 /* There had better have been a barrier there. Delete it. */
764 barrier = NEXT_INSN (BB_END (b));
765 if (barrier && BARRIER_P (barrier))
766 delete_insn (barrier);
769 /* Scramble the insn chain. */
770 reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
772 /* Restore the real end of b. */
773 BB_END (b) = real_b_end;
775 if (dump_file)
776 fprintf (dump_file, "Moved block %d after %d and merged.\n",
777 b->index, a->index);
779 /* Now blocks A and B are contiguous. Merge them. */
780 merge_blocks (a, b);
783 /* Attempt to merge basic blocks that are potentially non-adjacent.
784 Return NULL iff the attempt failed, otherwise return basic block
785 where cleanup_cfg should continue. Because the merging commonly
786 moves basic block away or introduces another optimization
787 possibility, return basic block just before B so cleanup_cfg don't
788 need to iterate.
790 It may be good idea to return basic block before C in the case
791 C has been moved after B and originally appeared earlier in the
792 insn sequence, but we have no information available about the
793 relative ordering of these two. Hopefully it is not too common. */
795 static basic_block
796 merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
798 basic_block next;
800 /* If we are partitioning hot/cold basic blocks, we don't want to
801 mess up unconditional or indirect jumps that cross between hot
802 and cold sections.
804 Basic block partitioning may result in some jumps that appear to
805 be optimizable (or blocks that appear to be mergeable), but which really
806 must be left untouched (they are required to make it safely across
807 partition boundaries). See the comments at the top of
808 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
810 if (BB_PARTITION (b) != BB_PARTITION (c))
811 return NULL;
813 /* If B has a fallthru edge to C, no need to move anything. */
814 if (e->flags & EDGE_FALLTHRU)
816 int b_index = b->index, c_index = c->index;
818 /* Protect the loop latches. */
819 if (current_loops && c->loop_father->latch == c)
820 return NULL;
822 merge_blocks (b, c);
823 update_forwarder_flag (b);
825 if (dump_file)
826 fprintf (dump_file, "Merged %d and %d without moving.\n",
827 b_index, c_index);
829 return b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? b : b->prev_bb;
832 /* Otherwise we will need to move code around. Do that only if expensive
833 transformations are allowed. */
834 else if (mode & CLEANUP_EXPENSIVE)
836 edge tmp_edge, b_fallthru_edge;
837 bool c_has_outgoing_fallthru;
838 bool b_has_incoming_fallthru;
840 /* Avoid overactive code motion, as the forwarder blocks should be
841 eliminated by edge redirection instead. One exception might have
842 been if B is a forwarder block and C has no fallthru edge, but
843 that should be cleaned up by bb-reorder instead. */
844 if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
845 return NULL;
847 /* We must make sure to not munge nesting of lexical blocks,
848 and loop notes. This is done by squeezing out all the notes
849 and leaving them there to lie. Not ideal, but functional. */
851 tmp_edge = find_fallthru_edge (c->succs);
852 c_has_outgoing_fallthru = (tmp_edge != NULL);
854 tmp_edge = find_fallthru_edge (b->preds);
855 b_has_incoming_fallthru = (tmp_edge != NULL);
856 b_fallthru_edge = tmp_edge;
857 next = b->prev_bb;
858 if (next == c)
859 next = next->prev_bb;
861 /* Otherwise, we're going to try to move C after B. If C does
862 not have an outgoing fallthru, then it can be moved
863 immediately after B without introducing or modifying jumps. */
864 if (! c_has_outgoing_fallthru)
866 merge_blocks_move_successor_nojumps (b, c);
867 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
870 /* If B does not have an incoming fallthru, then it can be moved
871 immediately before C without introducing or modifying jumps.
872 C cannot be the first block, so we do not have to worry about
873 accessing a non-existent block. */
875 if (b_has_incoming_fallthru)
877 basic_block bb;
879 if (b_fallthru_edge->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
880 return NULL;
881 bb = force_nonfallthru (b_fallthru_edge);
882 if (bb)
883 notice_new_block (bb);
886 merge_blocks_move_predecessor_nojumps (b, c);
887 return next == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? next->next_bb : next;
890 return NULL;
894 /* Removes the memory attributes of MEM expression
895 if they are not equal. */
897 static void
898 merge_memattrs (rtx x, rtx y)
900 int i;
901 int j;
902 enum rtx_code code;
903 const char *fmt;
905 if (x == y)
906 return;
907 if (x == 0 || y == 0)
908 return;
910 code = GET_CODE (x);
912 if (code != GET_CODE (y))
913 return;
915 if (GET_MODE (x) != GET_MODE (y))
916 return;
918 if (code == MEM && !mem_attrs_eq_p (MEM_ATTRS (x), MEM_ATTRS (y)))
920 if (! MEM_ATTRS (x))
921 MEM_ATTRS (y) = 0;
922 else if (! MEM_ATTRS (y))
923 MEM_ATTRS (x) = 0;
924 else
926 HOST_WIDE_INT mem_size;
928 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
930 set_mem_alias_set (x, 0);
931 set_mem_alias_set (y, 0);
934 if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
936 set_mem_expr (x, 0);
937 set_mem_expr (y, 0);
938 clear_mem_offset (x);
939 clear_mem_offset (y);
941 else if (MEM_OFFSET_KNOWN_P (x) != MEM_OFFSET_KNOWN_P (y)
942 || (MEM_OFFSET_KNOWN_P (x)
943 && MEM_OFFSET (x) != MEM_OFFSET (y)))
945 clear_mem_offset (x);
946 clear_mem_offset (y);
949 if (MEM_SIZE_KNOWN_P (x) && MEM_SIZE_KNOWN_P (y))
951 mem_size = MAX (MEM_SIZE (x), MEM_SIZE (y));
952 set_mem_size (x, mem_size);
953 set_mem_size (y, mem_size);
955 else
957 clear_mem_size (x);
958 clear_mem_size (y);
961 set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
962 set_mem_align (y, MEM_ALIGN (x));
965 if (code == MEM)
967 if (MEM_READONLY_P (x) != MEM_READONLY_P (y))
969 MEM_READONLY_P (x) = 0;
970 MEM_READONLY_P (y) = 0;
972 if (MEM_NOTRAP_P (x) != MEM_NOTRAP_P (y))
974 MEM_NOTRAP_P (x) = 0;
975 MEM_NOTRAP_P (y) = 0;
977 if (MEM_VOLATILE_P (x) != MEM_VOLATILE_P (y))
979 MEM_VOLATILE_P (x) = 1;
980 MEM_VOLATILE_P (y) = 1;
984 fmt = GET_RTX_FORMAT (code);
985 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
987 switch (fmt[i])
989 case 'E':
990 /* Two vectors must have the same length. */
991 if (XVECLEN (x, i) != XVECLEN (y, i))
992 return;
994 for (j = 0; j < XVECLEN (x, i); j++)
995 merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
997 break;
999 case 'e':
1000 merge_memattrs (XEXP (x, i), XEXP (y, i));
1003 return;
1007 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
1008 different single sets S1 and S2. */
1010 static bool
1011 equal_different_set_p (rtx p1, rtx s1, rtx p2, rtx s2)
1013 int i;
1014 rtx e1, e2;
1016 if (p1 == s1 && p2 == s2)
1017 return true;
1019 if (GET_CODE (p1) != PARALLEL || GET_CODE (p2) != PARALLEL)
1020 return false;
1022 if (XVECLEN (p1, 0) != XVECLEN (p2, 0))
1023 return false;
1025 for (i = 0; i < XVECLEN (p1, 0); i++)
1027 e1 = XVECEXP (p1, 0, i);
1028 e2 = XVECEXP (p2, 0, i);
1029 if (e1 == s1 && e2 == s2)
1030 continue;
1031 if (reload_completed
1032 ? rtx_renumbered_equal_p (e1, e2) : rtx_equal_p (e1, e2))
1033 continue;
1035 return false;
1038 return true;
1042 /* NOTE1 is the REG_EQUAL note, if any, attached to an insn
1043 that is a single_set with a SET_SRC of SRC1. Similarly
1044 for NOTE2/SRC2.
1046 So effectively NOTE1/NOTE2 are an alternate form of
1047 SRC1/SRC2 respectively.
1049 Return nonzero if SRC1 or NOTE1 has the same constant
1050 integer value as SRC2 or NOTE2. Else return zero. */
1051 static int
1052 values_equal_p (rtx note1, rtx note2, rtx src1, rtx src2)
1054 if (note1
1055 && note2
1056 && CONST_INT_P (XEXP (note1, 0))
1057 && rtx_equal_p (XEXP (note1, 0), XEXP (note2, 0)))
1058 return 1;
1060 if (!note1
1061 && !note2
1062 && CONST_INT_P (src1)
1063 && CONST_INT_P (src2)
1064 && rtx_equal_p (src1, src2))
1065 return 1;
1067 if (note1
1068 && CONST_INT_P (src2)
1069 && rtx_equal_p (XEXP (note1, 0), src2))
1070 return 1;
1072 if (note2
1073 && CONST_INT_P (src1)
1074 && rtx_equal_p (XEXP (note2, 0), src1))
1075 return 1;
1077 return 0;
1080 /* Examine register notes on I1 and I2 and return:
1081 - dir_forward if I1 can be replaced by I2, or
1082 - dir_backward if I2 can be replaced by I1, or
1083 - dir_both if both are the case. */
1085 static enum replace_direction
1086 can_replace_by (rtx_insn *i1, rtx_insn *i2)
1088 rtx s1, s2, d1, d2, src1, src2, note1, note2;
1089 bool c1, c2;
1091 /* Check for 2 sets. */
1092 s1 = single_set (i1);
1093 s2 = single_set (i2);
1094 if (s1 == NULL_RTX || s2 == NULL_RTX)
1095 return dir_none;
1097 /* Check that the 2 sets set the same dest. */
1098 d1 = SET_DEST (s1);
1099 d2 = SET_DEST (s2);
1100 if (!(reload_completed
1101 ? rtx_renumbered_equal_p (d1, d2) : rtx_equal_p (d1, d2)))
1102 return dir_none;
1104 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1105 set dest to the same value. */
1106 note1 = find_reg_equal_equiv_note (i1);
1107 note2 = find_reg_equal_equiv_note (i2);
1109 src1 = SET_SRC (s1);
1110 src2 = SET_SRC (s2);
1112 if (!values_equal_p (note1, note2, src1, src2))
1113 return dir_none;
1115 if (!equal_different_set_p (PATTERN (i1), s1, PATTERN (i2), s2))
1116 return dir_none;
1118 /* Although the 2 sets set dest to the same value, we cannot replace
1119 (set (dest) (const_int))
1121 (set (dest) (reg))
1122 because we don't know if the reg is live and has the same value at the
1123 location of replacement. */
1124 c1 = CONST_INT_P (src1);
1125 c2 = CONST_INT_P (src2);
1126 if (c1 && c2)
1127 return dir_both;
1128 else if (c2)
1129 return dir_forward;
1130 else if (c1)
1131 return dir_backward;
1133 return dir_none;
1136 /* Merges directions A and B. */
1138 static enum replace_direction
1139 merge_dir (enum replace_direction a, enum replace_direction b)
1141 /* Implements the following table:
1142 |bo fw bw no
1143 ---+-----------
1144 bo |bo fw bw no
1145 fw |-- fw no no
1146 bw |-- -- bw no
1147 no |-- -- -- no. */
1149 if (a == b)
1150 return a;
1152 if (a == dir_both)
1153 return b;
1154 if (b == dir_both)
1155 return a;
1157 return dir_none;
1160 /* Examine I1 and I2 and return:
1161 - dir_forward if I1 can be replaced by I2, or
1162 - dir_backward if I2 can be replaced by I1, or
1163 - dir_both if both are the case. */
1165 static enum replace_direction
1166 old_insns_match_p (int mode ATTRIBUTE_UNUSED, rtx_insn *i1, rtx_insn *i2)
1168 rtx p1, p2;
1170 /* Verify that I1 and I2 are equivalent. */
1171 if (GET_CODE (i1) != GET_CODE (i2))
1172 return dir_none;
1174 /* __builtin_unreachable() may lead to empty blocks (ending with
1175 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1176 if (NOTE_INSN_BASIC_BLOCK_P (i1) && NOTE_INSN_BASIC_BLOCK_P (i2))
1177 return dir_both;
1179 /* ??? Do not allow cross-jumping between different stack levels. */
1180 p1 = find_reg_note (i1, REG_ARGS_SIZE, NULL);
1181 p2 = find_reg_note (i2, REG_ARGS_SIZE, NULL);
1182 if (p1 && p2)
1184 p1 = XEXP (p1, 0);
1185 p2 = XEXP (p2, 0);
1186 if (!rtx_equal_p (p1, p2))
1187 return dir_none;
1189 /* ??? Worse, this adjustment had better be constant lest we
1190 have differing incoming stack levels. */
1191 if (!frame_pointer_needed
1192 && find_args_size_adjust (i1) == HOST_WIDE_INT_MIN)
1193 return dir_none;
1195 else if (p1 || p2)
1196 return dir_none;
1198 p1 = PATTERN (i1);
1199 p2 = PATTERN (i2);
1201 if (GET_CODE (p1) != GET_CODE (p2))
1202 return dir_none;
1204 /* If this is a CALL_INSN, compare register usage information.
1205 If we don't check this on stack register machines, the two
1206 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1207 numbers of stack registers in the same basic block.
1208 If we don't check this on machines with delay slots, a delay slot may
1209 be filled that clobbers a parameter expected by the subroutine.
1211 ??? We take the simple route for now and assume that if they're
1212 equal, they were constructed identically.
1214 Also check for identical exception regions. */
1216 if (CALL_P (i1))
1218 /* Ensure the same EH region. */
1219 rtx n1 = find_reg_note (i1, REG_EH_REGION, 0);
1220 rtx n2 = find_reg_note (i2, REG_EH_REGION, 0);
1222 if (!n1 && n2)
1223 return dir_none;
1225 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1226 return dir_none;
1228 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
1229 CALL_INSN_FUNCTION_USAGE (i2))
1230 || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2))
1231 return dir_none;
1233 /* For address sanitizer, never crossjump __asan_report_* builtins,
1234 otherwise errors might be reported on incorrect lines. */
1235 if (flag_sanitize & SANITIZE_ADDRESS)
1237 rtx call = get_call_rtx_from (i1);
1238 if (call && GET_CODE (XEXP (XEXP (call, 0), 0)) == SYMBOL_REF)
1240 rtx symbol = XEXP (XEXP (call, 0), 0);
1241 if (SYMBOL_REF_DECL (symbol)
1242 && TREE_CODE (SYMBOL_REF_DECL (symbol)) == FUNCTION_DECL)
1244 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol))
1245 == BUILT_IN_NORMAL)
1246 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1247 >= BUILT_IN_ASAN_REPORT_LOAD1
1248 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol))
1249 <= BUILT_IN_ASAN_STOREN)
1250 return dir_none;
1256 #ifdef STACK_REGS
1257 /* If cross_jump_death_matters is not 0, the insn's mode
1258 indicates whether or not the insn contains any stack-like
1259 regs. */
1261 if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
1263 /* If register stack conversion has already been done, then
1264 death notes must also be compared before it is certain that
1265 the two instruction streams match. */
1267 rtx note;
1268 HARD_REG_SET i1_regset, i2_regset;
1270 CLEAR_HARD_REG_SET (i1_regset);
1271 CLEAR_HARD_REG_SET (i2_regset);
1273 for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
1274 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1275 SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
1277 for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
1278 if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
1279 SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
1281 if (!hard_reg_set_equal_p (i1_regset, i2_regset))
1282 return dir_none;
1284 #endif
1286 if (reload_completed
1287 ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
1288 return dir_both;
1290 return can_replace_by (i1, i2);
1293 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1294 flow_find_head_matching_sequence, ensure the notes match. */
1296 static void
1297 merge_notes (rtx_insn *i1, rtx_insn *i2)
1299 /* If the merged insns have different REG_EQUAL notes, then
1300 remove them. */
1301 rtx equiv1 = find_reg_equal_equiv_note (i1);
1302 rtx equiv2 = find_reg_equal_equiv_note (i2);
1304 if (equiv1 && !equiv2)
1305 remove_note (i1, equiv1);
1306 else if (!equiv1 && equiv2)
1307 remove_note (i2, equiv2);
1308 else if (equiv1 && equiv2
1309 && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
1311 remove_note (i1, equiv1);
1312 remove_note (i2, equiv2);
1316 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1317 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1318 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1319 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1320 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1322 static void
1323 walk_to_nondebug_insn (rtx_insn **i1, basic_block *bb1, bool follow_fallthru,
1324 bool *did_fallthru)
1326 edge fallthru;
1328 *did_fallthru = false;
1330 /* Ignore notes. */
1331 while (!NONDEBUG_INSN_P (*i1))
1333 if (*i1 != BB_HEAD (*bb1))
1335 *i1 = PREV_INSN (*i1);
1336 continue;
1339 if (!follow_fallthru)
1340 return;
1342 fallthru = find_fallthru_edge ((*bb1)->preds);
1343 if (!fallthru || fallthru->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)
1344 || !single_succ_p (fallthru->src))
1345 return;
1347 *bb1 = fallthru->src;
1348 *i1 = BB_END (*bb1);
1349 *did_fallthru = true;
1353 /* Look through the insns at the end of BB1 and BB2 and find the longest
1354 sequence that are either equivalent, or allow forward or backward
1355 replacement. Store the first insns for that sequence in *F1 and *F2 and
1356 return the sequence length.
1358 DIR_P indicates the allowed replacement direction on function entry, and
1359 the actual replacement direction on function exit. If NULL, only equivalent
1360 sequences are allowed.
1362 To simplify callers of this function, if the blocks match exactly,
1363 store the head of the blocks in *F1 and *F2. */
1366 flow_find_cross_jump (basic_block bb1, basic_block bb2, rtx_insn **f1,
1367 rtx_insn **f2, enum replace_direction *dir_p)
1369 rtx_insn *i1, *i2, *last1, *last2, *afterlast1, *afterlast2;
1370 int ninsns = 0;
1371 enum replace_direction dir, last_dir, afterlast_dir;
1372 bool follow_fallthru, did_fallthru;
1374 if (dir_p)
1375 dir = *dir_p;
1376 else
1377 dir = dir_both;
1378 afterlast_dir = dir;
1379 last_dir = afterlast_dir;
1381 /* Skip simple jumps at the end of the blocks. Complex jumps still
1382 need to be compared for equivalence, which we'll do below. */
1384 i1 = BB_END (bb1);
1385 last1 = afterlast1 = last2 = afterlast2 = NULL;
1386 if (onlyjump_p (i1)
1387 || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
1389 last1 = i1;
1390 i1 = PREV_INSN (i1);
1393 i2 = BB_END (bb2);
1394 if (onlyjump_p (i2)
1395 || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
1397 last2 = i2;
1398 /* Count everything except for unconditional jump as insn.
1399 Don't count any jumps if dir_p is NULL. */
1400 if (!simplejump_p (i2) && !returnjump_p (i2) && last1 && dir_p)
1401 ninsns++;
1402 i2 = PREV_INSN (i2);
1405 while (true)
1407 /* In the following example, we can replace all jumps to C by jumps to A.
1409 This removes 4 duplicate insns.
1410 [bb A] insn1 [bb C] insn1
1411 insn2 insn2
1412 [bb B] insn3 insn3
1413 insn4 insn4
1414 jump_insn jump_insn
1416 We could also replace all jumps to A by jumps to C, but that leaves B
1417 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1418 step, all jumps to B would be replaced with jumps to the middle of C,
1419 achieving the same result with more effort.
1420 So we allow only the first possibility, which means that we don't allow
1421 fallthru in the block that's being replaced. */
1423 follow_fallthru = dir_p && dir != dir_forward;
1424 walk_to_nondebug_insn (&i1, &bb1, follow_fallthru, &did_fallthru);
1425 if (did_fallthru)
1426 dir = dir_backward;
1428 follow_fallthru = dir_p && dir != dir_backward;
1429 walk_to_nondebug_insn (&i2, &bb2, follow_fallthru, &did_fallthru);
1430 if (did_fallthru)
1431 dir = dir_forward;
1433 if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
1434 break;
1436 dir = merge_dir (dir, old_insns_match_p (0, i1, i2));
1437 if (dir == dir_none || (!dir_p && dir != dir_both))
1438 break;
1440 merge_memattrs (i1, i2);
1442 /* Don't begin a cross-jump with a NOTE insn. */
1443 if (INSN_P (i1))
1445 merge_notes (i1, i2);
1447 afterlast1 = last1, afterlast2 = last2;
1448 last1 = i1, last2 = i2;
1449 afterlast_dir = last_dir;
1450 last_dir = dir;
1451 if (active_insn_p (i1))
1452 ninsns++;
1455 i1 = PREV_INSN (i1);
1456 i2 = PREV_INSN (i2);
1459 /* Don't allow the insn after a compare to be shared by
1460 cross-jumping unless the compare is also shared. */
1461 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1462 && ! sets_cc0_p (last1))
1463 last1 = afterlast1, last2 = afterlast2, last_dir = afterlast_dir, ninsns--;
1465 /* Include preceding notes and labels in the cross-jump. One,
1466 this may bring us to the head of the blocks as requested above.
1467 Two, it keeps line number notes as matched as may be. */
1468 if (ninsns)
1470 bb1 = BLOCK_FOR_INSN (last1);
1471 while (last1 != BB_HEAD (bb1) && !NONDEBUG_INSN_P (PREV_INSN (last1)))
1472 last1 = PREV_INSN (last1);
1474 if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
1475 last1 = PREV_INSN (last1);
1477 bb2 = BLOCK_FOR_INSN (last2);
1478 while (last2 != BB_HEAD (bb2) && !NONDEBUG_INSN_P (PREV_INSN (last2)))
1479 last2 = PREV_INSN (last2);
1481 if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
1482 last2 = PREV_INSN (last2);
1484 *f1 = last1;
1485 *f2 = last2;
1488 if (dir_p)
1489 *dir_p = last_dir;
1490 return ninsns;
1493 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1494 the head of the two blocks. Do not include jumps at the end.
1495 If STOP_AFTER is nonzero, stop after finding that many matching
1496 instructions. If STOP_AFTER is zero, count all INSN_P insns, if it is
1497 non-zero, only count active insns. */
1500 flow_find_head_matching_sequence (basic_block bb1, basic_block bb2, rtx_insn **f1,
1501 rtx_insn **f2, int stop_after)
1503 rtx_insn *i1, *i2, *last1, *last2, *beforelast1, *beforelast2;
1504 int ninsns = 0;
1505 edge e;
1506 edge_iterator ei;
1507 int nehedges1 = 0, nehedges2 = 0;
1509 FOR_EACH_EDGE (e, ei, bb1->succs)
1510 if (e->flags & EDGE_EH)
1511 nehedges1++;
1512 FOR_EACH_EDGE (e, ei, bb2->succs)
1513 if (e->flags & EDGE_EH)
1514 nehedges2++;
1516 i1 = BB_HEAD (bb1);
1517 i2 = BB_HEAD (bb2);
1518 last1 = beforelast1 = last2 = beforelast2 = NULL;
1520 while (true)
1522 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1523 while (!NONDEBUG_INSN_P (i1) && i1 != BB_END (bb1))
1525 if (NOTE_P (i1) && NOTE_KIND (i1) == NOTE_INSN_EPILOGUE_BEG)
1526 break;
1527 i1 = NEXT_INSN (i1);
1530 while (!NONDEBUG_INSN_P (i2) && i2 != BB_END (bb2))
1532 if (NOTE_P (i2) && NOTE_KIND (i2) == NOTE_INSN_EPILOGUE_BEG)
1533 break;
1534 i2 = NEXT_INSN (i2);
1537 if ((i1 == BB_END (bb1) && !NONDEBUG_INSN_P (i1))
1538 || (i2 == BB_END (bb2) && !NONDEBUG_INSN_P (i2)))
1539 break;
1541 if (NOTE_P (i1) || NOTE_P (i2)
1542 || JUMP_P (i1) || JUMP_P (i2))
1543 break;
1545 /* A sanity check to make sure we're not merging insns with different
1546 effects on EH. If only one of them ends a basic block, it shouldn't
1547 have an EH edge; if both end a basic block, there should be the same
1548 number of EH edges. */
1549 if ((i1 == BB_END (bb1) && i2 != BB_END (bb2)
1550 && nehedges1 > 0)
1551 || (i2 == BB_END (bb2) && i1 != BB_END (bb1)
1552 && nehedges2 > 0)
1553 || (i1 == BB_END (bb1) && i2 == BB_END (bb2)
1554 && nehedges1 != nehedges2))
1555 break;
1557 if (old_insns_match_p (0, i1, i2) != dir_both)
1558 break;
1560 merge_memattrs (i1, i2);
1562 /* Don't begin a cross-jump with a NOTE insn. */
1563 if (INSN_P (i1))
1565 merge_notes (i1, i2);
1567 beforelast1 = last1, beforelast2 = last2;
1568 last1 = i1, last2 = i2;
1569 if (!stop_after || active_insn_p (i1))
1570 ninsns++;
1573 if (i1 == BB_END (bb1) || i2 == BB_END (bb2)
1574 || (stop_after > 0 && ninsns == stop_after))
1575 break;
1577 i1 = NEXT_INSN (i1);
1578 i2 = NEXT_INSN (i2);
1581 /* Don't allow a compare to be shared by cross-jumping unless the insn
1582 after the compare is also shared. */
1583 if (HAVE_cc0 && ninsns && reg_mentioned_p (cc0_rtx, last1)
1584 && sets_cc0_p (last1))
1585 last1 = beforelast1, last2 = beforelast2, ninsns--;
1587 if (ninsns)
1589 *f1 = last1;
1590 *f2 = last2;
1593 return ninsns;
1596 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1597 the branch instruction. This means that if we commonize the control
1598 flow before end of the basic block, the semantic remains unchanged.
1600 We may assume that there exists one edge with a common destination. */
1602 static bool
1603 outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
1605 int nehedges1 = 0, nehedges2 = 0;
1606 edge fallthru1 = 0, fallthru2 = 0;
1607 edge e1, e2;
1608 edge_iterator ei;
1610 /* If we performed shrink-wrapping, edges to the exit block can
1611 only be distinguished for JUMP_INSNs. The two paths may differ in
1612 whether they went through the prologue. Sibcalls are fine, we know
1613 that we either didn't need or inserted an epilogue before them. */
1614 if (crtl->shrink_wrapped
1615 && single_succ_p (bb1)
1616 && single_succ (bb1) == EXIT_BLOCK_PTR_FOR_FN (cfun)
1617 && !JUMP_P (BB_END (bb1))
1618 && !(CALL_P (BB_END (bb1)) && SIBLING_CALL_P (BB_END (bb1))))
1619 return false;
1621 /* If BB1 has only one successor, we may be looking at either an
1622 unconditional jump, or a fake edge to exit. */
1623 if (single_succ_p (bb1)
1624 && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1625 && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
1626 return (single_succ_p (bb2)
1627 && (single_succ_edge (bb2)->flags
1628 & (EDGE_COMPLEX | EDGE_FAKE)) == 0
1629 && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
1631 /* Match conditional jumps - this may get tricky when fallthru and branch
1632 edges are crossed. */
1633 if (EDGE_COUNT (bb1->succs) == 2
1634 && any_condjump_p (BB_END (bb1))
1635 && onlyjump_p (BB_END (bb1)))
1637 edge b1, f1, b2, f2;
1638 bool reverse, match;
1639 rtx set1, set2, cond1, cond2;
1640 enum rtx_code code1, code2;
1642 if (EDGE_COUNT (bb2->succs) != 2
1643 || !any_condjump_p (BB_END (bb2))
1644 || !onlyjump_p (BB_END (bb2)))
1645 return false;
1647 b1 = BRANCH_EDGE (bb1);
1648 b2 = BRANCH_EDGE (bb2);
1649 f1 = FALLTHRU_EDGE (bb1);
1650 f2 = FALLTHRU_EDGE (bb2);
1652 /* Get around possible forwarders on fallthru edges. Other cases
1653 should be optimized out already. */
1654 if (FORWARDER_BLOCK_P (f1->dest))
1655 f1 = single_succ_edge (f1->dest);
1657 if (FORWARDER_BLOCK_P (f2->dest))
1658 f2 = single_succ_edge (f2->dest);
1660 /* To simplify use of this function, return false if there are
1661 unneeded forwarder blocks. These will get eliminated later
1662 during cleanup_cfg. */
1663 if (FORWARDER_BLOCK_P (f1->dest)
1664 || FORWARDER_BLOCK_P (f2->dest)
1665 || FORWARDER_BLOCK_P (b1->dest)
1666 || FORWARDER_BLOCK_P (b2->dest))
1667 return false;
1669 if (f1->dest == f2->dest && b1->dest == b2->dest)
1670 reverse = false;
1671 else if (f1->dest == b2->dest && b1->dest == f2->dest)
1672 reverse = true;
1673 else
1674 return false;
1676 set1 = pc_set (BB_END (bb1));
1677 set2 = pc_set (BB_END (bb2));
1678 if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
1679 != (XEXP (SET_SRC (set2), 1) == pc_rtx))
1680 reverse = !reverse;
1682 cond1 = XEXP (SET_SRC (set1), 0);
1683 cond2 = XEXP (SET_SRC (set2), 0);
1684 code1 = GET_CODE (cond1);
1685 if (reverse)
1686 code2 = reversed_comparison_code (cond2, BB_END (bb2));
1687 else
1688 code2 = GET_CODE (cond2);
1690 if (code2 == UNKNOWN)
1691 return false;
1693 /* Verify codes and operands match. */
1694 match = ((code1 == code2
1695 && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
1696 && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
1697 || (code1 == swap_condition (code2)
1698 && rtx_renumbered_equal_p (XEXP (cond1, 1),
1699 XEXP (cond2, 0))
1700 && rtx_renumbered_equal_p (XEXP (cond1, 0),
1701 XEXP (cond2, 1))));
1703 /* If we return true, we will join the blocks. Which means that
1704 we will only have one branch prediction bit to work with. Thus
1705 we require the existing branches to have probabilities that are
1706 roughly similar. */
1707 if (match
1708 && optimize_bb_for_speed_p (bb1)
1709 && optimize_bb_for_speed_p (bb2))
1711 int prob2;
1713 if (b1->dest == b2->dest)
1714 prob2 = b2->probability;
1715 else
1716 /* Do not use f2 probability as f2 may be forwarded. */
1717 prob2 = REG_BR_PROB_BASE - b2->probability;
1719 /* Fail if the difference in probabilities is greater than 50%.
1720 This rules out two well-predicted branches with opposite
1721 outcomes. */
1722 if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
1724 if (dump_file)
1725 fprintf (dump_file,
1726 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1727 bb1->index, bb2->index, b1->probability, prob2);
1729 return false;
1733 if (dump_file && match)
1734 fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
1735 bb1->index, bb2->index);
1737 return match;
1740 /* Generic case - we are seeing a computed jump, table jump or trapping
1741 instruction. */
1743 /* Check whether there are tablejumps in the end of BB1 and BB2.
1744 Return true if they are identical. */
1746 rtx label1, label2;
1747 rtx_jump_table_data *table1, *table2;
1749 if (tablejump_p (BB_END (bb1), &label1, &table1)
1750 && tablejump_p (BB_END (bb2), &label2, &table2)
1751 && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
1753 /* The labels should never be the same rtx. If they really are same
1754 the jump tables are same too. So disable crossjumping of blocks BB1
1755 and BB2 because when deleting the common insns in the end of BB1
1756 by delete_basic_block () the jump table would be deleted too. */
1757 /* If LABEL2 is referenced in BB1->END do not do anything
1758 because we would loose information when replacing
1759 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1760 if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
1762 /* Set IDENTICAL to true when the tables are identical. */
1763 bool identical = false;
1764 rtx p1, p2;
1766 p1 = PATTERN (table1);
1767 p2 = PATTERN (table2);
1768 if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
1770 identical = true;
1772 else if (GET_CODE (p1) == ADDR_DIFF_VEC
1773 && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
1774 && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
1775 && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
1777 int i;
1779 identical = true;
1780 for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
1781 if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
1782 identical = false;
1785 if (identical)
1787 bool match;
1789 /* Temporarily replace references to LABEL1 with LABEL2
1790 in BB1->END so that we could compare the instructions. */
1791 replace_label_in_insn (BB_END (bb1), label1, label2, false);
1793 match = (old_insns_match_p (mode, BB_END (bb1), BB_END (bb2))
1794 == dir_both);
1795 if (dump_file && match)
1796 fprintf (dump_file,
1797 "Tablejumps in bb %i and %i match.\n",
1798 bb1->index, bb2->index);
1800 /* Set the original label in BB1->END because when deleting
1801 a block whose end is a tablejump, the tablejump referenced
1802 from the instruction is deleted too. */
1803 replace_label_in_insn (BB_END (bb1), label2, label1, false);
1805 return match;
1808 return false;
1812 /* Find the last non-debug non-note instruction in each bb, except
1813 stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
1814 handles that case specially. old_insns_match_p does not handle
1815 other types of instruction notes. */
1816 rtx_insn *last1 = BB_END (bb1);
1817 rtx_insn *last2 = BB_END (bb2);
1818 while (!NOTE_INSN_BASIC_BLOCK_P (last1) &&
1819 (DEBUG_INSN_P (last1) || NOTE_P (last1)))
1820 last1 = PREV_INSN (last1);
1821 while (!NOTE_INSN_BASIC_BLOCK_P (last2) &&
1822 (DEBUG_INSN_P (last2) || NOTE_P (last2)))
1823 last2 = PREV_INSN (last2);
1824 gcc_assert (last1 && last2);
1826 /* First ensure that the instructions match. There may be many outgoing
1827 edges so this test is generally cheaper. */
1828 if (old_insns_match_p (mode, last1, last2) != dir_both)
1829 return false;
1831 /* Search the outgoing edges, ensure that the counts do match, find possible
1832 fallthru and exception handling edges since these needs more
1833 validation. */
1834 if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
1835 return false;
1837 bool nonfakeedges = false;
1838 FOR_EACH_EDGE (e1, ei, bb1->succs)
1840 e2 = EDGE_SUCC (bb2, ei.index);
1842 if ((e1->flags & EDGE_FAKE) == 0)
1843 nonfakeedges = true;
1845 if (e1->flags & EDGE_EH)
1846 nehedges1++;
1848 if (e2->flags & EDGE_EH)
1849 nehedges2++;
1851 if (e1->flags & EDGE_FALLTHRU)
1852 fallthru1 = e1;
1853 if (e2->flags & EDGE_FALLTHRU)
1854 fallthru2 = e2;
1857 /* If number of edges of various types does not match, fail. */
1858 if (nehedges1 != nehedges2
1859 || (fallthru1 != 0) != (fallthru2 != 0))
1860 return false;
1862 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1863 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1864 attempt to optimize, as the two basic blocks might have different
1865 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1866 traps there should be REG_ARG_SIZE notes, they could be missing
1867 for __builtin_unreachable () uses though. */
1868 if (!nonfakeedges
1869 && !ACCUMULATE_OUTGOING_ARGS
1870 && (!INSN_P (last1)
1871 || !find_reg_note (last1, REG_ARGS_SIZE, NULL)))
1872 return false;
1874 /* fallthru edges must be forwarded to the same destination. */
1875 if (fallthru1)
1877 basic_block d1 = (forwarder_block_p (fallthru1->dest)
1878 ? single_succ (fallthru1->dest): fallthru1->dest);
1879 basic_block d2 = (forwarder_block_p (fallthru2->dest)
1880 ? single_succ (fallthru2->dest): fallthru2->dest);
1882 if (d1 != d2)
1883 return false;
1886 /* Ensure the same EH region. */
1888 rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
1889 rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
1891 if (!n1 && n2)
1892 return false;
1894 if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
1895 return false;
1898 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1899 version of sequence abstraction. */
1900 FOR_EACH_EDGE (e1, ei, bb2->succs)
1902 edge e2;
1903 edge_iterator ei;
1904 basic_block d1 = e1->dest;
1906 if (FORWARDER_BLOCK_P (d1))
1907 d1 = EDGE_SUCC (d1, 0)->dest;
1909 FOR_EACH_EDGE (e2, ei, bb1->succs)
1911 basic_block d2 = e2->dest;
1912 if (FORWARDER_BLOCK_P (d2))
1913 d2 = EDGE_SUCC (d2, 0)->dest;
1914 if (d1 == d2)
1915 break;
1918 if (!e2)
1919 return false;
1922 return true;
1925 /* Returns true if BB basic block has a preserve label. */
1927 static bool
1928 block_has_preserve_label (basic_block bb)
1930 return (bb
1931 && block_label (bb)
1932 && LABEL_PRESERVE_P (block_label (bb)));
1935 /* E1 and E2 are edges with the same destination block. Search their
1936 predecessors for common code. If found, redirect control flow from
1937 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1938 or the other way around (dir_backward). DIR specifies the allowed
1939 replacement direction. */
1941 static bool
1942 try_crossjump_to_edge (int mode, edge e1, edge e2,
1943 enum replace_direction dir)
1945 int nmatch;
1946 basic_block src1 = e1->src, src2 = e2->src;
1947 basic_block redirect_to, redirect_from, to_remove;
1948 basic_block osrc1, osrc2, redirect_edges_to, tmp;
1949 rtx_insn *newpos1, *newpos2;
1950 edge s;
1951 edge_iterator ei;
1953 newpos1 = newpos2 = NULL;
1955 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1956 to try this optimization.
1958 Basic block partitioning may result in some jumps that appear to
1959 be optimizable (or blocks that appear to be mergeable), but which really
1960 must be left untouched (they are required to make it safely across
1961 partition boundaries). See the comments at the top of
1962 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1964 if (crtl->has_bb_partition && reload_completed)
1965 return false;
1967 /* Search backward through forwarder blocks. We don't need to worry
1968 about multiple entry or chained forwarders, as they will be optimized
1969 away. We do this to look past the unconditional jump following a
1970 conditional jump that is required due to the current CFG shape. */
1971 if (single_pred_p (src1)
1972 && FORWARDER_BLOCK_P (src1))
1973 e1 = single_pred_edge (src1), src1 = e1->src;
1975 if (single_pred_p (src2)
1976 && FORWARDER_BLOCK_P (src2))
1977 e2 = single_pred_edge (src2), src2 = e2->src;
1979 /* Nothing to do if we reach ENTRY, or a common source block. */
1980 if (src1 == ENTRY_BLOCK_PTR_FOR_FN (cfun) || src2
1981 == ENTRY_BLOCK_PTR_FOR_FN (cfun))
1982 return false;
1983 if (src1 == src2)
1984 return false;
1986 /* Seeing more than 1 forwarder blocks would confuse us later... */
1987 if (FORWARDER_BLOCK_P (e1->dest)
1988 && FORWARDER_BLOCK_P (single_succ (e1->dest)))
1989 return false;
1991 if (FORWARDER_BLOCK_P (e2->dest)
1992 && FORWARDER_BLOCK_P (single_succ (e2->dest)))
1993 return false;
1995 /* Likewise with dead code (possibly newly created by the other optimizations
1996 of cfg_cleanup). */
1997 if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
1998 return false;
2000 /* Look for the common insn sequence, part the first ... */
2001 if (!outgoing_edges_match (mode, src1, src2))
2002 return false;
2004 /* ... and part the second. */
2005 nmatch = flow_find_cross_jump (src1, src2, &newpos1, &newpos2, &dir);
2007 osrc1 = src1;
2008 osrc2 = src2;
2009 if (newpos1 != NULL_RTX)
2010 src1 = BLOCK_FOR_INSN (newpos1);
2011 if (newpos2 != NULL_RTX)
2012 src2 = BLOCK_FOR_INSN (newpos2);
2014 if (dir == dir_backward)
2016 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
2017 SWAP (basic_block, osrc1, osrc2);
2018 SWAP (basic_block, src1, src2);
2019 SWAP (edge, e1, e2);
2020 SWAP (rtx_insn *, newpos1, newpos2);
2021 #undef SWAP
2024 /* Don't proceed with the crossjump unless we found a sufficient number
2025 of matching instructions or the 'from' block was totally matched
2026 (such that its predecessors will hopefully be redirected and the
2027 block removed). */
2028 if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
2029 && (newpos1 != BB_HEAD (src1)))
2030 return false;
2032 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
2033 if (block_has_preserve_label (e1->dest)
2034 && (e1->flags & EDGE_ABNORMAL))
2035 return false;
2037 /* Here we know that the insns in the end of SRC1 which are common with SRC2
2038 will be deleted.
2039 If we have tablejumps in the end of SRC1 and SRC2
2040 they have been already compared for equivalence in outgoing_edges_match ()
2041 so replace the references to TABLE1 by references to TABLE2. */
2043 rtx label1, label2;
2044 rtx_jump_table_data *table1, *table2;
2046 if (tablejump_p (BB_END (osrc1), &label1, &table1)
2047 && tablejump_p (BB_END (osrc2), &label2, &table2)
2048 && label1 != label2)
2050 rtx_insn *insn;
2052 /* Replace references to LABEL1 with LABEL2. */
2053 for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
2055 /* Do not replace the label in SRC1->END because when deleting
2056 a block whose end is a tablejump, the tablejump referenced
2057 from the instruction is deleted too. */
2058 if (insn != BB_END (osrc1))
2059 replace_label_in_insn (insn, label1, label2, true);
2064 /* Avoid splitting if possible. We must always split when SRC2 has
2065 EH predecessor edges, or we may end up with basic blocks with both
2066 normal and EH predecessor edges. */
2067 if (newpos2 == BB_HEAD (src2)
2068 && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
2069 redirect_to = src2;
2070 else
2072 if (newpos2 == BB_HEAD (src2))
2074 /* Skip possible basic block header. */
2075 if (LABEL_P (newpos2))
2076 newpos2 = NEXT_INSN (newpos2);
2077 while (DEBUG_INSN_P (newpos2))
2078 newpos2 = NEXT_INSN (newpos2);
2079 if (NOTE_P (newpos2))
2080 newpos2 = NEXT_INSN (newpos2);
2081 while (DEBUG_INSN_P (newpos2))
2082 newpos2 = NEXT_INSN (newpos2);
2085 if (dump_file)
2086 fprintf (dump_file, "Splitting bb %i before %i insns\n",
2087 src2->index, nmatch);
2088 redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
2091 if (dump_file)
2092 fprintf (dump_file,
2093 "Cross jumping from bb %i to bb %i; %i common insns\n",
2094 src1->index, src2->index, nmatch);
2096 /* We may have some registers visible through the block. */
2097 df_set_bb_dirty (redirect_to);
2099 if (osrc2 == src2)
2100 redirect_edges_to = redirect_to;
2101 else
2102 redirect_edges_to = osrc2;
2104 /* Recompute the frequencies and counts of outgoing edges. */
2105 FOR_EACH_EDGE (s, ei, redirect_edges_to->succs)
2107 edge s2;
2108 edge_iterator ei;
2109 basic_block d = s->dest;
2111 if (FORWARDER_BLOCK_P (d))
2112 d = single_succ (d);
2114 FOR_EACH_EDGE (s2, ei, src1->succs)
2116 basic_block d2 = s2->dest;
2117 if (FORWARDER_BLOCK_P (d2))
2118 d2 = single_succ (d2);
2119 if (d == d2)
2120 break;
2123 s->count += s2->count;
2125 /* Take care to update possible forwarder blocks. We verified
2126 that there is no more than one in the chain, so we can't run
2127 into infinite loop. */
2128 if (FORWARDER_BLOCK_P (s->dest))
2130 single_succ_edge (s->dest)->count += s2->count;
2131 s->dest->count += s2->count;
2132 s->dest->frequency += EDGE_FREQUENCY (s);
2135 if (FORWARDER_BLOCK_P (s2->dest))
2137 single_succ_edge (s2->dest)->count -= s2->count;
2138 if (single_succ_edge (s2->dest)->count < 0)
2139 single_succ_edge (s2->dest)->count = 0;
2140 s2->dest->count -= s2->count;
2141 s2->dest->frequency -= EDGE_FREQUENCY (s);
2142 if (s2->dest->frequency < 0)
2143 s2->dest->frequency = 0;
2144 if (s2->dest->count < 0)
2145 s2->dest->count = 0;
2148 if (!redirect_edges_to->frequency && !src1->frequency)
2149 s->probability = (s->probability + s2->probability) / 2;
2150 else
2151 s->probability
2152 = ((s->probability * redirect_edges_to->frequency +
2153 s2->probability * src1->frequency)
2154 / (redirect_edges_to->frequency + src1->frequency));
2157 /* Adjust count and frequency for the block. An earlier jump
2158 threading pass may have left the profile in an inconsistent
2159 state (see update_bb_profile_for_threading) so we must be
2160 prepared for overflows. */
2161 tmp = redirect_to;
2164 tmp->count += src1->count;
2165 tmp->frequency += src1->frequency;
2166 if (tmp->frequency > BB_FREQ_MAX)
2167 tmp->frequency = BB_FREQ_MAX;
2168 if (tmp == redirect_edges_to)
2169 break;
2170 tmp = find_fallthru_edge (tmp->succs)->dest;
2172 while (true);
2173 update_br_prob_note (redirect_edges_to);
2175 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2177 /* Skip possible basic block header. */
2178 if (LABEL_P (newpos1))
2179 newpos1 = NEXT_INSN (newpos1);
2181 while (DEBUG_INSN_P (newpos1))
2182 newpos1 = NEXT_INSN (newpos1);
2184 if (NOTE_INSN_BASIC_BLOCK_P (newpos1))
2185 newpos1 = NEXT_INSN (newpos1);
2187 while (DEBUG_INSN_P (newpos1))
2188 newpos1 = NEXT_INSN (newpos1);
2190 redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
2191 to_remove = single_succ (redirect_from);
2193 redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
2194 delete_basic_block (to_remove);
2196 update_forwarder_flag (redirect_from);
2197 if (redirect_to != src2)
2198 update_forwarder_flag (src2);
2200 return true;
2203 /* Search the predecessors of BB for common insn sequences. When found,
2204 share code between them by redirecting control flow. Return true if
2205 any changes made. */
2207 static bool
2208 try_crossjump_bb (int mode, basic_block bb)
2210 edge e, e2, fallthru;
2211 bool changed;
2212 unsigned max, ix, ix2;
2214 /* Nothing to do if there is not at least two incoming edges. */
2215 if (EDGE_COUNT (bb->preds) < 2)
2216 return false;
2218 /* Don't crossjump if this block ends in a computed jump,
2219 unless we are optimizing for size. */
2220 if (optimize_bb_for_size_p (bb)
2221 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2222 && computed_jump_p (BB_END (bb)))
2223 return false;
2225 /* If we are partitioning hot/cold basic blocks, we don't want to
2226 mess up unconditional or indirect jumps that cross between hot
2227 and cold sections.
2229 Basic block partitioning may result in some jumps that appear to
2230 be optimizable (or blocks that appear to be mergeable), but which really
2231 must be left untouched (they are required to make it safely across
2232 partition boundaries). See the comments at the top of
2233 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2235 if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
2236 BB_PARTITION (EDGE_PRED (bb, 1)->src)
2237 || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
2238 return false;
2240 /* It is always cheapest to redirect a block that ends in a branch to
2241 a block that falls through into BB, as that adds no branches to the
2242 program. We'll try that combination first. */
2243 fallthru = NULL;
2244 max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
2246 if (EDGE_COUNT (bb->preds) > max)
2247 return false;
2249 fallthru = find_fallthru_edge (bb->preds);
2251 changed = false;
2252 for (ix = 0; ix < EDGE_COUNT (bb->preds);)
2254 e = EDGE_PRED (bb, ix);
2255 ix++;
2257 /* As noted above, first try with the fallthru predecessor (or, a
2258 fallthru predecessor if we are in cfglayout mode). */
2259 if (fallthru)
2261 /* Don't combine the fallthru edge into anything else.
2262 If there is a match, we'll do it the other way around. */
2263 if (e == fallthru)
2264 continue;
2265 /* If nothing changed since the last attempt, there is nothing
2266 we can do. */
2267 if (!first_pass
2268 && !((e->src->flags & BB_MODIFIED)
2269 || (fallthru->src->flags & BB_MODIFIED)))
2270 continue;
2272 if (try_crossjump_to_edge (mode, e, fallthru, dir_forward))
2274 changed = true;
2275 ix = 0;
2276 continue;
2280 /* Non-obvious work limiting check: Recognize that we're going
2281 to call try_crossjump_bb on every basic block. So if we have
2282 two blocks with lots of outgoing edges (a switch) and they
2283 share lots of common destinations, then we would do the
2284 cross-jump check once for each common destination.
2286 Now, if the blocks actually are cross-jump candidates, then
2287 all of their destinations will be shared. Which means that
2288 we only need check them for cross-jump candidacy once. We
2289 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2290 choosing to do the check from the block for which the edge
2291 in question is the first successor of A. */
2292 if (EDGE_SUCC (e->src, 0) != e)
2293 continue;
2295 for (ix2 = 0; ix2 < EDGE_COUNT (bb->preds); ix2++)
2297 e2 = EDGE_PRED (bb, ix2);
2299 if (e2 == e)
2300 continue;
2302 /* We've already checked the fallthru edge above. */
2303 if (e2 == fallthru)
2304 continue;
2306 /* The "first successor" check above only prevents multiple
2307 checks of crossjump(A,B). In order to prevent redundant
2308 checks of crossjump(B,A), require that A be the block
2309 with the lowest index. */
2310 if (e->src->index > e2->src->index)
2311 continue;
2313 /* If nothing changed since the last attempt, there is nothing
2314 we can do. */
2315 if (!first_pass
2316 && !((e->src->flags & BB_MODIFIED)
2317 || (e2->src->flags & BB_MODIFIED)))
2318 continue;
2320 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2321 direction. */
2322 if (try_crossjump_to_edge (mode, e, e2, dir_both))
2324 changed = true;
2325 ix = 0;
2326 break;
2331 if (changed)
2332 crossjumps_occured = true;
2334 return changed;
2337 /* Search the successors of BB for common insn sequences. When found,
2338 share code between them by moving it across the basic block
2339 boundary. Return true if any changes made. */
2341 static bool
2342 try_head_merge_bb (basic_block bb)
2344 basic_block final_dest_bb = NULL;
2345 int max_match = INT_MAX;
2346 edge e0;
2347 rtx_insn **headptr, **currptr, **nextptr;
2348 bool changed, moveall;
2349 unsigned ix;
2350 rtx_insn *e0_last_head;
2351 rtx cond;
2352 rtx_insn *move_before;
2353 unsigned nedges = EDGE_COUNT (bb->succs);
2354 rtx_insn *jump = BB_END (bb);
2355 regset live, live_union;
2357 /* Nothing to do if there is not at least two outgoing edges. */
2358 if (nedges < 2)
2359 return false;
2361 /* Don't crossjump if this block ends in a computed jump,
2362 unless we are optimizing for size. */
2363 if (optimize_bb_for_size_p (bb)
2364 && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
2365 && computed_jump_p (BB_END (bb)))
2366 return false;
2368 cond = get_condition (jump, &move_before, true, false);
2369 if (cond == NULL_RTX)
2371 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2372 move_before = prev_nonnote_nondebug_insn (jump);
2373 else
2374 move_before = jump;
2377 for (ix = 0; ix < nedges; ix++)
2378 if (EDGE_SUCC (bb, ix)->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
2379 return false;
2381 for (ix = 0; ix < nedges; ix++)
2383 edge e = EDGE_SUCC (bb, ix);
2384 basic_block other_bb = e->dest;
2386 if (df_get_bb_dirty (other_bb))
2388 block_was_dirty = true;
2389 return false;
2392 if (e->flags & EDGE_ABNORMAL)
2393 return false;
2395 /* Normally, all destination blocks must only be reachable from this
2396 block, i.e. they must have one incoming edge.
2398 There is one special case we can handle, that of multiple consecutive
2399 jumps where the first jumps to one of the targets of the second jump.
2400 This happens frequently in switch statements for default labels.
2401 The structure is as follows:
2402 FINAL_DEST_BB
2403 ....
2404 if (cond) jump A;
2405 fall through
2407 jump with targets A, B, C, D...
2409 has two incoming edges, from FINAL_DEST_BB and BB
2411 In this case, we can try to move the insns through BB and into
2412 FINAL_DEST_BB. */
2413 if (EDGE_COUNT (other_bb->preds) != 1)
2415 edge incoming_edge, incoming_bb_other_edge;
2416 edge_iterator ei;
2418 if (final_dest_bb != NULL
2419 || EDGE_COUNT (other_bb->preds) != 2)
2420 return false;
2422 /* We must be able to move the insns across the whole block. */
2423 move_before = BB_HEAD (bb);
2424 while (!NONDEBUG_INSN_P (move_before))
2425 move_before = NEXT_INSN (move_before);
2427 if (EDGE_COUNT (bb->preds) != 1)
2428 return false;
2429 incoming_edge = EDGE_PRED (bb, 0);
2430 final_dest_bb = incoming_edge->src;
2431 if (EDGE_COUNT (final_dest_bb->succs) != 2)
2432 return false;
2433 FOR_EACH_EDGE (incoming_bb_other_edge, ei, final_dest_bb->succs)
2434 if (incoming_bb_other_edge != incoming_edge)
2435 break;
2436 if (incoming_bb_other_edge->dest != other_bb)
2437 return false;
2441 e0 = EDGE_SUCC (bb, 0);
2442 e0_last_head = NULL;
2443 changed = false;
2445 for (ix = 1; ix < nedges; ix++)
2447 edge e = EDGE_SUCC (bb, ix);
2448 rtx_insn *e0_last, *e_last;
2449 int nmatch;
2451 nmatch = flow_find_head_matching_sequence (e0->dest, e->dest,
2452 &e0_last, &e_last, 0);
2453 if (nmatch == 0)
2454 return false;
2456 if (nmatch < max_match)
2458 max_match = nmatch;
2459 e0_last_head = e0_last;
2463 /* If we matched an entire block, we probably have to avoid moving the
2464 last insn. */
2465 if (max_match > 0
2466 && e0_last_head == BB_END (e0->dest)
2467 && (find_reg_note (e0_last_head, REG_EH_REGION, 0)
2468 || control_flow_insn_p (e0_last_head)))
2470 max_match--;
2471 if (max_match == 0)
2472 return false;
2474 e0_last_head = prev_real_insn (e0_last_head);
2475 while (DEBUG_INSN_P (e0_last_head));
2478 if (max_match == 0)
2479 return false;
2481 /* We must find a union of the live registers at each of the end points. */
2482 live = BITMAP_ALLOC (NULL);
2483 live_union = BITMAP_ALLOC (NULL);
2485 currptr = XNEWVEC (rtx_insn *, nedges);
2486 headptr = XNEWVEC (rtx_insn *, nedges);
2487 nextptr = XNEWVEC (rtx_insn *, nedges);
2489 for (ix = 0; ix < nedges; ix++)
2491 int j;
2492 basic_block merge_bb = EDGE_SUCC (bb, ix)->dest;
2493 rtx_insn *head = BB_HEAD (merge_bb);
2495 while (!NONDEBUG_INSN_P (head))
2496 head = NEXT_INSN (head);
2497 headptr[ix] = head;
2498 currptr[ix] = head;
2500 /* Compute the end point and live information */
2501 for (j = 1; j < max_match; j++)
2503 head = NEXT_INSN (head);
2504 while (!NONDEBUG_INSN_P (head));
2505 simulate_backwards_to_point (merge_bb, live, head);
2506 IOR_REG_SET (live_union, live);
2509 /* If we're moving across two blocks, verify the validity of the
2510 first move, then adjust the target and let the loop below deal
2511 with the final move. */
2512 if (final_dest_bb != NULL)
2514 rtx_insn *move_upto;
2516 moveall = can_move_insns_across (currptr[0], e0_last_head, move_before,
2517 jump, e0->dest, live_union,
2518 NULL, &move_upto);
2519 if (!moveall)
2521 if (move_upto == NULL_RTX)
2522 goto out;
2524 while (e0_last_head != move_upto)
2526 df_simulate_one_insn_backwards (e0->dest, e0_last_head,
2527 live_union);
2528 e0_last_head = PREV_INSN (e0_last_head);
2531 if (e0_last_head == NULL_RTX)
2532 goto out;
2534 jump = BB_END (final_dest_bb);
2535 cond = get_condition (jump, &move_before, true, false);
2536 if (cond == NULL_RTX)
2538 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2539 move_before = prev_nonnote_nondebug_insn (jump);
2540 else
2541 move_before = jump;
2547 rtx_insn *move_upto;
2548 moveall = can_move_insns_across (currptr[0], e0_last_head,
2549 move_before, jump, e0->dest, live_union,
2550 NULL, &move_upto);
2551 if (!moveall && move_upto == NULL_RTX)
2553 if (jump == move_before)
2554 break;
2556 /* Try again, using a different insertion point. */
2557 move_before = jump;
2559 /* Don't try moving before a cc0 user, as that may invalidate
2560 the cc0. */
2561 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2562 break;
2564 continue;
2567 if (final_dest_bb && !moveall)
2568 /* We haven't checked whether a partial move would be OK for the first
2569 move, so we have to fail this case. */
2570 break;
2572 changed = true;
2573 for (;;)
2575 if (currptr[0] == move_upto)
2576 break;
2577 for (ix = 0; ix < nedges; ix++)
2579 rtx_insn *curr = currptr[ix];
2581 curr = NEXT_INSN (curr);
2582 while (!NONDEBUG_INSN_P (curr));
2583 currptr[ix] = curr;
2587 /* If we can't currently move all of the identical insns, remember
2588 each insn after the range that we'll merge. */
2589 if (!moveall)
2590 for (ix = 0; ix < nedges; ix++)
2592 rtx_insn *curr = currptr[ix];
2594 curr = NEXT_INSN (curr);
2595 while (!NONDEBUG_INSN_P (curr));
2596 nextptr[ix] = curr;
2599 reorder_insns (headptr[0], currptr[0], PREV_INSN (move_before));
2600 df_set_bb_dirty (EDGE_SUCC (bb, 0)->dest);
2601 if (final_dest_bb != NULL)
2602 df_set_bb_dirty (final_dest_bb);
2603 df_set_bb_dirty (bb);
2604 for (ix = 1; ix < nedges; ix++)
2606 df_set_bb_dirty (EDGE_SUCC (bb, ix)->dest);
2607 delete_insn_chain (headptr[ix], currptr[ix], false);
2609 if (!moveall)
2611 if (jump == move_before)
2612 break;
2614 /* For the unmerged insns, try a different insertion point. */
2615 move_before = jump;
2617 /* Don't try moving before a cc0 user, as that may invalidate
2618 the cc0. */
2619 if (HAVE_cc0 && reg_mentioned_p (cc0_rtx, jump))
2620 break;
2622 for (ix = 0; ix < nedges; ix++)
2623 currptr[ix] = headptr[ix] = nextptr[ix];
2626 while (!moveall);
2628 out:
2629 free (currptr);
2630 free (headptr);
2631 free (nextptr);
2633 crossjumps_occured |= changed;
2635 return changed;
2638 /* Return true if BB contains just bb note, or bb note followed
2639 by only DEBUG_INSNs. */
2641 static bool
2642 trivially_empty_bb_p (basic_block bb)
2644 rtx_insn *insn = BB_END (bb);
2646 while (1)
2648 if (insn == BB_HEAD (bb))
2649 return true;
2650 if (!DEBUG_INSN_P (insn))
2651 return false;
2652 insn = PREV_INSN (insn);
2656 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2657 instructions etc. Return nonzero if changes were made. */
2659 static bool
2660 try_optimize_cfg (int mode)
2662 bool changed_overall = false;
2663 bool changed;
2664 int iterations = 0;
2665 basic_block bb, b, next;
2667 if (mode & (CLEANUP_CROSSJUMP | CLEANUP_THREADING))
2668 clear_bb_flags ();
2670 crossjumps_occured = false;
2672 FOR_EACH_BB_FN (bb, cfun)
2673 update_forwarder_flag (bb);
2675 if (! targetm.cannot_modify_jumps_p ())
2677 first_pass = true;
2678 /* Attempt to merge blocks as made possible by edge removal. If
2679 a block has only one successor, and the successor has only
2680 one predecessor, they may be combined. */
2683 block_was_dirty = false;
2684 changed = false;
2685 iterations++;
2687 if (dump_file)
2688 fprintf (dump_file,
2689 "\n\ntry_optimize_cfg iteration %i\n\n",
2690 iterations);
2692 for (b = ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb; b
2693 != EXIT_BLOCK_PTR_FOR_FN (cfun);)
2695 basic_block c;
2696 edge s;
2697 bool changed_here = false;
2699 /* Delete trivially dead basic blocks. This is either
2700 blocks with no predecessors, or empty blocks with no
2701 successors. However if the empty block with no
2702 successors is the successor of the ENTRY_BLOCK, it is
2703 kept. This ensures that the ENTRY_BLOCK will have a
2704 successor which is a precondition for many RTL
2705 passes. Empty blocks may result from expanding
2706 __builtin_unreachable (). */
2707 if (EDGE_COUNT (b->preds) == 0
2708 || (EDGE_COUNT (b->succs) == 0
2709 && trivially_empty_bb_p (b)
2710 && single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))->dest
2711 != b))
2713 c = b->prev_bb;
2714 if (EDGE_COUNT (b->preds) > 0)
2716 edge e;
2717 edge_iterator ei;
2719 if (current_ir_type () == IR_RTL_CFGLAYOUT)
2721 if (BB_FOOTER (b)
2722 && BARRIER_P (BB_FOOTER (b)))
2723 FOR_EACH_EDGE (e, ei, b->preds)
2724 if ((e->flags & EDGE_FALLTHRU)
2725 && BB_FOOTER (e->src) == NULL)
2727 if (BB_FOOTER (b))
2729 BB_FOOTER (e->src) = BB_FOOTER (b);
2730 BB_FOOTER (b) = NULL;
2732 else
2734 start_sequence ();
2735 BB_FOOTER (e->src) = emit_barrier ();
2736 end_sequence ();
2740 else
2742 rtx_insn *last = get_last_bb_insn (b);
2743 if (last && BARRIER_P (last))
2744 FOR_EACH_EDGE (e, ei, b->preds)
2745 if ((e->flags & EDGE_FALLTHRU))
2746 emit_barrier_after (BB_END (e->src));
2749 delete_basic_block (b);
2750 changed = true;
2751 /* Avoid trying to remove the exit block. */
2752 b = (c == ENTRY_BLOCK_PTR_FOR_FN (cfun) ? c->next_bb : c);
2753 continue;
2756 /* Remove code labels no longer used. */
2757 if (single_pred_p (b)
2758 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2759 && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
2760 && LABEL_P (BB_HEAD (b))
2761 && !LABEL_PRESERVE_P (BB_HEAD (b))
2762 /* If the previous block ends with a branch to this
2763 block, we can't delete the label. Normally this
2764 is a condjump that is yet to be simplified, but
2765 if CASE_DROPS_THRU, this can be a tablejump with
2766 some element going to the same place as the
2767 default (fallthru). */
2768 && (single_pred (b) == ENTRY_BLOCK_PTR_FOR_FN (cfun)
2769 || !JUMP_P (BB_END (single_pred (b)))
2770 || ! label_is_jump_target_p (BB_HEAD (b),
2771 BB_END (single_pred (b)))))
2773 delete_insn (BB_HEAD (b));
2774 if (dump_file)
2775 fprintf (dump_file, "Deleted label in block %i.\n",
2776 b->index);
2779 /* If we fall through an empty block, we can remove it. */
2780 if (!(mode & (CLEANUP_CFGLAYOUT | CLEANUP_NO_INSN_DEL))
2781 && single_pred_p (b)
2782 && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
2783 && !LABEL_P (BB_HEAD (b))
2784 && FORWARDER_BLOCK_P (b)
2785 /* Note that forwarder_block_p true ensures that
2786 there is a successor for this block. */
2787 && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
2788 && n_basic_blocks_for_fn (cfun) > NUM_FIXED_BLOCKS + 1)
2790 if (dump_file)
2791 fprintf (dump_file,
2792 "Deleting fallthru block %i.\n",
2793 b->index);
2795 c = ((b->prev_bb == ENTRY_BLOCK_PTR_FOR_FN (cfun))
2796 ? b->next_bb : b->prev_bb);
2797 redirect_edge_succ_nodup (single_pred_edge (b),
2798 single_succ (b));
2799 delete_basic_block (b);
2800 changed = true;
2801 b = c;
2802 continue;
2805 /* Merge B with its single successor, if any. */
2806 if (single_succ_p (b)
2807 && (s = single_succ_edge (b))
2808 && !(s->flags & EDGE_COMPLEX)
2809 && (c = s->dest) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2810 && single_pred_p (c)
2811 && b != c)
2813 /* When not in cfg_layout mode use code aware of reordering
2814 INSN. This code possibly creates new basic blocks so it
2815 does not fit merge_blocks interface and is kept here in
2816 hope that it will become useless once more of compiler
2817 is transformed to use cfg_layout mode. */
2819 if ((mode & CLEANUP_CFGLAYOUT)
2820 && can_merge_blocks_p (b, c))
2822 merge_blocks (b, c);
2823 update_forwarder_flag (b);
2824 changed_here = true;
2826 else if (!(mode & CLEANUP_CFGLAYOUT)
2827 /* If the jump insn has side effects,
2828 we can't kill the edge. */
2829 && (!JUMP_P (BB_END (b))
2830 || (reload_completed
2831 ? simplejump_p (BB_END (b))
2832 : (onlyjump_p (BB_END (b))
2833 && !tablejump_p (BB_END (b),
2834 NULL, NULL))))
2835 && (next = merge_blocks_move (s, b, c, mode)))
2837 b = next;
2838 changed_here = true;
2842 /* Simplify branch over branch. */
2843 if ((mode & CLEANUP_EXPENSIVE)
2844 && !(mode & CLEANUP_CFGLAYOUT)
2845 && try_simplify_condjump (b))
2846 changed_here = true;
2848 /* If B has a single outgoing edge, but uses a
2849 non-trivial jump instruction without side-effects, we
2850 can either delete the jump entirely, or replace it
2851 with a simple unconditional jump. */
2852 if (single_succ_p (b)
2853 && single_succ (b) != EXIT_BLOCK_PTR_FOR_FN (cfun)
2854 && onlyjump_p (BB_END (b))
2855 && !CROSSING_JUMP_P (BB_END (b))
2856 && try_redirect_by_replacing_jump (single_succ_edge (b),
2857 single_succ (b),
2858 (mode & CLEANUP_CFGLAYOUT) != 0))
2860 update_forwarder_flag (b);
2861 changed_here = true;
2864 /* Simplify branch to branch. */
2865 if (try_forward_edges (mode, b))
2867 update_forwarder_flag (b);
2868 changed_here = true;
2871 /* Look for shared code between blocks. */
2872 if ((mode & CLEANUP_CROSSJUMP)
2873 && try_crossjump_bb (mode, b))
2874 changed_here = true;
2876 if ((mode & CLEANUP_CROSSJUMP)
2877 /* This can lengthen register lifetimes. Do it only after
2878 reload. */
2879 && reload_completed
2880 && try_head_merge_bb (b))
2881 changed_here = true;
2883 /* Don't get confused by the index shift caused by
2884 deleting blocks. */
2885 if (!changed_here)
2886 b = b->next_bb;
2887 else
2888 changed = true;
2891 if ((mode & CLEANUP_CROSSJUMP)
2892 && try_crossjump_bb (mode, EXIT_BLOCK_PTR_FOR_FN (cfun)))
2893 changed = true;
2895 if (block_was_dirty)
2897 /* This should only be set by head-merging. */
2898 gcc_assert (mode & CLEANUP_CROSSJUMP);
2899 df_analyze ();
2902 if (changed)
2904 /* Edge forwarding in particular can cause hot blocks previously
2905 reached by both hot and cold blocks to become dominated only
2906 by cold blocks. This will cause the verification below to fail,
2907 and lead to now cold code in the hot section. This is not easy
2908 to detect and fix during edge forwarding, and in some cases
2909 is only visible after newly unreachable blocks are deleted,
2910 which will be done in fixup_partitions. */
2911 fixup_partitions ();
2913 #ifdef ENABLE_CHECKING
2914 verify_flow_info ();
2915 #endif
2918 changed_overall |= changed;
2919 first_pass = false;
2921 while (changed);
2924 FOR_ALL_BB_FN (b, cfun)
2925 b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
2927 return changed_overall;
2930 /* Delete all unreachable basic blocks. */
2932 bool
2933 delete_unreachable_blocks (void)
2935 bool changed = false;
2936 basic_block b, prev_bb;
2938 find_unreachable_blocks ();
2940 /* When we're in GIMPLE mode and there may be debug insns, we should
2941 delete blocks in reverse dominator order, so as to get a chance
2942 to substitute all released DEFs into debug stmts. If we don't
2943 have dominators information, walking blocks backward gets us a
2944 better chance of retaining most debug information than
2945 otherwise. */
2946 if (MAY_HAVE_DEBUG_INSNS && current_ir_type () == IR_GIMPLE
2947 && dom_info_available_p (CDI_DOMINATORS))
2949 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
2950 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
2952 prev_bb = b->prev_bb;
2954 if (!(b->flags & BB_REACHABLE))
2956 /* Speed up the removal of blocks that don't dominate
2957 others. Walking backwards, this should be the common
2958 case. */
2959 if (!first_dom_son (CDI_DOMINATORS, b))
2960 delete_basic_block (b);
2961 else
2963 vec<basic_block> h
2964 = get_all_dominated_blocks (CDI_DOMINATORS, b);
2966 while (h.length ())
2968 b = h.pop ();
2970 prev_bb = b->prev_bb;
2972 gcc_assert (!(b->flags & BB_REACHABLE));
2974 delete_basic_block (b);
2977 h.release ();
2980 changed = true;
2984 else
2986 for (b = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb;
2987 b != ENTRY_BLOCK_PTR_FOR_FN (cfun); b = prev_bb)
2989 prev_bb = b->prev_bb;
2991 if (!(b->flags & BB_REACHABLE))
2993 delete_basic_block (b);
2994 changed = true;
2999 if (changed)
3000 tidy_fallthru_edges ();
3001 return changed;
3004 /* Delete any jump tables never referenced. We can't delete them at the
3005 time of removing tablejump insn as they are referenced by the preceding
3006 insns computing the destination, so we delay deleting and garbagecollect
3007 them once life information is computed. */
3008 void
3009 delete_dead_jumptables (void)
3011 basic_block bb;
3013 /* A dead jump table does not belong to any basic block. Scan insns
3014 between two adjacent basic blocks. */
3015 FOR_EACH_BB_FN (bb, cfun)
3017 rtx_insn *insn, *next;
3019 for (insn = NEXT_INSN (BB_END (bb));
3020 insn && !NOTE_INSN_BASIC_BLOCK_P (insn);
3021 insn = next)
3023 next = NEXT_INSN (insn);
3024 if (LABEL_P (insn)
3025 && LABEL_NUSES (insn) == LABEL_PRESERVE_P (insn)
3026 && JUMP_TABLE_DATA_P (next))
3028 rtx_insn *label = insn, *jump = next;
3030 if (dump_file)
3031 fprintf (dump_file, "Dead jumptable %i removed\n",
3032 INSN_UID (insn));
3034 next = NEXT_INSN (next);
3035 delete_insn (jump);
3036 delete_insn (label);
3043 /* Tidy the CFG by deleting unreachable code and whatnot. */
3045 bool
3046 cleanup_cfg (int mode)
3048 bool changed = false;
3050 /* Set the cfglayout mode flag here. We could update all the callers
3051 but that is just inconvenient, especially given that we eventually
3052 want to have cfglayout mode as the default. */
3053 if (current_ir_type () == IR_RTL_CFGLAYOUT)
3054 mode |= CLEANUP_CFGLAYOUT;
3056 timevar_push (TV_CLEANUP_CFG);
3057 if (delete_unreachable_blocks ())
3059 changed = true;
3060 /* We've possibly created trivially dead code. Cleanup it right
3061 now to introduce more opportunities for try_optimize_cfg. */
3062 if (!(mode & (CLEANUP_NO_INSN_DEL))
3063 && !reload_completed)
3064 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3067 compact_blocks ();
3069 /* To tail-merge blocks ending in the same noreturn function (e.g.
3070 a call to abort) we have to insert fake edges to exit. Do this
3071 here once. The fake edges do not interfere with any other CFG
3072 cleanups. */
3073 if (mode & CLEANUP_CROSSJUMP)
3074 add_noreturn_fake_exit_edges ();
3076 if (!dbg_cnt (cfg_cleanup))
3077 return changed;
3079 while (try_optimize_cfg (mode))
3081 delete_unreachable_blocks (), changed = true;
3082 if (!(mode & CLEANUP_NO_INSN_DEL))
3084 /* Try to remove some trivially dead insns when doing an expensive
3085 cleanup. But delete_trivially_dead_insns doesn't work after
3086 reload (it only handles pseudos) and run_fast_dce is too costly
3087 to run in every iteration.
3089 For effective cross jumping, we really want to run a fast DCE to
3090 clean up any dead conditions, or they get in the way of performing
3091 useful tail merges.
3093 Other transformations in cleanup_cfg are not so sensitive to dead
3094 code, so delete_trivially_dead_insns or even doing nothing at all
3095 is good enough. */
3096 if ((mode & CLEANUP_EXPENSIVE) && !reload_completed
3097 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3098 break;
3099 if ((mode & CLEANUP_CROSSJUMP) && crossjumps_occured)
3100 run_fast_dce ();
3102 else
3103 break;
3106 if (mode & CLEANUP_CROSSJUMP)
3107 remove_fake_exit_edges ();
3109 /* Don't call delete_dead_jumptables in cfglayout mode, because
3110 that function assumes that jump tables are in the insns stream.
3111 But we also don't _have_ to delete dead jumptables in cfglayout
3112 mode because we shouldn't even be looking at things that are
3113 not in a basic block. Dead jumptables are cleaned up when
3114 going out of cfglayout mode. */
3115 if (!(mode & CLEANUP_CFGLAYOUT))
3116 delete_dead_jumptables ();
3118 /* ??? We probably do this way too often. */
3119 if (current_loops
3120 && (changed
3121 || (mode & CLEANUP_CFG_CHANGED)))
3123 timevar_push (TV_REPAIR_LOOPS);
3124 /* The above doesn't preserve dominance info if available. */
3125 gcc_assert (!dom_info_available_p (CDI_DOMINATORS));
3126 calculate_dominance_info (CDI_DOMINATORS);
3127 fix_loop_structure (NULL);
3128 free_dominance_info (CDI_DOMINATORS);
3129 timevar_pop (TV_REPAIR_LOOPS);
3132 timevar_pop (TV_CLEANUP_CFG);
3134 return changed;
3137 namespace {
3139 const pass_data pass_data_jump =
3141 RTL_PASS, /* type */
3142 "jump", /* name */
3143 OPTGROUP_NONE, /* optinfo_flags */
3144 TV_JUMP, /* tv_id */
3145 0, /* properties_required */
3146 0, /* properties_provided */
3147 0, /* properties_destroyed */
3148 0, /* todo_flags_start */
3149 0, /* todo_flags_finish */
3152 class pass_jump : public rtl_opt_pass
3154 public:
3155 pass_jump (gcc::context *ctxt)
3156 : rtl_opt_pass (pass_data_jump, ctxt)
3159 /* opt_pass methods: */
3160 virtual unsigned int execute (function *);
3162 }; // class pass_jump
3164 unsigned int
3165 pass_jump::execute (function *)
3167 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3168 if (dump_file)
3169 dump_flow_info (dump_file, dump_flags);
3170 cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0)
3171 | (flag_thread_jumps ? CLEANUP_THREADING : 0));
3172 return 0;
3175 } // anon namespace
3177 rtl_opt_pass *
3178 make_pass_jump (gcc::context *ctxt)
3180 return new pass_jump (ctxt);
3183 namespace {
3185 const pass_data pass_data_jump2 =
3187 RTL_PASS, /* type */
3188 "jump2", /* name */
3189 OPTGROUP_NONE, /* optinfo_flags */
3190 TV_JUMP, /* tv_id */
3191 0, /* properties_required */
3192 0, /* properties_provided */
3193 0, /* properties_destroyed */
3194 0, /* todo_flags_start */
3195 0, /* todo_flags_finish */
3198 class pass_jump2 : public rtl_opt_pass
3200 public:
3201 pass_jump2 (gcc::context *ctxt)
3202 : rtl_opt_pass (pass_data_jump2, ctxt)
3205 /* opt_pass methods: */
3206 virtual unsigned int execute (function *)
3208 cleanup_cfg (flag_crossjumping ? CLEANUP_CROSSJUMP : 0);
3209 return 0;
3212 }; // class pass_jump2
3214 } // anon namespace
3216 rtl_opt_pass *
3217 make_pass_jump2 (gcc::context *ctxt)
3219 return new pass_jump2 (ctxt);