1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987-2014 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
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
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
28 - Cross jumping (tail merging)
29 - Conditional jump-around-simplejump simplification
30 - Basic block merging. */
34 #include "coretypes.h"
38 #include "hard-reg-set.h"
40 #include "insn-config.h"
43 #include "diagnostic-core.h"
48 #include "function.h" /* For inline functions in emit-rtl.h they need crtl. */
50 #include "tree-pass.h"
58 #define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
60 /* Set to true when we are running first pass of try_optimize_cfg loop. */
61 static bool first_pass
;
63 /* Set to true if crossjumps occurred in the latest run of try_optimize_cfg. */
64 static bool crossjumps_occured
;
66 /* Set to true if we couldn't run an optimization due to stale liveness
67 information; we should run df_analyze to enable more opportunities. */
68 static bool block_was_dirty
;
70 static bool try_crossjump_to_edge (int, edge
, edge
, enum replace_direction
);
71 static bool try_crossjump_bb (int, basic_block
);
72 static bool outgoing_edges_match (int, basic_block
, basic_block
);
73 static enum replace_direction
old_insns_match_p (int, rtx
, rtx
);
75 static void merge_blocks_move_predecessor_nojumps (basic_block
, basic_block
);
76 static void merge_blocks_move_successor_nojumps (basic_block
, basic_block
);
77 static bool try_optimize_cfg (int);
78 static bool try_simplify_condjump (basic_block
);
79 static bool try_forward_edges (int, basic_block
);
80 static edge
thread_jump (edge
, basic_block
);
81 static bool mark_effect (rtx
, bitmap
);
82 static void notice_new_block (basic_block
);
83 static void update_forwarder_flag (basic_block
);
84 static int mentions_nonequal_regs (rtx
*, void *);
85 static void merge_memattrs (rtx
, rtx
);
87 /* Set flags for newly created block. */
90 notice_new_block (basic_block bb
)
95 if (forwarder_block_p (bb
))
96 bb
->flags
|= BB_FORWARDER_BLOCK
;
99 /* Recompute forwarder flag after block has been modified. */
102 update_forwarder_flag (basic_block bb
)
104 if (forwarder_block_p (bb
))
105 bb
->flags
|= BB_FORWARDER_BLOCK
;
107 bb
->flags
&= ~BB_FORWARDER_BLOCK
;
110 /* Simplify a conditional jump around an unconditional jump.
111 Return true if something changed. */
114 try_simplify_condjump (basic_block cbranch_block
)
116 basic_block jump_block
, jump_dest_block
, cbranch_dest_block
;
117 edge cbranch_jump_edge
, cbranch_fallthru_edge
;
120 /* Verify that there are exactly two successors. */
121 if (EDGE_COUNT (cbranch_block
->succs
) != 2)
124 /* Verify that we've got a normal conditional branch at the end
126 cbranch_insn
= BB_END (cbranch_block
);
127 if (!any_condjump_p (cbranch_insn
))
130 cbranch_fallthru_edge
= FALLTHRU_EDGE (cbranch_block
);
131 cbranch_jump_edge
= BRANCH_EDGE (cbranch_block
);
133 /* The next block must not have multiple predecessors, must not
134 be the last block in the function, and must contain just the
135 unconditional jump. */
136 jump_block
= cbranch_fallthru_edge
->dest
;
137 if (!single_pred_p (jump_block
)
138 || jump_block
->next_bb
== EXIT_BLOCK_PTR_FOR_FN (cfun
)
139 || !FORWARDER_BLOCK_P (jump_block
))
141 jump_dest_block
= single_succ (jump_block
);
143 /* If we are partitioning hot/cold basic blocks, we don't want to
144 mess up unconditional or indirect jumps that cross between hot
147 Basic block partitioning may result in some jumps that appear to
148 be optimizable (or blocks that appear to be mergeable), but which really
149 must be left untouched (they are required to make it safely across
150 partition boundaries). See the comments at the top of
151 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
153 if (BB_PARTITION (jump_block
) != BB_PARTITION (jump_dest_block
)
154 || (cbranch_jump_edge
->flags
& EDGE_CROSSING
))
157 /* The conditional branch must target the block after the
158 unconditional branch. */
159 cbranch_dest_block
= cbranch_jump_edge
->dest
;
161 if (cbranch_dest_block
== EXIT_BLOCK_PTR_FOR_FN (cfun
)
162 || !can_fallthru (jump_block
, cbranch_dest_block
))
165 /* Invert the conditional branch. */
166 if (!invert_jump (cbranch_insn
, block_label (jump_dest_block
), 0))
170 fprintf (dump_file
, "Simplifying condjump %i around jump %i\n",
171 INSN_UID (cbranch_insn
), INSN_UID (BB_END (jump_block
)));
173 /* Success. Update the CFG to match. Note that after this point
174 the edge variable names appear backwards; the redirection is done
175 this way to preserve edge profile data. */
176 cbranch_jump_edge
= redirect_edge_succ_nodup (cbranch_jump_edge
,
178 cbranch_fallthru_edge
= redirect_edge_succ_nodup (cbranch_fallthru_edge
,
180 cbranch_jump_edge
->flags
|= EDGE_FALLTHRU
;
181 cbranch_fallthru_edge
->flags
&= ~EDGE_FALLTHRU
;
182 update_br_prob_note (cbranch_block
);
184 /* Delete the block with the unconditional jump, and clean up the mess. */
185 delete_basic_block (jump_block
);
186 tidy_fallthru_edge (cbranch_jump_edge
);
187 update_forwarder_flag (cbranch_block
);
192 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
193 on register. Used by jump threading. */
196 mark_effect (rtx exp
, regset nonequal
)
200 switch (GET_CODE (exp
))
202 /* In case we do clobber the register, mark it as equal, as we know the
203 value is dead so it don't have to match. */
205 if (REG_P (XEXP (exp
, 0)))
207 dest
= XEXP (exp
, 0);
208 regno
= REGNO (dest
);
209 if (HARD_REGISTER_NUM_P (regno
))
210 bitmap_clear_range (nonequal
, regno
,
211 hard_regno_nregs
[regno
][GET_MODE (dest
)]);
213 bitmap_clear_bit (nonequal
, regno
);
218 if (rtx_equal_for_cselib_p (SET_DEST (exp
), SET_SRC (exp
)))
220 dest
= SET_DEST (exp
);
225 regno
= REGNO (dest
);
226 if (HARD_REGISTER_NUM_P (regno
))
227 bitmap_set_range (nonequal
, regno
,
228 hard_regno_nregs
[regno
][GET_MODE (dest
)]);
230 bitmap_set_bit (nonequal
, regno
);
238 /* Return nonzero if X is a register set in regset DATA.
239 Called via for_each_rtx. */
241 mentions_nonequal_regs (rtx
*x
, void *data
)
243 regset nonequal
= (regset
) data
;
249 if (REGNO_REG_SET_P (nonequal
, regno
))
251 if (regno
< FIRST_PSEUDO_REGISTER
)
253 int n
= hard_regno_nregs
[regno
][GET_MODE (*x
)];
255 if (REGNO_REG_SET_P (nonequal
, regno
+ n
))
261 /* Attempt to prove that the basic block B will have no side effects and
262 always continues in the same edge if reached via E. Return the edge
263 if exist, NULL otherwise. */
266 thread_jump (edge e
, basic_block b
)
268 rtx set1
, set2
, cond1
, cond2
, insn
;
269 enum rtx_code code1
, code2
, reversed_code2
;
270 bool reverse1
= false;
274 reg_set_iterator rsi
;
276 if (b
->flags
& BB_NONTHREADABLE_BLOCK
)
279 /* At the moment, we do handle only conditional jumps, but later we may
280 want to extend this code to tablejumps and others. */
281 if (EDGE_COUNT (e
->src
->succs
) != 2)
283 if (EDGE_COUNT (b
->succs
) != 2)
285 b
->flags
|= BB_NONTHREADABLE_BLOCK
;
289 /* Second branch must end with onlyjump, as we will eliminate the jump. */
290 if (!any_condjump_p (BB_END (e
->src
)))
293 if (!any_condjump_p (BB_END (b
)) || !onlyjump_p (BB_END (b
)))
295 b
->flags
|= BB_NONTHREADABLE_BLOCK
;
299 set1
= pc_set (BB_END (e
->src
));
300 set2
= pc_set (BB_END (b
));
301 if (((e
->flags
& EDGE_FALLTHRU
) != 0)
302 != (XEXP (SET_SRC (set1
), 1) == pc_rtx
))
305 cond1
= XEXP (SET_SRC (set1
), 0);
306 cond2
= XEXP (SET_SRC (set2
), 0);
308 code1
= reversed_comparison_code (cond1
, BB_END (e
->src
));
310 code1
= GET_CODE (cond1
);
312 code2
= GET_CODE (cond2
);
313 reversed_code2
= reversed_comparison_code (cond2
, BB_END (b
));
315 if (!comparison_dominates_p (code1
, code2
)
316 && !comparison_dominates_p (code1
, reversed_code2
))
319 /* Ensure that the comparison operators are equivalent.
320 ??? This is far too pessimistic. We should allow swapped operands,
321 different CCmodes, or for example comparisons for interval, that
322 dominate even when operands are not equivalent. */
323 if (!rtx_equal_p (XEXP (cond1
, 0), XEXP (cond2
, 0))
324 || !rtx_equal_p (XEXP (cond1
, 1), XEXP (cond2
, 1)))
327 /* Short circuit cases where block B contains some side effects, as we can't
329 for (insn
= NEXT_INSN (BB_HEAD (b
)); insn
!= NEXT_INSN (BB_END (b
));
330 insn
= NEXT_INSN (insn
))
331 if (INSN_P (insn
) && side_effects_p (PATTERN (insn
)))
333 b
->flags
|= BB_NONTHREADABLE_BLOCK
;
339 /* First process all values computed in the source basic block. */
340 for (insn
= NEXT_INSN (BB_HEAD (e
->src
));
341 insn
!= NEXT_INSN (BB_END (e
->src
));
342 insn
= NEXT_INSN (insn
))
344 cselib_process_insn (insn
);
346 nonequal
= BITMAP_ALLOC (NULL
);
347 CLEAR_REG_SET (nonequal
);
349 /* Now assume that we've continued by the edge E to B and continue
350 processing as if it were same basic block.
351 Our goal is to prove that whole block is an NOOP. */
353 for (insn
= NEXT_INSN (BB_HEAD (b
));
354 insn
!= NEXT_INSN (BB_END (b
)) && !failed
;
355 insn
= NEXT_INSN (insn
))
359 rtx pat
= PATTERN (insn
);
361 if (GET_CODE (pat
) == PARALLEL
)
363 for (i
= 0; i
< (unsigned)XVECLEN (pat
, 0); i
++)
364 failed
|= mark_effect (XVECEXP (pat
, 0, i
), nonequal
);
367 failed
|= mark_effect (pat
, nonequal
);
370 cselib_process_insn (insn
);
373 /* Later we should clear nonequal of dead registers. So far we don't
374 have life information in cfg_cleanup. */
377 b
->flags
|= BB_NONTHREADABLE_BLOCK
;
381 /* cond2 must not mention any register that is not equal to the
383 if (for_each_rtx (&cond2
, mentions_nonequal_regs
, nonequal
))
386 EXECUTE_IF_SET_IN_REG_SET (nonequal
, 0, i
, rsi
)
389 BITMAP_FREE (nonequal
);
391 if ((comparison_dominates_p (code1
, code2
) != 0)
392 != (XEXP (SET_SRC (set2
), 1) == pc_rtx
))
393 return BRANCH_EDGE (b
);
395 return FALLTHRU_EDGE (b
);
398 BITMAP_FREE (nonequal
);
403 /* Attempt to forward edges leaving basic block B.
404 Return true if successful. */
407 try_forward_edges (int mode
, basic_block b
)
409 bool changed
= false;
411 edge e
, *threaded_edges
= NULL
;
413 /* If we are partitioning hot/cold basic blocks, we don't want to
414 mess up unconditional or indirect jumps that cross between hot
417 Basic block partitioning may result in some jumps that appear to
418 be optimizable (or blocks that appear to be mergeable), but which really
419 must be left untouched (they are required to make it safely across
420 partition boundaries). See the comments at the top of
421 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
423 if (JUMP_P (BB_END (b
)) && CROSSING_JUMP_P (BB_END (b
)))
426 for (ei
= ei_start (b
->succs
); (e
= ei_safe_edge (ei
)); )
428 basic_block target
, first
;
429 location_t goto_locus
;
431 bool threaded
= false;
432 int nthreaded_edges
= 0;
433 bool may_thread
= first_pass
|| (b
->flags
& BB_MODIFIED
) != 0;
435 /* Skip complex edges because we don't know how to update them.
437 Still handle fallthru edges, as we can succeed to forward fallthru
438 edge to the same place as the branch edge of conditional branch
439 and turn conditional branch to an unconditional branch. */
440 if (e
->flags
& EDGE_COMPLEX
)
446 target
= first
= e
->dest
;
447 counter
= NUM_FIXED_BLOCKS
;
448 goto_locus
= e
->goto_locus
;
450 /* If we are partitioning hot/cold basic_blocks, we don't want to mess
451 up jumps that cross between hot/cold sections.
453 Basic block partitioning may result in some jumps that appear
454 to be optimizable (or blocks that appear to be mergeable), but which
455 really must be left untouched (they are required to make it safely
456 across partition boundaries). See the comments at the top of
457 bb-reorder.c:partition_hot_cold_basic_blocks for complete
460 if (first
!= EXIT_BLOCK_PTR_FOR_FN (cfun
)
461 && JUMP_P (BB_END (first
))
462 && CROSSING_JUMP_P (BB_END (first
)))
465 while (counter
< n_basic_blocks_for_fn (cfun
))
467 basic_block new_target
= NULL
;
468 bool new_target_threaded
= false;
469 may_thread
|= (target
->flags
& BB_MODIFIED
) != 0;
471 if (FORWARDER_BLOCK_P (target
)
472 && !(single_succ_edge (target
)->flags
& EDGE_CROSSING
)
473 && single_succ (target
) != EXIT_BLOCK_PTR_FOR_FN (cfun
))
475 /* Bypass trivial infinite loops. */
476 new_target
= single_succ (target
);
477 if (target
== new_target
)
478 counter
= n_basic_blocks_for_fn (cfun
);
481 /* When not optimizing, ensure that edges or forwarder
482 blocks with different locus are not optimized out. */
483 location_t new_locus
= single_succ_edge (target
)->goto_locus
;
484 location_t locus
= goto_locus
;
486 if (LOCATION_LOCUS (new_locus
) != UNKNOWN_LOCATION
487 && LOCATION_LOCUS (locus
) != UNKNOWN_LOCATION
488 && new_locus
!= locus
)
492 if (LOCATION_LOCUS (new_locus
) != UNKNOWN_LOCATION
)
495 rtx last
= BB_END (target
);
496 if (DEBUG_INSN_P (last
))
497 last
= prev_nondebug_insn (last
);
498 if (last
&& INSN_P (last
))
499 new_locus
= INSN_LOCATION (last
);
501 new_locus
= UNKNOWN_LOCATION
;
503 if (LOCATION_LOCUS (new_locus
) != UNKNOWN_LOCATION
504 && LOCATION_LOCUS (locus
) != UNKNOWN_LOCATION
505 && new_locus
!= locus
)
509 if (LOCATION_LOCUS (new_locus
) != UNKNOWN_LOCATION
)
518 /* Allow to thread only over one edge at time to simplify updating
520 else if ((mode
& CLEANUP_THREADING
) && may_thread
)
522 edge t
= thread_jump (e
, target
);
526 threaded_edges
= XNEWVEC (edge
,
527 n_basic_blocks_for_fn (cfun
));
532 /* Detect an infinite loop across blocks not
533 including the start block. */
534 for (i
= 0; i
< nthreaded_edges
; ++i
)
535 if (threaded_edges
[i
] == t
)
537 if (i
< nthreaded_edges
)
539 counter
= n_basic_blocks_for_fn (cfun
);
544 /* Detect an infinite loop across the start block. */
548 gcc_assert (nthreaded_edges
549 < (n_basic_blocks_for_fn (cfun
)
550 - NUM_FIXED_BLOCKS
));
551 threaded_edges
[nthreaded_edges
++] = t
;
553 new_target
= t
->dest
;
554 new_target_threaded
= true;
563 threaded
|= new_target_threaded
;
566 if (counter
>= n_basic_blocks_for_fn (cfun
))
569 fprintf (dump_file
, "Infinite loop in BB %i.\n",
572 else if (target
== first
)
573 ; /* We didn't do anything. */
576 /* Save the values now, as the edge may get removed. */
577 gcov_type edge_count
= e
->count
;
578 int edge_probability
= e
->probability
;
582 e
->goto_locus
= goto_locus
;
584 /* Don't force if target is exit block. */
585 if (threaded
&& target
!= EXIT_BLOCK_PTR_FOR_FN (cfun
))
587 notice_new_block (redirect_edge_and_branch_force (e
, target
));
589 fprintf (dump_file
, "Conditionals threaded.\n");
591 else if (!redirect_edge_and_branch (e
, target
))
595 "Forwarding edge %i->%i to %i failed.\n",
596 b
->index
, e
->dest
->index
, target
->index
);
601 /* We successfully forwarded the edge. Now update profile
602 data: for each edge we traversed in the chain, remove
603 the original edge's execution count. */
604 edge_frequency
= apply_probability (b
->frequency
, edge_probability
);
610 if (!single_succ_p (first
))
612 gcc_assert (n
< nthreaded_edges
);
613 t
= threaded_edges
[n
++];
614 gcc_assert (t
->src
== first
);
615 update_bb_profile_for_threading (first
, edge_frequency
,
617 update_br_prob_note (first
);
621 first
->count
-= edge_count
;
622 if (first
->count
< 0)
624 first
->frequency
-= edge_frequency
;
625 if (first
->frequency
< 0)
626 first
->frequency
= 0;
627 /* It is possible that as the result of
628 threading we've removed edge as it is
629 threaded to the fallthru edge. Avoid
630 getting out of sync. */
631 if (n
< nthreaded_edges
632 && first
== threaded_edges
[n
]->src
)
634 t
= single_succ_edge (first
);
637 t
->count
-= edge_count
;
642 while (first
!= target
);
650 free (threaded_edges
);
655 /* Blocks A and B are to be merged into a single block. A has no incoming
656 fallthru edge, so it can be moved before B without adding or modifying
657 any jumps (aside from the jump from A to B). */
660 merge_blocks_move_predecessor_nojumps (basic_block a
, basic_block b
)
664 /* If we are partitioning hot/cold basic blocks, we don't want to
665 mess up unconditional or indirect jumps that cross between hot
668 Basic block partitioning may result in some jumps that appear to
669 be optimizable (or blocks that appear to be mergeable), but which really
670 must be left untouched (they are required to make it safely across
671 partition boundaries). See the comments at the top of
672 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
674 if (BB_PARTITION (a
) != BB_PARTITION (b
))
677 barrier
= next_nonnote_insn (BB_END (a
));
678 gcc_assert (BARRIER_P (barrier
));
679 delete_insn (barrier
);
681 /* Scramble the insn chain. */
682 if (BB_END (a
) != PREV_INSN (BB_HEAD (b
)))
683 reorder_insns_nobb (BB_HEAD (a
), BB_END (a
), PREV_INSN (BB_HEAD (b
)));
687 fprintf (dump_file
, "Moved block %d before %d and merged.\n",
690 /* Swap the records for the two blocks around. */
693 link_block (a
, b
->prev_bb
);
695 /* Now blocks A and B are contiguous. Merge them. */
699 /* Blocks A and B are to be merged into a single block. B has no outgoing
700 fallthru edge, so it can be moved after A without adding or modifying
701 any jumps (aside from the jump from A to B). */
704 merge_blocks_move_successor_nojumps (basic_block a
, basic_block b
)
706 rtx barrier
, real_b_end
;
709 /* If we are partitioning hot/cold basic blocks, we don't want to
710 mess up unconditional or indirect jumps that cross between hot
713 Basic block partitioning may result in some jumps that appear to
714 be optimizable (or blocks that appear to be mergeable), but which really
715 must be left untouched (they are required to make it safely across
716 partition boundaries). See the comments at the top of
717 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
719 if (BB_PARTITION (a
) != BB_PARTITION (b
))
722 real_b_end
= BB_END (b
);
724 /* If there is a jump table following block B temporarily add the jump table
725 to block B so that it will also be moved to the correct location. */
726 if (tablejump_p (BB_END (b
), &label
, &table
)
727 && prev_active_insn (label
) == BB_END (b
))
732 /* There had better have been a barrier there. Delete it. */
733 barrier
= NEXT_INSN (BB_END (b
));
734 if (barrier
&& BARRIER_P (barrier
))
735 delete_insn (barrier
);
738 /* Scramble the insn chain. */
739 reorder_insns_nobb (BB_HEAD (b
), BB_END (b
), BB_END (a
));
741 /* Restore the real end of b. */
742 BB_END (b
) = real_b_end
;
745 fprintf (dump_file
, "Moved block %d after %d and merged.\n",
748 /* Now blocks A and B are contiguous. Merge them. */
752 /* Attempt to merge basic blocks that are potentially non-adjacent.
753 Return NULL iff the attempt failed, otherwise return basic block
754 where cleanup_cfg should continue. Because the merging commonly
755 moves basic block away or introduces another optimization
756 possibility, return basic block just before B so cleanup_cfg don't
759 It may be good idea to return basic block before C in the case
760 C has been moved after B and originally appeared earlier in the
761 insn sequence, but we have no information available about the
762 relative ordering of these two. Hopefully it is not too common. */
765 merge_blocks_move (edge e
, basic_block b
, basic_block c
, int mode
)
769 /* If we are partitioning hot/cold basic blocks, we don't want to
770 mess up unconditional or indirect jumps that cross between hot
773 Basic block partitioning may result in some jumps that appear to
774 be optimizable (or blocks that appear to be mergeable), but which really
775 must be left untouched (they are required to make it safely across
776 partition boundaries). See the comments at the top of
777 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
779 if (BB_PARTITION (b
) != BB_PARTITION (c
))
782 /* If B has a fallthru edge to C, no need to move anything. */
783 if (e
->flags
& EDGE_FALLTHRU
)
785 int b_index
= b
->index
, c_index
= c
->index
;
787 /* Protect the loop latches. */
788 if (current_loops
&& c
->loop_father
->latch
== c
)
792 update_forwarder_flag (b
);
795 fprintf (dump_file
, "Merged %d and %d without moving.\n",
798 return b
->prev_bb
== ENTRY_BLOCK_PTR_FOR_FN (cfun
) ? b
: b
->prev_bb
;
801 /* Otherwise we will need to move code around. Do that only if expensive
802 transformations are allowed. */
803 else if (mode
& CLEANUP_EXPENSIVE
)
805 edge tmp_edge
, b_fallthru_edge
;
806 bool c_has_outgoing_fallthru
;
807 bool b_has_incoming_fallthru
;
809 /* Avoid overactive code motion, as the forwarder blocks should be
810 eliminated by edge redirection instead. One exception might have
811 been if B is a forwarder block and C has no fallthru edge, but
812 that should be cleaned up by bb-reorder instead. */
813 if (FORWARDER_BLOCK_P (b
) || FORWARDER_BLOCK_P (c
))
816 /* We must make sure to not munge nesting of lexical blocks,
817 and loop notes. This is done by squeezing out all the notes
818 and leaving them there to lie. Not ideal, but functional. */
820 tmp_edge
= find_fallthru_edge (c
->succs
);
821 c_has_outgoing_fallthru
= (tmp_edge
!= NULL
);
823 tmp_edge
= find_fallthru_edge (b
->preds
);
824 b_has_incoming_fallthru
= (tmp_edge
!= NULL
);
825 b_fallthru_edge
= tmp_edge
;
828 next
= next
->prev_bb
;
830 /* Otherwise, we're going to try to move C after B. If C does
831 not have an outgoing fallthru, then it can be moved
832 immediately after B without introducing or modifying jumps. */
833 if (! c_has_outgoing_fallthru
)
835 merge_blocks_move_successor_nojumps (b
, c
);
836 return next
== ENTRY_BLOCK_PTR_FOR_FN (cfun
) ? next
->next_bb
: next
;
839 /* If B does not have an incoming fallthru, then it can be moved
840 immediately before C without introducing or modifying jumps.
841 C cannot be the first block, so we do not have to worry about
842 accessing a non-existent block. */
844 if (b_has_incoming_fallthru
)
848 if (b_fallthru_edge
->src
== ENTRY_BLOCK_PTR_FOR_FN (cfun
))
850 bb
= force_nonfallthru (b_fallthru_edge
);
852 notice_new_block (bb
);
855 merge_blocks_move_predecessor_nojumps (b
, c
);
856 return next
== ENTRY_BLOCK_PTR_FOR_FN (cfun
) ? next
->next_bb
: next
;
863 /* Removes the memory attributes of MEM expression
864 if they are not equal. */
867 merge_memattrs (rtx x
, rtx y
)
876 if (x
== 0 || y
== 0)
881 if (code
!= GET_CODE (y
))
884 if (GET_MODE (x
) != GET_MODE (y
))
887 if (code
== MEM
&& !mem_attrs_eq_p (MEM_ATTRS (x
), MEM_ATTRS (y
)))
891 else if (! MEM_ATTRS (y
))
895 HOST_WIDE_INT mem_size
;
897 if (MEM_ALIAS_SET (x
) != MEM_ALIAS_SET (y
))
899 set_mem_alias_set (x
, 0);
900 set_mem_alias_set (y
, 0);
903 if (! mem_expr_equal_p (MEM_EXPR (x
), MEM_EXPR (y
)))
907 clear_mem_offset (x
);
908 clear_mem_offset (y
);
910 else if (MEM_OFFSET_KNOWN_P (x
) != MEM_OFFSET_KNOWN_P (y
)
911 || (MEM_OFFSET_KNOWN_P (x
)
912 && MEM_OFFSET (x
) != MEM_OFFSET (y
)))
914 clear_mem_offset (x
);
915 clear_mem_offset (y
);
918 if (MEM_SIZE_KNOWN_P (x
) && MEM_SIZE_KNOWN_P (y
))
920 mem_size
= MAX (MEM_SIZE (x
), MEM_SIZE (y
));
921 set_mem_size (x
, mem_size
);
922 set_mem_size (y
, mem_size
);
930 set_mem_align (x
, MIN (MEM_ALIGN (x
), MEM_ALIGN (y
)));
931 set_mem_align (y
, MEM_ALIGN (x
));
936 if (MEM_READONLY_P (x
) != MEM_READONLY_P (y
))
938 MEM_READONLY_P (x
) = 0;
939 MEM_READONLY_P (y
) = 0;
941 if (MEM_NOTRAP_P (x
) != MEM_NOTRAP_P (y
))
943 MEM_NOTRAP_P (x
) = 0;
944 MEM_NOTRAP_P (y
) = 0;
946 if (MEM_VOLATILE_P (x
) != MEM_VOLATILE_P (y
))
948 MEM_VOLATILE_P (x
) = 1;
949 MEM_VOLATILE_P (y
) = 1;
953 fmt
= GET_RTX_FORMAT (code
);
954 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
959 /* Two vectors must have the same length. */
960 if (XVECLEN (x
, i
) != XVECLEN (y
, i
))
963 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
964 merge_memattrs (XVECEXP (x
, i
, j
), XVECEXP (y
, i
, j
));
969 merge_memattrs (XEXP (x
, i
), XEXP (y
, i
));
976 /* Checks if patterns P1 and P2 are equivalent, apart from the possibly
977 different single sets S1 and S2. */
980 equal_different_set_p (rtx p1
, rtx s1
, rtx p2
, rtx s2
)
985 if (p1
== s1
&& p2
== s2
)
988 if (GET_CODE (p1
) != PARALLEL
|| GET_CODE (p2
) != PARALLEL
)
991 if (XVECLEN (p1
, 0) != XVECLEN (p2
, 0))
994 for (i
= 0; i
< XVECLEN (p1
, 0); i
++)
996 e1
= XVECEXP (p1
, 0, i
);
997 e2
= XVECEXP (p2
, 0, i
);
998 if (e1
== s1
&& e2
== s2
)
1000 if (reload_completed
1001 ? rtx_renumbered_equal_p (e1
, e2
) : rtx_equal_p (e1
, e2
))
1010 /* Examine register notes on I1 and I2 and return:
1011 - dir_forward if I1 can be replaced by I2, or
1012 - dir_backward if I2 can be replaced by I1, or
1013 - dir_both if both are the case. */
1015 static enum replace_direction
1016 can_replace_by (rtx i1
, rtx i2
)
1018 rtx s1
, s2
, d1
, d2
, src1
, src2
, note1
, note2
;
1021 /* Check for 2 sets. */
1022 s1
= single_set (i1
);
1023 s2
= single_set (i2
);
1024 if (s1
== NULL_RTX
|| s2
== NULL_RTX
)
1027 /* Check that the 2 sets set the same dest. */
1030 if (!(reload_completed
1031 ? rtx_renumbered_equal_p (d1
, d2
) : rtx_equal_p (d1
, d2
)))
1034 /* Find identical req_equiv or reg_equal note, which implies that the 2 sets
1035 set dest to the same value. */
1036 note1
= find_reg_equal_equiv_note (i1
);
1037 note2
= find_reg_equal_equiv_note (i2
);
1038 if (!note1
|| !note2
|| !rtx_equal_p (XEXP (note1
, 0), XEXP (note2
, 0))
1039 || !CONST_INT_P (XEXP (note1
, 0)))
1042 if (!equal_different_set_p (PATTERN (i1
), s1
, PATTERN (i2
), s2
))
1045 /* Although the 2 sets set dest to the same value, we cannot replace
1046 (set (dest) (const_int))
1049 because we don't know if the reg is live and has the same value at the
1050 location of replacement. */
1051 src1
= SET_SRC (s1
);
1052 src2
= SET_SRC (s2
);
1053 c1
= CONST_INT_P (src1
);
1054 c2
= CONST_INT_P (src2
);
1060 return dir_backward
;
1065 /* Merges directions A and B. */
1067 static enum replace_direction
1068 merge_dir (enum replace_direction a
, enum replace_direction b
)
1070 /* Implements the following table:
1089 /* Examine I1 and I2 and return:
1090 - dir_forward if I1 can be replaced by I2, or
1091 - dir_backward if I2 can be replaced by I1, or
1092 - dir_both if both are the case. */
1094 static enum replace_direction
1095 old_insns_match_p (int mode ATTRIBUTE_UNUSED
, rtx i1
, rtx i2
)
1099 /* Verify that I1 and I2 are equivalent. */
1100 if (GET_CODE (i1
) != GET_CODE (i2
))
1103 /* __builtin_unreachable() may lead to empty blocks (ending with
1104 NOTE_INSN_BASIC_BLOCK). They may be crossjumped. */
1105 if (NOTE_INSN_BASIC_BLOCK_P (i1
) && NOTE_INSN_BASIC_BLOCK_P (i2
))
1108 /* ??? Do not allow cross-jumping between different stack levels. */
1109 p1
= find_reg_note (i1
, REG_ARGS_SIZE
, NULL
);
1110 p2
= find_reg_note (i2
, REG_ARGS_SIZE
, NULL
);
1115 if (!rtx_equal_p (p1
, p2
))
1118 /* ??? Worse, this adjustment had better be constant lest we
1119 have differing incoming stack levels. */
1120 if (!frame_pointer_needed
1121 && find_args_size_adjust (i1
) == HOST_WIDE_INT_MIN
)
1130 if (GET_CODE (p1
) != GET_CODE (p2
))
1133 /* If this is a CALL_INSN, compare register usage information.
1134 If we don't check this on stack register machines, the two
1135 CALL_INSNs might be merged leaving reg-stack.c with mismatching
1136 numbers of stack registers in the same basic block.
1137 If we don't check this on machines with delay slots, a delay slot may
1138 be filled that clobbers a parameter expected by the subroutine.
1140 ??? We take the simple route for now and assume that if they're
1141 equal, they were constructed identically.
1143 Also check for identical exception regions. */
1147 /* Ensure the same EH region. */
1148 rtx n1
= find_reg_note (i1
, REG_EH_REGION
, 0);
1149 rtx n2
= find_reg_note (i2
, REG_EH_REGION
, 0);
1154 if (n1
&& (!n2
|| XEXP (n1
, 0) != XEXP (n2
, 0)))
1157 if (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1
),
1158 CALL_INSN_FUNCTION_USAGE (i2
))
1159 || SIBLING_CALL_P (i1
) != SIBLING_CALL_P (i2
))
1162 /* For address sanitizer, never crossjump __asan_report_* builtins,
1163 otherwise errors might be reported on incorrect lines. */
1164 if (flag_sanitize
& SANITIZE_ADDRESS
)
1166 rtx call
= get_call_rtx_from (i1
);
1167 if (call
&& GET_CODE (XEXP (XEXP (call
, 0), 0)) == SYMBOL_REF
)
1169 rtx symbol
= XEXP (XEXP (call
, 0), 0);
1170 if (SYMBOL_REF_DECL (symbol
)
1171 && TREE_CODE (SYMBOL_REF_DECL (symbol
)) == FUNCTION_DECL
)
1173 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol
))
1175 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol
))
1176 >= BUILT_IN_ASAN_REPORT_LOAD1
1177 && DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol
))
1178 <= BUILT_IN_ASAN_STOREN
)
1186 /* If cross_jump_death_matters is not 0, the insn's mode
1187 indicates whether or not the insn contains any stack-like
1190 if ((mode
& CLEANUP_POST_REGSTACK
) && stack_regs_mentioned (i1
))
1192 /* If register stack conversion has already been done, then
1193 death notes must also be compared before it is certain that
1194 the two instruction streams match. */
1197 HARD_REG_SET i1_regset
, i2_regset
;
1199 CLEAR_HARD_REG_SET (i1_regset
);
1200 CLEAR_HARD_REG_SET (i2_regset
);
1202 for (note
= REG_NOTES (i1
); note
; note
= XEXP (note
, 1))
1203 if (REG_NOTE_KIND (note
) == REG_DEAD
&& STACK_REG_P (XEXP (note
, 0)))
1204 SET_HARD_REG_BIT (i1_regset
, REGNO (XEXP (note
, 0)));
1206 for (note
= REG_NOTES (i2
); note
; note
= XEXP (note
, 1))
1207 if (REG_NOTE_KIND (note
) == REG_DEAD
&& STACK_REG_P (XEXP (note
, 0)))
1208 SET_HARD_REG_BIT (i2_regset
, REGNO (XEXP (note
, 0)));
1210 if (!hard_reg_set_equal_p (i1_regset
, i2_regset
))
1215 if (reload_completed
1216 ? rtx_renumbered_equal_p (p1
, p2
) : rtx_equal_p (p1
, p2
))
1219 return can_replace_by (i1
, i2
);
1222 /* When comparing insns I1 and I2 in flow_find_cross_jump or
1223 flow_find_head_matching_sequence, ensure the notes match. */
1226 merge_notes (rtx i1
, rtx i2
)
1228 /* If the merged insns have different REG_EQUAL notes, then
1230 rtx equiv1
= find_reg_equal_equiv_note (i1
);
1231 rtx equiv2
= find_reg_equal_equiv_note (i2
);
1233 if (equiv1
&& !equiv2
)
1234 remove_note (i1
, equiv1
);
1235 else if (!equiv1
&& equiv2
)
1236 remove_note (i2
, equiv2
);
1237 else if (equiv1
&& equiv2
1238 && !rtx_equal_p (XEXP (equiv1
, 0), XEXP (equiv2
, 0)))
1240 remove_note (i1
, equiv1
);
1241 remove_note (i2
, equiv2
);
1245 /* Walks from I1 in BB1 backward till the next non-debug insn, and returns the
1246 resulting insn in I1, and the corresponding bb in BB1. At the head of a
1247 bb, if there is a predecessor bb that reaches this bb via fallthru, and
1248 FOLLOW_FALLTHRU, walks further in the predecessor bb and registers this in
1249 DID_FALLTHRU. Otherwise, stops at the head of the bb. */
1252 walk_to_nondebug_insn (rtx
*i1
, basic_block
*bb1
, bool follow_fallthru
,
1257 *did_fallthru
= false;
1260 while (!NONDEBUG_INSN_P (*i1
))
1262 if (*i1
!= BB_HEAD (*bb1
))
1264 *i1
= PREV_INSN (*i1
);
1268 if (!follow_fallthru
)
1271 fallthru
= find_fallthru_edge ((*bb1
)->preds
);
1272 if (!fallthru
|| fallthru
->src
== ENTRY_BLOCK_PTR_FOR_FN (cfun
)
1273 || !single_succ_p (fallthru
->src
))
1276 *bb1
= fallthru
->src
;
1277 *i1
= BB_END (*bb1
);
1278 *did_fallthru
= true;
1282 /* Look through the insns at the end of BB1 and BB2 and find the longest
1283 sequence that are either equivalent, or allow forward or backward
1284 replacement. Store the first insns for that sequence in *F1 and *F2 and
1285 return the sequence length.
1287 DIR_P indicates the allowed replacement direction on function entry, and
1288 the actual replacement direction on function exit. If NULL, only equivalent
1289 sequences are allowed.
1291 To simplify callers of this function, if the blocks match exactly,
1292 store the head of the blocks in *F1 and *F2. */
1295 flow_find_cross_jump (basic_block bb1
, basic_block bb2
, rtx
*f1
, rtx
*f2
,
1296 enum replace_direction
*dir_p
)
1298 rtx i1
, i2
, last1
, last2
, afterlast1
, afterlast2
;
1300 enum replace_direction dir
, last_dir
, afterlast_dir
;
1301 bool follow_fallthru
, did_fallthru
;
1307 afterlast_dir
= dir
;
1308 last_dir
= afterlast_dir
;
1310 /* Skip simple jumps at the end of the blocks. Complex jumps still
1311 need to be compared for equivalence, which we'll do below. */
1314 last1
= afterlast1
= last2
= afterlast2
= NULL_RTX
;
1316 || (returnjump_p (i1
) && !side_effects_p (PATTERN (i1
))))
1319 i1
= PREV_INSN (i1
);
1324 || (returnjump_p (i2
) && !side_effects_p (PATTERN (i2
))))
1327 /* Count everything except for unconditional jump as insn.
1328 Don't count any jumps if dir_p is NULL. */
1329 if (!simplejump_p (i2
) && !returnjump_p (i2
) && last1
&& dir_p
)
1331 i2
= PREV_INSN (i2
);
1336 /* In the following example, we can replace all jumps to C by jumps to A.
1338 This removes 4 duplicate insns.
1339 [bb A] insn1 [bb C] insn1
1345 We could also replace all jumps to A by jumps to C, but that leaves B
1346 alive, and removes only 2 duplicate insns. In a subsequent crossjump
1347 step, all jumps to B would be replaced with jumps to the middle of C,
1348 achieving the same result with more effort.
1349 So we allow only the first possibility, which means that we don't allow
1350 fallthru in the block that's being replaced. */
1352 follow_fallthru
= dir_p
&& dir
!= dir_forward
;
1353 walk_to_nondebug_insn (&i1
, &bb1
, follow_fallthru
, &did_fallthru
);
1357 follow_fallthru
= dir_p
&& dir
!= dir_backward
;
1358 walk_to_nondebug_insn (&i2
, &bb2
, follow_fallthru
, &did_fallthru
);
1362 if (i1
== BB_HEAD (bb1
) || i2
== BB_HEAD (bb2
))
1365 dir
= merge_dir (dir
, old_insns_match_p (0, i1
, i2
));
1366 if (dir
== dir_none
|| (!dir_p
&& dir
!= dir_both
))
1369 merge_memattrs (i1
, i2
);
1371 /* Don't begin a cross-jump with a NOTE insn. */
1374 merge_notes (i1
, i2
);
1376 afterlast1
= last1
, afterlast2
= last2
;
1377 last1
= i1
, last2
= i2
;
1378 afterlast_dir
= last_dir
;
1380 if (active_insn_p (i1
))
1384 i1
= PREV_INSN (i1
);
1385 i2
= PREV_INSN (i2
);
1389 /* Don't allow the insn after a compare to be shared by
1390 cross-jumping unless the compare is also shared. */
1391 if (ninsns
&& reg_mentioned_p (cc0_rtx
, last1
) && ! sets_cc0_p (last1
))
1392 last1
= afterlast1
, last2
= afterlast2
, last_dir
= afterlast_dir
, ninsns
--;
1395 /* Include preceding notes and labels in the cross-jump. One,
1396 this may bring us to the head of the blocks as requested above.
1397 Two, it keeps line number notes as matched as may be. */
1400 bb1
= BLOCK_FOR_INSN (last1
);
1401 while (last1
!= BB_HEAD (bb1
) && !NONDEBUG_INSN_P (PREV_INSN (last1
)))
1402 last1
= PREV_INSN (last1
);
1404 if (last1
!= BB_HEAD (bb1
) && LABEL_P (PREV_INSN (last1
)))
1405 last1
= PREV_INSN (last1
);
1407 bb2
= BLOCK_FOR_INSN (last2
);
1408 while (last2
!= BB_HEAD (bb2
) && !NONDEBUG_INSN_P (PREV_INSN (last2
)))
1409 last2
= PREV_INSN (last2
);
1411 if (last2
!= BB_HEAD (bb2
) && LABEL_P (PREV_INSN (last2
)))
1412 last2
= PREV_INSN (last2
);
1423 /* Like flow_find_cross_jump, except start looking for a matching sequence from
1424 the head of the two blocks. Do not include jumps at the end.
1425 If STOP_AFTER is nonzero, stop after finding that many matching
1426 instructions. If STOP_AFTER is zero, count all INSN_P insns, if it is
1427 non-zero, only count active insns. */
1430 flow_find_head_matching_sequence (basic_block bb1
, basic_block bb2
, rtx
*f1
,
1431 rtx
*f2
, int stop_after
)
1433 rtx i1
, i2
, last1
, last2
, beforelast1
, beforelast2
;
1437 int nehedges1
= 0, nehedges2
= 0;
1439 FOR_EACH_EDGE (e
, ei
, bb1
->succs
)
1440 if (e
->flags
& EDGE_EH
)
1442 FOR_EACH_EDGE (e
, ei
, bb2
->succs
)
1443 if (e
->flags
& EDGE_EH
)
1448 last1
= beforelast1
= last2
= beforelast2
= NULL_RTX
;
1452 /* Ignore notes, except NOTE_INSN_EPILOGUE_BEG. */
1453 while (!NONDEBUG_INSN_P (i1
) && i1
!= BB_END (bb1
))
1455 if (NOTE_P (i1
) && NOTE_KIND (i1
) == NOTE_INSN_EPILOGUE_BEG
)
1457 i1
= NEXT_INSN (i1
);
1460 while (!NONDEBUG_INSN_P (i2
) && i2
!= BB_END (bb2
))
1462 if (NOTE_P (i2
) && NOTE_KIND (i2
) == NOTE_INSN_EPILOGUE_BEG
)
1464 i2
= NEXT_INSN (i2
);
1467 if ((i1
== BB_END (bb1
) && !NONDEBUG_INSN_P (i1
))
1468 || (i2
== BB_END (bb2
) && !NONDEBUG_INSN_P (i2
)))
1471 if (NOTE_P (i1
) || NOTE_P (i2
)
1472 || JUMP_P (i1
) || JUMP_P (i2
))
1475 /* A sanity check to make sure we're not merging insns with different
1476 effects on EH. If only one of them ends a basic block, it shouldn't
1477 have an EH edge; if both end a basic block, there should be the same
1478 number of EH edges. */
1479 if ((i1
== BB_END (bb1
) && i2
!= BB_END (bb2
)
1481 || (i2
== BB_END (bb2
) && i1
!= BB_END (bb1
)
1483 || (i1
== BB_END (bb1
) && i2
== BB_END (bb2
)
1484 && nehedges1
!= nehedges2
))
1487 if (old_insns_match_p (0, i1
, i2
) != dir_both
)
1490 merge_memattrs (i1
, i2
);
1492 /* Don't begin a cross-jump with a NOTE insn. */
1495 merge_notes (i1
, i2
);
1497 beforelast1
= last1
, beforelast2
= last2
;
1498 last1
= i1
, last2
= i2
;
1499 if (!stop_after
|| active_insn_p (i1
))
1503 if (i1
== BB_END (bb1
) || i2
== BB_END (bb2
)
1504 || (stop_after
> 0 && ninsns
== stop_after
))
1507 i1
= NEXT_INSN (i1
);
1508 i2
= NEXT_INSN (i2
);
1512 /* Don't allow a compare to be shared by cross-jumping unless the insn
1513 after the compare is also shared. */
1514 if (ninsns
&& reg_mentioned_p (cc0_rtx
, last1
) && sets_cc0_p (last1
))
1515 last1
= beforelast1
, last2
= beforelast2
, ninsns
--;
1527 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1528 the branch instruction. This means that if we commonize the control
1529 flow before end of the basic block, the semantic remains unchanged.
1531 We may assume that there exists one edge with a common destination. */
1534 outgoing_edges_match (int mode
, basic_block bb1
, basic_block bb2
)
1536 int nehedges1
= 0, nehedges2
= 0;
1537 edge fallthru1
= 0, fallthru2
= 0;
1541 /* If we performed shrink-wrapping, edges to the exit block can
1542 only be distinguished for JUMP_INSNs. The two paths may differ in
1543 whether they went through the prologue. Sibcalls are fine, we know
1544 that we either didn't need or inserted an epilogue before them. */
1545 if (crtl
->shrink_wrapped
1546 && single_succ_p (bb1
)
1547 && single_succ (bb1
) == EXIT_BLOCK_PTR_FOR_FN (cfun
)
1548 && !JUMP_P (BB_END (bb1
))
1549 && !(CALL_P (BB_END (bb1
)) && SIBLING_CALL_P (BB_END (bb1
))))
1552 /* If BB1 has only one successor, we may be looking at either an
1553 unconditional jump, or a fake edge to exit. */
1554 if (single_succ_p (bb1
)
1555 && (single_succ_edge (bb1
)->flags
& (EDGE_COMPLEX
| EDGE_FAKE
)) == 0
1556 && (!JUMP_P (BB_END (bb1
)) || simplejump_p (BB_END (bb1
))))
1557 return (single_succ_p (bb2
)
1558 && (single_succ_edge (bb2
)->flags
1559 & (EDGE_COMPLEX
| EDGE_FAKE
)) == 0
1560 && (!JUMP_P (BB_END (bb2
)) || simplejump_p (BB_END (bb2
))));
1562 /* Match conditional jumps - this may get tricky when fallthru and branch
1563 edges are crossed. */
1564 if (EDGE_COUNT (bb1
->succs
) == 2
1565 && any_condjump_p (BB_END (bb1
))
1566 && onlyjump_p (BB_END (bb1
)))
1568 edge b1
, f1
, b2
, f2
;
1569 bool reverse
, match
;
1570 rtx set1
, set2
, cond1
, cond2
;
1571 enum rtx_code code1
, code2
;
1573 if (EDGE_COUNT (bb2
->succs
) != 2
1574 || !any_condjump_p (BB_END (bb2
))
1575 || !onlyjump_p (BB_END (bb2
)))
1578 b1
= BRANCH_EDGE (bb1
);
1579 b2
= BRANCH_EDGE (bb2
);
1580 f1
= FALLTHRU_EDGE (bb1
);
1581 f2
= FALLTHRU_EDGE (bb2
);
1583 /* Get around possible forwarders on fallthru edges. Other cases
1584 should be optimized out already. */
1585 if (FORWARDER_BLOCK_P (f1
->dest
))
1586 f1
= single_succ_edge (f1
->dest
);
1588 if (FORWARDER_BLOCK_P (f2
->dest
))
1589 f2
= single_succ_edge (f2
->dest
);
1591 /* To simplify use of this function, return false if there are
1592 unneeded forwarder blocks. These will get eliminated later
1593 during cleanup_cfg. */
1594 if (FORWARDER_BLOCK_P (f1
->dest
)
1595 || FORWARDER_BLOCK_P (f2
->dest
)
1596 || FORWARDER_BLOCK_P (b1
->dest
)
1597 || FORWARDER_BLOCK_P (b2
->dest
))
1600 if (f1
->dest
== f2
->dest
&& b1
->dest
== b2
->dest
)
1602 else if (f1
->dest
== b2
->dest
&& b1
->dest
== f2
->dest
)
1607 set1
= pc_set (BB_END (bb1
));
1608 set2
= pc_set (BB_END (bb2
));
1609 if ((XEXP (SET_SRC (set1
), 1) == pc_rtx
)
1610 != (XEXP (SET_SRC (set2
), 1) == pc_rtx
))
1613 cond1
= XEXP (SET_SRC (set1
), 0);
1614 cond2
= XEXP (SET_SRC (set2
), 0);
1615 code1
= GET_CODE (cond1
);
1617 code2
= reversed_comparison_code (cond2
, BB_END (bb2
));
1619 code2
= GET_CODE (cond2
);
1621 if (code2
== UNKNOWN
)
1624 /* Verify codes and operands match. */
1625 match
= ((code1
== code2
1626 && rtx_renumbered_equal_p (XEXP (cond1
, 0), XEXP (cond2
, 0))
1627 && rtx_renumbered_equal_p (XEXP (cond1
, 1), XEXP (cond2
, 1)))
1628 || (code1
== swap_condition (code2
)
1629 && rtx_renumbered_equal_p (XEXP (cond1
, 1),
1631 && rtx_renumbered_equal_p (XEXP (cond1
, 0),
1634 /* If we return true, we will join the blocks. Which means that
1635 we will only have one branch prediction bit to work with. Thus
1636 we require the existing branches to have probabilities that are
1639 && optimize_bb_for_speed_p (bb1
)
1640 && optimize_bb_for_speed_p (bb2
))
1644 if (b1
->dest
== b2
->dest
)
1645 prob2
= b2
->probability
;
1647 /* Do not use f2 probability as f2 may be forwarded. */
1648 prob2
= REG_BR_PROB_BASE
- b2
->probability
;
1650 /* Fail if the difference in probabilities is greater than 50%.
1651 This rules out two well-predicted branches with opposite
1653 if (abs (b1
->probability
- prob2
) > REG_BR_PROB_BASE
/ 2)
1657 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1658 bb1
->index
, bb2
->index
, b1
->probability
, prob2
);
1664 if (dump_file
&& match
)
1665 fprintf (dump_file
, "Conditionals in bb %i and %i match.\n",
1666 bb1
->index
, bb2
->index
);
1671 /* Generic case - we are seeing a computed jump, table jump or trapping
1674 /* Check whether there are tablejumps in the end of BB1 and BB2.
1675 Return true if they are identical. */
1680 if (tablejump_p (BB_END (bb1
), &label1
, &table1
)
1681 && tablejump_p (BB_END (bb2
), &label2
, &table2
)
1682 && GET_CODE (PATTERN (table1
)) == GET_CODE (PATTERN (table2
)))
1684 /* The labels should never be the same rtx. If they really are same
1685 the jump tables are same too. So disable crossjumping of blocks BB1
1686 and BB2 because when deleting the common insns in the end of BB1
1687 by delete_basic_block () the jump table would be deleted too. */
1688 /* If LABEL2 is referenced in BB1->END do not do anything
1689 because we would loose information when replacing
1690 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1691 if (label1
!= label2
&& !rtx_referenced_p (label2
, BB_END (bb1
)))
1693 /* Set IDENTICAL to true when the tables are identical. */
1694 bool identical
= false;
1697 p1
= PATTERN (table1
);
1698 p2
= PATTERN (table2
);
1699 if (GET_CODE (p1
) == ADDR_VEC
&& rtx_equal_p (p1
, p2
))
1703 else if (GET_CODE (p1
) == ADDR_DIFF_VEC
1704 && (XVECLEN (p1
, 1) == XVECLEN (p2
, 1))
1705 && rtx_equal_p (XEXP (p1
, 2), XEXP (p2
, 2))
1706 && rtx_equal_p (XEXP (p1
, 3), XEXP (p2
, 3)))
1711 for (i
= XVECLEN (p1
, 1) - 1; i
>= 0 && identical
; i
--)
1712 if (!rtx_equal_p (XVECEXP (p1
, 1, i
), XVECEXP (p2
, 1, i
)))
1718 replace_label_data rr
;
1721 /* Temporarily replace references to LABEL1 with LABEL2
1722 in BB1->END so that we could compare the instructions. */
1725 rr
.update_label_nuses
= false;
1726 for_each_rtx (&BB_END (bb1
), replace_label
, &rr
);
1728 match
= (old_insns_match_p (mode
, BB_END (bb1
), BB_END (bb2
))
1730 if (dump_file
&& match
)
1732 "Tablejumps in bb %i and %i match.\n",
1733 bb1
->index
, bb2
->index
);
1735 /* Set the original label in BB1->END because when deleting
1736 a block whose end is a tablejump, the tablejump referenced
1737 from the instruction is deleted too. */
1740 for_each_rtx (&BB_END (bb1
), replace_label
, &rr
);
1749 /* Find the last non-debug non-note instruction in each bb, except
1750 stop when we see the NOTE_INSN_BASIC_BLOCK, as old_insns_match_p
1751 handles that case specially. old_insns_match_p does not handle
1752 other types of instruction notes. */
1753 rtx last1
= BB_END (bb1
);
1754 rtx last2
= BB_END (bb2
);
1755 while (!NOTE_INSN_BASIC_BLOCK_P (last1
) &&
1756 (DEBUG_INSN_P (last1
) || NOTE_P (last1
)))
1757 last1
= PREV_INSN (last1
);
1758 while (!NOTE_INSN_BASIC_BLOCK_P (last2
) &&
1759 (DEBUG_INSN_P (last2
) || NOTE_P (last2
)))
1760 last2
= PREV_INSN (last2
);
1761 gcc_assert (last1
&& last2
);
1763 /* First ensure that the instructions match. There may be many outgoing
1764 edges so this test is generally cheaper. */
1765 if (old_insns_match_p (mode
, last1
, last2
) != dir_both
)
1768 /* Search the outgoing edges, ensure that the counts do match, find possible
1769 fallthru and exception handling edges since these needs more
1771 if (EDGE_COUNT (bb1
->succs
) != EDGE_COUNT (bb2
->succs
))
1774 bool nonfakeedges
= false;
1775 FOR_EACH_EDGE (e1
, ei
, bb1
->succs
)
1777 e2
= EDGE_SUCC (bb2
, ei
.index
);
1779 if ((e1
->flags
& EDGE_FAKE
) == 0)
1780 nonfakeedges
= true;
1782 if (e1
->flags
& EDGE_EH
)
1785 if (e2
->flags
& EDGE_EH
)
1788 if (e1
->flags
& EDGE_FALLTHRU
)
1790 if (e2
->flags
& EDGE_FALLTHRU
)
1794 /* If number of edges of various types does not match, fail. */
1795 if (nehedges1
!= nehedges2
1796 || (fallthru1
!= 0) != (fallthru2
!= 0))
1799 /* If !ACCUMULATE_OUTGOING_ARGS, bb1 (and bb2) have no successors
1800 and the last real insn doesn't have REG_ARGS_SIZE note, don't
1801 attempt to optimize, as the two basic blocks might have different
1802 REG_ARGS_SIZE depths. For noreturn calls and unconditional
1803 traps there should be REG_ARG_SIZE notes, they could be missing
1804 for __builtin_unreachable () uses though. */
1806 && !ACCUMULATE_OUTGOING_ARGS
1808 || !find_reg_note (last1
, REG_ARGS_SIZE
, NULL
)))
1811 /* fallthru edges must be forwarded to the same destination. */
1814 basic_block d1
= (forwarder_block_p (fallthru1
->dest
)
1815 ? single_succ (fallthru1
->dest
): fallthru1
->dest
);
1816 basic_block d2
= (forwarder_block_p (fallthru2
->dest
)
1817 ? single_succ (fallthru2
->dest
): fallthru2
->dest
);
1823 /* Ensure the same EH region. */
1825 rtx n1
= find_reg_note (BB_END (bb1
), REG_EH_REGION
, 0);
1826 rtx n2
= find_reg_note (BB_END (bb2
), REG_EH_REGION
, 0);
1831 if (n1
&& (!n2
|| XEXP (n1
, 0) != XEXP (n2
, 0)))
1835 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1836 version of sequence abstraction. */
1837 FOR_EACH_EDGE (e1
, ei
, bb2
->succs
)
1841 basic_block d1
= e1
->dest
;
1843 if (FORWARDER_BLOCK_P (d1
))
1844 d1
= EDGE_SUCC (d1
, 0)->dest
;
1846 FOR_EACH_EDGE (e2
, ei
, bb1
->succs
)
1848 basic_block d2
= e2
->dest
;
1849 if (FORWARDER_BLOCK_P (d2
))
1850 d2
= EDGE_SUCC (d2
, 0)->dest
;
1862 /* Returns true if BB basic block has a preserve label. */
1865 block_has_preserve_label (basic_block bb
)
1869 && LABEL_PRESERVE_P (block_label (bb
)));
1872 /* E1 and E2 are edges with the same destination block. Search their
1873 predecessors for common code. If found, redirect control flow from
1874 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC (dir_forward),
1875 or the other way around (dir_backward). DIR specifies the allowed
1876 replacement direction. */
1879 try_crossjump_to_edge (int mode
, edge e1
, edge e2
,
1880 enum replace_direction dir
)
1883 basic_block src1
= e1
->src
, src2
= e2
->src
;
1884 basic_block redirect_to
, redirect_from
, to_remove
;
1885 basic_block osrc1
, osrc2
, redirect_edges_to
, tmp
;
1886 rtx newpos1
, newpos2
;
1890 newpos1
= newpos2
= NULL_RTX
;
1892 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1893 to try this optimization.
1895 Basic block partitioning may result in some jumps that appear to
1896 be optimizable (or blocks that appear to be mergeable), but which really
1897 must be left untouched (they are required to make it safely across
1898 partition boundaries). See the comments at the top of
1899 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1901 if (crtl
->has_bb_partition
&& reload_completed
)
1904 /* Search backward through forwarder blocks. We don't need to worry
1905 about multiple entry or chained forwarders, as they will be optimized
1906 away. We do this to look past the unconditional jump following a
1907 conditional jump that is required due to the current CFG shape. */
1908 if (single_pred_p (src1
)
1909 && FORWARDER_BLOCK_P (src1
))
1910 e1
= single_pred_edge (src1
), src1
= e1
->src
;
1912 if (single_pred_p (src2
)
1913 && FORWARDER_BLOCK_P (src2
))
1914 e2
= single_pred_edge (src2
), src2
= e2
->src
;
1916 /* Nothing to do if we reach ENTRY, or a common source block. */
1917 if (src1
== ENTRY_BLOCK_PTR_FOR_FN (cfun
) || src2
1918 == ENTRY_BLOCK_PTR_FOR_FN (cfun
))
1923 /* Seeing more than 1 forwarder blocks would confuse us later... */
1924 if (FORWARDER_BLOCK_P (e1
->dest
)
1925 && FORWARDER_BLOCK_P (single_succ (e1
->dest
)))
1928 if (FORWARDER_BLOCK_P (e2
->dest
)
1929 && FORWARDER_BLOCK_P (single_succ (e2
->dest
)))
1932 /* Likewise with dead code (possibly newly created by the other optimizations
1934 if (EDGE_COUNT (src1
->preds
) == 0 || EDGE_COUNT (src2
->preds
) == 0)
1937 /* Look for the common insn sequence, part the first ... */
1938 if (!outgoing_edges_match (mode
, src1
, src2
))
1941 /* ... and part the second. */
1942 nmatch
= flow_find_cross_jump (src1
, src2
, &newpos1
, &newpos2
, &dir
);
1946 if (newpos1
!= NULL_RTX
)
1947 src1
= BLOCK_FOR_INSN (newpos1
);
1948 if (newpos2
!= NULL_RTX
)
1949 src2
= BLOCK_FOR_INSN (newpos2
);
1951 if (dir
== dir_backward
)
1953 #define SWAP(T, X, Y) do { T tmp = (X); (X) = (Y); (Y) = tmp; } while (0)
1954 SWAP (basic_block
, osrc1
, osrc2
);
1955 SWAP (basic_block
, src1
, src2
);
1956 SWAP (edge
, e1
, e2
);
1957 SWAP (rtx
, newpos1
, newpos2
);
1961 /* Don't proceed with the crossjump unless we found a sufficient number
1962 of matching instructions or the 'from' block was totally matched
1963 (such that its predecessors will hopefully be redirected and the
1965 if ((nmatch
< PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS
))
1966 && (newpos1
!= BB_HEAD (src1
)))
1969 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
1970 if (block_has_preserve_label (e1
->dest
)
1971 && (e1
->flags
& EDGE_ABNORMAL
))
1974 /* Here we know that the insns in the end of SRC1 which are common with SRC2
1976 If we have tablejumps in the end of SRC1 and SRC2
1977 they have been already compared for equivalence in outgoing_edges_match ()
1978 so replace the references to TABLE1 by references to TABLE2. */
1983 if (tablejump_p (BB_END (osrc1
), &label1
, &table1
)
1984 && tablejump_p (BB_END (osrc2
), &label2
, &table2
)
1985 && label1
!= label2
)
1987 replace_label_data rr
;
1990 /* Replace references to LABEL1 with LABEL2. */
1993 rr
.update_label_nuses
= true;
1994 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
1996 /* Do not replace the label in SRC1->END because when deleting
1997 a block whose end is a tablejump, the tablejump referenced
1998 from the instruction is deleted too. */
1999 if (insn
!= BB_END (osrc1
))
2000 for_each_rtx (&insn
, replace_label
, &rr
);
2005 /* Avoid splitting if possible. We must always split when SRC2 has
2006 EH predecessor edges, or we may end up with basic blocks with both
2007 normal and EH predecessor edges. */
2008 if (newpos2
== BB_HEAD (src2
)
2009 && !(EDGE_PRED (src2
, 0)->flags
& EDGE_EH
))
2013 if (newpos2
== BB_HEAD (src2
))
2015 /* Skip possible basic block header. */
2016 if (LABEL_P (newpos2
))
2017 newpos2
= NEXT_INSN (newpos2
);
2018 while (DEBUG_INSN_P (newpos2
))
2019 newpos2
= NEXT_INSN (newpos2
);
2020 if (NOTE_P (newpos2
))
2021 newpos2
= NEXT_INSN (newpos2
);
2022 while (DEBUG_INSN_P (newpos2
))
2023 newpos2
= NEXT_INSN (newpos2
);
2027 fprintf (dump_file
, "Splitting bb %i before %i insns\n",
2028 src2
->index
, nmatch
);
2029 redirect_to
= split_block (src2
, PREV_INSN (newpos2
))->dest
;
2034 "Cross jumping from bb %i to bb %i; %i common insns\n",
2035 src1
->index
, src2
->index
, nmatch
);
2037 /* We may have some registers visible through the block. */
2038 df_set_bb_dirty (redirect_to
);
2041 redirect_edges_to
= redirect_to
;
2043 redirect_edges_to
= osrc2
;
2045 /* Recompute the frequencies and counts of outgoing edges. */
2046 FOR_EACH_EDGE (s
, ei
, redirect_edges_to
->succs
)
2050 basic_block d
= s
->dest
;
2052 if (FORWARDER_BLOCK_P (d
))
2053 d
= single_succ (d
);
2055 FOR_EACH_EDGE (s2
, ei
, src1
->succs
)
2057 basic_block d2
= s2
->dest
;
2058 if (FORWARDER_BLOCK_P (d2
))
2059 d2
= single_succ (d2
);
2064 s
->count
+= s2
->count
;
2066 /* Take care to update possible forwarder blocks. We verified
2067 that there is no more than one in the chain, so we can't run
2068 into infinite loop. */
2069 if (FORWARDER_BLOCK_P (s
->dest
))
2071 single_succ_edge (s
->dest
)->count
+= s2
->count
;
2072 s
->dest
->count
+= s2
->count
;
2073 s
->dest
->frequency
+= EDGE_FREQUENCY (s
);
2076 if (FORWARDER_BLOCK_P (s2
->dest
))
2078 single_succ_edge (s2
->dest
)->count
-= s2
->count
;
2079 if (single_succ_edge (s2
->dest
)->count
< 0)
2080 single_succ_edge (s2
->dest
)->count
= 0;
2081 s2
->dest
->count
-= s2
->count
;
2082 s2
->dest
->frequency
-= EDGE_FREQUENCY (s
);
2083 if (s2
->dest
->frequency
< 0)
2084 s2
->dest
->frequency
= 0;
2085 if (s2
->dest
->count
< 0)
2086 s2
->dest
->count
= 0;
2089 if (!redirect_edges_to
->frequency
&& !src1
->frequency
)
2090 s
->probability
= (s
->probability
+ s2
->probability
) / 2;
2093 = ((s
->probability
* redirect_edges_to
->frequency
+
2094 s2
->probability
* src1
->frequency
)
2095 / (redirect_edges_to
->frequency
+ src1
->frequency
));
2098 /* Adjust count and frequency for the block. An earlier jump
2099 threading pass may have left the profile in an inconsistent
2100 state (see update_bb_profile_for_threading) so we must be
2101 prepared for overflows. */
2105 tmp
->count
+= src1
->count
;
2106 tmp
->frequency
+= src1
->frequency
;
2107 if (tmp
->frequency
> BB_FREQ_MAX
)
2108 tmp
->frequency
= BB_FREQ_MAX
;
2109 if (tmp
== redirect_edges_to
)
2111 tmp
= find_fallthru_edge (tmp
->succs
)->dest
;
2114 update_br_prob_note (redirect_edges_to
);
2116 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
2118 /* Skip possible basic block header. */
2119 if (LABEL_P (newpos1
))
2120 newpos1
= NEXT_INSN (newpos1
);
2122 while (DEBUG_INSN_P (newpos1
))
2123 newpos1
= NEXT_INSN (newpos1
);
2125 if (NOTE_INSN_BASIC_BLOCK_P (newpos1
))
2126 newpos1
= NEXT_INSN (newpos1
);
2128 while (DEBUG_INSN_P (newpos1
))
2129 newpos1
= NEXT_INSN (newpos1
);
2131 redirect_from
= split_block (src1
, PREV_INSN (newpos1
))->src
;
2132 to_remove
= single_succ (redirect_from
);
2134 redirect_edge_and_branch_force (single_succ_edge (redirect_from
), redirect_to
);
2135 delete_basic_block (to_remove
);
2137 update_forwarder_flag (redirect_from
);
2138 if (redirect_to
!= src2
)
2139 update_forwarder_flag (src2
);
2144 /* Search the predecessors of BB for common insn sequences. When found,
2145 share code between them by redirecting control flow. Return true if
2146 any changes made. */
2149 try_crossjump_bb (int mode
, basic_block bb
)
2151 edge e
, e2
, fallthru
;
2153 unsigned max
, ix
, ix2
;
2155 /* Nothing to do if there is not at least two incoming edges. */
2156 if (EDGE_COUNT (bb
->preds
) < 2)
2159 /* Don't crossjump if this block ends in a computed jump,
2160 unless we are optimizing for size. */
2161 if (optimize_bb_for_size_p (bb
)
2162 && bb
!= EXIT_BLOCK_PTR_FOR_FN (cfun
)
2163 && computed_jump_p (BB_END (bb
)))
2166 /* If we are partitioning hot/cold basic blocks, we don't want to
2167 mess up unconditional or indirect jumps that cross between hot
2170 Basic block partitioning may result in some jumps that appear to
2171 be optimizable (or blocks that appear to be mergeable), but which really
2172 must be left untouched (they are required to make it safely across
2173 partition boundaries). See the comments at the top of
2174 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
2176 if (BB_PARTITION (EDGE_PRED (bb
, 0)->src
) !=
2177 BB_PARTITION (EDGE_PRED (bb
, 1)->src
)
2178 || (EDGE_PRED (bb
, 0)->flags
& EDGE_CROSSING
))
2181 /* It is always cheapest to redirect a block that ends in a branch to
2182 a block that falls through into BB, as that adds no branches to the
2183 program. We'll try that combination first. */
2185 max
= PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES
);
2187 if (EDGE_COUNT (bb
->preds
) > max
)
2190 fallthru
= find_fallthru_edge (bb
->preds
);
2193 for (ix
= 0; ix
< EDGE_COUNT (bb
->preds
);)
2195 e
= EDGE_PRED (bb
, ix
);
2198 /* As noted above, first try with the fallthru predecessor (or, a
2199 fallthru predecessor if we are in cfglayout mode). */
2202 /* Don't combine the fallthru edge into anything else.
2203 If there is a match, we'll do it the other way around. */
2206 /* If nothing changed since the last attempt, there is nothing
2209 && !((e
->src
->flags
& BB_MODIFIED
)
2210 || (fallthru
->src
->flags
& BB_MODIFIED
)))
2213 if (try_crossjump_to_edge (mode
, e
, fallthru
, dir_forward
))
2221 /* Non-obvious work limiting check: Recognize that we're going
2222 to call try_crossjump_bb on every basic block. So if we have
2223 two blocks with lots of outgoing edges (a switch) and they
2224 share lots of common destinations, then we would do the
2225 cross-jump check once for each common destination.
2227 Now, if the blocks actually are cross-jump candidates, then
2228 all of their destinations will be shared. Which means that
2229 we only need check them for cross-jump candidacy once. We
2230 can eliminate redundant checks of crossjump(A,B) by arbitrarily
2231 choosing to do the check from the block for which the edge
2232 in question is the first successor of A. */
2233 if (EDGE_SUCC (e
->src
, 0) != e
)
2236 for (ix2
= 0; ix2
< EDGE_COUNT (bb
->preds
); ix2
++)
2238 e2
= EDGE_PRED (bb
, ix2
);
2243 /* We've already checked the fallthru edge above. */
2247 /* The "first successor" check above only prevents multiple
2248 checks of crossjump(A,B). In order to prevent redundant
2249 checks of crossjump(B,A), require that A be the block
2250 with the lowest index. */
2251 if (e
->src
->index
> e2
->src
->index
)
2254 /* If nothing changed since the last attempt, there is nothing
2257 && !((e
->src
->flags
& BB_MODIFIED
)
2258 || (e2
->src
->flags
& BB_MODIFIED
)))
2261 /* Both e and e2 are not fallthru edges, so we can crossjump in either
2263 if (try_crossjump_to_edge (mode
, e
, e2
, dir_both
))
2273 crossjumps_occured
= true;
2278 /* Search the successors of BB for common insn sequences. When found,
2279 share code between them by moving it across the basic block
2280 boundary. Return true if any changes made. */
2283 try_head_merge_bb (basic_block bb
)
2285 basic_block final_dest_bb
= NULL
;
2286 int max_match
= INT_MAX
;
2288 rtx
*headptr
, *currptr
, *nextptr
;
2289 bool changed
, moveall
;
2291 rtx e0_last_head
, cond
, move_before
;
2292 unsigned nedges
= EDGE_COUNT (bb
->succs
);
2293 rtx jump
= BB_END (bb
);
2294 regset live
, live_union
;
2296 /* Nothing to do if there is not at least two outgoing edges. */
2300 /* Don't crossjump if this block ends in a computed jump,
2301 unless we are optimizing for size. */
2302 if (optimize_bb_for_size_p (bb
)
2303 && bb
!= EXIT_BLOCK_PTR_FOR_FN (cfun
)
2304 && computed_jump_p (BB_END (bb
)))
2307 cond
= get_condition (jump
, &move_before
, true, false);
2308 if (cond
== NULL_RTX
)
2311 if (reg_mentioned_p (cc0_rtx
, jump
))
2312 move_before
= prev_nonnote_nondebug_insn (jump
);
2318 for (ix
= 0; ix
< nedges
; ix
++)
2319 if (EDGE_SUCC (bb
, ix
)->dest
== EXIT_BLOCK_PTR_FOR_FN (cfun
))
2322 for (ix
= 0; ix
< nedges
; ix
++)
2324 edge e
= EDGE_SUCC (bb
, ix
);
2325 basic_block other_bb
= e
->dest
;
2327 if (df_get_bb_dirty (other_bb
))
2329 block_was_dirty
= true;
2333 if (e
->flags
& EDGE_ABNORMAL
)
2336 /* Normally, all destination blocks must only be reachable from this
2337 block, i.e. they must have one incoming edge.
2339 There is one special case we can handle, that of multiple consecutive
2340 jumps where the first jumps to one of the targets of the second jump.
2341 This happens frequently in switch statements for default labels.
2342 The structure is as follows:
2348 jump with targets A, B, C, D...
2350 has two incoming edges, from FINAL_DEST_BB and BB
2352 In this case, we can try to move the insns through BB and into
2354 if (EDGE_COUNT (other_bb
->preds
) != 1)
2356 edge incoming_edge
, incoming_bb_other_edge
;
2359 if (final_dest_bb
!= NULL
2360 || EDGE_COUNT (other_bb
->preds
) != 2)
2363 /* We must be able to move the insns across the whole block. */
2364 move_before
= BB_HEAD (bb
);
2365 while (!NONDEBUG_INSN_P (move_before
))
2366 move_before
= NEXT_INSN (move_before
);
2368 if (EDGE_COUNT (bb
->preds
) != 1)
2370 incoming_edge
= EDGE_PRED (bb
, 0);
2371 final_dest_bb
= incoming_edge
->src
;
2372 if (EDGE_COUNT (final_dest_bb
->succs
) != 2)
2374 FOR_EACH_EDGE (incoming_bb_other_edge
, ei
, final_dest_bb
->succs
)
2375 if (incoming_bb_other_edge
!= incoming_edge
)
2377 if (incoming_bb_other_edge
->dest
!= other_bb
)
2382 e0
= EDGE_SUCC (bb
, 0);
2383 e0_last_head
= NULL_RTX
;
2386 for (ix
= 1; ix
< nedges
; ix
++)
2388 edge e
= EDGE_SUCC (bb
, ix
);
2389 rtx e0_last
, e_last
;
2392 nmatch
= flow_find_head_matching_sequence (e0
->dest
, e
->dest
,
2393 &e0_last
, &e_last
, 0);
2397 if (nmatch
< max_match
)
2400 e0_last_head
= e0_last
;
2404 /* If we matched an entire block, we probably have to avoid moving the
2407 && e0_last_head
== BB_END (e0
->dest
)
2408 && (find_reg_note (e0_last_head
, REG_EH_REGION
, 0)
2409 || control_flow_insn_p (e0_last_head
)))
2415 e0_last_head
= prev_real_insn (e0_last_head
);
2416 while (DEBUG_INSN_P (e0_last_head
));
2422 /* We must find a union of the live registers at each of the end points. */
2423 live
= BITMAP_ALLOC (NULL
);
2424 live_union
= BITMAP_ALLOC (NULL
);
2426 currptr
= XNEWVEC (rtx
, nedges
);
2427 headptr
= XNEWVEC (rtx
, nedges
);
2428 nextptr
= XNEWVEC (rtx
, nedges
);
2430 for (ix
= 0; ix
< nedges
; ix
++)
2433 basic_block merge_bb
= EDGE_SUCC (bb
, ix
)->dest
;
2434 rtx head
= BB_HEAD (merge_bb
);
2436 while (!NONDEBUG_INSN_P (head
))
2437 head
= NEXT_INSN (head
);
2441 /* Compute the end point and live information */
2442 for (j
= 1; j
< max_match
; j
++)
2444 head
= NEXT_INSN (head
);
2445 while (!NONDEBUG_INSN_P (head
));
2446 simulate_backwards_to_point (merge_bb
, live
, head
);
2447 IOR_REG_SET (live_union
, live
);
2450 /* If we're moving across two blocks, verify the validity of the
2451 first move, then adjust the target and let the loop below deal
2452 with the final move. */
2453 if (final_dest_bb
!= NULL
)
2457 moveall
= can_move_insns_across (currptr
[0], e0_last_head
, move_before
,
2458 jump
, e0
->dest
, live_union
,
2462 if (move_upto
== NULL_RTX
)
2465 while (e0_last_head
!= move_upto
)
2467 df_simulate_one_insn_backwards (e0
->dest
, e0_last_head
,
2469 e0_last_head
= PREV_INSN (e0_last_head
);
2472 if (e0_last_head
== NULL_RTX
)
2475 jump
= BB_END (final_dest_bb
);
2476 cond
= get_condition (jump
, &move_before
, true, false);
2477 if (cond
== NULL_RTX
)
2480 if (reg_mentioned_p (cc0_rtx
, jump
))
2481 move_before
= prev_nonnote_nondebug_insn (jump
);
2491 moveall
= can_move_insns_across (currptr
[0], e0_last_head
,
2492 move_before
, jump
, e0
->dest
, live_union
,
2494 if (!moveall
&& move_upto
== NULL_RTX
)
2496 if (jump
== move_before
)
2499 /* Try again, using a different insertion point. */
2503 /* Don't try moving before a cc0 user, as that may invalidate
2505 if (reg_mentioned_p (cc0_rtx
, jump
))
2512 if (final_dest_bb
&& !moveall
)
2513 /* We haven't checked whether a partial move would be OK for the first
2514 move, so we have to fail this case. */
2520 if (currptr
[0] == move_upto
)
2522 for (ix
= 0; ix
< nedges
; ix
++)
2524 rtx curr
= currptr
[ix
];
2526 curr
= NEXT_INSN (curr
);
2527 while (!NONDEBUG_INSN_P (curr
));
2532 /* If we can't currently move all of the identical insns, remember
2533 each insn after the range that we'll merge. */
2535 for (ix
= 0; ix
< nedges
; ix
++)
2537 rtx curr
= currptr
[ix
];
2539 curr
= NEXT_INSN (curr
);
2540 while (!NONDEBUG_INSN_P (curr
));
2544 reorder_insns (headptr
[0], currptr
[0], PREV_INSN (move_before
));
2545 df_set_bb_dirty (EDGE_SUCC (bb
, 0)->dest
);
2546 if (final_dest_bb
!= NULL
)
2547 df_set_bb_dirty (final_dest_bb
);
2548 df_set_bb_dirty (bb
);
2549 for (ix
= 1; ix
< nedges
; ix
++)
2551 df_set_bb_dirty (EDGE_SUCC (bb
, ix
)->dest
);
2552 delete_insn_chain (headptr
[ix
], currptr
[ix
], false);
2556 if (jump
== move_before
)
2559 /* For the unmerged insns, try a different insertion point. */
2563 /* Don't try moving before a cc0 user, as that may invalidate
2565 if (reg_mentioned_p (cc0_rtx
, jump
))
2569 for (ix
= 0; ix
< nedges
; ix
++)
2570 currptr
[ix
] = headptr
[ix
] = nextptr
[ix
];
2580 crossjumps_occured
|= changed
;
2585 /* Return true if BB contains just bb note, or bb note followed
2586 by only DEBUG_INSNs. */
2589 trivially_empty_bb_p (basic_block bb
)
2591 rtx insn
= BB_END (bb
);
2595 if (insn
== BB_HEAD (bb
))
2597 if (!DEBUG_INSN_P (insn
))
2599 insn
= PREV_INSN (insn
);
2603 /* Do simple CFG optimizations - basic block merging, simplifying of jump
2604 instructions etc. Return nonzero if changes were made. */
2607 try_optimize_cfg (int mode
)
2609 bool changed_overall
= false;
2612 basic_block bb
, b
, next
;
2614 if (mode
& (CLEANUP_CROSSJUMP
| CLEANUP_THREADING
))
2617 crossjumps_occured
= false;
2619 FOR_EACH_BB_FN (bb
, cfun
)
2620 update_forwarder_flag (bb
);
2622 if (! targetm
.cannot_modify_jumps_p ())
2625 /* Attempt to merge blocks as made possible by edge removal. If
2626 a block has only one successor, and the successor has only
2627 one predecessor, they may be combined. */
2630 block_was_dirty
= false;
2636 "\n\ntry_optimize_cfg iteration %i\n\n",
2639 for (b
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->next_bb
; b
2640 != EXIT_BLOCK_PTR_FOR_FN (cfun
);)
2644 bool changed_here
= false;
2646 /* Delete trivially dead basic blocks. This is either
2647 blocks with no predecessors, or empty blocks with no
2648 successors. However if the empty block with no
2649 successors is the successor of the ENTRY_BLOCK, it is
2650 kept. This ensures that the ENTRY_BLOCK will have a
2651 successor which is a precondition for many RTL
2652 passes. Empty blocks may result from expanding
2653 __builtin_unreachable (). */
2654 if (EDGE_COUNT (b
->preds
) == 0
2655 || (EDGE_COUNT (b
->succs
) == 0
2656 && trivially_empty_bb_p (b
)
2657 && single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun
))->dest
2661 if (EDGE_COUNT (b
->preds
) > 0)
2666 if (current_ir_type () == IR_RTL_CFGLAYOUT
)
2669 && BARRIER_P (BB_FOOTER (b
)))
2670 FOR_EACH_EDGE (e
, ei
, b
->preds
)
2671 if ((e
->flags
& EDGE_FALLTHRU
)
2672 && BB_FOOTER (e
->src
) == NULL
)
2676 BB_FOOTER (e
->src
) = BB_FOOTER (b
);
2677 BB_FOOTER (b
) = NULL
;
2682 BB_FOOTER (e
->src
) = emit_barrier ();
2689 rtx last
= get_last_bb_insn (b
);
2690 if (last
&& BARRIER_P (last
))
2691 FOR_EACH_EDGE (e
, ei
, b
->preds
)
2692 if ((e
->flags
& EDGE_FALLTHRU
))
2693 emit_barrier_after (BB_END (e
->src
));
2696 delete_basic_block (b
);
2698 /* Avoid trying to remove the exit block. */
2699 b
= (c
== ENTRY_BLOCK_PTR_FOR_FN (cfun
) ? c
->next_bb
: c
);
2703 /* Remove code labels no longer used. */
2704 if (single_pred_p (b
)
2705 && (single_pred_edge (b
)->flags
& EDGE_FALLTHRU
)
2706 && !(single_pred_edge (b
)->flags
& EDGE_COMPLEX
)
2707 && LABEL_P (BB_HEAD (b
))
2708 /* If the previous block ends with a branch to this
2709 block, we can't delete the label. Normally this
2710 is a condjump that is yet to be simplified, but
2711 if CASE_DROPS_THRU, this can be a tablejump with
2712 some element going to the same place as the
2713 default (fallthru). */
2714 && (single_pred (b
) == ENTRY_BLOCK_PTR_FOR_FN (cfun
)
2715 || !JUMP_P (BB_END (single_pred (b
)))
2716 || ! label_is_jump_target_p (BB_HEAD (b
),
2717 BB_END (single_pred (b
)))))
2719 delete_insn (BB_HEAD (b
));
2721 fprintf (dump_file
, "Deleted label in block %i.\n",
2725 /* If we fall through an empty block, we can remove it. */
2726 if (!(mode
& (CLEANUP_CFGLAYOUT
| CLEANUP_NO_INSN_DEL
))
2727 && single_pred_p (b
)
2728 && (single_pred_edge (b
)->flags
& EDGE_FALLTHRU
)
2729 && !LABEL_P (BB_HEAD (b
))
2730 && FORWARDER_BLOCK_P (b
)
2731 /* Note that forwarder_block_p true ensures that
2732 there is a successor for this block. */
2733 && (single_succ_edge (b
)->flags
& EDGE_FALLTHRU
)
2734 && n_basic_blocks_for_fn (cfun
) > NUM_FIXED_BLOCKS
+ 1)
2738 "Deleting fallthru block %i.\n",
2741 c
= ((b
->prev_bb
== ENTRY_BLOCK_PTR_FOR_FN (cfun
))
2742 ? b
->next_bb
: b
->prev_bb
);
2743 redirect_edge_succ_nodup (single_pred_edge (b
),
2745 delete_basic_block (b
);
2751 /* Merge B with its single successor, if any. */
2752 if (single_succ_p (b
)
2753 && (s
= single_succ_edge (b
))
2754 && !(s
->flags
& EDGE_COMPLEX
)
2755 && (c
= s
->dest
) != EXIT_BLOCK_PTR_FOR_FN (cfun
)
2756 && single_pred_p (c
)
2759 /* When not in cfg_layout mode use code aware of reordering
2760 INSN. This code possibly creates new basic blocks so it
2761 does not fit merge_blocks interface and is kept here in
2762 hope that it will become useless once more of compiler
2763 is transformed to use cfg_layout mode. */
2765 if ((mode
& CLEANUP_CFGLAYOUT
)
2766 && can_merge_blocks_p (b
, c
))
2768 merge_blocks (b
, c
);
2769 update_forwarder_flag (b
);
2770 changed_here
= true;
2772 else if (!(mode
& CLEANUP_CFGLAYOUT
)
2773 /* If the jump insn has side effects,
2774 we can't kill the edge. */
2775 && (!JUMP_P (BB_END (b
))
2776 || (reload_completed
2777 ? simplejump_p (BB_END (b
))
2778 : (onlyjump_p (BB_END (b
))
2779 && !tablejump_p (BB_END (b
),
2781 && (next
= merge_blocks_move (s
, b
, c
, mode
)))
2784 changed_here
= true;
2788 /* Simplify branch over branch. */
2789 if ((mode
& CLEANUP_EXPENSIVE
)
2790 && !(mode
& CLEANUP_CFGLAYOUT
)
2791 && try_simplify_condjump (b
))
2792 changed_here
= true;
2794 /* If B has a single outgoing edge, but uses a
2795 non-trivial jump instruction without side-effects, we
2796 can either delete the jump entirely, or replace it
2797 with a simple unconditional jump. */
2798 if (single_succ_p (b
)
2799 && single_succ (b
) != EXIT_BLOCK_PTR_FOR_FN (cfun
)
2800 && onlyjump_p (BB_END (b
))
2801 && !CROSSING_JUMP_P (BB_END (b
))
2802 && try_redirect_by_replacing_jump (single_succ_edge (b
),
2804 (mode
& CLEANUP_CFGLAYOUT
) != 0))
2806 update_forwarder_flag (b
);
2807 changed_here
= true;
2810 /* Simplify branch to branch. */
2811 if (try_forward_edges (mode
, b
))
2813 update_forwarder_flag (b
);
2814 changed_here
= true;
2817 /* Look for shared code between blocks. */
2818 if ((mode
& CLEANUP_CROSSJUMP
)
2819 && try_crossjump_bb (mode
, b
))
2820 changed_here
= true;
2822 if ((mode
& CLEANUP_CROSSJUMP
)
2823 /* This can lengthen register lifetimes. Do it only after
2826 && try_head_merge_bb (b
))
2827 changed_here
= true;
2829 /* Don't get confused by the index shift caused by
2837 if ((mode
& CLEANUP_CROSSJUMP
)
2838 && try_crossjump_bb (mode
, EXIT_BLOCK_PTR_FOR_FN (cfun
)))
2841 if (block_was_dirty
)
2843 /* This should only be set by head-merging. */
2844 gcc_assert (mode
& CLEANUP_CROSSJUMP
);
2850 /* Edge forwarding in particular can cause hot blocks previously
2851 reached by both hot and cold blocks to become dominated only
2852 by cold blocks. This will cause the verification below to fail,
2853 and lead to now cold code in the hot section. This is not easy
2854 to detect and fix during edge forwarding, and in some cases
2855 is only visible after newly unreachable blocks are deleted,
2856 which will be done in fixup_partitions. */
2857 fixup_partitions ();
2859 #ifdef ENABLE_CHECKING
2860 verify_flow_info ();
2864 changed_overall
|= changed
;
2870 FOR_ALL_BB_FN (b
, cfun
)
2871 b
->flags
&= ~(BB_FORWARDER_BLOCK
| BB_NONTHREADABLE_BLOCK
);
2873 return changed_overall
;
2876 /* Delete all unreachable basic blocks. */
2879 delete_unreachable_blocks (void)
2881 bool changed
= false;
2882 basic_block b
, prev_bb
;
2884 find_unreachable_blocks ();
2886 /* When we're in GIMPLE mode and there may be debug insns, we should
2887 delete blocks in reverse dominator order, so as to get a chance
2888 to substitute all released DEFs into debug stmts. If we don't
2889 have dominators information, walking blocks backward gets us a
2890 better chance of retaining most debug information than
2892 if (MAY_HAVE_DEBUG_INSNS
&& current_ir_type () == IR_GIMPLE
2893 && dom_info_available_p (CDI_DOMINATORS
))
2895 for (b
= EXIT_BLOCK_PTR_FOR_FN (cfun
)->prev_bb
;
2896 b
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
); b
= prev_bb
)
2898 prev_bb
= b
->prev_bb
;
2900 if (!(b
->flags
& BB_REACHABLE
))
2902 /* Speed up the removal of blocks that don't dominate
2903 others. Walking backwards, this should be the common
2905 if (!first_dom_son (CDI_DOMINATORS
, b
))
2906 delete_basic_block (b
);
2910 = get_all_dominated_blocks (CDI_DOMINATORS
, b
);
2916 prev_bb
= b
->prev_bb
;
2918 gcc_assert (!(b
->flags
& BB_REACHABLE
));
2920 delete_basic_block (b
);
2932 for (b
= EXIT_BLOCK_PTR_FOR_FN (cfun
)->prev_bb
;
2933 b
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
); b
= prev_bb
)
2935 prev_bb
= b
->prev_bb
;
2937 if (!(b
->flags
& BB_REACHABLE
))
2939 delete_basic_block (b
);
2946 tidy_fallthru_edges ();
2950 /* Delete any jump tables never referenced. We can't delete them at the
2951 time of removing tablejump insn as they are referenced by the preceding
2952 insns computing the destination, so we delay deleting and garbagecollect
2953 them once life information is computed. */
2955 delete_dead_jumptables (void)
2959 /* A dead jump table does not belong to any basic block. Scan insns
2960 between two adjacent basic blocks. */
2961 FOR_EACH_BB_FN (bb
, cfun
)
2965 for (insn
= NEXT_INSN (BB_END (bb
));
2966 insn
&& !NOTE_INSN_BASIC_BLOCK_P (insn
);
2969 next
= NEXT_INSN (insn
);
2971 && LABEL_NUSES (insn
) == LABEL_PRESERVE_P (insn
)
2972 && JUMP_TABLE_DATA_P (next
))
2974 rtx label
= insn
, jump
= next
;
2977 fprintf (dump_file
, "Dead jumptable %i removed\n",
2980 next
= NEXT_INSN (next
);
2982 delete_insn (label
);
2989 /* Tidy the CFG by deleting unreachable code and whatnot. */
2992 cleanup_cfg (int mode
)
2994 bool changed
= false;
2996 /* Set the cfglayout mode flag here. We could update all the callers
2997 but that is just inconvenient, especially given that we eventually
2998 want to have cfglayout mode as the default. */
2999 if (current_ir_type () == IR_RTL_CFGLAYOUT
)
3000 mode
|= CLEANUP_CFGLAYOUT
;
3002 timevar_push (TV_CLEANUP_CFG
);
3003 if (delete_unreachable_blocks ())
3006 /* We've possibly created trivially dead code. Cleanup it right
3007 now to introduce more opportunities for try_optimize_cfg. */
3008 if (!(mode
& (CLEANUP_NO_INSN_DEL
))
3009 && !reload_completed
)
3010 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3015 /* To tail-merge blocks ending in the same noreturn function (e.g.
3016 a call to abort) we have to insert fake edges to exit. Do this
3017 here once. The fake edges do not interfere with any other CFG
3019 if (mode
& CLEANUP_CROSSJUMP
)
3020 add_noreturn_fake_exit_edges ();
3022 if (!dbg_cnt (cfg_cleanup
))
3025 while (try_optimize_cfg (mode
))
3027 delete_unreachable_blocks (), changed
= true;
3028 if (!(mode
& CLEANUP_NO_INSN_DEL
))
3030 /* Try to remove some trivially dead insns when doing an expensive
3031 cleanup. But delete_trivially_dead_insns doesn't work after
3032 reload (it only handles pseudos) and run_fast_dce is too costly
3033 to run in every iteration.
3035 For effective cross jumping, we really want to run a fast DCE to
3036 clean up any dead conditions, or they get in the way of performing
3039 Other transformations in cleanup_cfg are not so sensitive to dead
3040 code, so delete_trivially_dead_insns or even doing nothing at all
3042 if ((mode
& CLEANUP_EXPENSIVE
) && !reload_completed
3043 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
3045 if ((mode
& CLEANUP_CROSSJUMP
) && crossjumps_occured
)
3052 if (mode
& CLEANUP_CROSSJUMP
)
3053 remove_fake_exit_edges ();
3055 /* Don't call delete_dead_jumptables in cfglayout mode, because
3056 that function assumes that jump tables are in the insns stream.
3057 But we also don't _have_ to delete dead jumptables in cfglayout
3058 mode because we shouldn't even be looking at things that are
3059 not in a basic block. Dead jumptables are cleaned up when
3060 going out of cfglayout mode. */
3061 if (!(mode
& CLEANUP_CFGLAYOUT
))
3062 delete_dead_jumptables ();
3064 /* ??? We probably do this way too often. */
3067 || (mode
& CLEANUP_CFG_CHANGED
)))
3069 timevar_push (TV_REPAIR_LOOPS
);
3070 /* The above doesn't preserve dominance info if available. */
3071 gcc_assert (!dom_info_available_p (CDI_DOMINATORS
));
3072 calculate_dominance_info (CDI_DOMINATORS
);
3073 fix_loop_structure (NULL
);
3074 free_dominance_info (CDI_DOMINATORS
);
3075 timevar_pop (TV_REPAIR_LOOPS
);
3078 timevar_pop (TV_CLEANUP_CFG
);
3085 const pass_data pass_data_jump
=
3087 RTL_PASS
, /* type */
3089 OPTGROUP_NONE
, /* optinfo_flags */
3090 TV_JUMP
, /* tv_id */
3091 0, /* properties_required */
3092 0, /* properties_provided */
3093 0, /* properties_destroyed */
3094 0, /* todo_flags_start */
3095 0, /* todo_flags_finish */
3098 class pass_jump
: public rtl_opt_pass
3101 pass_jump (gcc::context
*ctxt
)
3102 : rtl_opt_pass (pass_data_jump
, ctxt
)
3105 /* opt_pass methods: */
3106 virtual unsigned int execute (function
*);
3108 }; // class pass_jump
3111 pass_jump::execute (function
*)
3113 delete_trivially_dead_insns (get_insns (), max_reg_num ());
3115 dump_flow_info (dump_file
, dump_flags
);
3116 cleanup_cfg ((optimize
? CLEANUP_EXPENSIVE
: 0)
3117 | (flag_thread_jumps
? CLEANUP_THREADING
: 0));
3124 make_pass_jump (gcc::context
*ctxt
)
3126 return new pass_jump (ctxt
);
3131 const pass_data pass_data_jump2
=
3133 RTL_PASS
, /* type */
3135 OPTGROUP_NONE
, /* optinfo_flags */
3136 TV_JUMP
, /* tv_id */
3137 0, /* properties_required */
3138 0, /* properties_provided */
3139 0, /* properties_destroyed */
3140 0, /* todo_flags_start */
3141 0, /* todo_flags_finish */
3144 class pass_jump2
: public rtl_opt_pass
3147 pass_jump2 (gcc::context
*ctxt
)
3148 : rtl_opt_pass (pass_data_jump2
, ctxt
)
3151 /* opt_pass methods: */
3152 virtual unsigned int execute (function
*)
3154 cleanup_cfg (flag_crossjumping
? CLEANUP_CROSSJUMP
: 0);
3158 }; // class pass_jump2
3163 make_pass_jump2 (gcc::context
*ctxt
)
3165 return new pass_jump2 (ctxt
);