1 /* Control flow optimization code for GNU compiler.
2 Copyright (C) 1987, 1988, 1992, 1993, 1994, 1995, 1996, 1997, 1998,
3 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007
4 Free Software Foundation, Inc.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 /* This file contains optimizer of the control flow. The main entry point is
23 cleanup_cfg. Following optimizations are performed:
25 - Unreachable blocks removal
26 - Edge forwarding (edge to the forwarder block is forwarded to its
27 successor. Simplification of the branch instruction is performed by
28 underlying infrastructure so branch can be converted to simplejump or
30 - Cross jumping (tail merging)
31 - Conditional jump-around-simplejump simplification
32 - Basic block merging. */
36 #include "coretypes.h"
39 #include "hard-reg-set.h"
43 #include "insn-config.h"
51 #include "cfglayout.h"
53 #include "tree-pass.h"
59 #define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
61 /* Set to true when we are running first pass of try_optimize_cfg loop. */
62 static bool first_pass
;
64 /* Set to true if crossjumps occured in the latest run of try_optimize_cfg. */
65 static bool crossjumps_occured
;
67 static bool try_crossjump_to_edge (int, edge
, edge
);
68 static bool try_crossjump_bb (int, basic_block
);
69 static bool outgoing_edges_match (int, basic_block
, basic_block
);
70 static int flow_find_cross_jump (int, basic_block
, basic_block
, rtx
*, rtx
*);
71 static bool old_insns_match_p (int, rtx
, rtx
);
73 static void merge_blocks_move_predecessor_nojumps (basic_block
, basic_block
);
74 static void merge_blocks_move_successor_nojumps (basic_block
, basic_block
);
75 static bool try_optimize_cfg (int);
76 static bool try_simplify_condjump (basic_block
);
77 static bool try_forward_edges (int, basic_block
);
78 static edge
thread_jump (edge
, basic_block
);
79 static bool mark_effect (rtx
, bitmap
);
80 static void notice_new_block (basic_block
);
81 static void update_forwarder_flag (basic_block
);
82 static int mentions_nonequal_regs (rtx
*, void *);
83 static void merge_memattrs (rtx
, rtx
);
85 /* Set flags for newly created block. */
88 notice_new_block (basic_block bb
)
93 if (forwarder_block_p (bb
))
94 bb
->flags
|= BB_FORWARDER_BLOCK
;
97 /* Recompute forwarder flag after block has been modified. */
100 update_forwarder_flag (basic_block bb
)
102 if (forwarder_block_p (bb
))
103 bb
->flags
|= BB_FORWARDER_BLOCK
;
105 bb
->flags
&= ~BB_FORWARDER_BLOCK
;
108 /* Simplify a conditional jump around an unconditional jump.
109 Return true if something changed. */
112 try_simplify_condjump (basic_block cbranch_block
)
114 basic_block jump_block
, jump_dest_block
, cbranch_dest_block
;
115 edge cbranch_jump_edge
, cbranch_fallthru_edge
;
118 /* Verify that there are exactly two successors. */
119 if (EDGE_COUNT (cbranch_block
->succs
) != 2)
122 /* Verify that we've got a normal conditional branch at the end
124 cbranch_insn
= BB_END (cbranch_block
);
125 if (!any_condjump_p (cbranch_insn
))
128 cbranch_fallthru_edge
= FALLTHRU_EDGE (cbranch_block
);
129 cbranch_jump_edge
= BRANCH_EDGE (cbranch_block
);
131 /* The next block must not have multiple predecessors, must not
132 be the last block in the function, and must contain just the
133 unconditional jump. */
134 jump_block
= cbranch_fallthru_edge
->dest
;
135 if (!single_pred_p (jump_block
)
136 || jump_block
->next_bb
== EXIT_BLOCK_PTR
137 || !FORWARDER_BLOCK_P (jump_block
))
139 jump_dest_block
= single_succ (jump_block
);
141 /* If we are partitioning hot/cold basic blocks, we don't want to
142 mess up unconditional or indirect jumps that cross between hot
145 Basic block partitioning may result in some jumps that appear to
146 be optimizable (or blocks that appear to be mergeable), but which really
147 must be left untouched (they are required to make it safely across
148 partition boundaries). See the comments at the top of
149 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
151 if (BB_PARTITION (jump_block
) != BB_PARTITION (jump_dest_block
)
152 || (cbranch_jump_edge
->flags
& EDGE_CROSSING
))
155 /* The conditional branch must target the block after the
156 unconditional branch. */
157 cbranch_dest_block
= cbranch_jump_edge
->dest
;
159 if (cbranch_dest_block
== EXIT_BLOCK_PTR
160 || !can_fallthru (jump_block
, cbranch_dest_block
))
163 /* Invert the conditional branch. */
164 if (!invert_jump (cbranch_insn
, block_label (jump_dest_block
), 0))
168 fprintf (dump_file
, "Simplifying condjump %i around jump %i\n",
169 INSN_UID (cbranch_insn
), INSN_UID (BB_END (jump_block
)));
171 /* Success. Update the CFG to match. Note that after this point
172 the edge variable names appear backwards; the redirection is done
173 this way to preserve edge profile data. */
174 cbranch_jump_edge
= redirect_edge_succ_nodup (cbranch_jump_edge
,
176 cbranch_fallthru_edge
= redirect_edge_succ_nodup (cbranch_fallthru_edge
,
178 cbranch_jump_edge
->flags
|= EDGE_FALLTHRU
;
179 cbranch_fallthru_edge
->flags
&= ~EDGE_FALLTHRU
;
180 update_br_prob_note (cbranch_block
);
182 /* Delete the block with the unconditional jump, and clean up the mess. */
183 delete_basic_block (jump_block
);
184 tidy_fallthru_edge (cbranch_jump_edge
);
185 update_forwarder_flag (cbranch_block
);
190 /* Attempt to prove that operation is NOOP using CSElib or mark the effect
191 on register. Used by jump threading. */
194 mark_effect (rtx exp
, regset nonequal
)
198 switch (GET_CODE (exp
))
200 /* In case we do clobber the register, mark it as equal, as we know the
201 value is dead so it don't have to match. */
203 if (REG_P (XEXP (exp
, 0)))
205 dest
= XEXP (exp
, 0);
206 regno
= REGNO (dest
);
207 CLEAR_REGNO_REG_SET (nonequal
, regno
);
208 if (regno
< FIRST_PSEUDO_REGISTER
)
210 int n
= hard_regno_nregs
[regno
][GET_MODE (dest
)];
212 CLEAR_REGNO_REG_SET (nonequal
, regno
+ n
);
218 if (rtx_equal_for_cselib_p (SET_DEST (exp
), SET_SRC (exp
)))
220 dest
= SET_DEST (exp
);
225 regno
= REGNO (dest
);
226 SET_REGNO_REG_SET (nonequal
, regno
);
227 if (regno
< FIRST_PSEUDO_REGISTER
)
229 int n
= hard_regno_nregs
[regno
][GET_MODE (dest
)];
231 SET_REGNO_REG_SET (nonequal
, regno
+ n
);
240 /* Return nonzero if X is a register set in regset DATA.
241 Called via for_each_rtx. */
243 mentions_nonequal_regs (rtx
*x
, void *data
)
245 regset nonequal
= (regset
) data
;
251 if (REGNO_REG_SET_P (nonequal
, regno
))
253 if (regno
< FIRST_PSEUDO_REGISTER
)
255 int n
= hard_regno_nregs
[regno
][GET_MODE (*x
)];
257 if (REGNO_REG_SET_P (nonequal
, regno
+ n
))
263 /* Attempt to prove that the basic block B will have no side effects and
264 always continues in the same edge if reached via E. Return the edge
265 if exist, NULL otherwise. */
268 thread_jump (edge e
, basic_block b
)
270 rtx set1
, set2
, cond1
, cond2
, insn
;
271 enum rtx_code code1
, code2
, reversed_code2
;
272 bool reverse1
= false;
276 reg_set_iterator rsi
;
278 if (b
->flags
& BB_NONTHREADABLE_BLOCK
)
281 /* At the moment, we do handle only conditional jumps, but later we may
282 want to extend this code to tablejumps and others. */
283 if (EDGE_COUNT (e
->src
->succs
) != 2)
285 if (EDGE_COUNT (b
->succs
) != 2)
287 b
->flags
|= BB_NONTHREADABLE_BLOCK
;
291 /* Second branch must end with onlyjump, as we will eliminate the jump. */
292 if (!any_condjump_p (BB_END (e
->src
)))
295 if (!any_condjump_p (BB_END (b
)) || !onlyjump_p (BB_END (b
)))
297 b
->flags
|= BB_NONTHREADABLE_BLOCK
;
301 set1
= pc_set (BB_END (e
->src
));
302 set2
= pc_set (BB_END (b
));
303 if (((e
->flags
& EDGE_FALLTHRU
) != 0)
304 != (XEXP (SET_SRC (set1
), 1) == pc_rtx
))
307 cond1
= XEXP (SET_SRC (set1
), 0);
308 cond2
= XEXP (SET_SRC (set2
), 0);
310 code1
= reversed_comparison_code (cond1
, BB_END (e
->src
));
312 code1
= GET_CODE (cond1
);
314 code2
= GET_CODE (cond2
);
315 reversed_code2
= reversed_comparison_code (cond2
, BB_END (b
));
317 if (!comparison_dominates_p (code1
, code2
)
318 && !comparison_dominates_p (code1
, reversed_code2
))
321 /* Ensure that the comparison operators are equivalent.
322 ??? This is far too pessimistic. We should allow swapped operands,
323 different CCmodes, or for example comparisons for interval, that
324 dominate even when operands are not equivalent. */
325 if (!rtx_equal_p (XEXP (cond1
, 0), XEXP (cond2
, 0))
326 || !rtx_equal_p (XEXP (cond1
, 1), XEXP (cond2
, 1)))
329 /* Short circuit cases where block B contains some side effects, as we can't
331 for (insn
= NEXT_INSN (BB_HEAD (b
)); insn
!= NEXT_INSN (BB_END (b
));
332 insn
= NEXT_INSN (insn
))
333 if (INSN_P (insn
) && side_effects_p (PATTERN (insn
)))
335 b
->flags
|= BB_NONTHREADABLE_BLOCK
;
341 /* First process all values computed in the source basic block. */
342 for (insn
= NEXT_INSN (BB_HEAD (e
->src
));
343 insn
!= NEXT_INSN (BB_END (e
->src
));
344 insn
= NEXT_INSN (insn
))
346 cselib_process_insn (insn
);
348 nonequal
= BITMAP_ALLOC (NULL
);
349 CLEAR_REG_SET (nonequal
);
351 /* Now assume that we've continued by the edge E to B and continue
352 processing as if it were same basic block.
353 Our goal is to prove that whole block is an NOOP. */
355 for (insn
= NEXT_INSN (BB_HEAD (b
));
356 insn
!= NEXT_INSN (BB_END (b
)) && !failed
;
357 insn
= NEXT_INSN (insn
))
361 rtx pat
= PATTERN (insn
);
363 if (GET_CODE (pat
) == PARALLEL
)
365 for (i
= 0; i
< (unsigned)XVECLEN (pat
, 0); i
++)
366 failed
|= mark_effect (XVECEXP (pat
, 0, i
), nonequal
);
369 failed
|= mark_effect (pat
, nonequal
);
372 cselib_process_insn (insn
);
375 /* Later we should clear nonequal of dead registers. So far we don't
376 have life information in cfg_cleanup. */
379 b
->flags
|= BB_NONTHREADABLE_BLOCK
;
383 /* cond2 must not mention any register that is not equal to the
385 if (for_each_rtx (&cond2
, mentions_nonequal_regs
, nonequal
))
388 EXECUTE_IF_SET_IN_REG_SET (nonequal
, 0, i
, rsi
)
391 BITMAP_FREE (nonequal
);
393 if ((comparison_dominates_p (code1
, code2
) != 0)
394 != (XEXP (SET_SRC (set2
), 1) == pc_rtx
))
395 return BRANCH_EDGE (b
);
397 return FALLTHRU_EDGE (b
);
400 BITMAP_FREE (nonequal
);
405 /* Attempt to forward edges leaving basic block B.
406 Return true if successful. */
409 try_forward_edges (int mode
, basic_block b
)
411 bool changed
= false;
413 edge e
, *threaded_edges
= NULL
;
415 /* If we are partitioning hot/cold basic blocks, we don't want to
416 mess up unconditional or indirect jumps that cross between hot
419 Basic block partitioning may result in some jumps that appear to
420 be optimizable (or blocks that appear to be mergeable), but which really m
421 ust be left untouched (they are required to make it safely across
422 partition boundaries). See the comments at the top of
423 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
425 if (find_reg_note (BB_END (b
), REG_CROSSING_JUMP
, NULL_RTX
))
428 for (ei
= ei_start (b
->succs
); (e
= ei_safe_edge (ei
)); )
430 basic_block target
, first
;
432 bool threaded
= false;
433 int nthreaded_edges
= 0;
434 bool may_thread
= first_pass
| df_get_bb_dirty (b
);
436 /* Skip complex edges because we don't know how to update them.
438 Still handle fallthru edges, as we can succeed to forward fallthru
439 edge to the same place as the branch edge of conditional branch
440 and turn conditional branch to an unconditional branch. */
441 if (e
->flags
& EDGE_COMPLEX
)
447 target
= first
= e
->dest
;
448 counter
= NUM_FIXED_BLOCKS
;
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
461 && find_reg_note (BB_END (first
), REG_CROSSING_JUMP
, NULL_RTX
))
464 while (counter
< n_basic_blocks
)
466 basic_block new_target
= NULL
;
467 bool new_target_threaded
= false;
468 may_thread
|= df_get_bb_dirty (target
);
470 if (FORWARDER_BLOCK_P (target
)
471 && !(single_succ_edge (target
)->flags
& EDGE_CROSSING
)
472 && single_succ (target
) != EXIT_BLOCK_PTR
)
474 /* Bypass trivial infinite loops. */
475 new_target
= single_succ (target
);
476 if (target
== new_target
)
477 counter
= n_basic_blocks
;
480 /* Allow to thread only over one edge at time to simplify updating
482 else if ((mode
& CLEANUP_THREADING
) && may_thread
)
484 edge t
= thread_jump (e
, target
);
488 threaded_edges
= XNEWVEC (edge
, n_basic_blocks
);
493 /* Detect an infinite loop across blocks not
494 including the start block. */
495 for (i
= 0; i
< nthreaded_edges
; ++i
)
496 if (threaded_edges
[i
] == t
)
498 if (i
< nthreaded_edges
)
500 counter
= n_basic_blocks
;
505 /* Detect an infinite loop across the start block. */
509 gcc_assert (nthreaded_edges
< n_basic_blocks
- NUM_FIXED_BLOCKS
);
510 threaded_edges
[nthreaded_edges
++] = t
;
512 new_target
= t
->dest
;
513 new_target_threaded
= true;
522 threaded
|= new_target_threaded
;
525 if (counter
>= n_basic_blocks
)
528 fprintf (dump_file
, "Infinite loop in BB %i.\n",
531 else if (target
== first
)
532 ; /* We didn't do anything. */
535 /* Save the values now, as the edge may get removed. */
536 gcov_type edge_count
= e
->count
;
537 int edge_probability
= e
->probability
;
541 /* Don't force if target is exit block. */
542 if (threaded
&& target
!= EXIT_BLOCK_PTR
)
544 notice_new_block (redirect_edge_and_branch_force (e
, target
));
546 fprintf (dump_file
, "Conditionals threaded.\n");
548 else if (!redirect_edge_and_branch (e
, target
))
552 "Forwarding edge %i->%i to %i failed.\n",
553 b
->index
, e
->dest
->index
, target
->index
);
558 /* We successfully forwarded the edge. Now update profile
559 data: for each edge we traversed in the chain, remove
560 the original edge's execution count. */
561 edge_frequency
= ((edge_probability
* b
->frequency
562 + REG_BR_PROB_BASE
/ 2)
565 if (!FORWARDER_BLOCK_P (b
) && forwarder_block_p (b
))
566 b
->flags
|= BB_FORWARDER_BLOCK
;
572 if (!single_succ_p (first
))
574 gcc_assert (n
< nthreaded_edges
);
575 t
= threaded_edges
[n
++];
576 gcc_assert (t
->src
== first
);
577 update_bb_profile_for_threading (first
, edge_frequency
,
579 update_br_prob_note (first
);
583 first
->count
-= edge_count
;
584 if (first
->count
< 0)
586 first
->frequency
-= edge_frequency
;
587 if (first
->frequency
< 0)
588 first
->frequency
= 0;
589 /* It is possible that as the result of
590 threading we've removed edge as it is
591 threaded to the fallthru edge. Avoid
592 getting out of sync. */
593 if (n
< nthreaded_edges
594 && first
== threaded_edges
[n
]->src
)
596 t
= single_succ_edge (first
);
599 t
->count
-= edge_count
;
604 while (first
!= target
);
613 free (threaded_edges
);
618 /* Blocks A and B are to be merged into a single block. A has no incoming
619 fallthru edge, so it can be moved before B without adding or modifying
620 any jumps (aside from the jump from A to B). */
623 merge_blocks_move_predecessor_nojumps (basic_block a
, basic_block b
)
627 /* If we are partitioning hot/cold basic blocks, we don't want to
628 mess up unconditional or indirect jumps that cross between hot
631 Basic block partitioning may result in some jumps that appear to
632 be optimizable (or blocks that appear to be mergeable), but which really
633 must be left untouched (they are required to make it safely across
634 partition boundaries). See the comments at the top of
635 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
637 if (BB_PARTITION (a
) != BB_PARTITION (b
))
640 barrier
= next_nonnote_insn (BB_END (a
));
641 gcc_assert (BARRIER_P (barrier
));
642 delete_insn (barrier
);
644 /* Scramble the insn chain. */
645 if (BB_END (a
) != PREV_INSN (BB_HEAD (b
)))
646 reorder_insns_nobb (BB_HEAD (a
), BB_END (a
), PREV_INSN (BB_HEAD (b
)));
650 fprintf (dump_file
, "Moved block %d before %d and merged.\n",
653 /* Swap the records for the two blocks around. */
656 link_block (a
, b
->prev_bb
);
658 /* Now blocks A and B are contiguous. Merge them. */
662 /* Blocks A and B are to be merged into a single block. B has no outgoing
663 fallthru edge, so it can be moved after A without adding or modifying
664 any jumps (aside from the jump from A to B). */
667 merge_blocks_move_successor_nojumps (basic_block a
, basic_block b
)
669 rtx barrier
, real_b_end
;
672 /* If we are partitioning hot/cold basic blocks, we don't want to
673 mess up unconditional or indirect jumps that cross between hot
676 Basic block partitioning may result in some jumps that appear to
677 be optimizable (or blocks that appear to be mergeable), but which really
678 must be left untouched (they are required to make it safely across
679 partition boundaries). See the comments at the top of
680 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
682 if (BB_PARTITION (a
) != BB_PARTITION (b
))
685 real_b_end
= BB_END (b
);
687 /* If there is a jump table following block B temporarily add the jump table
688 to block B so that it will also be moved to the correct location. */
689 if (tablejump_p (BB_END (b
), &label
, &table
)
690 && prev_active_insn (label
) == BB_END (b
))
695 /* There had better have been a barrier there. Delete it. */
696 barrier
= NEXT_INSN (BB_END (b
));
697 if (barrier
&& BARRIER_P (barrier
))
698 delete_insn (barrier
);
701 /* Scramble the insn chain. */
702 reorder_insns_nobb (BB_HEAD (b
), BB_END (b
), BB_END (a
));
704 /* Restore the real end of b. */
705 BB_END (b
) = real_b_end
;
708 fprintf (dump_file
, "Moved block %d after %d and merged.\n",
711 /* Now blocks A and B are contiguous. Merge them. */
715 /* Attempt to merge basic blocks that are potentially non-adjacent.
716 Return NULL iff the attempt failed, otherwise return basic block
717 where cleanup_cfg should continue. Because the merging commonly
718 moves basic block away or introduces another optimization
719 possibility, return basic block just before B so cleanup_cfg don't
722 It may be good idea to return basic block before C in the case
723 C has been moved after B and originally appeared earlier in the
724 insn sequence, but we have no information available about the
725 relative ordering of these two. Hopefully it is not too common. */
728 merge_blocks_move (edge e
, basic_block b
, basic_block c
, int mode
)
732 /* If we are partitioning hot/cold basic blocks, we don't want to
733 mess up unconditional or indirect jumps that cross between hot
736 Basic block partitioning may result in some jumps that appear to
737 be optimizable (or blocks that appear to be mergeable), but which really
738 must be left untouched (they are required to make it safely across
739 partition boundaries). See the comments at the top of
740 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
742 if (BB_PARTITION (b
) != BB_PARTITION (c
))
745 /* If B has a fallthru edge to C, no need to move anything. */
746 if (e
->flags
& EDGE_FALLTHRU
)
748 int b_index
= b
->index
, c_index
= c
->index
;
750 update_forwarder_flag (b
);
753 fprintf (dump_file
, "Merged %d and %d without moving.\n",
756 return b
->prev_bb
== ENTRY_BLOCK_PTR
? b
: b
->prev_bb
;
759 /* Otherwise we will need to move code around. Do that only if expensive
760 transformations are allowed. */
761 else if (mode
& CLEANUP_EXPENSIVE
)
763 edge tmp_edge
, b_fallthru_edge
;
764 bool c_has_outgoing_fallthru
;
765 bool b_has_incoming_fallthru
;
768 /* Avoid overactive code motion, as the forwarder blocks should be
769 eliminated by edge redirection instead. One exception might have
770 been if B is a forwarder block and C has no fallthru edge, but
771 that should be cleaned up by bb-reorder instead. */
772 if (FORWARDER_BLOCK_P (b
) || FORWARDER_BLOCK_P (c
))
775 /* We must make sure to not munge nesting of lexical blocks,
776 and loop notes. This is done by squeezing out all the notes
777 and leaving them there to lie. Not ideal, but functional. */
779 FOR_EACH_EDGE (tmp_edge
, ei
, c
->succs
)
780 if (tmp_edge
->flags
& EDGE_FALLTHRU
)
783 c_has_outgoing_fallthru
= (tmp_edge
!= NULL
);
785 FOR_EACH_EDGE (tmp_edge
, ei
, b
->preds
)
786 if (tmp_edge
->flags
& EDGE_FALLTHRU
)
789 b_has_incoming_fallthru
= (tmp_edge
!= NULL
);
790 b_fallthru_edge
= tmp_edge
;
793 next
= next
->prev_bb
;
795 /* Otherwise, we're going to try to move C after B. If C does
796 not have an outgoing fallthru, then it can be moved
797 immediately after B without introducing or modifying jumps. */
798 if (! c_has_outgoing_fallthru
)
800 merge_blocks_move_successor_nojumps (b
, c
);
801 return next
== ENTRY_BLOCK_PTR
? next
->next_bb
: next
;
804 /* If B does not have an incoming fallthru, then it can be moved
805 immediately before C without introducing or modifying jumps.
806 C cannot be the first block, so we do not have to worry about
807 accessing a non-existent block. */
809 if (b_has_incoming_fallthru
)
813 if (b_fallthru_edge
->src
== ENTRY_BLOCK_PTR
)
815 bb
= force_nonfallthru (b_fallthru_edge
);
817 notice_new_block (bb
);
820 merge_blocks_move_predecessor_nojumps (b
, c
);
821 return next
== ENTRY_BLOCK_PTR
? next
->next_bb
: next
;
828 /* Removes the memory attributes of MEM expression
829 if they are not equal. */
832 merge_memattrs (rtx x
, rtx y
)
841 if (x
== 0 || y
== 0)
846 if (code
!= GET_CODE (y
))
849 if (GET_MODE (x
) != GET_MODE (y
))
852 if (code
== MEM
&& MEM_ATTRS (x
) != MEM_ATTRS (y
))
856 else if (! MEM_ATTRS (y
))
862 if (MEM_ALIAS_SET (x
) != MEM_ALIAS_SET (y
))
864 set_mem_alias_set (x
, 0);
865 set_mem_alias_set (y
, 0);
868 if (! mem_expr_equal_p (MEM_EXPR (x
), MEM_EXPR (y
)))
872 set_mem_offset (x
, 0);
873 set_mem_offset (y
, 0);
875 else if (MEM_OFFSET (x
) != MEM_OFFSET (y
))
877 set_mem_offset (x
, 0);
878 set_mem_offset (y
, 0);
883 else if (!MEM_SIZE (y
))
886 mem_size
= GEN_INT (MAX (INTVAL (MEM_SIZE (x
)),
887 INTVAL (MEM_SIZE (y
))));
888 set_mem_size (x
, mem_size
);
889 set_mem_size (y
, mem_size
);
891 set_mem_align (x
, MIN (MEM_ALIGN (x
), MEM_ALIGN (y
)));
892 set_mem_align (y
, MEM_ALIGN (x
));
896 fmt
= GET_RTX_FORMAT (code
);
897 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
902 /* Two vectors must have the same length. */
903 if (XVECLEN (x
, i
) != XVECLEN (y
, i
))
906 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
907 merge_memattrs (XVECEXP (x
, i
, j
), XVECEXP (y
, i
, j
));
912 merge_memattrs (XEXP (x
, i
), XEXP (y
, i
));
919 /* Return true if I1 and I2 are equivalent and thus can be crossjumped. */
922 old_insns_match_p (int mode ATTRIBUTE_UNUSED
, rtx i1
, rtx i2
)
926 /* Verify that I1 and I2 are equivalent. */
927 if (GET_CODE (i1
) != GET_CODE (i2
))
933 if (GET_CODE (p1
) != GET_CODE (p2
))
936 /* If this is a CALL_INSN, compare register usage information.
937 If we don't check this on stack register machines, the two
938 CALL_INSNs might be merged leaving reg-stack.c with mismatching
939 numbers of stack registers in the same basic block.
940 If we don't check this on machines with delay slots, a delay slot may
941 be filled that clobbers a parameter expected by the subroutine.
943 ??? We take the simple route for now and assume that if they're
944 equal, they were constructed identically. */
947 && (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1
),
948 CALL_INSN_FUNCTION_USAGE (i2
))
949 || SIBLING_CALL_P (i1
) != SIBLING_CALL_P (i2
)))
953 /* If cross_jump_death_matters is not 0, the insn's mode
954 indicates whether or not the insn contains any stack-like
957 if ((mode
& CLEANUP_POST_REGSTACK
) && stack_regs_mentioned (i1
))
959 /* If register stack conversion has already been done, then
960 death notes must also be compared before it is certain that
961 the two instruction streams match. */
964 HARD_REG_SET i1_regset
, i2_regset
;
966 CLEAR_HARD_REG_SET (i1_regset
);
967 CLEAR_HARD_REG_SET (i2_regset
);
969 for (note
= REG_NOTES (i1
); note
; note
= XEXP (note
, 1))
970 if (REG_NOTE_KIND (note
) == REG_DEAD
&& STACK_REG_P (XEXP (note
, 0)))
971 SET_HARD_REG_BIT (i1_regset
, REGNO (XEXP (note
, 0)));
973 for (note
= REG_NOTES (i2
); note
; note
= XEXP (note
, 1))
974 if (REG_NOTE_KIND (note
) == REG_DEAD
&& STACK_REG_P (XEXP (note
, 0)))
975 SET_HARD_REG_BIT (i2_regset
, REGNO (XEXP (note
, 0)));
977 if (!hard_reg_set_equal_p (i1_regset
, i2_regset
))
983 ? rtx_renumbered_equal_p (p1
, p2
) : rtx_equal_p (p1
, p2
))
986 /* Do not do EQUIV substitution after reload. First, we're undoing the
987 work of reload_cse. Second, we may be undoing the work of the post-
988 reload splitting pass. */
989 /* ??? Possibly add a new phase switch variable that can be used by
990 targets to disallow the troublesome insns after splitting. */
991 if (!reload_completed
)
993 /* The following code helps take care of G++ cleanups. */
994 rtx equiv1
= find_reg_equal_equiv_note (i1
);
995 rtx equiv2
= find_reg_equal_equiv_note (i2
);
998 /* If the equivalences are not to a constant, they may
999 reference pseudos that no longer exist, so we can't
1001 && (! reload_completed
1002 || (CONSTANT_P (XEXP (equiv1
, 0))
1003 && rtx_equal_p (XEXP (equiv1
, 0), XEXP (equiv2
, 0)))))
1005 rtx s1
= single_set (i1
);
1006 rtx s2
= single_set (i2
);
1007 if (s1
!= 0 && s2
!= 0
1008 && rtx_renumbered_equal_p (SET_DEST (s1
), SET_DEST (s2
)))
1010 validate_change (i1
, &SET_SRC (s1
), XEXP (equiv1
, 0), 1);
1011 validate_change (i2
, &SET_SRC (s2
), XEXP (equiv2
, 0), 1);
1012 if (! rtx_renumbered_equal_p (p1
, p2
))
1014 else if (apply_change_group ())
1023 /* Look through the insns at the end of BB1 and BB2 and find the longest
1024 sequence that are equivalent. Store the first insns for that sequence
1025 in *F1 and *F2 and return the sequence length.
1027 To simplify callers of this function, if the blocks match exactly,
1028 store the head of the blocks in *F1 and *F2. */
1031 flow_find_cross_jump (int mode ATTRIBUTE_UNUSED
, basic_block bb1
,
1032 basic_block bb2
, rtx
*f1
, rtx
*f2
)
1034 rtx i1
, i2
, last1
, last2
, afterlast1
, afterlast2
;
1037 /* Skip simple jumps at the end of the blocks. Complex jumps still
1038 need to be compared for equivalence, which we'll do below. */
1041 last1
= afterlast1
= last2
= afterlast2
= NULL_RTX
;
1043 || (returnjump_p (i1
) && !side_effects_p (PATTERN (i1
))))
1046 i1
= PREV_INSN (i1
);
1051 || (returnjump_p (i2
) && !side_effects_p (PATTERN (i2
))))
1054 /* Count everything except for unconditional jump as insn. */
1055 if (!simplejump_p (i2
) && !returnjump_p (i2
) && last1
)
1057 i2
= PREV_INSN (i2
);
1063 while (!INSN_P (i1
) && i1
!= BB_HEAD (bb1
))
1064 i1
= PREV_INSN (i1
);
1066 while (!INSN_P (i2
) && i2
!= BB_HEAD (bb2
))
1067 i2
= PREV_INSN (i2
);
1069 if (i1
== BB_HEAD (bb1
) || i2
== BB_HEAD (bb2
))
1072 if (!old_insns_match_p (mode
, i1
, i2
))
1075 merge_memattrs (i1
, i2
);
1077 /* Don't begin a cross-jump with a NOTE insn. */
1080 /* If the merged insns have different REG_EQUAL notes, then
1082 rtx equiv1
= find_reg_equal_equiv_note (i1
);
1083 rtx equiv2
= find_reg_equal_equiv_note (i2
);
1085 if (equiv1
&& !equiv2
)
1086 remove_note (i1
, equiv1
);
1087 else if (!equiv1
&& equiv2
)
1088 remove_note (i2
, equiv2
);
1089 else if (equiv1
&& equiv2
1090 && !rtx_equal_p (XEXP (equiv1
, 0), XEXP (equiv2
, 0)))
1092 remove_note (i1
, equiv1
);
1093 remove_note (i2
, equiv2
);
1096 afterlast1
= last1
, afterlast2
= last2
;
1097 last1
= i1
, last2
= i2
;
1101 i1
= PREV_INSN (i1
);
1102 i2
= PREV_INSN (i2
);
1106 /* Don't allow the insn after a compare to be shared by
1107 cross-jumping unless the compare is also shared. */
1108 if (ninsns
&& reg_mentioned_p (cc0_rtx
, last1
) && ! sets_cc0_p (last1
))
1109 last1
= afterlast1
, last2
= afterlast2
, ninsns
--;
1112 /* Include preceding notes and labels in the cross-jump. One,
1113 this may bring us to the head of the blocks as requested above.
1114 Two, it keeps line number notes as matched as may be. */
1117 while (last1
!= BB_HEAD (bb1
) && !INSN_P (PREV_INSN (last1
)))
1118 last1
= PREV_INSN (last1
);
1120 if (last1
!= BB_HEAD (bb1
) && LABEL_P (PREV_INSN (last1
)))
1121 last1
= PREV_INSN (last1
);
1123 while (last2
!= BB_HEAD (bb2
) && !INSN_P (PREV_INSN (last2
)))
1124 last2
= PREV_INSN (last2
);
1126 if (last2
!= BB_HEAD (bb2
) && LABEL_P (PREV_INSN (last2
)))
1127 last2
= PREV_INSN (last2
);
1136 /* Return true iff the condbranches at the end of BB1 and BB2 match. */
1138 condjump_equiv_p (struct equiv_info
*info
, bool call_init
)
1140 basic_block bb1
= info
->x_block
;
1141 basic_block bb2
= info
->y_block
;
1142 edge b1
= BRANCH_EDGE (bb1
);
1143 edge b2
= BRANCH_EDGE (bb2
);
1144 edge f1
= FALLTHRU_EDGE (bb1
);
1145 edge f2
= FALLTHRU_EDGE (bb2
);
1146 bool reverse
, match
;
1147 rtx set1
, set2
, cond1
, cond2
;
1149 enum rtx_code code1
, code2
;
1151 /* Get around possible forwarders on fallthru edges. Other cases
1152 should be optimized out already. */
1153 if (FORWARDER_BLOCK_P (f1
->dest
))
1154 f1
= single_succ_edge (f1
->dest
);
1156 if (FORWARDER_BLOCK_P (f2
->dest
))
1157 f2
= single_succ_edge (f2
->dest
);
1159 /* To simplify use of this function, return false if there are
1160 unneeded forwarder blocks. These will get eliminated later
1161 during cleanup_cfg. */
1162 if (FORWARDER_BLOCK_P (f1
->dest
)
1163 || FORWARDER_BLOCK_P (f2
->dest
)
1164 || FORWARDER_BLOCK_P (b1
->dest
)
1165 || FORWARDER_BLOCK_P (b2
->dest
))
1168 if (f1
->dest
== f2
->dest
&& b1
->dest
== b2
->dest
)
1170 else if (f1
->dest
== b2
->dest
&& b1
->dest
== f2
->dest
)
1175 set1
= pc_set (BB_END (bb1
));
1176 set2
= pc_set (BB_END (bb2
));
1177 if ((XEXP (SET_SRC (set1
), 1) == pc_rtx
)
1178 != (XEXP (SET_SRC (set2
), 1) == pc_rtx
))
1181 src1
= SET_SRC (set1
);
1182 src2
= SET_SRC (set2
);
1183 cond1
= XEXP (src1
, 0);
1184 cond2
= XEXP (src2
, 0);
1185 code1
= GET_CODE (cond1
);
1187 code2
= reversed_comparison_code (cond2
, BB_END (bb2
));
1189 code2
= GET_CODE (cond2
);
1191 if (code2
== UNKNOWN
)
1194 if (call_init
&& !struct_equiv_init (STRUCT_EQUIV_START
| info
->mode
, info
))
1196 /* Make the sources of the pc sets unreadable so that when we call
1197 insns_match_p it won't process them.
1198 The death_notes_match_p from insns_match_p won't see the local registers
1199 used for the pc set, but that could only cause missed optimizations when
1200 there are actually condjumps that use stack registers. */
1201 SET_SRC (set1
) = pc_rtx
;
1202 SET_SRC (set2
) = pc_rtx
;
1203 /* Verify codes and operands match. */
1206 match
= (insns_match_p (BB_END (bb1
), BB_END (bb2
), info
)
1207 && rtx_equiv_p (&XEXP (cond1
, 0), XEXP (cond2
, 0), 1, info
)
1208 && rtx_equiv_p (&XEXP (cond1
, 1), XEXP (cond2
, 1), 1, info
));
1211 else if (code1
== swap_condition (code2
))
1213 match
= (insns_match_p (BB_END (bb1
), BB_END (bb2
), info
)
1214 && rtx_equiv_p (&XEXP (cond1
, 1), XEXP (cond2
, 0), 1, info
)
1215 && rtx_equiv_p (&XEXP (cond1
, 0), XEXP (cond2
, 1), 1, info
));
1220 SET_SRC (set1
) = src1
;
1221 SET_SRC (set2
) = src2
;
1222 match
&= verify_changes (0);
1224 /* If we return true, we will join the blocks. Which means that
1225 we will only have one branch prediction bit to work with. Thus
1226 we require the existing branches to have probabilities that are
1230 && maybe_hot_bb_p (bb1
)
1231 && maybe_hot_bb_p (bb2
))
1235 if (b1
->dest
== b2
->dest
)
1236 prob2
= b2
->probability
;
1238 /* Do not use f2 probability as f2 may be forwarded. */
1239 prob2
= REG_BR_PROB_BASE
- b2
->probability
;
1241 /* Fail if the difference in probabilities is greater than 50%.
1242 This rules out two well-predicted branches with opposite
1244 if (abs (b1
->probability
- prob2
) > REG_BR_PROB_BASE
/ 2)
1248 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1249 bb1
->index
, bb2
->index
, b1
->probability
, prob2
);
1255 if (dump_file
&& match
)
1256 fprintf (dump_file
, "Conditionals in bb %i and %i match.\n",
1257 bb1
->index
, bb2
->index
);
1264 /* Return true iff outgoing edges of BB1 and BB2 match, together with
1265 the branch instruction. This means that if we commonize the control
1266 flow before end of the basic block, the semantic remains unchanged.
1268 We may assume that there exists one edge with a common destination. */
1271 outgoing_edges_match (int mode
, basic_block bb1
, basic_block bb2
)
1273 int nehedges1
= 0, nehedges2
= 0;
1274 edge fallthru1
= 0, fallthru2
= 0;
1278 /* If BB1 has only one successor, we may be looking at either an
1279 unconditional jump, or a fake edge to exit. */
1280 if (single_succ_p (bb1
)
1281 && (single_succ_edge (bb1
)->flags
& (EDGE_COMPLEX
| EDGE_FAKE
)) == 0
1282 && (!JUMP_P (BB_END (bb1
)) || simplejump_p (BB_END (bb1
))))
1283 return (single_succ_p (bb2
)
1284 && (single_succ_edge (bb2
)->flags
1285 & (EDGE_COMPLEX
| EDGE_FAKE
)) == 0
1286 && (!JUMP_P (BB_END (bb2
)) || simplejump_p (BB_END (bb2
))));
1288 /* Match conditional jumps - this may get tricky when fallthru and branch
1289 edges are crossed. */
1290 if (EDGE_COUNT (bb1
->succs
) == 2
1291 && any_condjump_p (BB_END (bb1
))
1292 && onlyjump_p (BB_END (bb1
)))
1294 edge b1
, f1
, b2
, f2
;
1295 bool reverse
, match
;
1296 rtx set1
, set2
, cond1
, cond2
;
1297 enum rtx_code code1
, code2
;
1299 if (EDGE_COUNT (bb2
->succs
) != 2
1300 || !any_condjump_p (BB_END (bb2
))
1301 || !onlyjump_p (BB_END (bb2
)))
1304 b1
= BRANCH_EDGE (bb1
);
1305 b2
= BRANCH_EDGE (bb2
);
1306 f1
= FALLTHRU_EDGE (bb1
);
1307 f2
= FALLTHRU_EDGE (bb2
);
1309 /* Get around possible forwarders on fallthru edges. Other cases
1310 should be optimized out already. */
1311 if (FORWARDER_BLOCK_P (f1
->dest
))
1312 f1
= single_succ_edge (f1
->dest
);
1314 if (FORWARDER_BLOCK_P (f2
->dest
))
1315 f2
= single_succ_edge (f2
->dest
);
1317 /* To simplify use of this function, return false if there are
1318 unneeded forwarder blocks. These will get eliminated later
1319 during cleanup_cfg. */
1320 if (FORWARDER_BLOCK_P (f1
->dest
)
1321 || FORWARDER_BLOCK_P (f2
->dest
)
1322 || FORWARDER_BLOCK_P (b1
->dest
)
1323 || FORWARDER_BLOCK_P (b2
->dest
))
1326 if (f1
->dest
== f2
->dest
&& b1
->dest
== b2
->dest
)
1328 else if (f1
->dest
== b2
->dest
&& b1
->dest
== f2
->dest
)
1333 set1
= pc_set (BB_END (bb1
));
1334 set2
= pc_set (BB_END (bb2
));
1335 if ((XEXP (SET_SRC (set1
), 1) == pc_rtx
)
1336 != (XEXP (SET_SRC (set2
), 1) == pc_rtx
))
1339 cond1
= XEXP (SET_SRC (set1
), 0);
1340 cond2
= XEXP (SET_SRC (set2
), 0);
1341 code1
= GET_CODE (cond1
);
1343 code2
= reversed_comparison_code (cond2
, BB_END (bb2
));
1345 code2
= GET_CODE (cond2
);
1347 if (code2
== UNKNOWN
)
1350 /* Verify codes and operands match. */
1351 match
= ((code1
== code2
1352 && rtx_renumbered_equal_p (XEXP (cond1
, 0), XEXP (cond2
, 0))
1353 && rtx_renumbered_equal_p (XEXP (cond1
, 1), XEXP (cond2
, 1)))
1354 || (code1
== swap_condition (code2
)
1355 && rtx_renumbered_equal_p (XEXP (cond1
, 1),
1357 && rtx_renumbered_equal_p (XEXP (cond1
, 0),
1360 /* If we return true, we will join the blocks. Which means that
1361 we will only have one branch prediction bit to work with. Thus
1362 we require the existing branches to have probabilities that are
1366 && maybe_hot_bb_p (bb1
)
1367 && maybe_hot_bb_p (bb2
))
1371 if (b1
->dest
== b2
->dest
)
1372 prob2
= b2
->probability
;
1374 /* Do not use f2 probability as f2 may be forwarded. */
1375 prob2
= REG_BR_PROB_BASE
- b2
->probability
;
1377 /* Fail if the difference in probabilities is greater than 50%.
1378 This rules out two well-predicted branches with opposite
1380 if (abs (b1
->probability
- prob2
) > REG_BR_PROB_BASE
/ 2)
1384 "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
1385 bb1
->index
, bb2
->index
, b1
->probability
, prob2
);
1391 if (dump_file
&& match
)
1392 fprintf (dump_file
, "Conditionals in bb %i and %i match.\n",
1393 bb1
->index
, bb2
->index
);
1398 /* Generic case - we are seeing a computed jump, table jump or trapping
1401 /* Check whether there are tablejumps in the end of BB1 and BB2.
1402 Return true if they are identical. */
1407 if (tablejump_p (BB_END (bb1
), &label1
, &table1
)
1408 && tablejump_p (BB_END (bb2
), &label2
, &table2
)
1409 && GET_CODE (PATTERN (table1
)) == GET_CODE (PATTERN (table2
)))
1411 /* The labels should never be the same rtx. If they really are same
1412 the jump tables are same too. So disable crossjumping of blocks BB1
1413 and BB2 because when deleting the common insns in the end of BB1
1414 by delete_basic_block () the jump table would be deleted too. */
1415 /* If LABEL2 is referenced in BB1->END do not do anything
1416 because we would loose information when replacing
1417 LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
1418 if (label1
!= label2
&& !rtx_referenced_p (label2
, BB_END (bb1
)))
1420 /* Set IDENTICAL to true when the tables are identical. */
1421 bool identical
= false;
1424 p1
= PATTERN (table1
);
1425 p2
= PATTERN (table2
);
1426 if (GET_CODE (p1
) == ADDR_VEC
&& rtx_equal_p (p1
, p2
))
1430 else if (GET_CODE (p1
) == ADDR_DIFF_VEC
1431 && (XVECLEN (p1
, 1) == XVECLEN (p2
, 1))
1432 && rtx_equal_p (XEXP (p1
, 2), XEXP (p2
, 2))
1433 && rtx_equal_p (XEXP (p1
, 3), XEXP (p2
, 3)))
1438 for (i
= XVECLEN (p1
, 1) - 1; i
>= 0 && identical
; i
--)
1439 if (!rtx_equal_p (XVECEXP (p1
, 1, i
), XVECEXP (p2
, 1, i
)))
1445 replace_label_data rr
;
1448 /* Temporarily replace references to LABEL1 with LABEL2
1449 in BB1->END so that we could compare the instructions. */
1452 rr
.update_label_nuses
= false;
1453 for_each_rtx (&BB_END (bb1
), replace_label
, &rr
);
1455 match
= old_insns_match_p (mode
, BB_END (bb1
), BB_END (bb2
));
1456 if (dump_file
&& match
)
1458 "Tablejumps in bb %i and %i match.\n",
1459 bb1
->index
, bb2
->index
);
1461 /* Set the original label in BB1->END because when deleting
1462 a block whose end is a tablejump, the tablejump referenced
1463 from the instruction is deleted too. */
1466 for_each_rtx (&BB_END (bb1
), replace_label
, &rr
);
1475 /* First ensure that the instructions match. There may be many outgoing
1476 edges so this test is generally cheaper. */
1477 if (!old_insns_match_p (mode
, BB_END (bb1
), BB_END (bb2
)))
1480 /* Search the outgoing edges, ensure that the counts do match, find possible
1481 fallthru and exception handling edges since these needs more
1483 if (EDGE_COUNT (bb1
->succs
) != EDGE_COUNT (bb2
->succs
))
1486 FOR_EACH_EDGE (e1
, ei
, bb1
->succs
)
1488 e2
= EDGE_SUCC (bb2
, ei
.index
);
1490 if (e1
->flags
& EDGE_EH
)
1493 if (e2
->flags
& EDGE_EH
)
1496 if (e1
->flags
& EDGE_FALLTHRU
)
1498 if (e2
->flags
& EDGE_FALLTHRU
)
1502 /* If number of edges of various types does not match, fail. */
1503 if (nehedges1
!= nehedges2
1504 || (fallthru1
!= 0) != (fallthru2
!= 0))
1507 /* fallthru edges must be forwarded to the same destination. */
1510 basic_block d1
= (forwarder_block_p (fallthru1
->dest
)
1511 ? single_succ (fallthru1
->dest
): fallthru1
->dest
);
1512 basic_block d2
= (forwarder_block_p (fallthru2
->dest
)
1513 ? single_succ (fallthru2
->dest
): fallthru2
->dest
);
1519 /* Ensure the same EH region. */
1521 rtx n1
= find_reg_note (BB_END (bb1
), REG_EH_REGION
, 0);
1522 rtx n2
= find_reg_note (BB_END (bb2
), REG_EH_REGION
, 0);
1527 if (n1
&& (!n2
|| XEXP (n1
, 0) != XEXP (n2
, 0)))
1531 /* The same checks as in try_crossjump_to_edge. It is required for RTL
1532 version of sequence abstraction. */
1533 FOR_EACH_EDGE (e1
, ei
, bb2
->succs
)
1537 basic_block d1
= e1
->dest
;
1539 if (FORWARDER_BLOCK_P (d1
))
1540 d1
= EDGE_SUCC (d1
, 0)->dest
;
1542 FOR_EACH_EDGE (e2
, ei
, bb1
->succs
)
1544 basic_block d2
= e2
->dest
;
1545 if (FORWARDER_BLOCK_P (d2
))
1546 d2
= EDGE_SUCC (d2
, 0)->dest
;
1558 /* Returns true if BB basic block has a preserve label. */
1561 block_has_preserve_label (basic_block bb
)
1565 && LABEL_PRESERVE_P (block_label (bb
)));
1568 /* E1 and E2 are edges with the same destination block. Search their
1569 predecessors for common code. If found, redirect control flow from
1570 (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC. */
1573 try_crossjump_to_edge (int mode
, edge e1
, edge e2
)
1576 basic_block src1
= e1
->src
, src2
= e2
->src
;
1577 basic_block redirect_to
, redirect_from
, to_remove
;
1578 rtx newpos1
, newpos2
;
1582 newpos1
= newpos2
= NULL_RTX
;
1584 /* If we have partitioned hot/cold basic blocks, it is a bad idea
1585 to try this optimization.
1587 Basic block partitioning may result in some jumps that appear to
1588 be optimizable (or blocks that appear to be mergeable), but which really
1589 must be left untouched (they are required to make it safely across
1590 partition boundaries). See the comments at the top of
1591 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1593 if (flag_reorder_blocks_and_partition
&& reload_completed
)
1596 /* Search backward through forwarder blocks. We don't need to worry
1597 about multiple entry or chained forwarders, as they will be optimized
1598 away. We do this to look past the unconditional jump following a
1599 conditional jump that is required due to the current CFG shape. */
1600 if (single_pred_p (src1
)
1601 && FORWARDER_BLOCK_P (src1
))
1602 e1
= single_pred_edge (src1
), src1
= e1
->src
;
1604 if (single_pred_p (src2
)
1605 && FORWARDER_BLOCK_P (src2
))
1606 e2
= single_pred_edge (src2
), src2
= e2
->src
;
1608 /* Nothing to do if we reach ENTRY, or a common source block. */
1609 if (src1
== ENTRY_BLOCK_PTR
|| src2
== ENTRY_BLOCK_PTR
)
1614 /* Seeing more than 1 forwarder blocks would confuse us later... */
1615 if (FORWARDER_BLOCK_P (e1
->dest
)
1616 && FORWARDER_BLOCK_P (single_succ (e1
->dest
)))
1619 if (FORWARDER_BLOCK_P (e2
->dest
)
1620 && FORWARDER_BLOCK_P (single_succ (e2
->dest
)))
1623 /* Likewise with dead code (possibly newly created by the other optimizations
1625 if (EDGE_COUNT (src1
->preds
) == 0 || EDGE_COUNT (src2
->preds
) == 0)
1628 /* Look for the common insn sequence, part the first ... */
1629 if (!outgoing_edges_match (mode
, src1
, src2
))
1632 /* ... and part the second. */
1633 nmatch
= flow_find_cross_jump (mode
, src1
, src2
, &newpos1
, &newpos2
);
1635 /* Don't proceed with the crossjump unless we found a sufficient number
1636 of matching instructions or the 'from' block was totally matched
1637 (such that its predecessors will hopefully be redirected and the
1639 if ((nmatch
< PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS
))
1640 && (newpos1
!= BB_HEAD (src1
)))
1643 /* Avoid deleting preserve label when redirecting ABNORMAL edges. */
1644 if (block_has_preserve_label (e1
->dest
)
1645 && (e1
->flags
& EDGE_ABNORMAL
))
1648 /* Here we know that the insns in the end of SRC1 which are common with SRC2
1650 If we have tablejumps in the end of SRC1 and SRC2
1651 they have been already compared for equivalence in outgoing_edges_match ()
1652 so replace the references to TABLE1 by references to TABLE2. */
1657 if (tablejump_p (BB_END (src1
), &label1
, &table1
)
1658 && tablejump_p (BB_END (src2
), &label2
, &table2
)
1659 && label1
!= label2
)
1661 replace_label_data rr
;
1664 /* Replace references to LABEL1 with LABEL2. */
1667 rr
.update_label_nuses
= true;
1668 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
1670 /* Do not replace the label in SRC1->END because when deleting
1671 a block whose end is a tablejump, the tablejump referenced
1672 from the instruction is deleted too. */
1673 if (insn
!= BB_END (src1
))
1674 for_each_rtx (&insn
, replace_label
, &rr
);
1679 /* Avoid splitting if possible. We must always split when SRC2 has
1680 EH predecessor edges, or we may end up with basic blocks with both
1681 normal and EH predecessor edges. */
1682 if (newpos2
== BB_HEAD (src2
)
1683 && !(EDGE_PRED (src2
, 0)->flags
& EDGE_EH
))
1687 if (newpos2
== BB_HEAD (src2
))
1689 /* Skip possible basic block header. */
1690 if (LABEL_P (newpos2
))
1691 newpos2
= NEXT_INSN (newpos2
);
1692 if (NOTE_P (newpos2
))
1693 newpos2
= NEXT_INSN (newpos2
);
1697 fprintf (dump_file
, "Splitting bb %i before %i insns\n",
1698 src2
->index
, nmatch
);
1699 redirect_to
= split_block (src2
, PREV_INSN (newpos2
))->dest
;
1704 "Cross jumping from bb %i to bb %i; %i common insns\n",
1705 src1
->index
, src2
->index
, nmatch
);
1707 /* We may have some registers visible through the block. */
1708 df_set_bb_dirty (redirect_to
);
1710 /* Recompute the frequencies and counts of outgoing edges. */
1711 FOR_EACH_EDGE (s
, ei
, redirect_to
->succs
)
1715 basic_block d
= s
->dest
;
1717 if (FORWARDER_BLOCK_P (d
))
1718 d
= single_succ (d
);
1720 FOR_EACH_EDGE (s2
, ei
, src1
->succs
)
1722 basic_block d2
= s2
->dest
;
1723 if (FORWARDER_BLOCK_P (d2
))
1724 d2
= single_succ (d2
);
1729 s
->count
+= s2
->count
;
1731 /* Take care to update possible forwarder blocks. We verified
1732 that there is no more than one in the chain, so we can't run
1733 into infinite loop. */
1734 if (FORWARDER_BLOCK_P (s
->dest
))
1736 single_succ_edge (s
->dest
)->count
+= s2
->count
;
1737 s
->dest
->count
+= s2
->count
;
1738 s
->dest
->frequency
+= EDGE_FREQUENCY (s
);
1741 if (FORWARDER_BLOCK_P (s2
->dest
))
1743 single_succ_edge (s2
->dest
)->count
-= s2
->count
;
1744 if (single_succ_edge (s2
->dest
)->count
< 0)
1745 single_succ_edge (s2
->dest
)->count
= 0;
1746 s2
->dest
->count
-= s2
->count
;
1747 s2
->dest
->frequency
-= EDGE_FREQUENCY (s
);
1748 if (s2
->dest
->frequency
< 0)
1749 s2
->dest
->frequency
= 0;
1750 if (s2
->dest
->count
< 0)
1751 s2
->dest
->count
= 0;
1754 if (!redirect_to
->frequency
&& !src1
->frequency
)
1755 s
->probability
= (s
->probability
+ s2
->probability
) / 2;
1758 = ((s
->probability
* redirect_to
->frequency
+
1759 s2
->probability
* src1
->frequency
)
1760 / (redirect_to
->frequency
+ src1
->frequency
));
1763 /* Adjust count and frequency for the block. An earlier jump
1764 threading pass may have left the profile in an inconsistent
1765 state (see update_bb_profile_for_threading) so we must be
1766 prepared for overflows. */
1767 redirect_to
->count
+= src1
->count
;
1768 redirect_to
->frequency
+= src1
->frequency
;
1769 if (redirect_to
->frequency
> BB_FREQ_MAX
)
1770 redirect_to
->frequency
= BB_FREQ_MAX
;
1771 update_br_prob_note (redirect_to
);
1773 /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
1775 /* Skip possible basic block header. */
1776 if (LABEL_P (newpos1
))
1777 newpos1
= NEXT_INSN (newpos1
);
1779 if (NOTE_P (newpos1
))
1780 newpos1
= NEXT_INSN (newpos1
);
1782 redirect_from
= split_block (src1
, PREV_INSN (newpos1
))->src
;
1783 to_remove
= single_succ (redirect_from
);
1785 redirect_edge_and_branch_force (single_succ_edge (redirect_from
), redirect_to
);
1786 delete_basic_block (to_remove
);
1788 update_forwarder_flag (redirect_from
);
1789 if (redirect_to
!= src2
)
1790 update_forwarder_flag (src2
);
1795 /* Search the predecessors of BB for common insn sequences. When found,
1796 share code between them by redirecting control flow. Return true if
1797 any changes made. */
1800 try_crossjump_bb (int mode
, basic_block bb
)
1802 edge e
, e2
, fallthru
;
1804 unsigned max
, ix
, ix2
;
1805 basic_block ev
, ev2
;
1808 /* Nothing to do if there is not at least two incoming edges. */
1809 if (EDGE_COUNT (bb
->preds
) < 2)
1812 /* Don't crossjump if this block ends in a computed jump,
1813 unless we are optimizing for size. */
1815 && bb
!= EXIT_BLOCK_PTR
1816 && computed_jump_p (BB_END (bb
)))
1819 /* If we are partitioning hot/cold basic blocks, we don't want to
1820 mess up unconditional or indirect jumps that cross between hot
1823 Basic block partitioning may result in some jumps that appear to
1824 be optimizable (or blocks that appear to be mergeable), but which really
1825 must be left untouched (they are required to make it safely across
1826 partition boundaries). See the comments at the top of
1827 bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
1829 if (BB_PARTITION (EDGE_PRED (bb
, 0)->src
) !=
1830 BB_PARTITION (EDGE_PRED (bb
, 1)->src
)
1831 || (EDGE_PRED (bb
, 0)->flags
& EDGE_CROSSING
))
1834 /* It is always cheapest to redirect a block that ends in a branch to
1835 a block that falls through into BB, as that adds no branches to the
1836 program. We'll try that combination first. */
1838 max
= PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES
);
1840 if (EDGE_COUNT (bb
->preds
) > max
)
1843 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1845 if (e
->flags
& EDGE_FALLTHRU
)
1853 for (ix
= 0, ev
= bb
; ix
< EDGE_COUNT (ev
->preds
); )
1855 e
= EDGE_PRED (ev
, ix
);
1858 /* As noted above, first try with the fallthru predecessor (or, a
1859 fallthru predecessor if we are in cfglayout mode). */
1862 /* Don't combine the fallthru edge into anything else.
1863 If there is a match, we'll do it the other way around. */
1866 /* If nothing changed since the last attempt, there is nothing
1869 && (!(df_get_bb_dirty (e
->src
))
1870 && !(df_get_bb_dirty (fallthru
->src
))))
1873 if (try_crossjump_to_edge (mode
, e
, fallthru
))
1882 /* Non-obvious work limiting check: Recognize that we're going
1883 to call try_crossjump_bb on every basic block. So if we have
1884 two blocks with lots of outgoing edges (a switch) and they
1885 share lots of common destinations, then we would do the
1886 cross-jump check once for each common destination.
1888 Now, if the blocks actually are cross-jump candidates, then
1889 all of their destinations will be shared. Which means that
1890 we only need check them for cross-jump candidacy once. We
1891 can eliminate redundant checks of crossjump(A,B) by arbitrarily
1892 choosing to do the check from the block for which the edge
1893 in question is the first successor of A. */
1894 if (EDGE_SUCC (e
->src
, 0) != e
)
1897 for (ix2
= 0, ev2
= bb
; ix2
< EDGE_COUNT (ev2
->preds
); )
1899 e2
= EDGE_PRED (ev2
, ix2
);
1905 /* We've already checked the fallthru edge above. */
1909 /* The "first successor" check above only prevents multiple
1910 checks of crossjump(A,B). In order to prevent redundant
1911 checks of crossjump(B,A), require that A be the block
1912 with the lowest index. */
1913 if (e
->src
->index
> e2
->src
->index
)
1916 /* If nothing changed since the last attempt, there is nothing
1919 && (!(df_get_bb_dirty (e
->src
))
1920 && !(df_get_bb_dirty (e2
->src
))))
1923 if (try_crossjump_to_edge (mode
, e
, e2
))
1934 crossjumps_occured
= true;
1939 /* Do simple CFG optimizations - basic block merging, simplifying of jump
1940 instructions etc. Return nonzero if changes were made. */
1943 try_optimize_cfg (int mode
)
1945 bool changed_overall
= false;
1948 basic_block bb
, b
, next
;
1950 if (mode
& (CLEANUP_CROSSJUMP
| CLEANUP_THREADING
))
1953 crossjumps_occured
= false;
1956 update_forwarder_flag (bb
);
1958 if (! targetm
.cannot_modify_jumps_p ())
1961 /* Attempt to merge blocks as made possible by edge removal. If
1962 a block has only one successor, and the successor has only
1963 one predecessor, they may be combined. */
1971 "\n\ntry_optimize_cfg iteration %i\n\n",
1974 for (b
= ENTRY_BLOCK_PTR
->next_bb
; b
!= EXIT_BLOCK_PTR
;)
1978 bool changed_here
= false;
1980 /* Delete trivially dead basic blocks. */
1981 if (EDGE_COUNT (b
->preds
) == 0)
1985 fprintf (dump_file
, "Deleting block %i.\n",
1988 delete_basic_block (b
);
1989 if (!(mode
& CLEANUP_CFGLAYOUT
))
1991 /* Avoid trying to remove ENTRY_BLOCK_PTR. */
1992 b
= (c
== ENTRY_BLOCK_PTR
? c
->next_bb
: c
);
1996 /* Remove code labels no longer used. */
1997 if (single_pred_p (b
)
1998 && (single_pred_edge (b
)->flags
& EDGE_FALLTHRU
)
1999 && !(single_pred_edge (b
)->flags
& EDGE_COMPLEX
)
2000 && LABEL_P (BB_HEAD (b
))
2001 /* If the previous block ends with a branch to this
2002 block, we can't delete the label. Normally this
2003 is a condjump that is yet to be simplified, but
2004 if CASE_DROPS_THRU, this can be a tablejump with
2005 some element going to the same place as the
2006 default (fallthru). */
2007 && (single_pred (b
) == ENTRY_BLOCK_PTR
2008 || !JUMP_P (BB_END (single_pred (b
)))
2009 || ! label_is_jump_target_p (BB_HEAD (b
),
2010 BB_END (single_pred (b
)))))
2012 rtx label
= BB_HEAD (b
);
2014 delete_insn_chain (label
, label
, false);
2015 /* If the case label is undeletable, move it after the
2016 BASIC_BLOCK note. */
2017 if (NOTE_KIND (BB_HEAD (b
)) == NOTE_INSN_DELETED_LABEL
)
2019 rtx bb_note
= NEXT_INSN (BB_HEAD (b
));
2021 reorder_insns_nobb (label
, label
, bb_note
);
2022 BB_HEAD (b
) = bb_note
;
2023 if (BB_END (b
) == bb_note
)
2027 fprintf (dump_file
, "Deleted label in block %i.\n",
2031 /* If we fall through an empty block, we can remove it. */
2032 if (!(mode
& CLEANUP_CFGLAYOUT
)
2033 && single_pred_p (b
)
2034 && (single_pred_edge (b
)->flags
& EDGE_FALLTHRU
)
2035 && !LABEL_P (BB_HEAD (b
))
2036 && FORWARDER_BLOCK_P (b
)
2037 /* Note that forwarder_block_p true ensures that
2038 there is a successor for this block. */
2039 && (single_succ_edge (b
)->flags
& EDGE_FALLTHRU
)
2040 && n_basic_blocks
> NUM_FIXED_BLOCKS
+ 1)
2044 "Deleting fallthru block %i.\n",
2047 c
= b
->prev_bb
== ENTRY_BLOCK_PTR
? b
->next_bb
: b
->prev_bb
;
2048 redirect_edge_succ_nodup (single_pred_edge (b
),
2050 delete_basic_block (b
);
2055 if (single_succ_p (b
)
2056 && (s
= single_succ_edge (b
))
2057 && !(s
->flags
& EDGE_COMPLEX
)
2058 && (c
= s
->dest
) != EXIT_BLOCK_PTR
2059 && single_pred_p (c
)
2062 /* When not in cfg_layout mode use code aware of reordering
2063 INSN. This code possibly creates new basic blocks so it
2064 does not fit merge_blocks interface and is kept here in
2065 hope that it will become useless once more of compiler
2066 is transformed to use cfg_layout mode. */
2068 if ((mode
& CLEANUP_CFGLAYOUT
)
2069 && can_merge_blocks_p (b
, c
))
2071 merge_blocks (b
, c
);
2072 update_forwarder_flag (b
);
2073 changed_here
= true;
2075 else if (!(mode
& CLEANUP_CFGLAYOUT
)
2076 /* If the jump insn has side effects,
2077 we can't kill the edge. */
2078 && (!JUMP_P (BB_END (b
))
2079 || (reload_completed
2080 ? simplejump_p (BB_END (b
))
2081 : (onlyjump_p (BB_END (b
))
2082 && !tablejump_p (BB_END (b
),
2084 && (next
= merge_blocks_move (s
, b
, c
, mode
)))
2087 changed_here
= true;
2091 /* Simplify branch over branch. */
2092 if ((mode
& CLEANUP_EXPENSIVE
)
2093 && !(mode
& CLEANUP_CFGLAYOUT
)
2094 && try_simplify_condjump (b
))
2095 changed_here
= true;
2097 /* If B has a single outgoing edge, but uses a
2098 non-trivial jump instruction without side-effects, we
2099 can either delete the jump entirely, or replace it
2100 with a simple unconditional jump. */
2101 if (single_succ_p (b
)
2102 && single_succ (b
) != EXIT_BLOCK_PTR
2103 && onlyjump_p (BB_END (b
))
2104 && !find_reg_note (BB_END (b
), REG_CROSSING_JUMP
, NULL_RTX
)
2105 && try_redirect_by_replacing_jump (single_succ_edge (b
),
2107 (mode
& CLEANUP_CFGLAYOUT
) != 0))
2109 update_forwarder_flag (b
);
2110 changed_here
= true;
2113 /* Simplify branch to branch. */
2114 if (try_forward_edges (mode
, b
))
2115 changed_here
= true;
2117 /* Look for shared code between blocks. */
2118 if ((mode
& CLEANUP_CROSSJUMP
)
2119 && try_crossjump_bb (mode
, b
))
2120 changed_here
= true;
2122 /* Don't get confused by the index shift caused by
2130 if ((mode
& CLEANUP_CROSSJUMP
)
2131 && try_crossjump_bb (mode
, EXIT_BLOCK_PTR
))
2134 #ifdef ENABLE_CHECKING
2136 verify_flow_info ();
2139 changed_overall
|= changed
;
2146 b
->flags
&= ~(BB_FORWARDER_BLOCK
| BB_NONTHREADABLE_BLOCK
);
2148 return changed_overall
;
2151 /* Delete all unreachable basic blocks. */
2154 delete_unreachable_blocks (void)
2156 bool changed
= false;
2157 basic_block b
, next_bb
;
2159 find_unreachable_blocks ();
2161 /* Delete all unreachable basic blocks. */
2163 for (b
= ENTRY_BLOCK_PTR
->next_bb
; b
!= EXIT_BLOCK_PTR
; b
= next_bb
)
2165 next_bb
= b
->next_bb
;
2167 if (!(b
->flags
& BB_REACHABLE
))
2169 delete_basic_block (b
);
2175 tidy_fallthru_edges ();
2179 /* Delete any jump tables never referenced. We can't delete them at the
2180 time of removing tablejump insn as they are referenced by the preceding
2181 insns computing the destination, so we delay deleting and garbagecollect
2182 them once life information is computed. */
2184 delete_dead_jumptables (void)
2188 /* A dead jump table does not belong to any basic block. Scan insns
2189 between two adjacent basic blocks. */
2194 for (insn
= NEXT_INSN (BB_END (bb
));
2195 insn
&& !NOTE_INSN_BASIC_BLOCK_P (insn
);
2198 next
= NEXT_INSN (insn
);
2200 && LABEL_NUSES (insn
) == LABEL_PRESERVE_P (insn
)
2202 && (GET_CODE (PATTERN (next
)) == ADDR_VEC
2203 || GET_CODE (PATTERN (next
)) == ADDR_DIFF_VEC
))
2205 rtx label
= insn
, jump
= next
;
2208 fprintf (dump_file
, "Dead jumptable %i removed\n",
2211 next
= NEXT_INSN (next
);
2213 delete_insn (label
);
2220 /* Tidy the CFG by deleting unreachable code and whatnot. */
2223 cleanup_cfg (int mode
)
2225 bool changed
= false;
2227 /* Set the cfglayout mode flag here. We could update all the callers
2228 but that is just inconvenient, especially given that we eventually
2229 want to have cfglayout mode as the default. */
2230 if (current_ir_type () == IR_RTL_CFGLAYOUT
)
2231 mode
|= CLEANUP_CFGLAYOUT
;
2233 timevar_push (TV_CLEANUP_CFG
);
2234 if (delete_unreachable_blocks ())
2237 /* We've possibly created trivially dead code. Cleanup it right
2238 now to introduce more opportunities for try_optimize_cfg. */
2239 if (!(mode
& (CLEANUP_NO_INSN_DEL
))
2240 && !reload_completed
)
2241 delete_trivially_dead_insns (get_insns (), max_reg_num ());
2246 /* To tail-merge blocks ending in the same noreturn function (e.g.
2247 a call to abort) we have to insert fake edges to exit. Do this
2248 here once. The fake edges do not interfere with any other CFG
2250 if (mode
& CLEANUP_CROSSJUMP
)
2251 add_noreturn_fake_exit_edges ();
2253 while (try_optimize_cfg (mode
))
2255 delete_unreachable_blocks (), changed
= true;
2256 if (!(mode
& CLEANUP_NO_INSN_DEL
))
2258 /* Try to remove some trivially dead insns when doing an expensive
2259 cleanup. But delete_trivially_dead_insns doesn't work after
2260 reload (it only handles pseudos) and run_fast_dce is too costly
2261 to run in every iteration.
2263 For effective cross jumping, we really want to run a fast DCE to
2264 clean up any dead conditions, or they get in the way of performing
2267 Other transformations in cleanup_cfg are not so sensitive to dead
2268 code, so delete_trivially_dead_insns or even doing nothing at all
2270 if ((mode
& CLEANUP_EXPENSIVE
) && !reload_completed
2271 && !delete_trivially_dead_insns (get_insns (), max_reg_num ()))
2273 else if ((mode
& CLEANUP_CROSSJUMP
)
2274 && crossjumps_occured
)
2281 if (mode
& CLEANUP_CROSSJUMP
)
2282 remove_fake_exit_edges ();
2284 /* Don't call delete_dead_jumptables in cfglayout mode, because
2285 that function assumes that jump tables are in the insns stream.
2286 But we also don't _have_ to delete dead jumptables in cfglayout
2287 mode because we shouldn't even be looking at things that are
2288 not in a basic block. Dead jumptables are cleaned up when
2289 going out of cfglayout mode. */
2290 if (!(mode
& CLEANUP_CFGLAYOUT
))
2291 delete_dead_jumptables ();
2293 timevar_pop (TV_CLEANUP_CFG
);
2299 rest_of_handle_jump (void)
2301 delete_unreachable_blocks ();
2303 if (cfun
->tail_call_emit
)
2304 fixup_tail_calls ();
2308 struct tree_opt_pass pass_jump
=
2310 "sibling", /* name */
2312 rest_of_handle_jump
, /* execute */
2315 0, /* static_pass_number */
2316 TV_JUMP
, /* tv_id */
2317 0, /* properties_required */
2318 0, /* properties_provided */
2319 0, /* properties_destroyed */
2320 TODO_ggc_collect
, /* todo_flags_start */
2321 TODO_verify_flow
, /* todo_flags_finish */
2327 rest_of_handle_jump2 (void)
2329 delete_trivially_dead_insns (get_insns (), max_reg_num ());
2331 dump_flow_info (dump_file
, dump_flags
);
2332 cleanup_cfg ((optimize
? CLEANUP_EXPENSIVE
: 0)
2333 | (flag_thread_jumps
? CLEANUP_THREADING
: 0));
2338 struct tree_opt_pass pass_jump2
=
2342 rest_of_handle_jump2
, /* execute */
2345 0, /* static_pass_number */
2346 TV_JUMP
, /* tv_id */
2347 0, /* properties_required */
2348 0, /* properties_provided */
2349 0, /* properties_destroyed */
2350 TODO_ggc_collect
, /* todo_flags_start */
2351 TODO_dump_func
| TODO_verify_rtl_sharing
,/* todo_flags_finish */