1 /* Tail merging for gimple.
2 Copyright (C) 2011-2013 Free Software Foundation, Inc.
3 Contributed by Tom de Vries (tom@codesourcery.com)
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3, or (at your option)
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
26 gimple representation of gcc/testsuite/gcc.dg/pr43864.c at
28 hprofStartupp (charD.1 * outputFileNameD.2600, charD.1 * ctxD.2601)
30 struct FILED.1638 * fpD.2605;
31 charD.1 fileNameD.2604[1000];
33 const charD.1 * restrict outputFileName.0D.3914;
36 # PRED: ENTRY [100.0%] (fallthru,exec)
37 # PT = nonlocal { D.3926 } (restr)
38 outputFileName.0D.3914_3
39 = (const charD.1 * restrict) outputFileNameD.2600_2(D);
40 # .MEMD.3923_13 = VDEF <.MEMD.3923_12(D)>
41 # USE = nonlocal null { fileNameD.2604 D.3926 } (restr)
42 # CLB = nonlocal null { fileNameD.2604 D.3926 } (restr)
43 sprintfD.759 (&fileNameD.2604, outputFileName.0D.3914_3);
44 # .MEMD.3923_14 = VDEF <.MEMD.3923_13>
45 # USE = nonlocal null { fileNameD.2604 D.3926 } (restr)
46 # CLB = nonlocal null { fileNameD.2604 D.3926 } (restr)
47 D.3915_4 = accessD.2606 (&fileNameD.2604, 1);
52 # SUCC: 3 [10.0%] (true,exec) 4 [90.0%] (false,exec)
55 # PRED: 2 [10.0%] (true,exec)
56 # .MEMD.3923_15 = VDEF <.MEMD.3923_14>
57 # USE = nonlocal null { fileNameD.2604 D.3926 } (restr)
58 # CLB = nonlocal null { fileNameD.2604 D.3926 } (restr)
59 freeD.898 (ctxD.2601_5(D));
61 # SUCC: 7 [100.0%] (fallthru,exec)
64 # PRED: 2 [90.0%] (false,exec)
65 # .MEMD.3923_16 = VDEF <.MEMD.3923_14>
66 # PT = nonlocal escaped
67 # USE = nonlocal null { fileNameD.2604 D.3926 } (restr)
68 # CLB = nonlocal null { fileNameD.2604 D.3926 } (restr)
69 fpD.2605_8 = fopenD.1805 (&fileNameD.2604[0], 0B);
74 # SUCC: 5 [1.9%] (true,exec) 6 [98.1%] (false,exec)
77 # PRED: 4 [1.9%] (true,exec)
78 # .MEMD.3923_17 = VDEF <.MEMD.3923_16>
79 # USE = nonlocal null { fileNameD.2604 D.3926 } (restr)
80 # CLB = nonlocal null { fileNameD.2604 D.3926 } (restr)
81 freeD.898 (ctxD.2601_5(D));
83 # SUCC: 7 [100.0%] (fallthru,exec)
86 # PRED: 4 [98.1%] (false,exec)
87 # .MEMD.3923_18 = VDEF <.MEMD.3923_16>
88 # USE = nonlocal null { fileNameD.2604 D.3926 } (restr)
89 # CLB = nonlocal null { fileNameD.2604 D.3926 } (restr)
90 fooD.2599 (outputFileNameD.2600_2(D), fpD.2605_8);
91 # SUCC: 7 [100.0%] (fallthru,exec)
94 # PRED: 3 [100.0%] (fallthru,exec) 5 [100.0%] (fallthru,exec)
95 6 [100.0%] (fallthru,exec)
98 # ctxD.2601_1 = PHI <0B(3), 0B(5), ctxD.2601_5(D)(6)>
99 # .MEMD.3923_11 = PHI <.MEMD.3923_15(3), .MEMD.3923_17(5),
101 # VUSE <.MEMD.3923_11>
103 # SUCC: EXIT [100.0%]
106 bb 3 and bb 5 can be merged. The blocks have different predecessors, but the
107 same successors, and the same operations.
112 A technique called tail merging (or cross jumping) can fix the example
113 above. For a block, we look for common code at the end (the tail) of the
114 predecessor blocks, and insert jumps from one block to the other.
115 The example is a special case for tail merging, in that 2 whole blocks
116 can be merged, rather than just the end parts of it.
117 We currently only focus on whole block merging, so in that sense
118 calling this pass tail merge is a bit of a misnomer.
120 We distinguish 2 kinds of situations in which blocks can be merged:
121 - same operations, same predecessors. The successor edges coming from one
122 block are redirected to come from the other block.
123 - same operations, same successors. The predecessor edges entering one block
124 are redirected to enter the other block. Note that this operation might
125 involve introducing phi operations.
127 For efficient implementation, we would like to value numbers the blocks, and
128 have a comparison operator that tells us whether the blocks are equal.
129 Besides being runtime efficient, block value numbering should also abstract
130 from irrelevant differences in order of operations, much like normal value
131 numbering abstracts from irrelevant order of operations.
133 For the first situation (same_operations, same predecessors), normal value
134 numbering fits well. We can calculate a block value number based on the
135 value numbers of the defs and vdefs.
137 For the second situation (same operations, same successors), this approach
138 doesn't work so well. We can illustrate this using the example. The calls
139 to free use different vdefs: MEMD.3923_16 and MEMD.3923_14, and these will
140 remain different in value numbering, since they represent different memory
141 states. So the resulting vdefs of the frees will be different in value
142 numbering, so the block value numbers will be different.
144 The reason why we call the blocks equal is not because they define the same
145 values, but because uses in the blocks use (possibly different) defs in the
146 same way. To be able to detect this efficiently, we need to do some kind of
147 reverse value numbering, meaning number the uses rather than the defs, and
148 calculate a block value number based on the value number of the uses.
149 Ideally, a block comparison operator will also indicate which phis are needed
152 For the moment, we don't do block value numbering, but we do insn-by-insn
153 matching, using scc value numbers to match operations with results, and
154 structural comparison otherwise, while ignoring vop mismatches.
159 1. The pass first determines all groups of blocks with the same successor
161 2. Within each group, it tries to determine clusters of equal basic blocks.
162 3. The clusters are applied.
163 4. The same successor groups are updated.
164 5. This process is repeated from 2 onwards, until no more changes.
170 - handles only 'same operations, same successors'.
171 It handles same predecessors as a special subcase though.
172 - does not implement the reverse value numbering and block value numbering.
173 - improve memory allocation: use garbage collected memory, obstacks,
174 allocpools where appropriate.
175 - no insertion of gimple_reg phis, We only introduce vop-phis.
176 - handle blocks with gimple_reg phi_nodes.
180 This 'pass' is not a stand-alone gimple pass, but runs as part of
181 pass_pre, in order to share the value numbering.
186 - ftree-tail-merge. On at -O2. We may have to enable it only at -Os. */
190 #include "coretypes.h"
194 #include "basic-block.h"
196 #include "function.h"
198 #include "gimple-iterator.h"
199 #include "gimple-ssa.h"
200 #include "tree-cfg.h"
201 #include "tree-phinodes.h"
202 #include "ssa-iterators.h"
203 #include "tree-into-ssa.h"
204 #include "tree-ssa-alias.h"
206 #include "hash-table.h"
207 #include "gimple-pretty-print.h"
208 #include "tree-ssa-sccvn.h"
209 #include "tree-dump.h"
211 #include "tree-pass.h"
213 /* Describes a group of bbs with the same successors. The successor bbs are
214 cached in succs, and the successor edge flags are cached in succ_flags.
215 If a bb has the EDGE_TRUE/VALSE_VALUE flags swapped compared to succ_flags,
216 it's marked in inverse.
217 Additionally, the hash value for the struct is cached in hashval, and
218 in_worklist indicates whether it's currently part of worklist. */
222 /* The bbs that have the same successor bbs. */
224 /* The successor bbs. */
226 /* Indicates whether the EDGE_TRUE/FALSE_VALUEs of succ_flags are swapped for
229 /* The edge flags for each of the successor bbs. */
231 /* Indicates whether the struct is currently in the worklist. */
233 /* The hash value of the struct. */
236 /* hash_table support. */
237 typedef same_succ_def value_type
;
238 typedef same_succ_def compare_type
;
239 static inline hashval_t
hash (const value_type
*);
240 static int equal (const value_type
*, const compare_type
*);
241 static void remove (value_type
*);
243 typedef struct same_succ_def
*same_succ
;
244 typedef const struct same_succ_def
*const_same_succ
;
246 /* hash routine for hash_table support, returns hashval of E. */
249 same_succ_def::hash (const value_type
*e
)
254 /* A group of bbs where 1 bb from bbs can replace the other bbs. */
256 struct bb_cluster_def
258 /* The bbs in the cluster. */
260 /* The preds of the bbs in the cluster. */
262 /* Index in all_clusters vector. */
264 /* The bb to replace the cluster with. */
267 typedef struct bb_cluster_def
*bb_cluster
;
268 typedef const struct bb_cluster_def
*const_bb_cluster
;
274 /* The number of non-debug statements in the bb. */
276 /* The same_succ that this bb is a member of. */
277 same_succ bb_same_succ
;
278 /* The cluster that this bb is a member of. */
280 /* The vop state at the exit of a bb. This is shortlived data, used to
281 communicate data between update_block_by and update_vuses. */
283 /* The bb that either contains or is dominated by the dependencies of the
288 /* Macros to access the fields of struct aux_bb_info. */
290 #define BB_SIZE(bb) (((struct aux_bb_info *)bb->aux)->size)
291 #define BB_SAME_SUCC(bb) (((struct aux_bb_info *)bb->aux)->bb_same_succ)
292 #define BB_CLUSTER(bb) (((struct aux_bb_info *)bb->aux)->cluster)
293 #define BB_VOP_AT_EXIT(bb) (((struct aux_bb_info *)bb->aux)->vop_at_exit)
294 #define BB_DEP_BB(bb) (((struct aux_bb_info *)bb->aux)->dep_bb)
296 /* Returns true if the only effect a statement STMT has, is to define locally
300 stmt_local_def (gimple stmt
)
302 basic_block bb
, def_bb
;
303 imm_use_iterator iter
;
308 if (gimple_has_side_effects (stmt
)
309 || gimple_vdef (stmt
) != NULL_TREE
)
312 def_p
= SINGLE_SSA_DEF_OPERAND (stmt
, SSA_OP_DEF
);
316 val
= DEF_FROM_PTR (def_p
);
317 if (val
== NULL_TREE
|| TREE_CODE (val
) != SSA_NAME
)
320 def_bb
= gimple_bb (stmt
);
322 FOR_EACH_IMM_USE_FAST (use_p
, iter
, val
)
324 if (is_gimple_debug (USE_STMT (use_p
)))
326 bb
= gimple_bb (USE_STMT (use_p
));
330 if (gimple_code (USE_STMT (use_p
)) == GIMPLE_PHI
331 && EDGE_PRED (bb
, PHI_ARG_INDEX_FROM_USE (use_p
))->src
== def_bb
)
340 /* Let GSI skip forwards over local defs. */
343 gsi_advance_fw_nondebug_nonlocal (gimple_stmt_iterator
*gsi
)
349 if (gsi_end_p (*gsi
))
351 stmt
= gsi_stmt (*gsi
);
352 if (!stmt_local_def (stmt
))
354 gsi_next_nondebug (gsi
);
358 /* VAL1 and VAL2 are either:
359 - uses in BB1 and BB2, or
360 - phi alternatives for BB1 and BB2.
361 Return true if the uses have the same gvn value. */
364 gvn_uses_equal (tree val1
, tree val2
)
366 gcc_checking_assert (val1
!= NULL_TREE
&& val2
!= NULL_TREE
);
371 if (vn_valueize (val1
) != vn_valueize (val2
))
374 return ((TREE_CODE (val1
) == SSA_NAME
|| CONSTANT_CLASS_P (val1
))
375 && (TREE_CODE (val2
) == SSA_NAME
|| CONSTANT_CLASS_P (val2
)));
378 /* Prints E to FILE. */
381 same_succ_print (FILE *file
, const same_succ e
)
384 bitmap_print (file
, e
->bbs
, "bbs:", "\n");
385 bitmap_print (file
, e
->succs
, "succs:", "\n");
386 bitmap_print (file
, e
->inverse
, "inverse:", "\n");
387 fprintf (file
, "flags:");
388 for (i
= 0; i
< e
->succ_flags
.length (); ++i
)
389 fprintf (file
, " %x", e
->succ_flags
[i
]);
390 fprintf (file
, "\n");
393 /* Prints same_succ VE to VFILE. */
396 ssa_same_succ_print_traverse (same_succ
*pe
, FILE *file
)
398 const same_succ e
= *pe
;
399 same_succ_print (file
, e
);
403 /* Update BB_DEP_BB (USE_BB), given a use of VAL in USE_BB. */
406 update_dep_bb (basic_block use_bb
, tree val
)
411 if (TREE_CODE (val
) != SSA_NAME
)
414 /* Skip use of global def. */
415 if (SSA_NAME_IS_DEFAULT_DEF (val
))
418 /* Skip use of local def. */
419 dep_bb
= gimple_bb (SSA_NAME_DEF_STMT (val
));
420 if (dep_bb
== use_bb
)
423 if (BB_DEP_BB (use_bb
) == NULL
424 || dominated_by_p (CDI_DOMINATORS
, dep_bb
, BB_DEP_BB (use_bb
)))
425 BB_DEP_BB (use_bb
) = dep_bb
;
428 /* Update BB_DEP_BB, given the dependencies in STMT. */
431 stmt_update_dep_bb (gimple stmt
)
436 FOR_EACH_SSA_USE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
437 update_dep_bb (gimple_bb (stmt
), USE_FROM_PTR (use
));
440 /* Calculates hash value for same_succ VE. */
443 same_succ_hash (const_same_succ e
)
445 hashval_t hashval
= bitmap_hash (e
->succs
);
448 unsigned int first
= bitmap_first_set_bit (e
->bbs
);
449 basic_block bb
= BASIC_BLOCK (first
);
451 gimple_stmt_iterator gsi
;
457 for (gsi
= gsi_start_nondebug_bb (bb
);
458 !gsi_end_p (gsi
); gsi_next_nondebug (&gsi
))
460 stmt
= gsi_stmt (gsi
);
461 stmt_update_dep_bb (stmt
);
462 if (stmt_local_def (stmt
))
466 hashval
= iterative_hash_hashval_t (gimple_code (stmt
), hashval
);
467 if (is_gimple_assign (stmt
))
468 hashval
= iterative_hash_hashval_t (gimple_assign_rhs_code (stmt
),
470 if (!is_gimple_call (stmt
))
472 if (gimple_call_internal_p (stmt
))
473 hashval
= iterative_hash_hashval_t
474 ((hashval_t
) gimple_call_internal_fn (stmt
), hashval
);
476 hashval
= iterative_hash_expr (gimple_call_fn (stmt
), hashval
);
477 for (i
= 0; i
< gimple_call_num_args (stmt
); i
++)
479 arg
= gimple_call_arg (stmt
, i
);
480 arg
= vn_valueize (arg
);
481 hashval
= iterative_hash_expr (arg
, hashval
);
485 hashval
= iterative_hash_hashval_t (size
, hashval
);
488 for (i
= 0; i
< e
->succ_flags
.length (); ++i
)
490 flags
= e
->succ_flags
[i
];
491 flags
= flags
& ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
492 hashval
= iterative_hash_hashval_t (flags
, hashval
);
495 EXECUTE_IF_SET_IN_BITMAP (e
->succs
, 0, s
, bs
)
497 int n
= find_edge (bb
, BASIC_BLOCK (s
))->dest_idx
;
498 for (gsi
= gsi_start_phis (BASIC_BLOCK (s
)); !gsi_end_p (gsi
);
501 gimple phi
= gsi_stmt (gsi
);
502 tree lhs
= gimple_phi_result (phi
);
503 tree val
= gimple_phi_arg_def (phi
, n
);
505 if (virtual_operand_p (lhs
))
507 update_dep_bb (bb
, val
);
514 /* Returns true if E1 and E2 have 2 successors, and if the successor flags
515 are inverse for the EDGE_TRUE_VALUE and EDGE_FALSE_VALUE flags, and equal for
516 the other edge flags. */
519 inverse_flags (const_same_succ e1
, const_same_succ e2
)
521 int f1a
, f1b
, f2a
, f2b
;
522 int mask
= ~(EDGE_TRUE_VALUE
| EDGE_FALSE_VALUE
);
524 if (e1
->succ_flags
.length () != 2)
527 f1a
= e1
->succ_flags
[0];
528 f1b
= e1
->succ_flags
[1];
529 f2a
= e2
->succ_flags
[0];
530 f2b
= e2
->succ_flags
[1];
532 if (f1a
== f2a
&& f1b
== f2b
)
535 return (f1a
& mask
) == (f2a
& mask
) && (f1b
& mask
) == (f2b
& mask
);
538 /* Compares SAME_SUCCs E1 and E2. */
541 same_succ_def::equal (const value_type
*e1
, const compare_type
*e2
)
543 unsigned int i
, first1
, first2
;
544 gimple_stmt_iterator gsi1
, gsi2
;
546 basic_block bb1
, bb2
;
548 if (e1
->hashval
!= e2
->hashval
)
551 if (e1
->succ_flags
.length () != e2
->succ_flags
.length ())
554 if (!bitmap_equal_p (e1
->succs
, e2
->succs
))
557 if (!inverse_flags (e1
, e2
))
559 for (i
= 0; i
< e1
->succ_flags
.length (); ++i
)
560 if (e1
->succ_flags
[i
] != e1
->succ_flags
[i
])
564 first1
= bitmap_first_set_bit (e1
->bbs
);
565 first2
= bitmap_first_set_bit (e2
->bbs
);
567 bb1
= BASIC_BLOCK (first1
);
568 bb2
= BASIC_BLOCK (first2
);
570 if (BB_SIZE (bb1
) != BB_SIZE (bb2
))
573 gsi1
= gsi_start_nondebug_bb (bb1
);
574 gsi2
= gsi_start_nondebug_bb (bb2
);
575 gsi_advance_fw_nondebug_nonlocal (&gsi1
);
576 gsi_advance_fw_nondebug_nonlocal (&gsi2
);
577 while (!(gsi_end_p (gsi1
) || gsi_end_p (gsi2
)))
579 s1
= gsi_stmt (gsi1
);
580 s2
= gsi_stmt (gsi2
);
581 if (gimple_code (s1
) != gimple_code (s2
))
583 if (is_gimple_call (s1
) && !gimple_call_same_target_p (s1
, s2
))
585 gsi_next_nondebug (&gsi1
);
586 gsi_next_nondebug (&gsi2
);
587 gsi_advance_fw_nondebug_nonlocal (&gsi1
);
588 gsi_advance_fw_nondebug_nonlocal (&gsi2
);
594 /* Alloc and init a new SAME_SUCC. */
597 same_succ_alloc (void)
599 same_succ same
= XNEW (struct same_succ_def
);
601 same
->bbs
= BITMAP_ALLOC (NULL
);
602 same
->succs
= BITMAP_ALLOC (NULL
);
603 same
->inverse
= BITMAP_ALLOC (NULL
);
604 same
->succ_flags
.create (10);
605 same
->in_worklist
= false;
610 /* Delete same_succ E. */
613 same_succ_def::remove (same_succ e
)
615 BITMAP_FREE (e
->bbs
);
616 BITMAP_FREE (e
->succs
);
617 BITMAP_FREE (e
->inverse
);
618 e
->succ_flags
.release ();
623 /* Reset same_succ SAME. */
626 same_succ_reset (same_succ same
)
628 bitmap_clear (same
->bbs
);
629 bitmap_clear (same
->succs
);
630 bitmap_clear (same
->inverse
);
631 same
->succ_flags
.truncate (0);
634 static hash_table
<same_succ_def
> same_succ_htab
;
636 /* Array that is used to store the edge flags for a successor. */
638 static int *same_succ_edge_flags
;
640 /* Bitmap that is used to mark bbs that are recently deleted. */
642 static bitmap deleted_bbs
;
644 /* Bitmap that is used to mark predecessors of bbs that are
647 static bitmap deleted_bb_preds
;
649 /* Prints same_succ_htab to stderr. */
651 extern void debug_same_succ (void);
653 debug_same_succ ( void)
655 same_succ_htab
.traverse
<FILE *, ssa_same_succ_print_traverse
> (stderr
);
659 /* Vector of bbs to process. */
661 static vec
<same_succ
> worklist
;
663 /* Prints worklist to FILE. */
666 print_worklist (FILE *file
)
669 for (i
= 0; i
< worklist
.length (); ++i
)
670 same_succ_print (file
, worklist
[i
]);
673 /* Adds SAME to worklist. */
676 add_to_worklist (same_succ same
)
678 if (same
->in_worklist
)
681 if (bitmap_count_bits (same
->bbs
) < 2)
684 same
->in_worklist
= true;
685 worklist
.safe_push (same
);
688 /* Add BB to same_succ_htab. */
691 find_same_succ_bb (basic_block bb
, same_succ
*same_p
)
695 same_succ same
= *same_p
;
701 /* Be conservative with loop structure. It's not evident that this test
702 is sufficient. Before tail-merge, we've just called
703 loop_optimizer_finalize, and LOOPS_MAY_HAVE_MULTIPLE_LATCHES is now
704 set, so there's no guarantee that the loop->latch value is still valid.
705 But we assume that, since we've forced LOOPS_HAVE_SIMPLE_LATCHES at the
706 start of pre, we've kept that property intact throughout pre, and are
707 keeping it throughout tail-merge using this test. */
708 || bb
->loop_father
->latch
== bb
)
710 bitmap_set_bit (same
->bbs
, bb
->index
);
711 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
713 int index
= e
->dest
->index
;
714 bitmap_set_bit (same
->succs
, index
);
715 same_succ_edge_flags
[index
] = e
->flags
;
717 EXECUTE_IF_SET_IN_BITMAP (same
->succs
, 0, j
, bj
)
718 same
->succ_flags
.safe_push (same_succ_edge_flags
[j
]);
720 same
->hashval
= same_succ_hash (same
);
722 slot
= same_succ_htab
.find_slot_with_hash (same
, same
->hashval
, INSERT
);
726 BB_SAME_SUCC (bb
) = same
;
727 add_to_worklist (same
);
732 bitmap_set_bit ((*slot
)->bbs
, bb
->index
);
733 BB_SAME_SUCC (bb
) = *slot
;
734 add_to_worklist (*slot
);
735 if (inverse_flags (same
, *slot
))
736 bitmap_set_bit ((*slot
)->inverse
, bb
->index
);
737 same_succ_reset (same
);
741 /* Find bbs with same successors. */
744 find_same_succ (void)
746 same_succ same
= same_succ_alloc ();
751 find_same_succ_bb (bb
, &same
);
753 same
= same_succ_alloc ();
756 same_succ_def::remove (same
);
759 /* Initializes worklist administration. */
764 alloc_aux_for_blocks (sizeof (struct aux_bb_info
));
765 same_succ_htab
.create (n_basic_blocks
);
766 same_succ_edge_flags
= XCNEWVEC (int, last_basic_block
);
767 deleted_bbs
= BITMAP_ALLOC (NULL
);
768 deleted_bb_preds
= BITMAP_ALLOC (NULL
);
769 worklist
.create (n_basic_blocks
);
772 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
774 fprintf (dump_file
, "initial worklist:\n");
775 print_worklist (dump_file
);
779 /* Deletes worklist administration. */
782 delete_worklist (void)
784 free_aux_for_blocks ();
785 same_succ_htab
.dispose ();
786 XDELETEVEC (same_succ_edge_flags
);
787 same_succ_edge_flags
= NULL
;
788 BITMAP_FREE (deleted_bbs
);
789 BITMAP_FREE (deleted_bb_preds
);
793 /* Mark BB as deleted, and mark its predecessors. */
796 mark_basic_block_deleted (basic_block bb
)
801 bitmap_set_bit (deleted_bbs
, bb
->index
);
803 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
804 bitmap_set_bit (deleted_bb_preds
, e
->src
->index
);
807 /* Removes BB from its corresponding same_succ. */
810 same_succ_flush_bb (basic_block bb
)
812 same_succ same
= BB_SAME_SUCC (bb
);
813 BB_SAME_SUCC (bb
) = NULL
;
814 if (bitmap_single_bit_set_p (same
->bbs
))
815 same_succ_htab
.remove_elt_with_hash (same
, same
->hashval
);
817 bitmap_clear_bit (same
->bbs
, bb
->index
);
820 /* Removes all bbs in BBS from their corresponding same_succ. */
823 same_succ_flush_bbs (bitmap bbs
)
828 EXECUTE_IF_SET_IN_BITMAP (bbs
, 0, i
, bi
)
829 same_succ_flush_bb (BASIC_BLOCK (i
));
832 /* Release the last vdef in BB, either normal or phi result. */
835 release_last_vdef (basic_block bb
)
837 gimple_stmt_iterator i
;
839 for (i
= gsi_last_bb (bb
); !gsi_end_p (i
); gsi_prev_nondebug (&i
))
841 gimple stmt
= gsi_stmt (i
);
842 if (gimple_vdef (stmt
) == NULL_TREE
)
845 mark_virtual_operand_for_renaming (gimple_vdef (stmt
));
849 for (i
= gsi_start_phis (bb
); !gsi_end_p (i
); gsi_next (&i
))
851 gimple phi
= gsi_stmt (i
);
852 tree res
= gimple_phi_result (phi
);
854 if (!virtual_operand_p (res
))
857 mark_virtual_phi_result_for_renaming (phi
);
863 /* For deleted_bb_preds, find bbs with same successors. */
866 update_worklist (void)
873 bitmap_and_compl_into (deleted_bb_preds
, deleted_bbs
);
874 bitmap_clear (deleted_bbs
);
876 bitmap_clear_bit (deleted_bb_preds
, ENTRY_BLOCK
);
877 same_succ_flush_bbs (deleted_bb_preds
);
879 same
= same_succ_alloc ();
880 EXECUTE_IF_SET_IN_BITMAP (deleted_bb_preds
, 0, i
, bi
)
882 bb
= BASIC_BLOCK (i
);
883 gcc_assert (bb
!= NULL
);
884 find_same_succ_bb (bb
, &same
);
886 same
= same_succ_alloc ();
888 same_succ_def::remove (same
);
889 bitmap_clear (deleted_bb_preds
);
892 /* Prints cluster C to FILE. */
895 print_cluster (FILE *file
, bb_cluster c
)
899 bitmap_print (file
, c
->bbs
, "bbs:", "\n");
900 bitmap_print (file
, c
->preds
, "preds:", "\n");
903 /* Prints cluster C to stderr. */
905 extern void debug_cluster (bb_cluster
);
907 debug_cluster (bb_cluster c
)
909 print_cluster (stderr
, c
);
912 /* Update C->rep_bb, given that BB is added to the cluster. */
915 update_rep_bb (bb_cluster c
, basic_block bb
)
918 if (c
->rep_bb
== NULL
)
924 /* Current needs no deps, keep it. */
925 if (BB_DEP_BB (c
->rep_bb
) == NULL
)
928 /* Bb needs no deps, change rep_bb. */
929 if (BB_DEP_BB (bb
) == NULL
)
935 /* Bb needs last deps earlier than current, change rep_bb. A potential
936 problem with this, is that the first deps might also be earlier, which
937 would mean we prefer longer lifetimes for the deps. To be able to check
938 for this, we would have to trace BB_FIRST_DEP_BB as well, besides
939 BB_DEP_BB, which is really BB_LAST_DEP_BB.
940 The benefit of choosing the bb with last deps earlier, is that it can
941 potentially be used as replacement for more bbs. */
942 if (dominated_by_p (CDI_DOMINATORS
, BB_DEP_BB (c
->rep_bb
), BB_DEP_BB (bb
)))
946 /* Add BB to cluster C. Sets BB in C->bbs, and preds of BB in C->preds. */
949 add_bb_to_cluster (bb_cluster c
, basic_block bb
)
954 bitmap_set_bit (c
->bbs
, bb
->index
);
956 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
957 bitmap_set_bit (c
->preds
, e
->src
->index
);
959 update_rep_bb (c
, bb
);
962 /* Allocate and init new cluster. */
968 c
= XCNEW (struct bb_cluster_def
);
969 c
->bbs
= BITMAP_ALLOC (NULL
);
970 c
->preds
= BITMAP_ALLOC (NULL
);
975 /* Delete clusters. */
978 delete_cluster (bb_cluster c
)
982 BITMAP_FREE (c
->bbs
);
983 BITMAP_FREE (c
->preds
);
988 /* Array that contains all clusters. */
990 static vec
<bb_cluster
> all_clusters
;
992 /* Allocate all cluster vectors. */
995 alloc_cluster_vectors (void)
997 all_clusters
.create (n_basic_blocks
);
1000 /* Reset all cluster vectors. */
1003 reset_cluster_vectors (void)
1007 for (i
= 0; i
< all_clusters
.length (); ++i
)
1008 delete_cluster (all_clusters
[i
]);
1009 all_clusters
.truncate (0);
1011 BB_CLUSTER (bb
) = NULL
;
1014 /* Delete all cluster vectors. */
1017 delete_cluster_vectors (void)
1020 for (i
= 0; i
< all_clusters
.length (); ++i
)
1021 delete_cluster (all_clusters
[i
]);
1022 all_clusters
.release ();
1025 /* Merge cluster C2 into C1. */
1028 merge_clusters (bb_cluster c1
, bb_cluster c2
)
1030 bitmap_ior_into (c1
->bbs
, c2
->bbs
);
1031 bitmap_ior_into (c1
->preds
, c2
->preds
);
1034 /* Register equivalence of BB1 and BB2 (members of cluster C). Store c in
1035 all_clusters, or merge c with existing cluster. */
1038 set_cluster (basic_block bb1
, basic_block bb2
)
1040 basic_block merge_bb
, other_bb
;
1041 bb_cluster merge
, old
, c
;
1043 if (BB_CLUSTER (bb1
) == NULL
&& BB_CLUSTER (bb2
) == NULL
)
1046 add_bb_to_cluster (c
, bb1
);
1047 add_bb_to_cluster (c
, bb2
);
1048 BB_CLUSTER (bb1
) = c
;
1049 BB_CLUSTER (bb2
) = c
;
1050 c
->index
= all_clusters
.length ();
1051 all_clusters
.safe_push (c
);
1053 else if (BB_CLUSTER (bb1
) == NULL
|| BB_CLUSTER (bb2
) == NULL
)
1055 merge_bb
= BB_CLUSTER (bb1
) == NULL
? bb2
: bb1
;
1056 other_bb
= BB_CLUSTER (bb1
) == NULL
? bb1
: bb2
;
1057 merge
= BB_CLUSTER (merge_bb
);
1058 add_bb_to_cluster (merge
, other_bb
);
1059 BB_CLUSTER (other_bb
) = merge
;
1061 else if (BB_CLUSTER (bb1
) != BB_CLUSTER (bb2
))
1066 old
= BB_CLUSTER (bb2
);
1067 merge
= BB_CLUSTER (bb1
);
1068 merge_clusters (merge
, old
);
1069 EXECUTE_IF_SET_IN_BITMAP (old
->bbs
, 0, i
, bi
)
1070 BB_CLUSTER (BASIC_BLOCK (i
)) = merge
;
1071 all_clusters
[old
->index
] = NULL
;
1072 update_rep_bb (merge
, old
->rep_bb
);
1073 delete_cluster (old
);
1079 /* Return true if gimple statements S1 and S2 are equal. Gimple_bb (s1) and
1080 gimple_bb (s2) are members of SAME_SUCC. */
1083 gimple_equal_p (same_succ same_succ
, gimple s1
, gimple s2
)
1087 basic_block bb1
= gimple_bb (s1
), bb2
= gimple_bb (s2
);
1090 enum tree_code code1
, code2
;
1092 if (gimple_code (s1
) != gimple_code (s2
))
1095 switch (gimple_code (s1
))
1098 if (gimple_call_num_args (s1
) != gimple_call_num_args (s2
))
1100 if (!gimple_call_same_target_p (s1
, s2
))
1103 /* Eventually, we'll significantly complicate the CFG by adding
1104 back edges to properly model the effects of transaction restart.
1105 For the bulk of optimization this does not matter, but what we
1106 cannot recover from is tail merging blocks between two separate
1107 transactions. Avoid that by making commit not match. */
1108 if (gimple_call_builtin_p (s1
, BUILT_IN_TM_COMMIT
))
1111 for (i
= 0; i
< gimple_call_num_args (s1
); ++i
)
1113 t1
= gimple_call_arg (s1
, i
);
1114 t2
= gimple_call_arg (s2
, i
);
1115 if (operand_equal_p (t1
, t2
, 0))
1117 if (gvn_uses_equal (t1
, t2
))
1122 lhs1
= gimple_get_lhs (s1
);
1123 lhs2
= gimple_get_lhs (s2
);
1124 if (lhs1
== NULL_TREE
&& lhs2
== NULL_TREE
)
1126 if (lhs1
== NULL_TREE
|| lhs2
== NULL_TREE
)
1128 if (TREE_CODE (lhs1
) == SSA_NAME
&& TREE_CODE (lhs2
) == SSA_NAME
)
1129 return vn_valueize (lhs1
) == vn_valueize (lhs2
);
1130 return operand_equal_p (lhs1
, lhs2
, 0);
1133 lhs1
= gimple_get_lhs (s1
);
1134 lhs2
= gimple_get_lhs (s2
);
1135 if (TREE_CODE (lhs1
) != SSA_NAME
1136 && TREE_CODE (lhs2
) != SSA_NAME
)
1137 return (vn_valueize (gimple_vdef (s1
))
1138 == vn_valueize (gimple_vdef (s2
)));
1139 else if (TREE_CODE (lhs1
) == SSA_NAME
1140 && TREE_CODE (lhs2
) == SSA_NAME
)
1141 return vn_valueize (lhs1
) == vn_valueize (lhs2
);
1145 t1
= gimple_cond_lhs (s1
);
1146 t2
= gimple_cond_lhs (s2
);
1147 if (!operand_equal_p (t1
, t2
, 0)
1148 && !gvn_uses_equal (t1
, t2
))
1151 t1
= gimple_cond_rhs (s1
);
1152 t2
= gimple_cond_rhs (s2
);
1153 if (!operand_equal_p (t1
, t2
, 0)
1154 && !gvn_uses_equal (t1
, t2
))
1157 code1
= gimple_expr_code (s1
);
1158 code2
= gimple_expr_code (s2
);
1159 inv_cond
= (bitmap_bit_p (same_succ
->inverse
, bb1
->index
)
1160 != bitmap_bit_p (same_succ
->inverse
, bb2
->index
));
1164 = HONOR_NANS (TYPE_MODE (TREE_TYPE (gimple_cond_lhs (s1
))));
1165 code2
= invert_tree_comparison (code2
, honor_nans
);
1167 return code1
== code2
;
1174 /* Let GSI skip backwards over local defs. Return the earliest vuse in VUSE.
1175 Return true in VUSE_ESCAPED if the vuse influenced a SSA_OP_DEF of one of the
1176 processed statements. */
1179 gsi_advance_bw_nondebug_nonlocal (gimple_stmt_iterator
*gsi
, tree
*vuse
,
1187 if (gsi_end_p (*gsi
))
1189 stmt
= gsi_stmt (*gsi
);
1191 lvuse
= gimple_vuse (stmt
);
1192 if (lvuse
!= NULL_TREE
)
1195 if (!ZERO_SSA_OPERANDS (stmt
, SSA_OP_DEF
))
1196 *vuse_escaped
= true;
1199 if (!stmt_local_def (stmt
))
1201 gsi_prev_nondebug (gsi
);
1205 /* Determines whether BB1 and BB2 (members of same_succ) are duplicates. If so,
1209 find_duplicate (same_succ same_succ
, basic_block bb1
, basic_block bb2
)
1211 gimple_stmt_iterator gsi1
= gsi_last_nondebug_bb (bb1
);
1212 gimple_stmt_iterator gsi2
= gsi_last_nondebug_bb (bb2
);
1213 tree vuse1
= NULL_TREE
, vuse2
= NULL_TREE
;
1214 bool vuse_escaped
= false;
1216 gsi_advance_bw_nondebug_nonlocal (&gsi1
, &vuse1
, &vuse_escaped
);
1217 gsi_advance_bw_nondebug_nonlocal (&gsi2
, &vuse2
, &vuse_escaped
);
1219 while (!gsi_end_p (gsi1
) && !gsi_end_p (gsi2
))
1221 gimple stmt1
= gsi_stmt (gsi1
);
1222 gimple stmt2
= gsi_stmt (gsi2
);
1224 if (!gimple_equal_p (same_succ
, stmt1
, stmt2
))
1227 // We cannot tail-merge the builtins that end transactions.
1228 // ??? The alternative being unsharing of BBs in the tm_init pass.
1230 && is_gimple_call (stmt1
)
1231 && (gimple_call_flags (stmt1
) & ECF_TM_BUILTIN
)
1232 && is_tm_ending_fndecl (gimple_call_fndecl (stmt1
)))
1235 gsi_prev_nondebug (&gsi1
);
1236 gsi_prev_nondebug (&gsi2
);
1237 gsi_advance_bw_nondebug_nonlocal (&gsi1
, &vuse1
, &vuse_escaped
);
1238 gsi_advance_bw_nondebug_nonlocal (&gsi2
, &vuse2
, &vuse_escaped
);
1241 if (!(gsi_end_p (gsi1
) && gsi_end_p (gsi2
)))
1244 /* If the incoming vuses are not the same, and the vuse escaped into an
1245 SSA_OP_DEF, then merging the 2 blocks will change the value of the def,
1246 which potentially means the semantics of one of the blocks will be changed.
1247 TODO: make this check more precise. */
1248 if (vuse_escaped
&& vuse1
!= vuse2
)
1252 fprintf (dump_file
, "find_duplicates: <bb %d> duplicate of <bb %d>\n",
1253 bb1
->index
, bb2
->index
);
1255 set_cluster (bb1
, bb2
);
1258 /* Returns whether for all phis in DEST the phi alternatives for E1 and
1262 same_phi_alternatives_1 (basic_block dest
, edge e1
, edge e2
)
1264 int n1
= e1
->dest_idx
, n2
= e2
->dest_idx
;
1265 gimple_stmt_iterator gsi
;
1267 for (gsi
= gsi_start_phis (dest
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1269 gimple phi
= gsi_stmt (gsi
);
1270 tree lhs
= gimple_phi_result (phi
);
1271 tree val1
= gimple_phi_arg_def (phi
, n1
);
1272 tree val2
= gimple_phi_arg_def (phi
, n2
);
1274 if (virtual_operand_p (lhs
))
1277 if (operand_equal_for_phi_arg_p (val1
, val2
))
1279 if (gvn_uses_equal (val1
, val2
))
1288 /* Returns whether for all successors of BB1 and BB2 (members of SAME_SUCC), the
1289 phi alternatives for BB1 and BB2 are equal. */
1292 same_phi_alternatives (same_succ same_succ
, basic_block bb1
, basic_block bb2
)
1299 EXECUTE_IF_SET_IN_BITMAP (same_succ
->succs
, 0, s
, bs
)
1301 succ
= BASIC_BLOCK (s
);
1302 e1
= find_edge (bb1
, succ
);
1303 e2
= find_edge (bb2
, succ
);
1304 if (e1
->flags
& EDGE_COMPLEX
1305 || e2
->flags
& EDGE_COMPLEX
)
1308 /* For all phis in bb, the phi alternatives for e1 and e2 need to have
1310 if (!same_phi_alternatives_1 (succ
, e1
, e2
))
1317 /* Return true if BB has non-vop phis. */
1320 bb_has_non_vop_phi (basic_block bb
)
1322 gimple_seq phis
= phi_nodes (bb
);
1328 if (!gimple_seq_singleton_p (phis
))
1331 phi
= gimple_seq_first_stmt (phis
);
1332 return !virtual_operand_p (gimple_phi_result (phi
));
1335 /* Returns true if redirecting the incoming edges of FROM to TO maintains the
1336 invariant that uses in FROM are dominates by their defs. */
1339 deps_ok_for_redirect_from_bb_to_bb (basic_block from
, basic_block to
)
1341 basic_block cd
, dep_bb
= BB_DEP_BB (to
);
1344 bitmap from_preds
= BITMAP_ALLOC (NULL
);
1349 FOR_EACH_EDGE (e
, ei
, from
->preds
)
1350 bitmap_set_bit (from_preds
, e
->src
->index
);
1351 cd
= nearest_common_dominator_for_set (CDI_DOMINATORS
, from_preds
);
1352 BITMAP_FREE (from_preds
);
1354 return dominated_by_p (CDI_DOMINATORS
, dep_bb
, cd
);
1357 /* Returns true if replacing BB1 (or its replacement bb) by BB2 (or its
1358 replacement bb) and vice versa maintains the invariant that uses in the
1359 replacement are dominates by their defs. */
1362 deps_ok_for_redirect (basic_block bb1
, basic_block bb2
)
1364 if (BB_CLUSTER (bb1
) != NULL
)
1365 bb1
= BB_CLUSTER (bb1
)->rep_bb
;
1367 if (BB_CLUSTER (bb2
) != NULL
)
1368 bb2
= BB_CLUSTER (bb2
)->rep_bb
;
1370 return (deps_ok_for_redirect_from_bb_to_bb (bb1
, bb2
)
1371 && deps_ok_for_redirect_from_bb_to_bb (bb2
, bb1
));
1374 /* Within SAME_SUCC->bbs, find clusters of bbs which can be merged. */
1377 find_clusters_1 (same_succ same_succ
)
1379 basic_block bb1
, bb2
;
1381 bitmap_iterator bi
, bj
;
1383 int max_comparisons
= PARAM_VALUE (PARAM_MAX_TAIL_MERGE_COMPARISONS
);
1385 EXECUTE_IF_SET_IN_BITMAP (same_succ
->bbs
, 0, i
, bi
)
1387 bb1
= BASIC_BLOCK (i
);
1389 /* TODO: handle blocks with phi-nodes. We'll have to find corresponding
1390 phi-nodes in bb1 and bb2, with the same alternatives for the same
1392 if (bb_has_non_vop_phi (bb1
))
1396 EXECUTE_IF_SET_IN_BITMAP (same_succ
->bbs
, i
+ 1, j
, bj
)
1398 bb2
= BASIC_BLOCK (j
);
1400 if (bb_has_non_vop_phi (bb2
))
1403 if (BB_CLUSTER (bb1
) != NULL
&& BB_CLUSTER (bb1
) == BB_CLUSTER (bb2
))
1406 /* Limit quadratic behaviour. */
1408 if (nr_comparisons
> max_comparisons
)
1411 /* This is a conservative dependency check. We could test more
1412 precise for allowed replacement direction. */
1413 if (!deps_ok_for_redirect (bb1
, bb2
))
1416 if (!(same_phi_alternatives (same_succ
, bb1
, bb2
)))
1419 find_duplicate (same_succ
, bb1
, bb2
);
1424 /* Find clusters of bbs which can be merged. */
1427 find_clusters (void)
1431 while (!worklist
.is_empty ())
1433 same
= worklist
.pop ();
1434 same
->in_worklist
= false;
1435 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1437 fprintf (dump_file
, "processing worklist entry\n");
1438 same_succ_print (dump_file
, same
);
1440 find_clusters_1 (same
);
1444 /* Returns the vop phi of BB, if any. */
1447 vop_phi (basic_block bb
)
1450 gimple_stmt_iterator gsi
;
1451 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1453 stmt
= gsi_stmt (gsi
);
1454 if (! virtual_operand_p (gimple_phi_result (stmt
)))
1461 /* Redirect all edges from BB1 to BB2, removes BB1 and marks it as removed. */
1464 replace_block_by (basic_block bb1
, basic_block bb2
)
1472 bb2_phi
= vop_phi (bb2
);
1474 /* Mark the basic block as deleted. */
1475 mark_basic_block_deleted (bb1
);
1477 /* Redirect the incoming edges of bb1 to bb2. */
1478 for (i
= EDGE_COUNT (bb1
->preds
); i
> 0 ; --i
)
1480 pred_edge
= EDGE_PRED (bb1
, i
- 1);
1481 pred_edge
= redirect_edge_and_branch (pred_edge
, bb2
);
1482 gcc_assert (pred_edge
!= NULL
);
1484 if (bb2_phi
== NULL
)
1487 /* The phi might have run out of capacity when the redirect added an
1488 argument, which means it could have been replaced. Refresh it. */
1489 bb2_phi
= vop_phi (bb2
);
1491 add_phi_arg (bb2_phi
, SSA_NAME_VAR (gimple_phi_result (bb2_phi
)),
1492 pred_edge
, UNKNOWN_LOCATION
);
1495 bb2
->frequency
+= bb1
->frequency
;
1496 if (bb2
->frequency
> BB_FREQ_MAX
)
1497 bb2
->frequency
= BB_FREQ_MAX
;
1499 bb2
->count
+= bb1
->count
;
1501 /* Merge the outgoing edge counts from bb1 onto bb2. */
1502 gcov_type out_sum
= 0;
1503 FOR_EACH_EDGE (e1
, ei
, bb1
->succs
)
1505 e2
= find_edge (bb2
, e1
->dest
);
1507 e2
->count
+= e1
->count
;
1508 out_sum
+= e2
->count
;
1510 /* Recompute the edge probabilities from the new merged edge count.
1511 Use the sum of the new merged edge counts computed above instead
1512 of bb2's merged count, in case there are profile count insanities
1513 making the bb count inconsistent with the edge weights. */
1514 FOR_EACH_EDGE (e2
, ei
, bb2
->succs
)
1516 e2
->probability
= GCOV_COMPUTE_SCALE (e2
->count
, out_sum
);
1519 /* Do updates that use bb1, before deleting bb1. */
1520 release_last_vdef (bb1
);
1521 same_succ_flush_bb (bb1
);
1523 delete_basic_block (bb1
);
1526 /* Bbs for which update_debug_stmt need to be called. */
1528 static bitmap update_bbs
;
1530 /* For each cluster in all_clusters, merge all cluster->bbs. Returns
1531 number of bbs removed. */
1534 apply_clusters (void)
1536 basic_block bb1
, bb2
;
1540 int nr_bbs_removed
= 0;
1542 for (i
= 0; i
< all_clusters
.length (); ++i
)
1544 c
= all_clusters
[i
];
1549 bitmap_set_bit (update_bbs
, bb2
->index
);
1551 bitmap_clear_bit (c
->bbs
, bb2
->index
);
1552 EXECUTE_IF_SET_IN_BITMAP (c
->bbs
, 0, j
, bj
)
1554 bb1
= BASIC_BLOCK (j
);
1555 bitmap_clear_bit (update_bbs
, bb1
->index
);
1557 replace_block_by (bb1
, bb2
);
1562 return nr_bbs_removed
;
1565 /* Resets debug statement STMT if it has uses that are not dominated by their
1569 update_debug_stmt (gimple stmt
)
1571 use_operand_p use_p
;
1573 basic_block bbdef
, bbuse
;
1577 if (!gimple_debug_bind_p (stmt
))
1580 bbuse
= gimple_bb (stmt
);
1581 FOR_EACH_PHI_OR_STMT_USE (use_p
, stmt
, oi
, SSA_OP_USE
)
1583 name
= USE_FROM_PTR (use_p
);
1584 gcc_assert (TREE_CODE (name
) == SSA_NAME
);
1586 def_stmt
= SSA_NAME_DEF_STMT (name
);
1587 gcc_assert (def_stmt
!= NULL
);
1589 bbdef
= gimple_bb (def_stmt
);
1590 if (bbdef
== NULL
|| bbuse
== bbdef
1591 || dominated_by_p (CDI_DOMINATORS
, bbuse
, bbdef
))
1594 gimple_debug_bind_reset_value (stmt
);
1599 /* Resets all debug statements that have uses that are not
1600 dominated by their defs. */
1603 update_debug_stmts (void)
1609 EXECUTE_IF_SET_IN_BITMAP (update_bbs
, 0, i
, bi
)
1612 gimple_stmt_iterator gsi
;
1614 bb
= BASIC_BLOCK (i
);
1615 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1617 stmt
= gsi_stmt (gsi
);
1618 if (!is_gimple_debug (stmt
))
1620 update_debug_stmt (stmt
);
1625 /* Runs tail merge optimization. */
1628 tail_merge_optimize (unsigned int todo
)
1630 int nr_bbs_removed_total
= 0;
1632 bool loop_entered
= false;
1633 int iteration_nr
= 0;
1634 int max_iterations
= PARAM_VALUE (PARAM_MAX_TAIL_MERGE_ITERATIONS
);
1636 if (!flag_tree_tail_merge
1637 || max_iterations
== 0
1638 /* We try to be conservative with respect to loop structure, since:
1639 - the cases where tail-merging could both affect loop structure and be
1640 beneficial are rare,
1641 - it prevents us from having to fixup the loops using
1642 loops_state_set (LOOPS_NEED_FIXUP), and
1643 - keeping loop structure may allow us to simplify the pass.
1644 In order to be conservative, we need loop information. In rare cases
1645 (about 7 test-cases in the g++ testsuite) there is none (because
1646 loop_optimizer_finalize has been called before tail-merge, and
1647 PROP_loops is not set), so we bail out. */
1648 || current_loops
== NULL
)
1651 timevar_push (TV_TREE_TAIL_MERGE
);
1653 if (!dom_info_available_p (CDI_DOMINATORS
))
1655 /* PRE can leave us with unreachable blocks, remove them now. */
1656 delete_unreachable_blocks ();
1657 calculate_dominance_info (CDI_DOMINATORS
);
1661 while (!worklist
.is_empty ())
1665 loop_entered
= true;
1666 alloc_cluster_vectors ();
1667 update_bbs
= BITMAP_ALLOC (NULL
);
1670 reset_cluster_vectors ();
1673 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1674 fprintf (dump_file
, "worklist iteration #%d\n", iteration_nr
);
1677 gcc_assert (worklist
.is_empty ());
1678 if (all_clusters
.is_empty ())
1681 nr_bbs_removed
= apply_clusters ();
1682 nr_bbs_removed_total
+= nr_bbs_removed
;
1683 if (nr_bbs_removed
== 0)
1686 free_dominance_info (CDI_DOMINATORS
);
1688 if (iteration_nr
== max_iterations
)
1691 calculate_dominance_info (CDI_DOMINATORS
);
1695 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1696 fprintf (dump_file
, "htab collision / search: %f\n",
1697 same_succ_htab
.collisions ());
1699 if (nr_bbs_removed_total
> 0)
1701 if (MAY_HAVE_DEBUG_STMTS
)
1703 calculate_dominance_info (CDI_DOMINATORS
);
1704 update_debug_stmts ();
1707 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1709 fprintf (dump_file
, "Before TODOs.\n");
1710 dump_function_to_file (current_function_decl
, dump_file
, dump_flags
);
1713 todo
|= (TODO_verify_ssa
| TODO_verify_stmts
| TODO_verify_flow
);
1714 mark_virtual_operands_for_renaming (cfun
);
1720 delete_cluster_vectors ();
1721 BITMAP_FREE (update_bbs
);
1724 timevar_pop (TV_TREE_TAIL_MERGE
);