1 /* Coalesce SSA_NAMES together for the out-of-ssa pass.
2 Copyright (C) 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011
3 Free Software Foundation, Inc.
4 Contributed by Andrew MacLeod <amacleod@redhat.com>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3, or (at your option)
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
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/>. */
24 #include "coretypes.h"
28 #include "tree-pretty-print.h"
30 #include "tree-flow.h"
32 #include "tree-dump.h"
33 #include "tree-ssa-live.h"
34 #include "diagnostic-core.h"
37 /* This set of routines implements a coalesce_list. This is an object which
38 is used to track pairs of ssa_names which are desirable to coalesce
39 together to avoid copies. Costs are associated with each pair, and when
40 all desired information has been collected, the object can be used to
41 order the pairs for processing. */
43 /* This structure defines a pair entry. */
45 typedef struct coalesce_pair
51 typedef const struct coalesce_pair
*const_coalesce_pair_p
;
53 typedef struct cost_one_pair_d
57 struct cost_one_pair_d
*next
;
60 /* This structure maintains the list of coalesce pairs. */
62 typedef struct coalesce_list_d
64 htab_t list
; /* Hash table. */
65 coalesce_pair_p
*sorted
; /* List when sorted. */
66 int num_sorted
; /* Number in the sorted list. */
67 cost_one_pair_p cost_one_list
;/* Single use coalesces with cost 1. */
70 #define NO_BEST_COALESCE -1
71 #define MUST_COALESCE_COST INT_MAX
74 /* Return cost of execution of copy instruction with FREQUENCY. */
77 coalesce_cost (int frequency
, bool optimize_for_size
)
79 /* Base costs on BB frequencies bounded by 1. */
85 if (optimize_for_size
)
92 /* Return the cost of executing a copy instruction in basic block BB. */
95 coalesce_cost_bb (basic_block bb
)
97 return coalesce_cost (bb
->frequency
, optimize_bb_for_size_p (bb
));
101 /* Return the cost of executing a copy instruction on edge E. */
104 coalesce_cost_edge (edge e
)
108 /* Inserting copy on critical edge costs more than inserting it elsewhere. */
109 if (EDGE_CRITICAL_P (e
))
111 if (e
->flags
& EDGE_ABNORMAL
)
112 return MUST_COALESCE_COST
;
113 if (e
->flags
& EDGE_EH
)
117 FOR_EACH_EDGE (e2
, ei
, e
->dest
->preds
)
120 /* Putting code on EH edge that leads to BB
121 with multiple predecestors imply splitting of
125 /* If there are multiple EH predecestors, we
126 also copy EH regions and produce separate
127 landing pad. This is expensive. */
128 if (e2
->flags
& EDGE_EH
)
136 return coalesce_cost (EDGE_FREQUENCY (e
),
137 optimize_edge_for_size_p (e
)) * mult
;
141 /* Retrieve a pair to coalesce from the cost_one_list in CL. Returns the
142 2 elements via P1 and P2. 1 is returned by the function if there is a pair,
143 NO_BEST_COALESCE is returned if there aren't any. */
146 pop_cost_one_pair (coalesce_list_p cl
, int *p1
, int *p2
)
150 ptr
= cl
->cost_one_list
;
152 return NO_BEST_COALESCE
;
154 *p1
= ptr
->first_element
;
155 *p2
= ptr
->second_element
;
156 cl
->cost_one_list
= ptr
->next
;
163 /* Retrieve the most expensive remaining pair to coalesce from CL. Returns the
164 2 elements via P1 and P2. Their calculated cost is returned by the function.
165 NO_BEST_COALESCE is returned if the coalesce list is empty. */
168 pop_best_coalesce (coalesce_list_p cl
, int *p1
, int *p2
)
170 coalesce_pair_p node
;
173 if (cl
->sorted
== NULL
)
174 return pop_cost_one_pair (cl
, p1
, p2
);
176 if (cl
->num_sorted
== 0)
177 return pop_cost_one_pair (cl
, p1
, p2
);
179 node
= cl
->sorted
[--(cl
->num_sorted
)];
180 *p1
= node
->first_element
;
181 *p2
= node
->second_element
;
189 #define COALESCE_HASH_FN(R1, R2) ((R2) * ((R2) - 1) / 2 + (R1))
191 /* Hash function for coalesce list. Calculate hash for PAIR. */
194 coalesce_pair_map_hash (const void *pair
)
196 hashval_t a
= (hashval_t
)(((const_coalesce_pair_p
)pair
)->first_element
);
197 hashval_t b
= (hashval_t
)(((const_coalesce_pair_p
)pair
)->second_element
);
199 return COALESCE_HASH_FN (a
,b
);
203 /* Equality function for coalesce list hash table. Compare PAIR1 and PAIR2,
204 returning TRUE if the two pairs are equivalent. */
207 coalesce_pair_map_eq (const void *pair1
, const void *pair2
)
209 const_coalesce_pair_p
const p1
= (const_coalesce_pair_p
) pair1
;
210 const_coalesce_pair_p
const p2
= (const_coalesce_pair_p
) pair2
;
212 return (p1
->first_element
== p2
->first_element
213 && p1
->second_element
== p2
->second_element
);
217 /* Create a new empty coalesce list object and return it. */
219 static inline coalesce_list_p
220 create_coalesce_list (void)
222 coalesce_list_p list
;
223 unsigned size
= num_ssa_names
* 3;
228 list
= (coalesce_list_p
) xmalloc (sizeof (struct coalesce_list_d
));
229 list
->list
= htab_create (size
, coalesce_pair_map_hash
,
230 coalesce_pair_map_eq
, NULL
);
232 list
->num_sorted
= 0;
233 list
->cost_one_list
= NULL
;
238 /* Delete coalesce list CL. */
241 delete_coalesce_list (coalesce_list_p cl
)
243 gcc_assert (cl
->cost_one_list
== NULL
);
244 htab_delete (cl
->list
);
246 gcc_assert (cl
->num_sorted
== 0);
251 /* Find a matching coalesce pair object in CL for the pair P1 and P2. If
252 one isn't found, return NULL if CREATE is false, otherwise create a new
253 coalesce pair object and return it. */
255 static coalesce_pair_p
256 find_coalesce_pair (coalesce_list_p cl
, int p1
, int p2
, bool create
)
258 struct coalesce_pair p
;
262 /* Normalize so that p1 is the smaller value. */
265 p
.first_element
= p2
;
266 p
.second_element
= p1
;
270 p
.first_element
= p1
;
271 p
.second_element
= p2
;
274 hash
= coalesce_pair_map_hash (&p
);
275 slot
= htab_find_slot_with_hash (cl
->list
, &p
, hash
,
276 create
? INSERT
: NO_INSERT
);
282 struct coalesce_pair
* pair
= XNEW (struct coalesce_pair
);
283 gcc_assert (cl
->sorted
== NULL
);
284 pair
->first_element
= p
.first_element
;
285 pair
->second_element
= p
.second_element
;
287 *slot
= (void *)pair
;
290 return (struct coalesce_pair
*) *slot
;
294 add_cost_one_coalesce (coalesce_list_p cl
, int p1
, int p2
)
296 cost_one_pair_p pair
;
298 pair
= XNEW (struct cost_one_pair_d
);
299 pair
->first_element
= p1
;
300 pair
->second_element
= p2
;
301 pair
->next
= cl
->cost_one_list
;
302 cl
->cost_one_list
= pair
;
306 /* Add a coalesce between P1 and P2 in list CL with a cost of VALUE. */
309 add_coalesce (coalesce_list_p cl
, int p1
, int p2
, int value
)
311 coalesce_pair_p node
;
313 gcc_assert (cl
->sorted
== NULL
);
317 node
= find_coalesce_pair (cl
, p1
, p2
, true);
319 /* Once the value is at least MUST_COALESCE_COST - 1, leave it that way. */
320 if (node
->cost
< MUST_COALESCE_COST
- 1)
322 if (value
< MUST_COALESCE_COST
- 1)
330 /* Comparison function to allow qsort to sort P1 and P2 in Ascending order. */
333 compare_pairs (const void *p1
, const void *p2
)
335 const_coalesce_pair_p
const *const pp1
= (const_coalesce_pair_p
const *) p1
;
336 const_coalesce_pair_p
const *const pp2
= (const_coalesce_pair_p
const *) p2
;
339 result
= (* pp1
)->cost
- (* pp2
)->cost
;
340 /* Since qsort does not guarantee stability we use the elements
341 as a secondary key. This provides us with independence from
342 the host's implementation of the sorting algorithm. */
345 result
= (* pp2
)->first_element
- (* pp1
)->first_element
;
347 result
= (* pp2
)->second_element
- (* pp1
)->second_element
;
354 /* Return the number of unique coalesce pairs in CL. */
357 num_coalesce_pairs (coalesce_list_p cl
)
359 return htab_elements (cl
->list
);
363 /* Iterator over hash table pairs. */
367 } coalesce_pair_iterator
;
370 /* Return first partition pair from list CL, initializing iterator ITER. */
372 static inline coalesce_pair_p
373 first_coalesce_pair (coalesce_list_p cl
, coalesce_pair_iterator
*iter
)
375 coalesce_pair_p pair
;
377 pair
= (coalesce_pair_p
) first_htab_element (&(iter
->hti
), cl
->list
);
382 /* Return TRUE if there are no more partitions in for ITER to process. */
385 end_coalesce_pair_p (coalesce_pair_iterator
*iter
)
387 return end_htab_p (&(iter
->hti
));
391 /* Return the next partition pair to be visited by ITER. */
393 static inline coalesce_pair_p
394 next_coalesce_pair (coalesce_pair_iterator
*iter
)
396 coalesce_pair_p pair
;
398 pair
= (coalesce_pair_p
) next_htab_element (&(iter
->hti
));
403 /* Iterate over CL using ITER, returning values in PAIR. */
405 #define FOR_EACH_PARTITION_PAIR(PAIR, ITER, CL) \
406 for ((PAIR) = first_coalesce_pair ((CL), &(ITER)); \
407 !end_coalesce_pair_p (&(ITER)); \
408 (PAIR) = next_coalesce_pair (&(ITER)))
411 /* Prepare CL for removal of preferred pairs. When finished they are sorted
412 in order from most important coalesce to least important. */
415 sort_coalesce_list (coalesce_list_p cl
)
419 coalesce_pair_iterator ppi
;
421 gcc_assert (cl
->sorted
== NULL
);
423 num
= num_coalesce_pairs (cl
);
424 cl
->num_sorted
= num
;
428 /* Allocate a vector for the pair pointers. */
429 cl
->sorted
= XNEWVEC (coalesce_pair_p
, num
);
431 /* Populate the vector with pointers to the pairs. */
433 FOR_EACH_PARTITION_PAIR (p
, ppi
, cl
)
435 gcc_assert (x
== num
);
437 /* Already sorted. */
441 /* If there are only 2, just pick swap them if the order isn't correct. */
444 if (cl
->sorted
[0]->cost
> cl
->sorted
[1]->cost
)
447 cl
->sorted
[0] = cl
->sorted
[1];
453 /* Only call qsort if there are more than 2 items. */
455 qsort (cl
->sorted
, num
, sizeof (coalesce_pair_p
), compare_pairs
);
459 /* Send debug info for coalesce list CL to file F. */
462 dump_coalesce_list (FILE *f
, coalesce_list_p cl
)
464 coalesce_pair_p node
;
465 coalesce_pair_iterator ppi
;
469 if (cl
->sorted
== NULL
)
471 fprintf (f
, "Coalesce List:\n");
472 FOR_EACH_PARTITION_PAIR (node
, ppi
, cl
)
474 tree var1
= ssa_name (node
->first_element
);
475 tree var2
= ssa_name (node
->second_element
);
476 print_generic_expr (f
, var1
, TDF_SLIM
);
477 fprintf (f
, " <-> ");
478 print_generic_expr (f
, var2
, TDF_SLIM
);
479 fprintf (f
, " (%1d), ", node
->cost
);
485 fprintf (f
, "Sorted Coalesce list:\n");
486 for (x
= cl
->num_sorted
- 1 ; x
>=0; x
--)
488 node
= cl
->sorted
[x
];
489 fprintf (f
, "(%d) ", node
->cost
);
490 var
= ssa_name (node
->first_element
);
491 print_generic_expr (f
, var
, TDF_SLIM
);
492 fprintf (f
, " <-> ");
493 var
= ssa_name (node
->second_element
);
494 print_generic_expr (f
, var
, TDF_SLIM
);
501 /* This represents a conflict graph. Implemented as an array of bitmaps.
502 A full matrix is used for conflicts rather than just upper triangular form.
503 this make sit much simpler and faster to perform conflict merges. */
505 typedef struct ssa_conflicts_d
512 /* Return an empty new conflict graph for SIZE elements. */
514 static inline ssa_conflicts_p
515 ssa_conflicts_new (unsigned size
)
519 ptr
= XNEW (struct ssa_conflicts_d
);
520 ptr
->conflicts
= XCNEWVEC (bitmap
, size
);
526 /* Free storage for conflict graph PTR. */
529 ssa_conflicts_delete (ssa_conflicts_p ptr
)
532 for (x
= 0; x
< ptr
->size
; x
++)
533 if (ptr
->conflicts
[x
])
534 BITMAP_FREE (ptr
->conflicts
[x
]);
536 free (ptr
->conflicts
);
541 /* Test if elements X and Y conflict in graph PTR. */
544 ssa_conflicts_test_p (ssa_conflicts_p ptr
, unsigned x
, unsigned y
)
548 gcc_checking_assert (x
< ptr
->size
);
549 gcc_checking_assert (y
< ptr
->size
);
550 gcc_checking_assert (x
!= y
);
552 b
= ptr
->conflicts
[x
];
554 /* Avoid the lookup if Y has no conflicts. */
555 return ptr
->conflicts
[y
] ? bitmap_bit_p (b
, y
) : false;
561 /* Add a conflict with Y to the bitmap for X in graph PTR. */
564 ssa_conflicts_add_one (ssa_conflicts_p ptr
, unsigned x
, unsigned y
)
566 /* If there are no conflicts yet, allocate the bitmap and set bit. */
567 if (!ptr
->conflicts
[x
])
568 ptr
->conflicts
[x
] = BITMAP_ALLOC (NULL
);
569 bitmap_set_bit (ptr
->conflicts
[x
], y
);
573 /* Add conflicts between X and Y in graph PTR. */
576 ssa_conflicts_add (ssa_conflicts_p ptr
, unsigned x
, unsigned y
)
578 gcc_checking_assert (x
< ptr
->size
);
579 gcc_checking_assert (y
< ptr
->size
);
580 gcc_checking_assert (x
!= y
);
581 ssa_conflicts_add_one (ptr
, x
, y
);
582 ssa_conflicts_add_one (ptr
, y
, x
);
586 /* Merge all Y's conflict into X in graph PTR. */
589 ssa_conflicts_merge (ssa_conflicts_p ptr
, unsigned x
, unsigned y
)
595 if (!(ptr
->conflicts
[y
]))
598 /* Add a conflict between X and every one Y has. If the bitmap doesn't
599 exist, then it has already been coalesced, and we don't need to add a
601 EXECUTE_IF_SET_IN_BITMAP (ptr
->conflicts
[y
], 0, z
, bi
)
602 if (ptr
->conflicts
[z
])
603 bitmap_set_bit (ptr
->conflicts
[z
], x
);
605 if (ptr
->conflicts
[x
])
607 /* If X has conflicts, add Y's to X. */
608 bitmap_ior_into (ptr
->conflicts
[x
], ptr
->conflicts
[y
]);
609 BITMAP_FREE (ptr
->conflicts
[y
]);
613 /* If X has no conflicts, simply use Y's. */
614 ptr
->conflicts
[x
] = ptr
->conflicts
[y
];
615 ptr
->conflicts
[y
] = NULL
;
620 /* Dump a conflicts graph. */
623 ssa_conflicts_dump (FILE *file
, ssa_conflicts_p ptr
)
627 fprintf (file
, "\nConflict graph:\n");
629 for (x
= 0; x
< ptr
->size
; x
++)
630 if (ptr
->conflicts
[x
])
632 fprintf (dump_file
, "%d: ", x
);
633 dump_bitmap (file
, ptr
->conflicts
[x
]);
638 /* This structure is used to efficiently record the current status of live
639 SSA_NAMES when building a conflict graph.
640 LIVE_BASE_VAR has a bit set for each base variable which has at least one
642 LIVE_BASE_PARTITIONS is an array of bitmaps using the basevar table as an
643 index, and is used to track what partitions of each base variable are
644 live. This makes it easy to add conflicts between just live partitions
645 with the same base variable.
646 The values in LIVE_BASE_PARTITIONS are only valid if the base variable is
647 marked as being live. This delays clearing of these bitmaps until
648 they are actually needed again. */
650 typedef struct live_track_d
652 bitmap live_base_var
; /* Indicates if a basevar is live. */
653 bitmap
*live_base_partitions
; /* Live partitions for each basevar. */
654 var_map map
; /* Var_map being used for partition mapping. */
658 /* This routine will create a new live track structure based on the partitions
662 new_live_track (var_map map
)
667 /* Make sure there is a partition view in place. */
668 gcc_assert (map
->partition_to_base_index
!= NULL
);
670 ptr
= (live_track_p
) xmalloc (sizeof (struct live_track_d
));
672 lim
= num_basevars (map
);
673 ptr
->live_base_partitions
= (bitmap
*) xmalloc(sizeof (bitmap
*) * lim
);
674 ptr
->live_base_var
= BITMAP_ALLOC (NULL
);
675 for (x
= 0; x
< lim
; x
++)
676 ptr
->live_base_partitions
[x
] = BITMAP_ALLOC (NULL
);
681 /* This routine will free the memory associated with PTR. */
684 delete_live_track (live_track_p ptr
)
688 lim
= num_basevars (ptr
->map
);
689 for (x
= 0; x
< lim
; x
++)
690 BITMAP_FREE (ptr
->live_base_partitions
[x
]);
691 BITMAP_FREE (ptr
->live_base_var
);
692 free (ptr
->live_base_partitions
);
697 /* This function will remove PARTITION from the live list in PTR. */
700 live_track_remove_partition (live_track_p ptr
, int partition
)
704 root
= basevar_index (ptr
->map
, partition
);
705 bitmap_clear_bit (ptr
->live_base_partitions
[root
], partition
);
706 /* If the element list is empty, make the base variable not live either. */
707 if (bitmap_empty_p (ptr
->live_base_partitions
[root
]))
708 bitmap_clear_bit (ptr
->live_base_var
, root
);
712 /* This function will adds PARTITION to the live list in PTR. */
715 live_track_add_partition (live_track_p ptr
, int partition
)
719 root
= basevar_index (ptr
->map
, partition
);
720 /* If this base var wasn't live before, it is now. Clear the element list
721 since it was delayed until needed. */
722 if (bitmap_set_bit (ptr
->live_base_var
, root
))
723 bitmap_clear (ptr
->live_base_partitions
[root
]);
724 bitmap_set_bit (ptr
->live_base_partitions
[root
], partition
);
729 /* Clear the live bit for VAR in PTR. */
732 live_track_clear_var (live_track_p ptr
, tree var
)
736 p
= var_to_partition (ptr
->map
, var
);
737 if (p
!= NO_PARTITION
)
738 live_track_remove_partition (ptr
, p
);
742 /* Return TRUE if VAR is live in PTR. */
745 live_track_live_p (live_track_p ptr
, tree var
)
749 p
= var_to_partition (ptr
->map
, var
);
750 if (p
!= NO_PARTITION
)
752 root
= basevar_index (ptr
->map
, p
);
753 if (bitmap_bit_p (ptr
->live_base_var
, root
))
754 return bitmap_bit_p (ptr
->live_base_partitions
[root
], p
);
760 /* This routine will add USE to PTR. USE will be marked as live in both the
761 ssa live map and the live bitmap for the root of USE. */
764 live_track_process_use (live_track_p ptr
, tree use
)
768 p
= var_to_partition (ptr
->map
, use
);
769 if (p
== NO_PARTITION
)
772 /* Mark as live in the appropriate live list. */
773 live_track_add_partition (ptr
, p
);
777 /* This routine will process a DEF in PTR. DEF will be removed from the live
778 lists, and if there are any other live partitions with the same base
779 variable, conflicts will be added to GRAPH. */
782 live_track_process_def (live_track_p ptr
, tree def
, ssa_conflicts_p graph
)
789 p
= var_to_partition (ptr
->map
, def
);
790 if (p
== NO_PARTITION
)
793 /* Clear the liveness bit. */
794 live_track_remove_partition (ptr
, p
);
796 /* If the bitmap isn't empty now, conflicts need to be added. */
797 root
= basevar_index (ptr
->map
, p
);
798 if (bitmap_bit_p (ptr
->live_base_var
, root
))
800 b
= ptr
->live_base_partitions
[root
];
801 EXECUTE_IF_SET_IN_BITMAP (b
, 0, x
, bi
)
802 ssa_conflicts_add (graph
, p
, x
);
807 /* Initialize PTR with the partitions set in INIT. */
810 live_track_init (live_track_p ptr
, bitmap init
)
815 /* Mark all live on exit partitions. */
816 EXECUTE_IF_SET_IN_BITMAP (init
, 0, p
, bi
)
817 live_track_add_partition (ptr
, p
);
821 /* This routine will clear all live partitions in PTR. */
824 live_track_clear_base_vars (live_track_p ptr
)
826 /* Simply clear the live base list. Anything marked as live in the element
827 lists will be cleared later if/when the base variable ever comes alive
829 bitmap_clear (ptr
->live_base_var
);
833 /* Build a conflict graph based on LIVEINFO. Any partitions which are in the
834 partition view of the var_map liveinfo is based on get entries in the
835 conflict graph. Only conflicts between ssa_name partitions with the same
836 base variable are added. */
838 static ssa_conflicts_p
839 build_ssa_conflict_graph (tree_live_info_p liveinfo
)
841 ssa_conflicts_p graph
;
847 map
= live_var_map (liveinfo
);
848 graph
= ssa_conflicts_new (num_var_partitions (map
));
850 live
= new_live_track (map
);
854 gimple_stmt_iterator gsi
;
856 /* Start with live on exit temporaries. */
857 live_track_init (live
, live_on_exit (liveinfo
, bb
));
859 for (gsi
= gsi_last_bb (bb
); !gsi_end_p (gsi
); gsi_prev (&gsi
))
862 gimple stmt
= gsi_stmt (gsi
);
864 /* A copy between 2 partitions does not introduce an interference
865 by itself. If they did, you would never be able to coalesce
866 two things which are copied. If the two variables really do
867 conflict, they will conflict elsewhere in the program.
869 This is handled by simply removing the SRC of the copy from the
870 live list, and processing the stmt normally. */
871 if (is_gimple_assign (stmt
))
873 tree lhs
= gimple_assign_lhs (stmt
);
874 tree rhs1
= gimple_assign_rhs1 (stmt
);
875 if (gimple_assign_copy_p (stmt
)
876 && TREE_CODE (lhs
) == SSA_NAME
877 && TREE_CODE (rhs1
) == SSA_NAME
)
878 live_track_clear_var (live
, rhs1
);
880 else if (is_gimple_debug (stmt
))
883 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_DEF
)
884 live_track_process_def (live
, var
, graph
);
886 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
887 live_track_process_use (live
, var
);
890 /* If result of a PHI is unused, looping over the statements will not
891 record any conflicts since the def was never live. Since the PHI node
892 is going to be translated out of SSA form, it will insert a copy.
893 There must be a conflict recorded between the result of the PHI and
894 any variables that are live. Otherwise the out-of-ssa translation
895 may create incorrect code. */
896 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
898 gimple phi
= gsi_stmt (gsi
);
899 tree result
= PHI_RESULT (phi
);
900 if (live_track_live_p (live
, result
))
901 live_track_process_def (live
, result
, graph
);
904 live_track_clear_base_vars (live
);
907 delete_live_track (live
);
912 /* Shortcut routine to print messages to file F of the form:
913 "STR1 EXPR1 STR2 EXPR2 STR3." */
916 print_exprs (FILE *f
, const char *str1
, tree expr1
, const char *str2
,
917 tree expr2
, const char *str3
)
919 fprintf (f
, "%s", str1
);
920 print_generic_expr (f
, expr1
, TDF_SLIM
);
921 fprintf (f
, "%s", str2
);
922 print_generic_expr (f
, expr2
, TDF_SLIM
);
923 fprintf (f
, "%s", str3
);
927 /* Called if a coalesce across and abnormal edge cannot be performed. PHI is
928 the phi node at fault, I is the argument index at fault. A message is
929 printed and compilation is then terminated. */
932 abnormal_corrupt (gimple phi
, int i
)
934 edge e
= gimple_phi_arg_edge (phi
, i
);
935 tree res
= gimple_phi_result (phi
);
936 tree arg
= gimple_phi_arg_def (phi
, i
);
938 fprintf (stderr
, " Corrupt SSA across abnormal edge BB%d->BB%d\n",
939 e
->src
->index
, e
->dest
->index
);
940 fprintf (stderr
, "Argument %d (", i
);
941 print_generic_expr (stderr
, arg
, TDF_SLIM
);
942 if (TREE_CODE (arg
) != SSA_NAME
)
943 fprintf (stderr
, ") is not an SSA_NAME.\n");
946 gcc_assert (SSA_NAME_VAR (res
) != SSA_NAME_VAR (arg
));
947 fprintf (stderr
, ") does not have the same base variable as the result ");
948 print_generic_stmt (stderr
, res
, TDF_SLIM
);
951 internal_error ("SSA corruption");
955 /* Print a failure to coalesce a MUST_COALESCE pair X and Y. */
958 fail_abnormal_edge_coalesce (int x
, int y
)
960 fprintf (stderr
, "\nUnable to coalesce ssa_names %d and %d",x
, y
);
961 fprintf (stderr
, " which are marked as MUST COALESCE.\n");
962 print_generic_expr (stderr
, ssa_name (x
), TDF_SLIM
);
963 fprintf (stderr
, " and ");
964 print_generic_stmt (stderr
, ssa_name (y
), TDF_SLIM
);
966 internal_error ("SSA corruption");
970 /* This function creates a var_map for the current function as well as creating
971 a coalesce list for use later in the out of ssa process. */
974 create_outofssa_var_map (coalesce_list_p cl
, bitmap used_in_copy
)
976 gimple_stmt_iterator gsi
;
986 #ifdef ENABLE_CHECKING
987 bitmap used_in_real_ops
;
988 bitmap used_in_virtual_ops
;
990 used_in_real_ops
= BITMAP_ALLOC (NULL
);
991 used_in_virtual_ops
= BITMAP_ALLOC (NULL
);
994 map
= init_var_map (num_ssa_names
);
1000 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1002 gimple phi
= gsi_stmt (gsi
);
1006 bool saw_copy
= false;
1008 res
= gimple_phi_result (phi
);
1009 ver
= SSA_NAME_VERSION (res
);
1010 register_ssa_partition (map
, res
);
1012 /* Register ssa_names and coalesces between the args and the result
1014 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
1016 edge e
= gimple_phi_arg_edge (phi
, i
);
1017 arg
= PHI_ARG_DEF (phi
, i
);
1018 if (TREE_CODE (arg
) == SSA_NAME
)
1019 register_ssa_partition (map
, arg
);
1020 if (TREE_CODE (arg
) == SSA_NAME
1021 && SSA_NAME_VAR (arg
) == SSA_NAME_VAR (res
))
1024 bitmap_set_bit (used_in_copy
, SSA_NAME_VERSION (arg
));
1025 if ((e
->flags
& EDGE_ABNORMAL
) == 0)
1027 int cost
= coalesce_cost_edge (e
);
1028 if (cost
== 1 && has_single_use (arg
))
1029 add_cost_one_coalesce (cl
, ver
, SSA_NAME_VERSION (arg
));
1031 add_coalesce (cl
, ver
, SSA_NAME_VERSION (arg
), cost
);
1035 if (e
->flags
& EDGE_ABNORMAL
)
1036 abnormal_corrupt (phi
, i
);
1039 bitmap_set_bit (used_in_copy
, ver
);
1042 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1044 stmt
= gsi_stmt (gsi
);
1046 if (is_gimple_debug (stmt
))
1049 /* Register USE and DEF operands in each statement. */
1050 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, (SSA_OP_DEF
|SSA_OP_USE
))
1051 register_ssa_partition (map
, var
);
1053 /* Check for copy coalesces. */
1054 switch (gimple_code (stmt
))
1058 tree lhs
= gimple_assign_lhs (stmt
);
1059 tree rhs1
= gimple_assign_rhs1 (stmt
);
1061 if (gimple_assign_copy_p (stmt
)
1062 && TREE_CODE (lhs
) == SSA_NAME
1063 && TREE_CODE (rhs1
) == SSA_NAME
1064 && SSA_NAME_VAR (lhs
) == SSA_NAME_VAR (rhs1
))
1066 v1
= SSA_NAME_VERSION (lhs
);
1067 v2
= SSA_NAME_VERSION (rhs1
);
1068 cost
= coalesce_cost_bb (bb
);
1069 add_coalesce (cl
, v1
, v2
, cost
);
1070 bitmap_set_bit (used_in_copy
, v1
);
1071 bitmap_set_bit (used_in_copy
, v2
);
1078 unsigned long noutputs
, i
;
1079 unsigned long ninputs
;
1080 tree
*outputs
, link
;
1081 noutputs
= gimple_asm_noutputs (stmt
);
1082 ninputs
= gimple_asm_ninputs (stmt
);
1083 outputs
= (tree
*) alloca (noutputs
* sizeof (tree
));
1084 for (i
= 0; i
< noutputs
; ++i
) {
1085 link
= gimple_asm_output_op (stmt
, i
);
1086 outputs
[i
] = TREE_VALUE (link
);
1089 for (i
= 0; i
< ninputs
; ++i
)
1091 const char *constraint
;
1094 unsigned long match
;
1096 link
= gimple_asm_input_op (stmt
, i
);
1098 = TREE_STRING_POINTER (TREE_VALUE (TREE_PURPOSE (link
)));
1099 input
= TREE_VALUE (link
);
1101 if (TREE_CODE (input
) != SSA_NAME
)
1104 match
= strtoul (constraint
, &end
, 10);
1105 if (match
>= noutputs
|| end
== constraint
)
1108 if (TREE_CODE (outputs
[match
]) != SSA_NAME
)
1111 v1
= SSA_NAME_VERSION (outputs
[match
]);
1112 v2
= SSA_NAME_VERSION (input
);
1114 if (SSA_NAME_VAR (outputs
[match
]) == SSA_NAME_VAR (input
))
1116 cost
= coalesce_cost (REG_BR_PROB_BASE
,
1117 optimize_bb_for_size_p (bb
));
1118 add_coalesce (cl
, v1
, v2
, cost
);
1119 bitmap_set_bit (used_in_copy
, v1
);
1120 bitmap_set_bit (used_in_copy
, v2
);
1130 #ifdef ENABLE_CHECKING
1131 /* Mark real uses and defs. */
1132 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, (SSA_OP_DEF
|SSA_OP_USE
))
1133 bitmap_set_bit (used_in_real_ops
, DECL_UID (SSA_NAME_VAR (var
)));
1135 /* Validate that virtual ops don't get used in funny ways. */
1136 if (gimple_vuse (stmt
))
1137 bitmap_set_bit (used_in_virtual_ops
,
1138 DECL_UID (SSA_NAME_VAR (gimple_vuse (stmt
))));
1139 #endif /* ENABLE_CHECKING */
1143 /* Now process result decls and live on entry variables for entry into
1144 the coalesce list. */
1146 for (i
= 1; i
< num_ssa_names
; i
++)
1149 if (var
!= NULL_TREE
&& is_gimple_reg (var
))
1151 /* Add coalesces between all the result decls. */
1152 if (TREE_CODE (SSA_NAME_VAR (var
)) == RESULT_DECL
)
1154 if (first
== NULL_TREE
)
1158 gcc_assert (SSA_NAME_VAR (var
) == SSA_NAME_VAR (first
));
1159 v1
= SSA_NAME_VERSION (first
);
1160 v2
= SSA_NAME_VERSION (var
);
1161 bitmap_set_bit (used_in_copy
, v1
);
1162 bitmap_set_bit (used_in_copy
, v2
);
1163 cost
= coalesce_cost_bb (EXIT_BLOCK_PTR
);
1164 add_coalesce (cl
, v1
, v2
, cost
);
1167 /* Mark any default_def variables as being in the coalesce list
1168 since they will have to be coalesced with the base variable. If
1169 not marked as present, they won't be in the coalesce view. */
1170 if (gimple_default_def (cfun
, SSA_NAME_VAR (var
)) == var
1171 && !has_zero_uses (var
))
1172 bitmap_set_bit (used_in_copy
, SSA_NAME_VERSION (var
));
1176 #if defined ENABLE_CHECKING
1179 bitmap both
= BITMAP_ALLOC (NULL
);
1180 bitmap_and (both
, used_in_real_ops
, used_in_virtual_ops
);
1181 if (!bitmap_empty_p (both
))
1185 EXECUTE_IF_SET_IN_BITMAP (both
, 0, i
, bi
)
1186 fprintf (stderr
, "Variable %s used in real and virtual operands\n",
1187 get_name (referenced_var (i
)));
1188 internal_error ("SSA corruption");
1191 BITMAP_FREE (used_in_real_ops
);
1192 BITMAP_FREE (used_in_virtual_ops
);
1201 /* Attempt to coalesce ssa versions X and Y together using the partition
1202 mapping in MAP and checking conflicts in GRAPH. Output any debug info to
1203 DEBUG, if it is nun-NULL. */
1206 attempt_coalesce (var_map map
, ssa_conflicts_p graph
, int x
, int y
,
1213 p1
= var_to_partition (map
, ssa_name (x
));
1214 p2
= var_to_partition (map
, ssa_name (y
));
1218 fprintf (debug
, "(%d)", x
);
1219 print_generic_expr (debug
, partition_to_var (map
, p1
), TDF_SLIM
);
1220 fprintf (debug
, " & (%d)", y
);
1221 print_generic_expr (debug
, partition_to_var (map
, p2
), TDF_SLIM
);
1227 fprintf (debug
, ": Already Coalesced.\n");
1232 fprintf (debug
, " [map: %d, %d] ", p1
, p2
);
1235 if (!ssa_conflicts_test_p (graph
, p1
, p2
))
1237 var1
= partition_to_var (map
, p1
);
1238 var2
= partition_to_var (map
, p2
);
1239 z
= var_union (map
, var1
, var2
);
1240 if (z
== NO_PARTITION
)
1243 fprintf (debug
, ": Unable to perform partition union.\n");
1247 /* z is the new combined partition. Remove the other partition from
1248 the list, and merge the conflicts. */
1250 ssa_conflicts_merge (graph
, p1
, p2
);
1252 ssa_conflicts_merge (graph
, p2
, p1
);
1255 fprintf (debug
, ": Success -> %d\n", z
);
1260 fprintf (debug
, ": Fail due to conflict\n");
1266 /* Attempt to Coalesce partitions in MAP which occur in the list CL using
1267 GRAPH. Debug output is sent to DEBUG if it is non-NULL. */
1270 coalesce_partitions (var_map map
, ssa_conflicts_p graph
, coalesce_list_p cl
,
1280 /* First, coalesce all the copies across abnormal edges. These are not placed
1281 in the coalesce list because they do not need to be sorted, and simply
1282 consume extra memory/compilation time in large programs. */
1286 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
1287 if (e
->flags
& EDGE_ABNORMAL
)
1289 gimple_stmt_iterator gsi
;
1290 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
);
1293 gimple phi
= gsi_stmt (gsi
);
1294 tree res
= PHI_RESULT (phi
);
1295 tree arg
= PHI_ARG_DEF (phi
, e
->dest_idx
);
1296 int v1
= SSA_NAME_VERSION (res
);
1297 int v2
= SSA_NAME_VERSION (arg
);
1299 if (SSA_NAME_VAR (arg
) != SSA_NAME_VAR (res
))
1300 abnormal_corrupt (phi
, e
->dest_idx
);
1303 fprintf (debug
, "Abnormal coalesce: ");
1305 if (!attempt_coalesce (map
, graph
, v1
, v2
, debug
))
1306 fail_abnormal_edge_coalesce (v1
, v2
);
1311 /* Now process the items in the coalesce list. */
1313 while ((cost
= pop_best_coalesce (cl
, &x
, &y
)) != NO_BEST_COALESCE
)
1315 var1
= ssa_name (x
);
1316 var2
= ssa_name (y
);
1318 /* Assert the coalesces have the same base variable. */
1319 gcc_assert (SSA_NAME_VAR (var1
) == SSA_NAME_VAR (var2
));
1322 fprintf (debug
, "Coalesce list: ");
1323 attempt_coalesce (map
, graph
, x
, y
, debug
);
1327 /* Returns a hash code for P. */
1330 hash_ssa_name_by_var (const void *p
)
1332 const_tree n
= (const_tree
) p
;
1333 return (hashval_t
) htab_hash_pointer (SSA_NAME_VAR (n
));
1336 /* Returns nonzero if P1 and P2 are equal. */
1339 eq_ssa_name_by_var (const void *p1
, const void *p2
)
1341 const_tree n1
= (const_tree
) p1
;
1342 const_tree n2
= (const_tree
) p2
;
1343 return SSA_NAME_VAR (n1
) == SSA_NAME_VAR (n2
);
1346 /* Reduce the number of copies by coalescing variables in the function. Return
1347 a partition map with the resulting coalesces. */
1350 coalesce_ssa_name (void)
1352 tree_live_info_p liveinfo
;
1353 ssa_conflicts_p graph
;
1355 bitmap used_in_copies
= BITMAP_ALLOC (NULL
);
1358 static htab_t ssa_name_hash
;
1360 cl
= create_coalesce_list ();
1361 map
= create_outofssa_var_map (cl
, used_in_copies
);
1363 /* We need to coalesce all names originating same SSA_NAME_VAR
1364 so debug info remains undisturbed. */
1367 ssa_name_hash
= htab_create (10, hash_ssa_name_by_var
,
1368 eq_ssa_name_by_var
, NULL
);
1369 for (i
= 1; i
< num_ssa_names
; i
++)
1371 tree a
= ssa_name (i
);
1375 && !DECL_IGNORED_P (SSA_NAME_VAR (a
))
1376 && (!has_zero_uses (a
) || !SSA_NAME_IS_DEFAULT_DEF (a
)))
1378 tree
*slot
= (tree
*) htab_find_slot (ssa_name_hash
, a
, INSERT
);
1384 add_coalesce (cl
, SSA_NAME_VERSION (a
), SSA_NAME_VERSION (*slot
),
1385 MUST_COALESCE_COST
- 1);
1386 bitmap_set_bit (used_in_copies
, SSA_NAME_VERSION (a
));
1387 bitmap_set_bit (used_in_copies
, SSA_NAME_VERSION (*slot
));
1391 htab_delete (ssa_name_hash
);
1393 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1394 dump_var_map (dump_file
, map
);
1396 /* Don't calculate live ranges for variables not in the coalesce list. */
1397 partition_view_bitmap (map
, used_in_copies
, true);
1398 BITMAP_FREE (used_in_copies
);
1400 if (num_var_partitions (map
) < 1)
1402 delete_coalesce_list (cl
);
1406 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1407 dump_var_map (dump_file
, map
);
1409 liveinfo
= calculate_live_ranges (map
);
1411 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1412 dump_live_info (dump_file
, liveinfo
, LIVEDUMP_ENTRY
);
1414 /* Build a conflict graph. */
1415 graph
= build_ssa_conflict_graph (liveinfo
);
1416 delete_tree_live_info (liveinfo
);
1417 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1418 ssa_conflicts_dump (dump_file
, graph
);
1420 sort_coalesce_list (cl
);
1422 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1424 fprintf (dump_file
, "\nAfter sorting:\n");
1425 dump_coalesce_list (dump_file
, cl
);
1428 /* First, coalesce all live on entry variables to their base variable.
1429 This will ensure the first use is coming from the correct location. */
1431 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1432 dump_var_map (dump_file
, map
);
1434 /* Now coalesce everything in the list. */
1435 coalesce_partitions (map
, graph
, cl
,
1436 ((dump_flags
& TDF_DETAILS
) ? dump_file
1439 delete_coalesce_list (cl
);
1440 ssa_conflicts_delete (graph
);