1 /* IRA allocation based on graph coloring.
2 Copyright (C) 2006-2017 Free Software Foundation, Inc.
3 Contributed by Vladimir Makarov <vmakarov@redhat.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
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/>. */
23 #include "coretypes.h"
32 #include "insn-config.h"
39 typedef struct allocno_hard_regs
*allocno_hard_regs_t
;
41 /* The structure contains information about hard registers can be
42 assigned to allocnos. Usually it is allocno profitable hard
43 registers but in some cases this set can be a bit different. Major
44 reason of the difference is a requirement to use hard register sets
45 that form a tree or a forest (set of trees), i.e. hard register set
46 of a node should contain hard register sets of its subnodes. */
47 struct allocno_hard_regs
49 /* Hard registers can be assigned to an allocno. */
51 /* Overall (spilling) cost of all allocnos with given register
56 typedef struct allocno_hard_regs_node
*allocno_hard_regs_node_t
;
58 /* A node representing allocno hard registers. Such nodes form a
59 forest (set of trees). Each subnode of given node in the forest
60 refers for hard register set (usually allocno profitable hard
61 register set) which is a subset of one referred from given
63 struct allocno_hard_regs_node
65 /* Set up number of the node in preorder traversing of the forest. */
67 /* Used for different calculation like finding conflict size of an
70 /* Used for calculation of conflict size of an allocno. The
71 conflict size of the allocno is maximal number of given allocno
72 hard registers needed for allocation of the conflicting allocnos.
73 Given allocno is trivially colored if this number plus the number
74 of hard registers needed for given allocno is not greater than
75 the number of given allocno hard register set. */
77 /* The number of hard registers given by member hard_regs. */
79 /* The following member is used to form the final forest. */
81 /* Pointer to the corresponding profitable hard registers. */
82 allocno_hard_regs_t hard_regs
;
83 /* Parent, first subnode, previous and next node with the same
84 parent in the forest. */
85 allocno_hard_regs_node_t parent
, first
, prev
, next
;
88 /* Info about changing hard reg costs of an allocno. */
89 struct update_cost_record
91 /* Hard regno for which we changed the cost. */
93 /* Divisor used when we changed the cost of HARD_REGNO. */
95 /* Next record for given allocno. */
96 struct update_cost_record
*next
;
99 /* To decrease footprint of ira_allocno structure we store all data
100 needed only for coloring in the following structure. */
101 struct allocno_color_data
103 /* TRUE value means that the allocno was not removed yet from the
104 conflicting graph during coloring. */
105 unsigned int in_graph_p
: 1;
106 /* TRUE if it is put on the stack to make other allocnos
108 unsigned int may_be_spilled_p
: 1;
109 /* TRUE if the allocno is trivially colorable. */
110 unsigned int colorable_p
: 1;
111 /* Number of hard registers of the allocno class really
112 available for the allocno allocation. It is number of the
113 profitable hard regs. */
114 int available_regs_num
;
115 /* Allocnos in a bucket (used in coloring) chained by the following
117 ira_allocno_t next_bucket_allocno
;
118 ira_allocno_t prev_bucket_allocno
;
119 /* Used for temporary purposes. */
121 /* Used to exclude repeated processing. */
123 /* Profitable hard regs available for this pseudo allocation. It
124 means that the set excludes unavailable hard regs and hard regs
125 conflicting with given pseudo. They should be of the allocno
127 HARD_REG_SET profitable_hard_regs
;
128 /* The allocno hard registers node. */
129 allocno_hard_regs_node_t hard_regs_node
;
130 /* Array of structures allocno_hard_regs_subnode representing
131 given allocno hard registers node (the 1st element in the array)
132 and all its subnodes in the tree (forest) of allocno hard
133 register nodes (see comments above). */
134 int hard_regs_subnodes_start
;
135 /* The length of the previous array. */
136 int hard_regs_subnodes_num
;
137 /* Records about updating allocno hard reg costs from copies. If
138 the allocno did not get expected hard register, these records are
139 used to restore original hard reg costs of allocnos connected to
140 this allocno by copies. */
141 struct update_cost_record
*update_cost_records
;
142 /* Threads. We collect allocnos connected by copies into threads
143 and try to assign hard regs to allocnos by threads. */
144 /* Allocno representing all thread. */
145 ira_allocno_t first_thread_allocno
;
146 /* Allocnos in thread forms a cycle list through the following
148 ira_allocno_t next_thread_allocno
;
149 /* All thread frequency. Defined only for first thread allocno. */
154 typedef struct allocno_color_data
*allocno_color_data_t
;
156 /* Container for storing allocno data concerning coloring. */
157 static allocno_color_data_t allocno_color_data
;
159 /* Macro to access the data concerning coloring. */
160 #define ALLOCNO_COLOR_DATA(a) ((allocno_color_data_t) ALLOCNO_ADD_DATA (a))
162 /* Used for finding allocno colorability to exclude repeated allocno
163 processing and for updating preferencing to exclude repeated
164 allocno processing during assignment. */
165 static int curr_allocno_process
;
167 /* This file contains code for regional graph coloring, spill/restore
168 code placement optimization, and code helping the reload pass to do
171 /* Bitmap of allocnos which should be colored. */
172 static bitmap coloring_allocno_bitmap
;
174 /* Bitmap of allocnos which should be taken into account during
175 coloring. In general case it contains allocnos from
176 coloring_allocno_bitmap plus other already colored conflicting
178 static bitmap consideration_allocno_bitmap
;
180 /* All allocnos sorted according their priorities. */
181 static ira_allocno_t
*sorted_allocnos
;
183 /* Vec representing the stack of allocnos used during coloring. */
184 static vec
<ira_allocno_t
> allocno_stack_vec
;
186 /* Helper for qsort comparison callbacks - return a positive integer if
187 X > Y, or a negative value otherwise. Use a conditional expression
188 instead of a difference computation to insulate from possible overflow
189 issues, e.g. X - Y < 0 for some X > 0 and Y < 0. */
190 #define SORTGT(x,y) (((x) > (y)) ? 1 : -1)
194 /* Definition of vector of allocno hard registers. */
196 /* Vector of unique allocno hard registers. */
197 static vec
<allocno_hard_regs_t
> allocno_hard_regs_vec
;
199 struct allocno_hard_regs_hasher
: nofree_ptr_hash
<allocno_hard_regs
>
201 static inline hashval_t
hash (const allocno_hard_regs
*);
202 static inline bool equal (const allocno_hard_regs
*,
203 const allocno_hard_regs
*);
206 /* Returns hash value for allocno hard registers V. */
208 allocno_hard_regs_hasher::hash (const allocno_hard_regs
*hv
)
210 return iterative_hash (&hv
->set
, sizeof (HARD_REG_SET
), 0);
213 /* Compares allocno hard registers V1 and V2. */
215 allocno_hard_regs_hasher::equal (const allocno_hard_regs
*hv1
,
216 const allocno_hard_regs
*hv2
)
218 return hard_reg_set_equal_p (hv1
->set
, hv2
->set
);
221 /* Hash table of unique allocno hard registers. */
222 static hash_table
<allocno_hard_regs_hasher
> *allocno_hard_regs_htab
;
224 /* Return allocno hard registers in the hash table equal to HV. */
225 static allocno_hard_regs_t
226 find_hard_regs (allocno_hard_regs_t hv
)
228 return allocno_hard_regs_htab
->find (hv
);
231 /* Insert allocno hard registers HV in the hash table (if it is not
232 there yet) and return the value which in the table. */
233 static allocno_hard_regs_t
234 insert_hard_regs (allocno_hard_regs_t hv
)
236 allocno_hard_regs
**slot
= allocno_hard_regs_htab
->find_slot (hv
, INSERT
);
243 /* Initialize data concerning allocno hard registers. */
245 init_allocno_hard_regs (void)
247 allocno_hard_regs_vec
.create (200);
248 allocno_hard_regs_htab
249 = new hash_table
<allocno_hard_regs_hasher
> (200);
252 /* Add (or update info about) allocno hard registers with SET and
254 static allocno_hard_regs_t
255 add_allocno_hard_regs (HARD_REG_SET set
, int64_t cost
)
257 struct allocno_hard_regs temp
;
258 allocno_hard_regs_t hv
;
260 gcc_assert (! hard_reg_set_empty_p (set
));
261 COPY_HARD_REG_SET (temp
.set
, set
);
262 if ((hv
= find_hard_regs (&temp
)) != NULL
)
266 hv
= ((struct allocno_hard_regs
*)
267 ira_allocate (sizeof (struct allocno_hard_regs
)));
268 COPY_HARD_REG_SET (hv
->set
, set
);
270 allocno_hard_regs_vec
.safe_push (hv
);
271 insert_hard_regs (hv
);
276 /* Finalize data concerning allocno hard registers. */
278 finish_allocno_hard_regs (void)
281 allocno_hard_regs_t hv
;
284 allocno_hard_regs_vec
.iterate (i
, &hv
);
287 delete allocno_hard_regs_htab
;
288 allocno_hard_regs_htab
= NULL
;
289 allocno_hard_regs_vec
.release ();
292 /* Sort hard regs according to their frequency of usage. */
294 allocno_hard_regs_compare (const void *v1p
, const void *v2p
)
296 allocno_hard_regs_t hv1
= *(const allocno_hard_regs_t
*) v1p
;
297 allocno_hard_regs_t hv2
= *(const allocno_hard_regs_t
*) v2p
;
299 if (hv2
->cost
> hv1
->cost
)
301 else if (hv2
->cost
< hv1
->cost
)
309 /* Used for finding a common ancestor of two allocno hard registers
310 nodes in the forest. We use the current value of
311 'node_check_tick' to mark all nodes from one node to the top and
312 then walking up from another node until we find a marked node.
314 It is also used to figure out allocno colorability as a mark that
315 we already reset value of member 'conflict_size' for the forest
316 node corresponding to the processed allocno. */
317 static int node_check_tick
;
319 /* Roots of the forest containing hard register sets can be assigned
321 static allocno_hard_regs_node_t hard_regs_roots
;
323 /* Definition of vector of allocno hard register nodes. */
325 /* Vector used to create the forest. */
326 static vec
<allocno_hard_regs_node_t
> hard_regs_node_vec
;
328 /* Create and return allocno hard registers node containing allocno
329 hard registers HV. */
330 static allocno_hard_regs_node_t
331 create_new_allocno_hard_regs_node (allocno_hard_regs_t hv
)
333 allocno_hard_regs_node_t new_node
;
335 new_node
= ((struct allocno_hard_regs_node
*)
336 ira_allocate (sizeof (struct allocno_hard_regs_node
)));
338 new_node
->hard_regs
= hv
;
339 new_node
->hard_regs_num
= hard_reg_set_size (hv
->set
);
340 new_node
->first
= NULL
;
341 new_node
->used_p
= false;
345 /* Add allocno hard registers node NEW_NODE to the forest on its level
348 add_new_allocno_hard_regs_node_to_forest (allocno_hard_regs_node_t
*roots
,
349 allocno_hard_regs_node_t new_node
)
351 new_node
->next
= *roots
;
352 if (new_node
->next
!= NULL
)
353 new_node
->next
->prev
= new_node
;
354 new_node
->prev
= NULL
;
358 /* Add allocno hard registers HV (or its best approximation if it is
359 not possible) to the forest on its level given by ROOTS. */
361 add_allocno_hard_regs_to_forest (allocno_hard_regs_node_t
*roots
,
362 allocno_hard_regs_t hv
)
364 unsigned int i
, start
;
365 allocno_hard_regs_node_t node
, prev
, new_node
;
366 HARD_REG_SET temp_set
;
367 allocno_hard_regs_t hv2
;
369 start
= hard_regs_node_vec
.length ();
370 for (node
= *roots
; node
!= NULL
; node
= node
->next
)
372 if (hard_reg_set_equal_p (hv
->set
, node
->hard_regs
->set
))
374 if (hard_reg_set_subset_p (hv
->set
, node
->hard_regs
->set
))
376 add_allocno_hard_regs_to_forest (&node
->first
, hv
);
379 if (hard_reg_set_subset_p (node
->hard_regs
->set
, hv
->set
))
380 hard_regs_node_vec
.safe_push (node
);
381 else if (hard_reg_set_intersect_p (hv
->set
, node
->hard_regs
->set
))
383 COPY_HARD_REG_SET (temp_set
, hv
->set
);
384 AND_HARD_REG_SET (temp_set
, node
->hard_regs
->set
);
385 hv2
= add_allocno_hard_regs (temp_set
, hv
->cost
);
386 add_allocno_hard_regs_to_forest (&node
->first
, hv2
);
389 if (hard_regs_node_vec
.length ()
392 /* Create a new node which contains nodes in hard_regs_node_vec. */
393 CLEAR_HARD_REG_SET (temp_set
);
395 i
< hard_regs_node_vec
.length ();
398 node
= hard_regs_node_vec
[i
];
399 IOR_HARD_REG_SET (temp_set
, node
->hard_regs
->set
);
401 hv
= add_allocno_hard_regs (temp_set
, hv
->cost
);
402 new_node
= create_new_allocno_hard_regs_node (hv
);
405 i
< hard_regs_node_vec
.length ();
408 node
= hard_regs_node_vec
[i
];
409 if (node
->prev
== NULL
)
412 node
->prev
->next
= node
->next
;
413 if (node
->next
!= NULL
)
414 node
->next
->prev
= node
->prev
;
416 new_node
->first
= node
;
423 add_new_allocno_hard_regs_node_to_forest (roots
, new_node
);
425 hard_regs_node_vec
.truncate (start
);
428 /* Add allocno hard registers nodes starting with the forest level
429 given by FIRST which contains biggest set inside SET. */
431 collect_allocno_hard_regs_cover (allocno_hard_regs_node_t first
,
434 allocno_hard_regs_node_t node
;
436 ira_assert (first
!= NULL
);
437 for (node
= first
; node
!= NULL
; node
= node
->next
)
438 if (hard_reg_set_subset_p (node
->hard_regs
->set
, set
))
439 hard_regs_node_vec
.safe_push (node
);
440 else if (hard_reg_set_intersect_p (set
, node
->hard_regs
->set
))
441 collect_allocno_hard_regs_cover (node
->first
, set
);
444 /* Set up field parent as PARENT in all allocno hard registers nodes
445 in forest given by FIRST. */
447 setup_allocno_hard_regs_nodes_parent (allocno_hard_regs_node_t first
,
448 allocno_hard_regs_node_t parent
)
450 allocno_hard_regs_node_t node
;
452 for (node
= first
; node
!= NULL
; node
= node
->next
)
454 node
->parent
= parent
;
455 setup_allocno_hard_regs_nodes_parent (node
->first
, node
);
459 /* Return allocno hard registers node which is a first common ancestor
460 node of FIRST and SECOND in the forest. */
461 static allocno_hard_regs_node_t
462 first_common_ancestor_node (allocno_hard_regs_node_t first
,
463 allocno_hard_regs_node_t second
)
465 allocno_hard_regs_node_t node
;
468 for (node
= first
; node
!= NULL
; node
= node
->parent
)
469 node
->check
= node_check_tick
;
470 for (node
= second
; node
!= NULL
; node
= node
->parent
)
471 if (node
->check
== node_check_tick
)
473 return first_common_ancestor_node (second
, first
);
476 /* Print hard reg set SET to F. */
478 print_hard_reg_set (FILE *f
, HARD_REG_SET set
, bool new_line_p
)
482 for (start
= -1, i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
484 if (TEST_HARD_REG_BIT (set
, i
))
486 if (i
== 0 || ! TEST_HARD_REG_BIT (set
, i
- 1))
490 && (i
== FIRST_PSEUDO_REGISTER
- 1 || ! TEST_HARD_REG_BIT (set
, i
)))
493 fprintf (f
, " %d", start
);
494 else if (start
== i
- 2)
495 fprintf (f
, " %d %d", start
, start
+ 1);
497 fprintf (f
, " %d-%d", start
, i
- 1);
505 /* Print allocno hard register subforest given by ROOTS and its LEVEL
508 print_hard_regs_subforest (FILE *f
, allocno_hard_regs_node_t roots
,
512 allocno_hard_regs_node_t node
;
514 for (node
= roots
; node
!= NULL
; node
= node
->next
)
517 for (i
= 0; i
< level
* 2; i
++)
519 fprintf (f
, "%d:(", node
->preorder_num
);
520 print_hard_reg_set (f
, node
->hard_regs
->set
, false);
521 fprintf (f
, ")@%" PRId64
"\n", node
->hard_regs
->cost
);
522 print_hard_regs_subforest (f
, node
->first
, level
+ 1);
526 /* Print the allocno hard register forest to F. */
528 print_hard_regs_forest (FILE *f
)
530 fprintf (f
, " Hard reg set forest:\n");
531 print_hard_regs_subforest (f
, hard_regs_roots
, 1);
534 /* Print the allocno hard register forest to stderr. */
536 ira_debug_hard_regs_forest (void)
538 print_hard_regs_forest (stderr
);
541 /* Remove unused allocno hard registers nodes from forest given by its
544 remove_unused_allocno_hard_regs_nodes (allocno_hard_regs_node_t
*roots
)
546 allocno_hard_regs_node_t node
, prev
, next
, last
;
548 for (prev
= NULL
, node
= *roots
; node
!= NULL
; node
= next
)
553 remove_unused_allocno_hard_regs_nodes (&node
->first
);
558 for (last
= node
->first
;
559 last
!= NULL
&& last
->next
!= NULL
;
565 *roots
= node
->first
;
567 prev
->next
= node
->first
;
587 /* Set up fields preorder_num starting with START_NUM in all allocno
588 hard registers nodes in forest given by FIRST. Return biggest set
589 PREORDER_NUM increased by 1. */
591 enumerate_allocno_hard_regs_nodes (allocno_hard_regs_node_t first
,
592 allocno_hard_regs_node_t parent
,
595 allocno_hard_regs_node_t node
;
597 for (node
= first
; node
!= NULL
; node
= node
->next
)
599 node
->preorder_num
= start_num
++;
600 node
->parent
= parent
;
601 start_num
= enumerate_allocno_hard_regs_nodes (node
->first
, node
,
607 /* Number of allocno hard registers nodes in the forest. */
608 static int allocno_hard_regs_nodes_num
;
610 /* Table preorder number of allocno hard registers node in the forest
611 -> the allocno hard registers node. */
612 static allocno_hard_regs_node_t
*allocno_hard_regs_nodes
;
615 typedef struct allocno_hard_regs_subnode
*allocno_hard_regs_subnode_t
;
617 /* The structure is used to describes all subnodes (not only immediate
618 ones) in the mentioned above tree for given allocno hard register
619 node. The usage of such data accelerates calculation of
620 colorability of given allocno. */
621 struct allocno_hard_regs_subnode
623 /* The conflict size of conflicting allocnos whose hard register
624 sets are equal sets (plus supersets if given node is given
625 allocno hard registers node) of one in the given node. */
626 int left_conflict_size
;
627 /* The summary conflict size of conflicting allocnos whose hard
628 register sets are strict subsets of one in the given node.
629 Overall conflict size is
630 left_conflict_subnodes_size
631 + MIN (max_node_impact - left_conflict_subnodes_size,
634 short left_conflict_subnodes_size
;
635 short max_node_impact
;
638 /* Container for hard regs subnodes of all allocnos. */
639 static allocno_hard_regs_subnode_t allocno_hard_regs_subnodes
;
641 /* Table (preorder number of allocno hard registers node in the
642 forest, preorder number of allocno hard registers subnode) -> index
643 of the subnode relative to the node. -1 if it is not a
645 static int *allocno_hard_regs_subnode_index
;
647 /* Setup arrays ALLOCNO_HARD_REGS_NODES and
648 ALLOCNO_HARD_REGS_SUBNODE_INDEX. */
650 setup_allocno_hard_regs_subnode_index (allocno_hard_regs_node_t first
)
652 allocno_hard_regs_node_t node
, parent
;
655 for (node
= first
; node
!= NULL
; node
= node
->next
)
657 allocno_hard_regs_nodes
[node
->preorder_num
] = node
;
658 for (parent
= node
; parent
!= NULL
; parent
= parent
->parent
)
660 index
= parent
->preorder_num
* allocno_hard_regs_nodes_num
;
661 allocno_hard_regs_subnode_index
[index
+ node
->preorder_num
]
662 = node
->preorder_num
- parent
->preorder_num
;
664 setup_allocno_hard_regs_subnode_index (node
->first
);
668 /* Count all allocno hard registers nodes in tree ROOT. */
670 get_allocno_hard_regs_subnodes_num (allocno_hard_regs_node_t root
)
674 for (root
= root
->first
; root
!= NULL
; root
= root
->next
)
675 len
+= get_allocno_hard_regs_subnodes_num (root
);
679 /* Build the forest of allocno hard registers nodes and assign each
680 allocno a node from the forest. */
682 form_allocno_hard_regs_nodes_forest (void)
684 unsigned int i
, j
, size
, len
;
687 allocno_hard_regs_t hv
;
690 allocno_hard_regs_node_t node
, allocno_hard_regs_node
;
691 allocno_color_data_t allocno_data
;
694 init_allocno_hard_regs ();
695 hard_regs_roots
= NULL
;
696 hard_regs_node_vec
.create (100);
697 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
698 if (! TEST_HARD_REG_BIT (ira_no_alloc_regs
, i
))
700 CLEAR_HARD_REG_SET (temp
);
701 SET_HARD_REG_BIT (temp
, i
);
702 hv
= add_allocno_hard_regs (temp
, 0);
703 node
= create_new_allocno_hard_regs_node (hv
);
704 add_new_allocno_hard_regs_node_to_forest (&hard_regs_roots
, node
);
706 start
= allocno_hard_regs_vec
.length ();
707 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap
, 0, i
, bi
)
710 allocno_data
= ALLOCNO_COLOR_DATA (a
);
712 if (hard_reg_set_empty_p (allocno_data
->profitable_hard_regs
))
714 hv
= (add_allocno_hard_regs
715 (allocno_data
->profitable_hard_regs
,
716 ALLOCNO_MEMORY_COST (a
) - ALLOCNO_CLASS_COST (a
)));
718 SET_HARD_REG_SET (temp
);
719 AND_COMPL_HARD_REG_SET (temp
, ira_no_alloc_regs
);
720 add_allocno_hard_regs (temp
, 0);
721 qsort (allocno_hard_regs_vec
.address () + start
,
722 allocno_hard_regs_vec
.length () - start
,
723 sizeof (allocno_hard_regs_t
), allocno_hard_regs_compare
);
725 allocno_hard_regs_vec
.iterate (i
, &hv
);
728 add_allocno_hard_regs_to_forest (&hard_regs_roots
, hv
);
729 ira_assert (hard_regs_node_vec
.length () == 0);
731 /* We need to set up parent fields for right work of
732 first_common_ancestor_node. */
733 setup_allocno_hard_regs_nodes_parent (hard_regs_roots
, NULL
);
734 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap
, 0, i
, bi
)
737 allocno_data
= ALLOCNO_COLOR_DATA (a
);
738 if (hard_reg_set_empty_p (allocno_data
->profitable_hard_regs
))
740 hard_regs_node_vec
.truncate (0);
741 collect_allocno_hard_regs_cover (hard_regs_roots
,
742 allocno_data
->profitable_hard_regs
);
743 allocno_hard_regs_node
= NULL
;
744 for (j
= 0; hard_regs_node_vec
.iterate (j
, &node
); j
++)
745 allocno_hard_regs_node
748 : first_common_ancestor_node (node
, allocno_hard_regs_node
));
749 /* That is a temporary storage. */
750 allocno_hard_regs_node
->used_p
= true;
751 allocno_data
->hard_regs_node
= allocno_hard_regs_node
;
753 ira_assert (hard_regs_roots
->next
== NULL
);
754 hard_regs_roots
->used_p
= true;
755 remove_unused_allocno_hard_regs_nodes (&hard_regs_roots
);
756 allocno_hard_regs_nodes_num
757 = enumerate_allocno_hard_regs_nodes (hard_regs_roots
, NULL
, 0);
758 allocno_hard_regs_nodes
759 = ((allocno_hard_regs_node_t
*)
760 ira_allocate (allocno_hard_regs_nodes_num
761 * sizeof (allocno_hard_regs_node_t
)));
762 size
= allocno_hard_regs_nodes_num
* allocno_hard_regs_nodes_num
;
763 allocno_hard_regs_subnode_index
764 = (int *) ira_allocate (size
* sizeof (int));
765 for (i
= 0; i
< size
; i
++)
766 allocno_hard_regs_subnode_index
[i
] = -1;
767 setup_allocno_hard_regs_subnode_index (hard_regs_roots
);
769 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap
, 0, i
, bi
)
772 allocno_data
= ALLOCNO_COLOR_DATA (a
);
773 if (hard_reg_set_empty_p (allocno_data
->profitable_hard_regs
))
775 len
= get_allocno_hard_regs_subnodes_num (allocno_data
->hard_regs_node
);
776 allocno_data
->hard_regs_subnodes_start
= start
;
777 allocno_data
->hard_regs_subnodes_num
= len
;
780 allocno_hard_regs_subnodes
781 = ((allocno_hard_regs_subnode_t
)
782 ira_allocate (sizeof (struct allocno_hard_regs_subnode
) * start
));
783 hard_regs_node_vec
.release ();
786 /* Free tree of allocno hard registers nodes given by its ROOT. */
788 finish_allocno_hard_regs_nodes_tree (allocno_hard_regs_node_t root
)
790 allocno_hard_regs_node_t child
, next
;
792 for (child
= root
->first
; child
!= NULL
; child
= next
)
795 finish_allocno_hard_regs_nodes_tree (child
);
800 /* Finish work with the forest of allocno hard registers nodes. */
802 finish_allocno_hard_regs_nodes_forest (void)
804 allocno_hard_regs_node_t node
, next
;
806 ira_free (allocno_hard_regs_subnodes
);
807 for (node
= hard_regs_roots
; node
!= NULL
; node
= next
)
810 finish_allocno_hard_regs_nodes_tree (node
);
812 ira_free (allocno_hard_regs_nodes
);
813 ira_free (allocno_hard_regs_subnode_index
);
814 finish_allocno_hard_regs ();
817 /* Set up left conflict sizes and left conflict subnodes sizes of hard
818 registers subnodes of allocno A. Return TRUE if allocno A is
819 trivially colorable. */
821 setup_left_conflict_sizes_p (ira_allocno_t a
)
823 int i
, k
, nobj
, start
;
824 int conflict_size
, left_conflict_subnodes_size
, node_preorder_num
;
825 allocno_color_data_t data
;
826 HARD_REG_SET profitable_hard_regs
;
827 allocno_hard_regs_subnode_t subnodes
;
828 allocno_hard_regs_node_t node
;
829 HARD_REG_SET node_set
;
831 nobj
= ALLOCNO_NUM_OBJECTS (a
);
832 data
= ALLOCNO_COLOR_DATA (a
);
833 subnodes
= allocno_hard_regs_subnodes
+ data
->hard_regs_subnodes_start
;
834 COPY_HARD_REG_SET (profitable_hard_regs
, data
->profitable_hard_regs
);
835 node
= data
->hard_regs_node
;
836 node_preorder_num
= node
->preorder_num
;
837 COPY_HARD_REG_SET (node_set
, node
->hard_regs
->set
);
839 for (k
= 0; k
< nobj
; k
++)
841 ira_object_t obj
= ALLOCNO_OBJECT (a
, k
);
842 ira_object_t conflict_obj
;
843 ira_object_conflict_iterator oci
;
845 FOR_EACH_OBJECT_CONFLICT (obj
, conflict_obj
, oci
)
848 ira_allocno_t conflict_a
= OBJECT_ALLOCNO (conflict_obj
);
849 allocno_hard_regs_node_t conflict_node
, temp_node
;
850 HARD_REG_SET conflict_node_set
;
851 allocno_color_data_t conflict_data
;
853 conflict_data
= ALLOCNO_COLOR_DATA (conflict_a
);
854 if (! ALLOCNO_COLOR_DATA (conflict_a
)->in_graph_p
855 || ! hard_reg_set_intersect_p (profitable_hard_regs
,
857 ->profitable_hard_regs
))
859 conflict_node
= conflict_data
->hard_regs_node
;
860 COPY_HARD_REG_SET (conflict_node_set
, conflict_node
->hard_regs
->set
);
861 if (hard_reg_set_subset_p (node_set
, conflict_node_set
))
865 ira_assert (hard_reg_set_subset_p (conflict_node_set
, node_set
));
866 temp_node
= conflict_node
;
868 if (temp_node
->check
!= node_check_tick
)
870 temp_node
->check
= node_check_tick
;
871 temp_node
->conflict_size
= 0;
873 size
= (ira_reg_class_max_nregs
874 [ALLOCNO_CLASS (conflict_a
)][ALLOCNO_MODE (conflict_a
)]);
875 if (ALLOCNO_NUM_OBJECTS (conflict_a
) > 1)
876 /* We will deal with the subwords individually. */
878 temp_node
->conflict_size
+= size
;
881 for (i
= 0; i
< data
->hard_regs_subnodes_num
; i
++)
883 allocno_hard_regs_node_t temp_node
;
885 temp_node
= allocno_hard_regs_nodes
[i
+ node_preorder_num
];
886 ira_assert (temp_node
->preorder_num
== i
+ node_preorder_num
);
887 subnodes
[i
].left_conflict_size
= (temp_node
->check
!= node_check_tick
888 ? 0 : temp_node
->conflict_size
);
889 if (hard_reg_set_subset_p (temp_node
->hard_regs
->set
,
890 profitable_hard_regs
))
891 subnodes
[i
].max_node_impact
= temp_node
->hard_regs_num
;
894 HARD_REG_SET temp_set
;
895 int j
, n
, hard_regno
;
896 enum reg_class aclass
;
898 COPY_HARD_REG_SET (temp_set
, temp_node
->hard_regs
->set
);
899 AND_HARD_REG_SET (temp_set
, profitable_hard_regs
);
900 aclass
= ALLOCNO_CLASS (a
);
901 for (n
= 0, j
= ira_class_hard_regs_num
[aclass
] - 1; j
>= 0; j
--)
903 hard_regno
= ira_class_hard_regs
[aclass
][j
];
904 if (TEST_HARD_REG_BIT (temp_set
, hard_regno
))
907 subnodes
[i
].max_node_impact
= n
;
909 subnodes
[i
].left_conflict_subnodes_size
= 0;
911 start
= node_preorder_num
* allocno_hard_regs_nodes_num
;
912 for (i
= data
->hard_regs_subnodes_num
- 1; i
> 0; i
--)
915 allocno_hard_regs_node_t parent
;
917 size
= (subnodes
[i
].left_conflict_subnodes_size
918 + MIN (subnodes
[i
].max_node_impact
919 - subnodes
[i
].left_conflict_subnodes_size
,
920 subnodes
[i
].left_conflict_size
));
921 parent
= allocno_hard_regs_nodes
[i
+ node_preorder_num
]->parent
;
922 gcc_checking_assert(parent
);
924 = allocno_hard_regs_subnode_index
[start
+ parent
->preorder_num
];
925 gcc_checking_assert(parent_i
>= 0);
926 subnodes
[parent_i
].left_conflict_subnodes_size
+= size
;
928 left_conflict_subnodes_size
= subnodes
[0].left_conflict_subnodes_size
;
930 = (left_conflict_subnodes_size
931 + MIN (subnodes
[0].max_node_impact
- left_conflict_subnodes_size
,
932 subnodes
[0].left_conflict_size
));
933 conflict_size
+= ira_reg_class_max_nregs
[ALLOCNO_CLASS (a
)][ALLOCNO_MODE (a
)];
934 data
->colorable_p
= conflict_size
<= data
->available_regs_num
;
935 return data
->colorable_p
;
938 /* Update left conflict sizes of hard registers subnodes of allocno A
939 after removing allocno REMOVED_A with SIZE from the conflict graph.
940 Return TRUE if A is trivially colorable. */
942 update_left_conflict_sizes_p (ira_allocno_t a
,
943 ira_allocno_t removed_a
, int size
)
945 int i
, conflict_size
, before_conflict_size
, diff
, start
;
946 int node_preorder_num
, parent_i
;
947 allocno_hard_regs_node_t node
, removed_node
, parent
;
948 allocno_hard_regs_subnode_t subnodes
;
949 allocno_color_data_t data
= ALLOCNO_COLOR_DATA (a
);
951 ira_assert (! data
->colorable_p
);
952 node
= data
->hard_regs_node
;
953 node_preorder_num
= node
->preorder_num
;
954 removed_node
= ALLOCNO_COLOR_DATA (removed_a
)->hard_regs_node
;
955 ira_assert (hard_reg_set_subset_p (removed_node
->hard_regs
->set
,
956 node
->hard_regs
->set
)
957 || hard_reg_set_subset_p (node
->hard_regs
->set
,
958 removed_node
->hard_regs
->set
));
959 start
= node_preorder_num
* allocno_hard_regs_nodes_num
;
960 i
= allocno_hard_regs_subnode_index
[start
+ removed_node
->preorder_num
];
963 subnodes
= allocno_hard_regs_subnodes
+ data
->hard_regs_subnodes_start
;
965 = (subnodes
[i
].left_conflict_subnodes_size
966 + MIN (subnodes
[i
].max_node_impact
967 - subnodes
[i
].left_conflict_subnodes_size
,
968 subnodes
[i
].left_conflict_size
));
969 subnodes
[i
].left_conflict_size
-= size
;
973 = (subnodes
[i
].left_conflict_subnodes_size
974 + MIN (subnodes
[i
].max_node_impact
975 - subnodes
[i
].left_conflict_subnodes_size
,
976 subnodes
[i
].left_conflict_size
));
977 if ((diff
= before_conflict_size
- conflict_size
) == 0)
979 ira_assert (conflict_size
< before_conflict_size
);
980 parent
= allocno_hard_regs_nodes
[i
+ node_preorder_num
]->parent
;
984 = allocno_hard_regs_subnode_index
[start
+ parent
->preorder_num
];
989 = (subnodes
[i
].left_conflict_subnodes_size
990 + MIN (subnodes
[i
].max_node_impact
991 - subnodes
[i
].left_conflict_subnodes_size
,
992 subnodes
[i
].left_conflict_size
));
993 subnodes
[i
].left_conflict_subnodes_size
-= diff
;
997 + ira_reg_class_max_nregs
[ALLOCNO_CLASS (a
)][ALLOCNO_MODE (a
)]
998 > data
->available_regs_num
))
1000 data
->colorable_p
= true;
1004 /* Return true if allocno A has empty profitable hard regs. */
1006 empty_profitable_hard_regs (ira_allocno_t a
)
1008 allocno_color_data_t data
= ALLOCNO_COLOR_DATA (a
);
1010 return hard_reg_set_empty_p (data
->profitable_hard_regs
);
1013 /* Set up profitable hard registers for each allocno being
1016 setup_profitable_hard_regs (void)
1019 int j
, k
, nobj
, hard_regno
, nregs
, class_size
;
1022 enum reg_class aclass
;
1024 allocno_color_data_t data
;
1026 /* Initial set up from allocno classes and explicitly conflicting
1028 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap
, 0, i
, bi
)
1030 a
= ira_allocnos
[i
];
1031 if ((aclass
= ALLOCNO_CLASS (a
)) == NO_REGS
)
1033 data
= ALLOCNO_COLOR_DATA (a
);
1034 if (ALLOCNO_UPDATED_HARD_REG_COSTS (a
) == NULL
1035 && ALLOCNO_CLASS_COST (a
) > ALLOCNO_MEMORY_COST (a
)
1036 /* Do not empty profitable regs for static chain pointer
1037 pseudo when non-local goto is used. */
1038 && ! non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a
)))
1039 CLEAR_HARD_REG_SET (data
->profitable_hard_regs
);
1042 mode
= ALLOCNO_MODE (a
);
1043 COPY_HARD_REG_SET (data
->profitable_hard_regs
,
1044 ira_useful_class_mode_regs
[aclass
][mode
]);
1045 nobj
= ALLOCNO_NUM_OBJECTS (a
);
1046 for (k
= 0; k
< nobj
; k
++)
1048 ira_object_t obj
= ALLOCNO_OBJECT (a
, k
);
1050 AND_COMPL_HARD_REG_SET (data
->profitable_hard_regs
,
1051 OBJECT_TOTAL_CONFLICT_HARD_REGS (obj
));
1055 /* Exclude hard regs already assigned for conflicting objects. */
1056 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap
, 0, i
, bi
)
1058 a
= ira_allocnos
[i
];
1059 if ((aclass
= ALLOCNO_CLASS (a
)) == NO_REGS
1060 || ! ALLOCNO_ASSIGNED_P (a
)
1061 || (hard_regno
= ALLOCNO_HARD_REGNO (a
)) < 0)
1063 mode
= ALLOCNO_MODE (a
);
1064 nregs
= hard_regno_nregs (hard_regno
, mode
);
1065 nobj
= ALLOCNO_NUM_OBJECTS (a
);
1066 for (k
= 0; k
< nobj
; k
++)
1068 ira_object_t obj
= ALLOCNO_OBJECT (a
, k
);
1069 ira_object_t conflict_obj
;
1070 ira_object_conflict_iterator oci
;
1072 FOR_EACH_OBJECT_CONFLICT (obj
, conflict_obj
, oci
)
1074 ira_allocno_t conflict_a
= OBJECT_ALLOCNO (conflict_obj
);
1076 /* We can process the conflict allocno repeatedly with
1078 if (nregs
== nobj
&& nregs
> 1)
1080 int num
= OBJECT_SUBWORD (conflict_obj
);
1082 if (REG_WORDS_BIG_ENDIAN
)
1084 (ALLOCNO_COLOR_DATA (conflict_a
)->profitable_hard_regs
,
1085 hard_regno
+ nobj
- num
- 1);
1088 (ALLOCNO_COLOR_DATA (conflict_a
)->profitable_hard_regs
,
1092 AND_COMPL_HARD_REG_SET
1093 (ALLOCNO_COLOR_DATA (conflict_a
)->profitable_hard_regs
,
1094 ira_reg_mode_hard_regset
[hard_regno
][mode
]);
1098 /* Exclude too costly hard regs. */
1099 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap
, 0, i
, bi
)
1101 int min_cost
= INT_MAX
;
1104 a
= ira_allocnos
[i
];
1105 if ((aclass
= ALLOCNO_CLASS (a
)) == NO_REGS
1106 || empty_profitable_hard_regs (a
))
1108 data
= ALLOCNO_COLOR_DATA (a
);
1109 mode
= ALLOCNO_MODE (a
);
1110 if ((costs
= ALLOCNO_UPDATED_HARD_REG_COSTS (a
)) != NULL
1111 || (costs
= ALLOCNO_HARD_REG_COSTS (a
)) != NULL
)
1113 class_size
= ira_class_hard_regs_num
[aclass
];
1114 for (j
= 0; j
< class_size
; j
++)
1116 hard_regno
= ira_class_hard_regs
[aclass
][j
];
1117 if (! TEST_HARD_REG_BIT (data
->profitable_hard_regs
,
1120 if (ALLOCNO_UPDATED_MEMORY_COST (a
) < costs
[j
]
1121 /* Do not remove HARD_REGNO for static chain pointer
1122 pseudo when non-local goto is used. */
1123 && ! non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a
)))
1124 CLEAR_HARD_REG_BIT (data
->profitable_hard_regs
,
1126 else if (min_cost
> costs
[j
])
1127 min_cost
= costs
[j
];
1130 else if (ALLOCNO_UPDATED_MEMORY_COST (a
)
1131 < ALLOCNO_UPDATED_CLASS_COST (a
)
1132 /* Do not empty profitable regs for static chain
1133 pointer pseudo when non-local goto is used. */
1134 && ! non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a
)))
1135 CLEAR_HARD_REG_SET (data
->profitable_hard_regs
);
1136 if (ALLOCNO_UPDATED_CLASS_COST (a
) > min_cost
)
1137 ALLOCNO_UPDATED_CLASS_COST (a
) = min_cost
;
1143 /* This page contains functions used to choose hard registers for
1146 /* Pool for update cost records. */
1147 static object_allocator
<update_cost_record
> update_cost_record_pool
1148 ("update cost records");
1150 /* Return new update cost record with given params. */
1151 static struct update_cost_record
*
1152 get_update_cost_record (int hard_regno
, int divisor
,
1153 struct update_cost_record
*next
)
1155 struct update_cost_record
*record
;
1157 record
= update_cost_record_pool
.allocate ();
1158 record
->hard_regno
= hard_regno
;
1159 record
->divisor
= divisor
;
1160 record
->next
= next
;
1164 /* Free memory for all records in LIST. */
1166 free_update_cost_record_list (struct update_cost_record
*list
)
1168 struct update_cost_record
*next
;
1170 while (list
!= NULL
)
1173 update_cost_record_pool
.remove (list
);
1178 /* Free memory allocated for all update cost records. */
1180 finish_update_cost_records (void)
1182 update_cost_record_pool
.release ();
1185 /* Array whose element value is TRUE if the corresponding hard
1186 register was already allocated for an allocno. */
1187 static bool allocated_hardreg_p
[FIRST_PSEUDO_REGISTER
];
1189 /* Describes one element in a queue of allocnos whose costs need to be
1190 updated. Each allocno in the queue is known to have an allocno
1192 struct update_cost_queue_elem
1194 /* This element is in the queue iff CHECK == update_cost_check. */
1197 /* COST_HOP_DIVISOR**N, where N is the length of the shortest path
1198 connecting this allocno to the one being allocated. */
1201 /* Allocno from which we are chaining costs of connected allocnos.
1202 It is used not go back in graph of allocnos connected by
1206 /* The next allocno in the queue, or null if this is the last element. */
1210 /* The first element in a queue of allocnos whose copy costs need to be
1211 updated. Null if the queue is empty. */
1212 static ira_allocno_t update_cost_queue
;
1214 /* The last element in the queue described by update_cost_queue.
1215 Not valid if update_cost_queue is null. */
1216 static struct update_cost_queue_elem
*update_cost_queue_tail
;
1218 /* A pool of elements in the queue described by update_cost_queue.
1219 Elements are indexed by ALLOCNO_NUM. */
1220 static struct update_cost_queue_elem
*update_cost_queue_elems
;
1222 /* The current value of update_costs_from_copies call count. */
1223 static int update_cost_check
;
1225 /* Allocate and initialize data necessary for function
1226 update_costs_from_copies. */
1228 initiate_cost_update (void)
1232 size
= ira_allocnos_num
* sizeof (struct update_cost_queue_elem
);
1233 update_cost_queue_elems
1234 = (struct update_cost_queue_elem
*) ira_allocate (size
);
1235 memset (update_cost_queue_elems
, 0, size
);
1236 update_cost_check
= 0;
1239 /* Deallocate data used by function update_costs_from_copies. */
1241 finish_cost_update (void)
1243 ira_free (update_cost_queue_elems
);
1244 finish_update_cost_records ();
1247 /* When we traverse allocnos to update hard register costs, the cost
1248 divisor will be multiplied by the following macro value for each
1249 hop from given allocno to directly connected allocnos. */
1250 #define COST_HOP_DIVISOR 4
1252 /* Start a new cost-updating pass. */
1254 start_update_cost (void)
1256 update_cost_check
++;
1257 update_cost_queue
= NULL
;
1260 /* Add (ALLOCNO, FROM, DIVISOR) to the end of update_cost_queue, unless
1261 ALLOCNO is already in the queue, or has NO_REGS class. */
1263 queue_update_cost (ira_allocno_t allocno
, ira_allocno_t from
, int divisor
)
1265 struct update_cost_queue_elem
*elem
;
1267 elem
= &update_cost_queue_elems
[ALLOCNO_NUM (allocno
)];
1268 if (elem
->check
!= update_cost_check
1269 && ALLOCNO_CLASS (allocno
) != NO_REGS
)
1271 elem
->check
= update_cost_check
;
1273 elem
->divisor
= divisor
;
1275 if (update_cost_queue
== NULL
)
1276 update_cost_queue
= allocno
;
1278 update_cost_queue_tail
->next
= allocno
;
1279 update_cost_queue_tail
= elem
;
1283 /* Try to remove the first element from update_cost_queue. Return
1284 false if the queue was empty, otherwise make (*ALLOCNO, *FROM,
1285 *DIVISOR) describe the removed element. */
1287 get_next_update_cost (ira_allocno_t
*allocno
, ira_allocno_t
*from
, int *divisor
)
1289 struct update_cost_queue_elem
*elem
;
1291 if (update_cost_queue
== NULL
)
1294 *allocno
= update_cost_queue
;
1295 elem
= &update_cost_queue_elems
[ALLOCNO_NUM (*allocno
)];
1297 *divisor
= elem
->divisor
;
1298 update_cost_queue
= elem
->next
;
1302 /* Increase costs of HARD_REGNO by UPDATE_COST and conflict cost by
1303 UPDATE_CONFLICT_COST for ALLOCNO. Return true if we really
1304 modified the cost. */
1306 update_allocno_cost (ira_allocno_t allocno
, int hard_regno
,
1307 int update_cost
, int update_conflict_cost
)
1310 enum reg_class aclass
= ALLOCNO_CLASS (allocno
);
1312 i
= ira_class_hard_reg_index
[aclass
][hard_regno
];
1315 ira_allocate_and_set_or_copy_costs
1316 (&ALLOCNO_UPDATED_HARD_REG_COSTS (allocno
), aclass
,
1317 ALLOCNO_UPDATED_CLASS_COST (allocno
),
1318 ALLOCNO_HARD_REG_COSTS (allocno
));
1319 ira_allocate_and_set_or_copy_costs
1320 (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (allocno
),
1321 aclass
, 0, ALLOCNO_CONFLICT_HARD_REG_COSTS (allocno
));
1322 ALLOCNO_UPDATED_HARD_REG_COSTS (allocno
)[i
] += update_cost
;
1323 ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (allocno
)[i
] += update_conflict_cost
;
1327 /* Update (decrease if DECR_P) HARD_REGNO cost of allocnos connected
1328 by copies to ALLOCNO to increase chances to remove some copies as
1329 the result of subsequent assignment. Record cost updates if
1330 RECORD_P is true. */
1332 update_costs_from_allocno (ira_allocno_t allocno
, int hard_regno
,
1333 int divisor
, bool decr_p
, bool record_p
)
1335 int cost
, update_cost
, update_conflict_cost
;
1337 enum reg_class rclass
, aclass
;
1338 ira_allocno_t another_allocno
, from
= NULL
;
1339 ira_copy_t cp
, next_cp
;
1341 rclass
= REGNO_REG_CLASS (hard_regno
);
1344 mode
= ALLOCNO_MODE (allocno
);
1345 ira_init_register_move_cost_if_necessary (mode
);
1346 for (cp
= ALLOCNO_COPIES (allocno
); cp
!= NULL
; cp
= next_cp
)
1348 if (cp
->first
== allocno
)
1350 next_cp
= cp
->next_first_allocno_copy
;
1351 another_allocno
= cp
->second
;
1353 else if (cp
->second
== allocno
)
1355 next_cp
= cp
->next_second_allocno_copy
;
1356 another_allocno
= cp
->first
;
1361 if (another_allocno
== from
)
1364 aclass
= ALLOCNO_CLASS (another_allocno
);
1365 if (! TEST_HARD_REG_BIT (reg_class_contents
[aclass
],
1367 || ALLOCNO_ASSIGNED_P (another_allocno
))
1370 if (GET_MODE_SIZE (ALLOCNO_MODE (cp
->second
)) < GET_MODE_SIZE (mode
))
1371 /* If we have different modes use the smallest one. It is
1372 a sub-register move. It is hard to predict what LRA
1373 will reload (the pseudo or its sub-register) but LRA
1374 will try to minimize the data movement. Also for some
1375 register classes bigger modes might be invalid,
1376 e.g. DImode for AREG on x86. For such cases the
1377 register move cost will be maximal. */
1378 mode
= ALLOCNO_MODE (cp
->second
);
1380 cost
= (cp
->second
== allocno
1381 ? ira_register_move_cost
[mode
][rclass
][aclass
]
1382 : ira_register_move_cost
[mode
][aclass
][rclass
]);
1386 update_conflict_cost
= update_cost
= cp
->freq
* cost
/ divisor
;
1388 if (ALLOCNO_COLOR_DATA (another_allocno
) != NULL
1389 && (ALLOCNO_COLOR_DATA (allocno
)->first_thread_allocno
1390 != ALLOCNO_COLOR_DATA (another_allocno
)->first_thread_allocno
))
1391 /* Decrease conflict cost of ANOTHER_ALLOCNO if it is not
1392 in the same allocation thread. */
1393 update_conflict_cost
/= COST_HOP_DIVISOR
;
1395 if (update_cost
== 0)
1398 if (! update_allocno_cost (another_allocno
, hard_regno
,
1399 update_cost
, update_conflict_cost
))
1401 queue_update_cost (another_allocno
, allocno
, divisor
* COST_HOP_DIVISOR
);
1402 if (record_p
&& ALLOCNO_COLOR_DATA (another_allocno
) != NULL
)
1403 ALLOCNO_COLOR_DATA (another_allocno
)->update_cost_records
1404 = get_update_cost_record (hard_regno
, divisor
,
1405 ALLOCNO_COLOR_DATA (another_allocno
)
1406 ->update_cost_records
);
1409 while (get_next_update_cost (&allocno
, &from
, &divisor
));
1412 /* Decrease preferred ALLOCNO hard register costs and costs of
1413 allocnos connected to ALLOCNO through copy. */
1415 update_costs_from_prefs (ira_allocno_t allocno
)
1419 start_update_cost ();
1420 for (pref
= ALLOCNO_PREFS (allocno
); pref
!= NULL
; pref
= pref
->next_pref
)
1421 update_costs_from_allocno (allocno
, pref
->hard_regno
,
1422 COST_HOP_DIVISOR
, true, true);
1425 /* Update (decrease if DECR_P) the cost of allocnos connected to
1426 ALLOCNO through copies to increase chances to remove some copies as
1427 the result of subsequent assignment. ALLOCNO was just assigned to
1428 a hard register. Record cost updates if RECORD_P is true. */
1430 update_costs_from_copies (ira_allocno_t allocno
, bool decr_p
, bool record_p
)
1434 hard_regno
= ALLOCNO_HARD_REGNO (allocno
);
1435 ira_assert (hard_regno
>= 0 && ALLOCNO_CLASS (allocno
) != NO_REGS
);
1436 start_update_cost ();
1437 update_costs_from_allocno (allocno
, hard_regno
, 1, decr_p
, record_p
);
1440 /* Restore costs of allocnos connected to ALLOCNO by copies as it was
1441 before updating costs of these allocnos from given allocno. This
1442 is a wise thing to do as if given allocno did not get an expected
1443 hard reg, using smaller cost of the hard reg for allocnos connected
1444 by copies to given allocno becomes actually misleading. Free all
1445 update cost records for ALLOCNO as we don't need them anymore. */
1447 restore_costs_from_copies (ira_allocno_t allocno
)
1449 struct update_cost_record
*records
, *curr
;
1451 if (ALLOCNO_COLOR_DATA (allocno
) == NULL
)
1453 records
= ALLOCNO_COLOR_DATA (allocno
)->update_cost_records
;
1454 start_update_cost ();
1455 for (curr
= records
; curr
!= NULL
; curr
= curr
->next
)
1456 update_costs_from_allocno (allocno
, curr
->hard_regno
,
1457 curr
->divisor
, true, false);
1458 free_update_cost_record_list (records
);
1459 ALLOCNO_COLOR_DATA (allocno
)->update_cost_records
= NULL
;
1462 /* This function updates COSTS (decrease if DECR_P) for hard_registers
1463 of ACLASS by conflict costs of the unassigned allocnos
1464 connected by copies with allocnos in update_cost_queue. This
1465 update increases chances to remove some copies. */
1467 update_conflict_hard_regno_costs (int *costs
, enum reg_class aclass
,
1470 int i
, cost
, class_size
, freq
, mult
, div
, divisor
;
1471 int index
, hard_regno
;
1472 int *conflict_costs
;
1474 enum reg_class another_aclass
;
1475 ira_allocno_t allocno
, another_allocno
, from
;
1476 ira_copy_t cp
, next_cp
;
1478 while (get_next_update_cost (&allocno
, &from
, &divisor
))
1479 for (cp
= ALLOCNO_COPIES (allocno
); cp
!= NULL
; cp
= next_cp
)
1481 if (cp
->first
== allocno
)
1483 next_cp
= cp
->next_first_allocno_copy
;
1484 another_allocno
= cp
->second
;
1486 else if (cp
->second
== allocno
)
1488 next_cp
= cp
->next_second_allocno_copy
;
1489 another_allocno
= cp
->first
;
1494 if (another_allocno
== from
)
1497 another_aclass
= ALLOCNO_CLASS (another_allocno
);
1498 if (! ira_reg_classes_intersect_p
[aclass
][another_aclass
]
1499 || ALLOCNO_ASSIGNED_P (another_allocno
)
1500 || ALLOCNO_COLOR_DATA (another_allocno
)->may_be_spilled_p
)
1502 class_size
= ira_class_hard_regs_num
[another_aclass
];
1503 ira_allocate_and_copy_costs
1504 (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (another_allocno
),
1505 another_aclass
, ALLOCNO_CONFLICT_HARD_REG_COSTS (another_allocno
));
1507 = ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (another_allocno
);
1508 if (conflict_costs
== NULL
)
1513 freq
= ALLOCNO_FREQ (another_allocno
);
1516 div
= freq
* divisor
;
1518 for (i
= class_size
- 1; i
>= 0; i
--)
1520 hard_regno
= ira_class_hard_regs
[another_aclass
][i
];
1521 ira_assert (hard_regno
>= 0);
1522 index
= ira_class_hard_reg_index
[aclass
][hard_regno
];
1525 cost
= (int) (((int64_t) conflict_costs
[i
] * mult
) / div
);
1531 costs
[index
] += cost
;
1534 /* Probably 5 hops will be enough. */
1536 && divisor
<= (COST_HOP_DIVISOR
1539 * COST_HOP_DIVISOR
))
1540 queue_update_cost (another_allocno
, allocno
, divisor
* COST_HOP_DIVISOR
);
1544 /* Set up conflicting (through CONFLICT_REGS) for each object of
1545 allocno A and the start allocno profitable regs (through
1546 START_PROFITABLE_REGS). Remember that the start profitable regs
1547 exclude hard regs which can not hold value of mode of allocno A.
1548 This covers mostly cases when multi-register value should be
1551 get_conflict_and_start_profitable_regs (ira_allocno_t a
, bool retry_p
,
1552 HARD_REG_SET
*conflict_regs
,
1553 HARD_REG_SET
*start_profitable_regs
)
1558 nwords
= ALLOCNO_NUM_OBJECTS (a
);
1559 for (i
= 0; i
< nwords
; i
++)
1561 obj
= ALLOCNO_OBJECT (a
, i
);
1562 COPY_HARD_REG_SET (conflict_regs
[i
],
1563 OBJECT_TOTAL_CONFLICT_HARD_REGS (obj
));
1567 COPY_HARD_REG_SET (*start_profitable_regs
,
1568 reg_class_contents
[ALLOCNO_CLASS (a
)]);
1569 AND_COMPL_HARD_REG_SET (*start_profitable_regs
,
1570 ira_prohibited_class_mode_regs
1571 [ALLOCNO_CLASS (a
)][ALLOCNO_MODE (a
)]);
1574 COPY_HARD_REG_SET (*start_profitable_regs
,
1575 ALLOCNO_COLOR_DATA (a
)->profitable_hard_regs
);
1578 /* Return true if HARD_REGNO is ok for assigning to allocno A with
1579 PROFITABLE_REGS and whose objects have CONFLICT_REGS. */
1581 check_hard_reg_p (ira_allocno_t a
, int hard_regno
,
1582 HARD_REG_SET
*conflict_regs
, HARD_REG_SET profitable_regs
)
1584 int j
, nwords
, nregs
;
1585 enum reg_class aclass
;
1588 aclass
= ALLOCNO_CLASS (a
);
1589 mode
= ALLOCNO_MODE (a
);
1590 if (TEST_HARD_REG_BIT (ira_prohibited_class_mode_regs
[aclass
][mode
],
1593 /* Checking only profitable hard regs. */
1594 if (! TEST_HARD_REG_BIT (profitable_regs
, hard_regno
))
1596 nregs
= hard_regno_nregs (hard_regno
, mode
);
1597 nwords
= ALLOCNO_NUM_OBJECTS (a
);
1598 for (j
= 0; j
< nregs
; j
++)
1601 int set_to_test_start
= 0, set_to_test_end
= nwords
;
1603 if (nregs
== nwords
)
1605 if (REG_WORDS_BIG_ENDIAN
)
1606 set_to_test_start
= nwords
- j
- 1;
1608 set_to_test_start
= j
;
1609 set_to_test_end
= set_to_test_start
+ 1;
1611 for (k
= set_to_test_start
; k
< set_to_test_end
; k
++)
1612 if (TEST_HARD_REG_BIT (conflict_regs
[k
], hard_regno
+ j
))
1614 if (k
!= set_to_test_end
)
1620 /* Return number of registers needed to be saved and restored at
1621 function prologue/epilogue if we allocate HARD_REGNO to hold value
1624 calculate_saved_nregs (int hard_regno
, machine_mode mode
)
1629 ira_assert (hard_regno
>= 0);
1630 for (i
= hard_regno_nregs (hard_regno
, mode
) - 1; i
>= 0; i
--)
1631 if (!allocated_hardreg_p
[hard_regno
+ i
]
1632 && !TEST_HARD_REG_BIT (call_used_reg_set
, hard_regno
+ i
)
1633 && !LOCAL_REGNO (hard_regno
+ i
))
1638 /* Choose a hard register for allocno A. If RETRY_P is TRUE, it means
1639 that the function called from function
1640 `ira_reassign_conflict_allocnos' and `allocno_reload_assign'. In
1641 this case some allocno data are not defined or updated and we
1642 should not touch these data. The function returns true if we
1643 managed to assign a hard register to the allocno.
1645 To assign a hard register, first of all we calculate all conflict
1646 hard registers which can come from conflicting allocnos with
1647 already assigned hard registers. After that we find first free
1648 hard register with the minimal cost. During hard register cost
1649 calculation we take conflict hard register costs into account to
1650 give a chance for conflicting allocnos to get a better hard
1651 register in the future.
1653 If the best hard register cost is bigger than cost of memory usage
1654 for the allocno, we don't assign a hard register to given allocno
1657 If we assign a hard register to the allocno, we update costs of the
1658 hard register for allocnos connected by copies to improve a chance
1659 to coalesce insns represented by the copies when we assign hard
1660 registers to the allocnos connected by the copies. */
1662 assign_hard_reg (ira_allocno_t a
, bool retry_p
)
1664 HARD_REG_SET conflicting_regs
[2], profitable_hard_regs
;
1665 int i
, j
, hard_regno
, best_hard_regno
, class_size
;
1666 int cost
, mem_cost
, min_cost
, full_cost
, min_full_cost
, nwords
, word
;
1668 enum reg_class aclass
;
1670 static int costs
[FIRST_PSEUDO_REGISTER
], full_costs
[FIRST_PSEUDO_REGISTER
];
1672 enum reg_class rclass
;
1675 bool no_stack_reg_p
;
1678 ira_assert (! ALLOCNO_ASSIGNED_P (a
));
1679 get_conflict_and_start_profitable_regs (a
, retry_p
,
1681 &profitable_hard_regs
);
1682 aclass
= ALLOCNO_CLASS (a
);
1683 class_size
= ira_class_hard_regs_num
[aclass
];
1684 best_hard_regno
= -1;
1685 memset (full_costs
, 0, sizeof (int) * class_size
);
1687 memset (costs
, 0, sizeof (int) * class_size
);
1688 memset (full_costs
, 0, sizeof (int) * class_size
);
1690 no_stack_reg_p
= false;
1693 start_update_cost ();
1694 mem_cost
+= ALLOCNO_UPDATED_MEMORY_COST (a
);
1696 ira_allocate_and_copy_costs (&ALLOCNO_UPDATED_HARD_REG_COSTS (a
),
1697 aclass
, ALLOCNO_HARD_REG_COSTS (a
));
1698 a_costs
= ALLOCNO_UPDATED_HARD_REG_COSTS (a
);
1700 no_stack_reg_p
= no_stack_reg_p
|| ALLOCNO_TOTAL_NO_STACK_REG_P (a
);
1702 cost
= ALLOCNO_UPDATED_CLASS_COST (a
);
1703 for (i
= 0; i
< class_size
; i
++)
1704 if (a_costs
!= NULL
)
1706 costs
[i
] += a_costs
[i
];
1707 full_costs
[i
] += a_costs
[i
];
1712 full_costs
[i
] += cost
;
1714 nwords
= ALLOCNO_NUM_OBJECTS (a
);
1715 curr_allocno_process
++;
1716 for (word
= 0; word
< nwords
; word
++)
1718 ira_object_t conflict_obj
;
1719 ira_object_t obj
= ALLOCNO_OBJECT (a
, word
);
1720 ira_object_conflict_iterator oci
;
1722 /* Take preferences of conflicting allocnos into account. */
1723 FOR_EACH_OBJECT_CONFLICT (obj
, conflict_obj
, oci
)
1725 ira_allocno_t conflict_a
= OBJECT_ALLOCNO (conflict_obj
);
1726 enum reg_class conflict_aclass
;
1727 allocno_color_data_t data
= ALLOCNO_COLOR_DATA (conflict_a
);
1729 /* Reload can give another class so we need to check all
1732 && ((!ALLOCNO_ASSIGNED_P (conflict_a
)
1733 || ALLOCNO_HARD_REGNO (conflict_a
) < 0)
1734 && !(hard_reg_set_intersect_p
1735 (profitable_hard_regs
,
1737 (conflict_a
)->profitable_hard_regs
))))
1739 /* All conflict allocnos are in consideration bitmap
1740 when retry_p is false. It might change in future and
1741 if it happens the assert will be broken. It means
1742 the code should be modified for the new
1744 ira_assert (bitmap_bit_p (consideration_allocno_bitmap
,
1745 ALLOCNO_NUM (conflict_a
)));
1748 conflict_aclass
= ALLOCNO_CLASS (conflict_a
);
1749 ira_assert (ira_reg_classes_intersect_p
1750 [aclass
][conflict_aclass
]);
1751 if (ALLOCNO_ASSIGNED_P (conflict_a
))
1753 hard_regno
= ALLOCNO_HARD_REGNO (conflict_a
);
1755 && (ira_hard_reg_set_intersection_p
1756 (hard_regno
, ALLOCNO_MODE (conflict_a
),
1757 reg_class_contents
[aclass
])))
1759 int n_objects
= ALLOCNO_NUM_OBJECTS (conflict_a
);
1762 mode
= ALLOCNO_MODE (conflict_a
);
1763 conflict_nregs
= hard_regno_nregs (hard_regno
, mode
);
1764 if (conflict_nregs
== n_objects
&& conflict_nregs
> 1)
1766 int num
= OBJECT_SUBWORD (conflict_obj
);
1768 if (REG_WORDS_BIG_ENDIAN
)
1769 SET_HARD_REG_BIT (conflicting_regs
[word
],
1770 hard_regno
+ n_objects
- num
- 1);
1772 SET_HARD_REG_BIT (conflicting_regs
[word
],
1777 (conflicting_regs
[word
],
1778 ira_reg_mode_hard_regset
[hard_regno
][mode
]);
1779 if (hard_reg_set_subset_p (profitable_hard_regs
,
1780 conflicting_regs
[word
]))
1785 && ! ALLOCNO_COLOR_DATA (conflict_a
)->may_be_spilled_p
1786 /* Don't process the conflict allocno twice. */
1787 && (ALLOCNO_COLOR_DATA (conflict_a
)->last_process
1788 != curr_allocno_process
))
1790 int k
, *conflict_costs
;
1792 ALLOCNO_COLOR_DATA (conflict_a
)->last_process
1793 = curr_allocno_process
;
1794 ira_allocate_and_copy_costs
1795 (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (conflict_a
),
1797 ALLOCNO_CONFLICT_HARD_REG_COSTS (conflict_a
));
1799 = ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (conflict_a
);
1800 if (conflict_costs
!= NULL
)
1801 for (j
= class_size
- 1; j
>= 0; j
--)
1803 hard_regno
= ira_class_hard_regs
[aclass
][j
];
1804 ira_assert (hard_regno
>= 0);
1805 k
= ira_class_hard_reg_index
[conflict_aclass
][hard_regno
];
1807 /* If HARD_REGNO is not available for CONFLICT_A,
1808 the conflict would be ignored, since HARD_REGNO
1809 will never be assigned to CONFLICT_A. */
1810 || !TEST_HARD_REG_BIT (data
->profitable_hard_regs
,
1813 full_costs
[j
] -= conflict_costs
[k
];
1815 queue_update_cost (conflict_a
, NULL
, COST_HOP_DIVISOR
);
1821 /* Take into account preferences of allocnos connected by copies to
1822 the conflict allocnos. */
1823 update_conflict_hard_regno_costs (full_costs
, aclass
, true);
1825 /* Take preferences of allocnos connected by copies into
1829 start_update_cost ();
1830 queue_update_cost (a
, NULL
, COST_HOP_DIVISOR
);
1831 update_conflict_hard_regno_costs (full_costs
, aclass
, false);
1833 min_cost
= min_full_cost
= INT_MAX
;
1834 /* We don't care about giving callee saved registers to allocnos no
1835 living through calls because call clobbered registers are
1836 allocated first (it is usual practice to put them first in
1837 REG_ALLOC_ORDER). */
1838 mode
= ALLOCNO_MODE (a
);
1839 for (i
= 0; i
< class_size
; i
++)
1841 hard_regno
= ira_class_hard_regs
[aclass
][i
];
1844 && FIRST_STACK_REG
<= hard_regno
&& hard_regno
<= LAST_STACK_REG
)
1847 if (! check_hard_reg_p (a
, hard_regno
,
1848 conflicting_regs
, profitable_hard_regs
))
1851 full_cost
= full_costs
[i
];
1852 if (!HONOR_REG_ALLOC_ORDER
)
1854 if ((saved_nregs
= calculate_saved_nregs (hard_regno
, mode
)) != 0)
1855 /* We need to save/restore the hard register in
1856 epilogue/prologue. Therefore we increase the cost. */
1858 rclass
= REGNO_REG_CLASS (hard_regno
);
1859 add_cost
= ((ira_memory_move_cost
[mode
][rclass
][0]
1860 + ira_memory_move_cost
[mode
][rclass
][1])
1861 * saved_nregs
/ hard_regno_nregs (hard_regno
,
1864 full_cost
+= add_cost
;
1867 if (min_cost
> cost
)
1869 if (min_full_cost
> full_cost
)
1871 min_full_cost
= full_cost
;
1872 best_hard_regno
= hard_regno
;
1873 ira_assert (hard_regno
>= 0);
1876 if (min_full_cost
> mem_cost
1877 /* Do not spill static chain pointer pseudo when non-local goto
1879 && ! non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a
)))
1881 if (! retry_p
&& internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
1882 fprintf (ira_dump_file
, "(memory is more profitable %d vs %d) ",
1883 mem_cost
, min_full_cost
);
1884 best_hard_regno
= -1;
1887 if (best_hard_regno
>= 0)
1889 for (i
= hard_regno_nregs (best_hard_regno
, mode
) - 1; i
>= 0; i
--)
1890 allocated_hardreg_p
[best_hard_regno
+ i
] = true;
1893 restore_costs_from_copies (a
);
1894 ALLOCNO_HARD_REGNO (a
) = best_hard_regno
;
1895 ALLOCNO_ASSIGNED_P (a
) = true;
1896 if (best_hard_regno
>= 0)
1897 update_costs_from_copies (a
, true, ! retry_p
);
1898 ira_assert (ALLOCNO_CLASS (a
) == aclass
);
1899 /* We don't need updated costs anymore. */
1900 ira_free_allocno_updated_costs (a
);
1901 return best_hard_regno
>= 0;
1906 /* An array used to sort copies. */
1907 static ira_copy_t
*sorted_copies
;
1909 /* Return TRUE if live ranges of allocnos A1 and A2 intersect. It is
1910 used to find a conflict for new allocnos or allocnos with the
1911 different allocno classes. */
1913 allocnos_conflict_by_live_ranges_p (ira_allocno_t a1
, ira_allocno_t a2
)
1917 int n1
= ALLOCNO_NUM_OBJECTS (a1
);
1918 int n2
= ALLOCNO_NUM_OBJECTS (a2
);
1922 reg1
= regno_reg_rtx
[ALLOCNO_REGNO (a1
)];
1923 reg2
= regno_reg_rtx
[ALLOCNO_REGNO (a2
)];
1924 if (reg1
!= NULL
&& reg2
!= NULL
1925 && ORIGINAL_REGNO (reg1
) == ORIGINAL_REGNO (reg2
))
1928 for (i
= 0; i
< n1
; i
++)
1930 ira_object_t c1
= ALLOCNO_OBJECT (a1
, i
);
1932 for (j
= 0; j
< n2
; j
++)
1934 ira_object_t c2
= ALLOCNO_OBJECT (a2
, j
);
1936 if (ira_live_ranges_intersect_p (OBJECT_LIVE_RANGES (c1
),
1937 OBJECT_LIVE_RANGES (c2
)))
1944 /* The function is used to sort copies according to their execution
1947 copy_freq_compare_func (const void *v1p
, const void *v2p
)
1949 ira_copy_t cp1
= *(const ira_copy_t
*) v1p
, cp2
= *(const ira_copy_t
*) v2p
;
1957 /* If frequencies are equal, sort by copies, so that the results of
1958 qsort leave nothing to chance. */
1959 return cp1
->num
- cp2
->num
;
1964 /* Return true if any allocno from thread of A1 conflicts with any
1965 allocno from thread A2. */
1967 allocno_thread_conflict_p (ira_allocno_t a1
, ira_allocno_t a2
)
1969 ira_allocno_t a
, conflict_a
;
1971 for (a
= ALLOCNO_COLOR_DATA (a2
)->next_thread_allocno
;;
1972 a
= ALLOCNO_COLOR_DATA (a
)->next_thread_allocno
)
1974 for (conflict_a
= ALLOCNO_COLOR_DATA (a1
)->next_thread_allocno
;;
1975 conflict_a
= ALLOCNO_COLOR_DATA (conflict_a
)->next_thread_allocno
)
1977 if (allocnos_conflict_by_live_ranges_p (a
, conflict_a
))
1979 if (conflict_a
== a1
)
1988 /* Merge two threads given correspondingly by their first allocnos T1
1989 and T2 (more accurately merging T2 into T1). */
1991 merge_threads (ira_allocno_t t1
, ira_allocno_t t2
)
1993 ira_allocno_t a
, next
, last
;
1995 gcc_assert (t1
!= t2
1996 && ALLOCNO_COLOR_DATA (t1
)->first_thread_allocno
== t1
1997 && ALLOCNO_COLOR_DATA (t2
)->first_thread_allocno
== t2
);
1998 for (last
= t2
, a
= ALLOCNO_COLOR_DATA (t2
)->next_thread_allocno
;;
1999 a
= ALLOCNO_COLOR_DATA (a
)->next_thread_allocno
)
2001 ALLOCNO_COLOR_DATA (a
)->first_thread_allocno
= t1
;
2006 next
= ALLOCNO_COLOR_DATA (t1
)->next_thread_allocno
;
2007 ALLOCNO_COLOR_DATA (t1
)->next_thread_allocno
= t2
;
2008 ALLOCNO_COLOR_DATA (last
)->next_thread_allocno
= next
;
2009 ALLOCNO_COLOR_DATA (t1
)->thread_freq
+= ALLOCNO_COLOR_DATA (t2
)->thread_freq
;
2012 /* Create threads by processing CP_NUM copies from sorted copies. We
2013 process the most expensive copies first. */
2015 form_threads_from_copies (int cp_num
)
2017 ira_allocno_t a
, thread1
, thread2
;
2021 qsort (sorted_copies
, cp_num
, sizeof (ira_copy_t
), copy_freq_compare_func
);
2022 /* Form threads processing copies, most frequently executed
2024 for (; cp_num
!= 0;)
2026 for (i
= 0; i
< cp_num
; i
++)
2028 cp
= sorted_copies
[i
];
2029 thread1
= ALLOCNO_COLOR_DATA (cp
->first
)->first_thread_allocno
;
2030 thread2
= ALLOCNO_COLOR_DATA (cp
->second
)->first_thread_allocno
;
2031 if (thread1
== thread2
)
2033 if (! allocno_thread_conflict_p (thread1
, thread2
))
2035 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
2038 " Forming thread by copy %d:a%dr%d-a%dr%d (freq=%d):\n",
2039 cp
->num
, ALLOCNO_NUM (cp
->first
), ALLOCNO_REGNO (cp
->first
),
2040 ALLOCNO_NUM (cp
->second
), ALLOCNO_REGNO (cp
->second
),
2042 merge_threads (thread1
, thread2
);
2043 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
2045 thread1
= ALLOCNO_COLOR_DATA (thread1
)->first_thread_allocno
;
2046 fprintf (ira_dump_file
, " Result (freq=%d): a%dr%d(%d)",
2047 ALLOCNO_COLOR_DATA (thread1
)->thread_freq
,
2048 ALLOCNO_NUM (thread1
), ALLOCNO_REGNO (thread1
),
2049 ALLOCNO_FREQ (thread1
));
2050 for (a
= ALLOCNO_COLOR_DATA (thread1
)->next_thread_allocno
;
2052 a
= ALLOCNO_COLOR_DATA (a
)->next_thread_allocno
)
2053 fprintf (ira_dump_file
, " a%dr%d(%d)",
2054 ALLOCNO_NUM (a
), ALLOCNO_REGNO (a
),
2056 fprintf (ira_dump_file
, "\n");
2062 /* Collect the rest of copies. */
2063 for (n
= 0; i
< cp_num
; i
++)
2065 cp
= sorted_copies
[i
];
2066 if (ALLOCNO_COLOR_DATA (cp
->first
)->first_thread_allocno
2067 != ALLOCNO_COLOR_DATA (cp
->second
)->first_thread_allocno
)
2068 sorted_copies
[n
++] = cp
;
2074 /* Create threads by processing copies of all alocnos from BUCKET. We
2075 process the most expensive copies first. */
2077 form_threads_from_bucket (ira_allocno_t bucket
)
2080 ira_copy_t cp
, next_cp
;
2083 for (a
= bucket
; a
!= NULL
; a
= ALLOCNO_COLOR_DATA (a
)->next_bucket_allocno
)
2085 for (cp
= ALLOCNO_COPIES (a
); cp
!= NULL
; cp
= next_cp
)
2089 next_cp
= cp
->next_first_allocno_copy
;
2090 sorted_copies
[cp_num
++] = cp
;
2092 else if (cp
->second
== a
)
2093 next_cp
= cp
->next_second_allocno_copy
;
2098 form_threads_from_copies (cp_num
);
2101 /* Create threads by processing copies of colorable allocno A. We
2102 process most expensive copies first. */
2104 form_threads_from_colorable_allocno (ira_allocno_t a
)
2106 ira_allocno_t another_a
;
2107 ira_copy_t cp
, next_cp
;
2110 for (cp
= ALLOCNO_COPIES (a
); cp
!= NULL
; cp
= next_cp
)
2114 next_cp
= cp
->next_first_allocno_copy
;
2115 another_a
= cp
->second
;
2117 else if (cp
->second
== a
)
2119 next_cp
= cp
->next_second_allocno_copy
;
2120 another_a
= cp
->first
;
2124 if ((! ALLOCNO_COLOR_DATA (another_a
)->in_graph_p
2125 && !ALLOCNO_COLOR_DATA (another_a
)->may_be_spilled_p
)
2126 || ALLOCNO_COLOR_DATA (another_a
)->colorable_p
)
2127 sorted_copies
[cp_num
++] = cp
;
2129 form_threads_from_copies (cp_num
);
2132 /* Form initial threads which contain only one allocno. */
2134 init_allocno_threads (void)
2140 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap
, 0, j
, bi
)
2142 a
= ira_allocnos
[j
];
2143 /* Set up initial thread data: */
2144 ALLOCNO_COLOR_DATA (a
)->first_thread_allocno
2145 = ALLOCNO_COLOR_DATA (a
)->next_thread_allocno
= a
;
2146 ALLOCNO_COLOR_DATA (a
)->thread_freq
= ALLOCNO_FREQ (a
);
2152 /* This page contains the allocator based on the Chaitin-Briggs algorithm. */
2154 /* Bucket of allocnos that can colored currently without spilling. */
2155 static ira_allocno_t colorable_allocno_bucket
;
2157 /* Bucket of allocnos that might be not colored currently without
2159 static ira_allocno_t uncolorable_allocno_bucket
;
2161 /* The current number of allocnos in the uncolorable_bucket. */
2162 static int uncolorable_allocnos_num
;
2164 /* Return the current spill priority of allocno A. The less the
2165 number, the more preferable the allocno for spilling. */
2167 allocno_spill_priority (ira_allocno_t a
)
2169 allocno_color_data_t data
= ALLOCNO_COLOR_DATA (a
);
2172 / (ALLOCNO_EXCESS_PRESSURE_POINTS_NUM (a
)
2173 * ira_reg_class_max_nregs
[ALLOCNO_CLASS (a
)][ALLOCNO_MODE (a
)]
2177 /* Add allocno A to bucket *BUCKET_PTR. A should be not in a bucket
2180 add_allocno_to_bucket (ira_allocno_t a
, ira_allocno_t
*bucket_ptr
)
2182 ira_allocno_t first_a
;
2183 allocno_color_data_t data
;
2185 if (bucket_ptr
== &uncolorable_allocno_bucket
2186 && ALLOCNO_CLASS (a
) != NO_REGS
)
2188 uncolorable_allocnos_num
++;
2189 ira_assert (uncolorable_allocnos_num
> 0);
2191 first_a
= *bucket_ptr
;
2192 data
= ALLOCNO_COLOR_DATA (a
);
2193 data
->next_bucket_allocno
= first_a
;
2194 data
->prev_bucket_allocno
= NULL
;
2195 if (first_a
!= NULL
)
2196 ALLOCNO_COLOR_DATA (first_a
)->prev_bucket_allocno
= a
;
2200 /* Compare two allocnos to define which allocno should be pushed first
2201 into the coloring stack. If the return is a negative number, the
2202 allocno given by the first parameter will be pushed first. In this
2203 case such allocno has less priority than the second one and the
2204 hard register will be assigned to it after assignment to the second
2205 one. As the result of such assignment order, the second allocno
2206 has a better chance to get the best hard register. */
2208 bucket_allocno_compare_func (const void *v1p
, const void *v2p
)
2210 ira_allocno_t a1
= *(const ira_allocno_t
*) v1p
;
2211 ira_allocno_t a2
= *(const ira_allocno_t
*) v2p
;
2212 int diff
, freq1
, freq2
, a1_num
, a2_num
;
2213 ira_allocno_t t1
= ALLOCNO_COLOR_DATA (a1
)->first_thread_allocno
;
2214 ira_allocno_t t2
= ALLOCNO_COLOR_DATA (a2
)->first_thread_allocno
;
2215 int cl1
= ALLOCNO_CLASS (a1
), cl2
= ALLOCNO_CLASS (a2
);
2217 freq1
= ALLOCNO_COLOR_DATA (t1
)->thread_freq
;
2218 freq2
= ALLOCNO_COLOR_DATA (t2
)->thread_freq
;
2219 if ((diff
= freq1
- freq2
) != 0)
2222 if ((diff
= ALLOCNO_NUM (t2
) - ALLOCNO_NUM (t1
)) != 0)
2225 /* Push pseudos requiring less hard registers first. It means that
2226 we will assign pseudos requiring more hard registers first
2227 avoiding creation small holes in free hard register file into
2228 which the pseudos requiring more hard registers can not fit. */
2229 if ((diff
= (ira_reg_class_max_nregs
[cl1
][ALLOCNO_MODE (a1
)]
2230 - ira_reg_class_max_nregs
[cl2
][ALLOCNO_MODE (a2
)])) != 0)
2233 freq1
= ALLOCNO_FREQ (a1
);
2234 freq2
= ALLOCNO_FREQ (a2
);
2235 if ((diff
= freq1
- freq2
) != 0)
2238 a1_num
= ALLOCNO_COLOR_DATA (a1
)->available_regs_num
;
2239 a2_num
= ALLOCNO_COLOR_DATA (a2
)->available_regs_num
;
2240 if ((diff
= a2_num
- a1_num
) != 0)
2242 return ALLOCNO_NUM (a2
) - ALLOCNO_NUM (a1
);
2245 /* Sort bucket *BUCKET_PTR and return the result through
2248 sort_bucket (ira_allocno_t
*bucket_ptr
,
2249 int (*compare_func
) (const void *, const void *))
2251 ira_allocno_t a
, head
;
2254 for (n
= 0, a
= *bucket_ptr
;
2256 a
= ALLOCNO_COLOR_DATA (a
)->next_bucket_allocno
)
2257 sorted_allocnos
[n
++] = a
;
2260 qsort (sorted_allocnos
, n
, sizeof (ira_allocno_t
), compare_func
);
2262 for (n
--; n
>= 0; n
--)
2264 a
= sorted_allocnos
[n
];
2265 ALLOCNO_COLOR_DATA (a
)->next_bucket_allocno
= head
;
2266 ALLOCNO_COLOR_DATA (a
)->prev_bucket_allocno
= NULL
;
2268 ALLOCNO_COLOR_DATA (head
)->prev_bucket_allocno
= a
;
2274 /* Add ALLOCNO to colorable bucket maintaining the order according
2275 their priority. ALLOCNO should be not in a bucket before the
2278 add_allocno_to_ordered_colorable_bucket (ira_allocno_t allocno
)
2280 ira_allocno_t before
, after
;
2282 form_threads_from_colorable_allocno (allocno
);
2283 for (before
= colorable_allocno_bucket
, after
= NULL
;
2286 before
= ALLOCNO_COLOR_DATA (before
)->next_bucket_allocno
)
2287 if (bucket_allocno_compare_func (&allocno
, &before
) < 0)
2289 ALLOCNO_COLOR_DATA (allocno
)->next_bucket_allocno
= before
;
2290 ALLOCNO_COLOR_DATA (allocno
)->prev_bucket_allocno
= after
;
2292 colorable_allocno_bucket
= allocno
;
2294 ALLOCNO_COLOR_DATA (after
)->next_bucket_allocno
= allocno
;
2296 ALLOCNO_COLOR_DATA (before
)->prev_bucket_allocno
= allocno
;
2299 /* Delete ALLOCNO from bucket *BUCKET_PTR. It should be there before
2302 delete_allocno_from_bucket (ira_allocno_t allocno
, ira_allocno_t
*bucket_ptr
)
2304 ira_allocno_t prev_allocno
, next_allocno
;
2306 if (bucket_ptr
== &uncolorable_allocno_bucket
2307 && ALLOCNO_CLASS (allocno
) != NO_REGS
)
2309 uncolorable_allocnos_num
--;
2310 ira_assert (uncolorable_allocnos_num
>= 0);
2312 prev_allocno
= ALLOCNO_COLOR_DATA (allocno
)->prev_bucket_allocno
;
2313 next_allocno
= ALLOCNO_COLOR_DATA (allocno
)->next_bucket_allocno
;
2314 if (prev_allocno
!= NULL
)
2315 ALLOCNO_COLOR_DATA (prev_allocno
)->next_bucket_allocno
= next_allocno
;
2318 ira_assert (*bucket_ptr
== allocno
);
2319 *bucket_ptr
= next_allocno
;
2321 if (next_allocno
!= NULL
)
2322 ALLOCNO_COLOR_DATA (next_allocno
)->prev_bucket_allocno
= prev_allocno
;
2325 /* Put allocno A onto the coloring stack without removing it from its
2326 bucket. Pushing allocno to the coloring stack can result in moving
2327 conflicting allocnos from the uncolorable bucket to the colorable
2330 push_allocno_to_stack (ira_allocno_t a
)
2332 enum reg_class aclass
;
2333 allocno_color_data_t data
, conflict_data
;
2334 int size
, i
, n
= ALLOCNO_NUM_OBJECTS (a
);
2336 data
= ALLOCNO_COLOR_DATA (a
);
2337 data
->in_graph_p
= false;
2338 allocno_stack_vec
.safe_push (a
);
2339 aclass
= ALLOCNO_CLASS (a
);
2340 if (aclass
== NO_REGS
)
2342 size
= ira_reg_class_max_nregs
[aclass
][ALLOCNO_MODE (a
)];
2345 /* We will deal with the subwords individually. */
2346 gcc_assert (size
== ALLOCNO_NUM_OBJECTS (a
));
2349 for (i
= 0; i
< n
; i
++)
2351 ira_object_t obj
= ALLOCNO_OBJECT (a
, i
);
2352 ira_object_t conflict_obj
;
2353 ira_object_conflict_iterator oci
;
2355 FOR_EACH_OBJECT_CONFLICT (obj
, conflict_obj
, oci
)
2357 ira_allocno_t conflict_a
= OBJECT_ALLOCNO (conflict_obj
);
2359 conflict_data
= ALLOCNO_COLOR_DATA (conflict_a
);
2360 if (conflict_data
->colorable_p
2361 || ! conflict_data
->in_graph_p
2362 || ALLOCNO_ASSIGNED_P (conflict_a
)
2363 || !(hard_reg_set_intersect_p
2364 (ALLOCNO_COLOR_DATA (a
)->profitable_hard_regs
,
2365 conflict_data
->profitable_hard_regs
)))
2367 ira_assert (bitmap_bit_p (coloring_allocno_bitmap
,
2368 ALLOCNO_NUM (conflict_a
)));
2369 if (update_left_conflict_sizes_p (conflict_a
, a
, size
))
2371 delete_allocno_from_bucket
2372 (conflict_a
, &uncolorable_allocno_bucket
);
2373 add_allocno_to_ordered_colorable_bucket (conflict_a
);
2374 if (internal_flag_ira_verbose
> 4 && ira_dump_file
!= NULL
)
2376 fprintf (ira_dump_file
, " Making");
2377 ira_print_expanded_allocno (conflict_a
);
2378 fprintf (ira_dump_file
, " colorable\n");
2386 /* Put ALLOCNO onto the coloring stack and remove it from its bucket.
2387 The allocno is in the colorable bucket if COLORABLE_P is TRUE. */
2389 remove_allocno_from_bucket_and_push (ira_allocno_t allocno
, bool colorable_p
)
2392 delete_allocno_from_bucket (allocno
, &colorable_allocno_bucket
);
2394 delete_allocno_from_bucket (allocno
, &uncolorable_allocno_bucket
);
2395 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
2397 fprintf (ira_dump_file
, " Pushing");
2398 ira_print_expanded_allocno (allocno
);
2400 fprintf (ira_dump_file
, "(cost %d)\n",
2401 ALLOCNO_COLOR_DATA (allocno
)->temp
);
2403 fprintf (ira_dump_file
, "(potential spill: %spri=%d, cost=%d)\n",
2404 ALLOCNO_BAD_SPILL_P (allocno
) ? "bad spill, " : "",
2405 allocno_spill_priority (allocno
),
2406 ALLOCNO_COLOR_DATA (allocno
)->temp
);
2409 ALLOCNO_COLOR_DATA (allocno
)->may_be_spilled_p
= true;
2410 push_allocno_to_stack (allocno
);
2413 /* Put all allocnos from colorable bucket onto the coloring stack. */
2415 push_only_colorable (void)
2417 form_threads_from_bucket (colorable_allocno_bucket
);
2418 sort_bucket (&colorable_allocno_bucket
, bucket_allocno_compare_func
);
2419 for (;colorable_allocno_bucket
!= NULL
;)
2420 remove_allocno_from_bucket_and_push (colorable_allocno_bucket
, true);
2423 /* Return the frequency of exit edges (if EXIT_P) or entry from/to the
2424 loop given by its LOOP_NODE. */
2426 ira_loop_edge_freq (ira_loop_tree_node_t loop_node
, int regno
, bool exit_p
)
2433 ira_assert (current_loops
!= NULL
&& loop_node
->loop
!= NULL
2434 && (regno
< 0 || regno
>= FIRST_PSEUDO_REGISTER
));
2438 FOR_EACH_EDGE (e
, ei
, loop_node
->loop
->header
->preds
)
2439 if (e
->src
!= loop_node
->loop
->latch
2441 || (bitmap_bit_p (df_get_live_out (e
->src
), regno
)
2442 && bitmap_bit_p (df_get_live_in (e
->dest
), regno
))))
2443 freq
+= EDGE_FREQUENCY (e
);
2447 edges
= get_loop_exit_edges (loop_node
->loop
);
2448 FOR_EACH_VEC_ELT (edges
, i
, e
)
2450 || (bitmap_bit_p (df_get_live_out (e
->src
), regno
)
2451 && bitmap_bit_p (df_get_live_in (e
->dest
), regno
)))
2452 freq
+= EDGE_FREQUENCY (e
);
2456 return REG_FREQ_FROM_EDGE_FREQ (freq
);
2459 /* Calculate and return the cost of putting allocno A into memory. */
2461 calculate_allocno_spill_cost (ira_allocno_t a
)
2465 enum reg_class rclass
;
2466 ira_allocno_t parent_allocno
;
2467 ira_loop_tree_node_t parent_node
, loop_node
;
2469 regno
= ALLOCNO_REGNO (a
);
2470 cost
= ALLOCNO_UPDATED_MEMORY_COST (a
) - ALLOCNO_UPDATED_CLASS_COST (a
);
2471 if (ALLOCNO_CAP (a
) != NULL
)
2473 loop_node
= ALLOCNO_LOOP_TREE_NODE (a
);
2474 if ((parent_node
= loop_node
->parent
) == NULL
)
2476 if ((parent_allocno
= parent_node
->regno_allocno_map
[regno
]) == NULL
)
2478 mode
= ALLOCNO_MODE (a
);
2479 rclass
= ALLOCNO_CLASS (a
);
2480 if (ALLOCNO_HARD_REGNO (parent_allocno
) < 0)
2481 cost
-= (ira_memory_move_cost
[mode
][rclass
][0]
2482 * ira_loop_edge_freq (loop_node
, regno
, true)
2483 + ira_memory_move_cost
[mode
][rclass
][1]
2484 * ira_loop_edge_freq (loop_node
, regno
, false));
2487 ira_init_register_move_cost_if_necessary (mode
);
2488 cost
+= ((ira_memory_move_cost
[mode
][rclass
][1]
2489 * ira_loop_edge_freq (loop_node
, regno
, true)
2490 + ira_memory_move_cost
[mode
][rclass
][0]
2491 * ira_loop_edge_freq (loop_node
, regno
, false))
2492 - (ira_register_move_cost
[mode
][rclass
][rclass
]
2493 * (ira_loop_edge_freq (loop_node
, regno
, false)
2494 + ira_loop_edge_freq (loop_node
, regno
, true))));
2499 /* Used for sorting allocnos for spilling. */
2501 allocno_spill_priority_compare (ira_allocno_t a1
, ira_allocno_t a2
)
2503 int pri1
, pri2
, diff
;
2505 /* Avoid spilling static chain pointer pseudo when non-local goto is
2507 if (non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a1
)))
2509 else if (non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a2
)))
2511 if (ALLOCNO_BAD_SPILL_P (a1
) && ! ALLOCNO_BAD_SPILL_P (a2
))
2513 if (ALLOCNO_BAD_SPILL_P (a2
) && ! ALLOCNO_BAD_SPILL_P (a1
))
2515 pri1
= allocno_spill_priority (a1
);
2516 pri2
= allocno_spill_priority (a2
);
2517 if ((diff
= pri1
- pri2
) != 0)
2520 = ALLOCNO_COLOR_DATA (a1
)->temp
- ALLOCNO_COLOR_DATA (a2
)->temp
) != 0)
2522 return ALLOCNO_NUM (a1
) - ALLOCNO_NUM (a2
);
2525 /* Used for sorting allocnos for spilling. */
2527 allocno_spill_sort_compare (const void *v1p
, const void *v2p
)
2529 ira_allocno_t p1
= *(const ira_allocno_t
*) v1p
;
2530 ira_allocno_t p2
= *(const ira_allocno_t
*) v2p
;
2532 return allocno_spill_priority_compare (p1
, p2
);
2535 /* Push allocnos to the coloring stack. The order of allocnos in the
2536 stack defines the order for the subsequent coloring. */
2538 push_allocnos_to_stack (void)
2543 /* Calculate uncolorable allocno spill costs. */
2544 for (a
= uncolorable_allocno_bucket
;
2546 a
= ALLOCNO_COLOR_DATA (a
)->next_bucket_allocno
)
2547 if (ALLOCNO_CLASS (a
) != NO_REGS
)
2549 cost
= calculate_allocno_spill_cost (a
);
2550 /* ??? Remove cost of copies between the coalesced
2552 ALLOCNO_COLOR_DATA (a
)->temp
= cost
;
2554 sort_bucket (&uncolorable_allocno_bucket
, allocno_spill_sort_compare
);
2557 push_only_colorable ();
2558 a
= uncolorable_allocno_bucket
;
2561 remove_allocno_from_bucket_and_push (a
, false);
2563 ira_assert (colorable_allocno_bucket
== NULL
2564 && uncolorable_allocno_bucket
== NULL
);
2565 ira_assert (uncolorable_allocnos_num
== 0);
2568 /* Pop the coloring stack and assign hard registers to the popped
2571 pop_allocnos_from_stack (void)
2573 ira_allocno_t allocno
;
2574 enum reg_class aclass
;
2576 for (;allocno_stack_vec
.length () != 0;)
2578 allocno
= allocno_stack_vec
.pop ();
2579 aclass
= ALLOCNO_CLASS (allocno
);
2580 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
2582 fprintf (ira_dump_file
, " Popping");
2583 ira_print_expanded_allocno (allocno
);
2584 fprintf (ira_dump_file
, " -- ");
2586 if (aclass
== NO_REGS
)
2588 ALLOCNO_HARD_REGNO (allocno
) = -1;
2589 ALLOCNO_ASSIGNED_P (allocno
) = true;
2590 ira_assert (ALLOCNO_UPDATED_HARD_REG_COSTS (allocno
) == NULL
);
2592 (ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (allocno
) == NULL
);
2593 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
2594 fprintf (ira_dump_file
, "assign memory\n");
2596 else if (assign_hard_reg (allocno
, false))
2598 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
2599 fprintf (ira_dump_file
, "assign reg %d\n",
2600 ALLOCNO_HARD_REGNO (allocno
));
2602 else if (ALLOCNO_ASSIGNED_P (allocno
))
2604 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
2605 fprintf (ira_dump_file
, "spill%s\n",
2606 ALLOCNO_COLOR_DATA (allocno
)->may_be_spilled_p
2609 ALLOCNO_COLOR_DATA (allocno
)->in_graph_p
= true;
2613 /* Set up number of available hard registers for allocno A. */
2615 setup_allocno_available_regs_num (ira_allocno_t a
)
2617 int i
, n
, hard_regno
, hard_regs_num
, nwords
;
2618 enum reg_class aclass
;
2619 allocno_color_data_t data
;
2621 aclass
= ALLOCNO_CLASS (a
);
2622 data
= ALLOCNO_COLOR_DATA (a
);
2623 data
->available_regs_num
= 0;
2624 if (aclass
== NO_REGS
)
2626 hard_regs_num
= ira_class_hard_regs_num
[aclass
];
2627 nwords
= ALLOCNO_NUM_OBJECTS (a
);
2628 for (n
= 0, i
= hard_regs_num
- 1; i
>= 0; i
--)
2630 hard_regno
= ira_class_hard_regs
[aclass
][i
];
2631 /* Checking only profitable hard regs. */
2632 if (TEST_HARD_REG_BIT (data
->profitable_hard_regs
, hard_regno
))
2635 data
->available_regs_num
= n
;
2636 if (internal_flag_ira_verbose
<= 2 || ira_dump_file
== NULL
)
2640 " Allocno a%dr%d of %s(%d) has %d avail. regs ",
2641 ALLOCNO_NUM (a
), ALLOCNO_REGNO (a
),
2642 reg_class_names
[aclass
], ira_class_hard_regs_num
[aclass
], n
);
2643 print_hard_reg_set (ira_dump_file
, data
->profitable_hard_regs
, false);
2644 fprintf (ira_dump_file
, ", %snode: ",
2645 hard_reg_set_equal_p (data
->profitable_hard_regs
,
2646 data
->hard_regs_node
->hard_regs
->set
)
2648 print_hard_reg_set (ira_dump_file
,
2649 data
->hard_regs_node
->hard_regs
->set
, false);
2650 for (i
= 0; i
< nwords
; i
++)
2652 ira_object_t obj
= ALLOCNO_OBJECT (a
, i
);
2657 fprintf (ira_dump_file
, ", ");
2658 fprintf (ira_dump_file
, " obj %d", i
);
2660 fprintf (ira_dump_file
, " (confl regs = ");
2661 print_hard_reg_set (ira_dump_file
, OBJECT_TOTAL_CONFLICT_HARD_REGS (obj
),
2663 fprintf (ira_dump_file
, ")");
2665 fprintf (ira_dump_file
, "\n");
2668 /* Put ALLOCNO in a bucket corresponding to its number and size of its
2669 conflicting allocnos and hard registers. */
2671 put_allocno_into_bucket (ira_allocno_t allocno
)
2673 ALLOCNO_COLOR_DATA (allocno
)->in_graph_p
= true;
2674 setup_allocno_available_regs_num (allocno
);
2675 if (setup_left_conflict_sizes_p (allocno
))
2676 add_allocno_to_bucket (allocno
, &colorable_allocno_bucket
);
2678 add_allocno_to_bucket (allocno
, &uncolorable_allocno_bucket
);
2681 /* Map: allocno number -> allocno priority. */
2682 static int *allocno_priorities
;
2684 /* Set up priorities for N allocnos in array
2685 CONSIDERATION_ALLOCNOS. */
2687 setup_allocno_priorities (ira_allocno_t
*consideration_allocnos
, int n
)
2689 int i
, length
, nrefs
, priority
, max_priority
, mult
;
2693 for (i
= 0; i
< n
; i
++)
2695 a
= consideration_allocnos
[i
];
2696 nrefs
= ALLOCNO_NREFS (a
);
2697 ira_assert (nrefs
>= 0);
2698 mult
= floor_log2 (ALLOCNO_NREFS (a
)) + 1;
2699 ira_assert (mult
>= 0);
2700 allocno_priorities
[ALLOCNO_NUM (a
)]
2703 * (ALLOCNO_MEMORY_COST (a
) - ALLOCNO_CLASS_COST (a
))
2704 * ira_reg_class_max_nregs
[ALLOCNO_CLASS (a
)][ALLOCNO_MODE (a
)]);
2706 priority
= -priority
;
2707 if (max_priority
< priority
)
2708 max_priority
= priority
;
2710 mult
= max_priority
== 0 ? 1 : INT_MAX
/ max_priority
;
2711 for (i
= 0; i
< n
; i
++)
2713 a
= consideration_allocnos
[i
];
2714 length
= ALLOCNO_EXCESS_PRESSURE_POINTS_NUM (a
);
2715 if (ALLOCNO_NUM_OBJECTS (a
) > 1)
2716 length
/= ALLOCNO_NUM_OBJECTS (a
);
2719 allocno_priorities
[ALLOCNO_NUM (a
)]
2720 = allocno_priorities
[ALLOCNO_NUM (a
)] * mult
/ length
;
2724 /* Sort allocnos according to the profit of usage of a hard register
2725 instead of memory for them. */
2727 allocno_cost_compare_func (const void *v1p
, const void *v2p
)
2729 ira_allocno_t p1
= *(const ira_allocno_t
*) v1p
;
2730 ira_allocno_t p2
= *(const ira_allocno_t
*) v2p
;
2733 c1
= ALLOCNO_UPDATED_MEMORY_COST (p1
) - ALLOCNO_UPDATED_CLASS_COST (p1
);
2734 c2
= ALLOCNO_UPDATED_MEMORY_COST (p2
) - ALLOCNO_UPDATED_CLASS_COST (p2
);
2738 /* If regs are equally good, sort by allocno numbers, so that the
2739 results of qsort leave nothing to chance. */
2740 return ALLOCNO_NUM (p1
) - ALLOCNO_NUM (p2
);
2743 /* Return savings on removed copies when ALLOCNO is assigned to
2746 allocno_copy_cost_saving (ira_allocno_t allocno
, int hard_regno
)
2749 machine_mode allocno_mode
= ALLOCNO_MODE (allocno
);
2750 enum reg_class rclass
;
2751 ira_copy_t cp
, next_cp
;
2753 rclass
= REGNO_REG_CLASS (hard_regno
);
2754 if (ira_reg_class_max_nregs
[rclass
][allocno_mode
]
2755 > ira_class_hard_regs_num
[rclass
])
2756 /* For the above condition the cost can be wrong. Use the allocno
2757 class in this case. */
2758 rclass
= ALLOCNO_CLASS (allocno
);
2759 for (cp
= ALLOCNO_COPIES (allocno
); cp
!= NULL
; cp
= next_cp
)
2761 if (cp
->first
== allocno
)
2763 next_cp
= cp
->next_first_allocno_copy
;
2764 if (ALLOCNO_HARD_REGNO (cp
->second
) != hard_regno
)
2767 else if (cp
->second
== allocno
)
2769 next_cp
= cp
->next_second_allocno_copy
;
2770 if (ALLOCNO_HARD_REGNO (cp
->first
) != hard_regno
)
2775 cost
+= cp
->freq
* ira_register_move_cost
[allocno_mode
][rclass
][rclass
];
2780 /* We used Chaitin-Briggs coloring to assign as many pseudos as
2781 possible to hard registers. Let us try to improve allocation with
2782 cost point of view. This function improves the allocation by
2783 spilling some allocnos and assigning the freed hard registers to
2784 other allocnos if it decreases the overall allocation cost. */
2786 improve_allocation (void)
2789 int j
, k
, n
, hregno
, conflict_hregno
, base_cost
, class_size
, word
, nwords
;
2790 int check
, spill_cost
, min_cost
, nregs
, conflict_nregs
, r
, best
;
2792 enum reg_class aclass
;
2795 int costs
[FIRST_PSEUDO_REGISTER
];
2796 HARD_REG_SET conflicting_regs
[2], profitable_hard_regs
;
2800 /* Don't bother to optimize the code with static chain pointer and
2801 non-local goto in order not to spill the chain pointer
2803 if (cfun
->static_chain_decl
&& crtl
->has_nonlocal_goto
)
2805 /* Clear counts used to process conflicting allocnos only once for
2807 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap
, 0, i
, bi
)
2808 ALLOCNO_COLOR_DATA (ira_allocnos
[i
])->temp
= 0;
2810 /* Process each allocno and try to assign a hard register to it by
2811 spilling some its conflicting allocnos. */
2812 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap
, 0, i
, bi
)
2814 a
= ira_allocnos
[i
];
2815 ALLOCNO_COLOR_DATA (a
)->temp
= 0;
2816 if (empty_profitable_hard_regs (a
))
2819 aclass
= ALLOCNO_CLASS (a
);
2820 allocno_costs
= ALLOCNO_HARD_REG_COSTS (a
);
2821 if ((hregno
= ALLOCNO_HARD_REGNO (a
)) < 0)
2822 base_cost
= ALLOCNO_UPDATED_MEMORY_COST (a
);
2823 else if (allocno_costs
== NULL
)
2824 /* It means that assigning a hard register is not profitable
2825 (we don't waste memory for hard register costs in this
2829 base_cost
= (allocno_costs
[ira_class_hard_reg_index
[aclass
][hregno
]]
2830 - allocno_copy_cost_saving (a
, hregno
));
2832 get_conflict_and_start_profitable_regs (a
, false,
2834 &profitable_hard_regs
);
2835 class_size
= ira_class_hard_regs_num
[aclass
];
2836 /* Set up cost improvement for usage of each profitable hard
2837 register for allocno A. */
2838 for (j
= 0; j
< class_size
; j
++)
2840 hregno
= ira_class_hard_regs
[aclass
][j
];
2841 if (! check_hard_reg_p (a
, hregno
,
2842 conflicting_regs
, profitable_hard_regs
))
2844 ira_assert (ira_class_hard_reg_index
[aclass
][hregno
] == j
);
2845 k
= allocno_costs
== NULL
? 0 : j
;
2846 costs
[hregno
] = (allocno_costs
== NULL
2847 ? ALLOCNO_UPDATED_CLASS_COST (a
) : allocno_costs
[k
]);
2848 costs
[hregno
] -= allocno_copy_cost_saving (a
, hregno
);
2849 costs
[hregno
] -= base_cost
;
2850 if (costs
[hregno
] < 0)
2854 /* There is no chance to improve the allocation cost by
2855 assigning hard register to allocno A even without spilling
2856 conflicting allocnos. */
2858 mode
= ALLOCNO_MODE (a
);
2859 nwords
= ALLOCNO_NUM_OBJECTS (a
);
2860 /* Process each allocno conflicting with A and update the cost
2861 improvement for profitable hard registers of A. To use a
2862 hard register for A we need to spill some conflicting
2863 allocnos and that creates penalty for the cost
2865 for (word
= 0; word
< nwords
; word
++)
2867 ira_object_t conflict_obj
;
2868 ira_object_t obj
= ALLOCNO_OBJECT (a
, word
);
2869 ira_object_conflict_iterator oci
;
2871 FOR_EACH_OBJECT_CONFLICT (obj
, conflict_obj
, oci
)
2873 ira_allocno_t conflict_a
= OBJECT_ALLOCNO (conflict_obj
);
2875 if (ALLOCNO_COLOR_DATA (conflict_a
)->temp
== check
)
2876 /* We already processed this conflicting allocno
2877 because we processed earlier another object of the
2878 conflicting allocno. */
2880 ALLOCNO_COLOR_DATA (conflict_a
)->temp
= check
;
2881 if ((conflict_hregno
= ALLOCNO_HARD_REGNO (conflict_a
)) < 0)
2883 spill_cost
= ALLOCNO_UPDATED_MEMORY_COST (conflict_a
);
2884 k
= (ira_class_hard_reg_index
2885 [ALLOCNO_CLASS (conflict_a
)][conflict_hregno
]);
2886 ira_assert (k
>= 0);
2887 if ((allocno_costs
= ALLOCNO_HARD_REG_COSTS (conflict_a
))
2889 spill_cost
-= allocno_costs
[k
];
2891 spill_cost
-= ALLOCNO_UPDATED_CLASS_COST (conflict_a
);
2893 += allocno_copy_cost_saving (conflict_a
, conflict_hregno
);
2894 conflict_nregs
= hard_regno_nregs (conflict_hregno
,
2895 ALLOCNO_MODE (conflict_a
));
2896 for (r
= conflict_hregno
;
2897 r
>= 0 && (int) end_hard_regno (mode
, r
) > conflict_hregno
;
2899 if (check_hard_reg_p (a
, r
,
2900 conflicting_regs
, profitable_hard_regs
))
2901 costs
[r
] += spill_cost
;
2902 for (r
= conflict_hregno
+ 1;
2903 r
< conflict_hregno
+ conflict_nregs
;
2905 if (check_hard_reg_p (a
, r
,
2906 conflicting_regs
, profitable_hard_regs
))
2907 costs
[r
] += spill_cost
;
2912 /* Now we choose hard register for A which results in highest
2913 allocation cost improvement. */
2914 for (j
= 0; j
< class_size
; j
++)
2916 hregno
= ira_class_hard_regs
[aclass
][j
];
2917 if (check_hard_reg_p (a
, hregno
,
2918 conflicting_regs
, profitable_hard_regs
)
2919 && min_cost
> costs
[hregno
])
2922 min_cost
= costs
[hregno
];
2926 /* We are in a situation when assigning any hard register to A
2927 by spilling some conflicting allocnos does not improve the
2930 nregs
= hard_regno_nregs (best
, mode
);
2931 /* Now spill conflicting allocnos which contain a hard register
2932 of A when we assign the best chosen hard register to it. */
2933 for (word
= 0; word
< nwords
; word
++)
2935 ira_object_t conflict_obj
;
2936 ira_object_t obj
= ALLOCNO_OBJECT (a
, word
);
2937 ira_object_conflict_iterator oci
;
2939 FOR_EACH_OBJECT_CONFLICT (obj
, conflict_obj
, oci
)
2941 ira_allocno_t conflict_a
= OBJECT_ALLOCNO (conflict_obj
);
2943 if ((conflict_hregno
= ALLOCNO_HARD_REGNO (conflict_a
)) < 0)
2945 conflict_nregs
= hard_regno_nregs (conflict_hregno
,
2946 ALLOCNO_MODE (conflict_a
));
2947 if (best
+ nregs
<= conflict_hregno
2948 || conflict_hregno
+ conflict_nregs
<= best
)
2949 /* No intersection. */
2951 ALLOCNO_HARD_REGNO (conflict_a
) = -1;
2952 sorted_allocnos
[n
++] = conflict_a
;
2953 if (internal_flag_ira_verbose
> 2 && ira_dump_file
!= NULL
)
2954 fprintf (ira_dump_file
, "Spilling a%dr%d for a%dr%d\n",
2955 ALLOCNO_NUM (conflict_a
), ALLOCNO_REGNO (conflict_a
),
2956 ALLOCNO_NUM (a
), ALLOCNO_REGNO (a
));
2959 /* Assign the best chosen hard register to A. */
2960 ALLOCNO_HARD_REGNO (a
) = best
;
2961 if (internal_flag_ira_verbose
> 2 && ira_dump_file
!= NULL
)
2962 fprintf (ira_dump_file
, "Assigning %d to a%dr%d\n",
2963 best
, ALLOCNO_NUM (a
), ALLOCNO_REGNO (a
));
2967 /* We spilled some allocnos to assign their hard registers to other
2968 allocnos. The spilled allocnos are now in array
2969 'sorted_allocnos'. There is still a possibility that some of the
2970 spilled allocnos can get hard registers. So let us try assign
2971 them hard registers again (just a reminder -- function
2972 'assign_hard_reg' assigns hard registers only if it is possible
2973 and profitable). We process the spilled allocnos with biggest
2974 benefit to get hard register first -- see function
2975 'allocno_cost_compare_func'. */
2976 qsort (sorted_allocnos
, n
, sizeof (ira_allocno_t
),
2977 allocno_cost_compare_func
);
2978 for (j
= 0; j
< n
; j
++)
2980 a
= sorted_allocnos
[j
];
2981 ALLOCNO_ASSIGNED_P (a
) = false;
2982 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
2984 fprintf (ira_dump_file
, " ");
2985 ira_print_expanded_allocno (a
);
2986 fprintf (ira_dump_file
, " -- ");
2988 if (assign_hard_reg (a
, false))
2990 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
2991 fprintf (ira_dump_file
, "assign hard reg %d\n",
2992 ALLOCNO_HARD_REGNO (a
));
2996 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
2997 fprintf (ira_dump_file
, "assign memory\n");
3002 /* Sort allocnos according to their priorities. */
3004 allocno_priority_compare_func (const void *v1p
, const void *v2p
)
3006 ira_allocno_t a1
= *(const ira_allocno_t
*) v1p
;
3007 ira_allocno_t a2
= *(const ira_allocno_t
*) v2p
;
3008 int pri1
, pri2
, diff
;
3010 /* Assign hard reg to static chain pointer pseudo first when
3011 non-local goto is used. */
3012 if ((diff
= (non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a2
))
3013 - non_spilled_static_chain_regno_p (ALLOCNO_REGNO (a1
)))) != 0)
3015 pri1
= allocno_priorities
[ALLOCNO_NUM (a1
)];
3016 pri2
= allocno_priorities
[ALLOCNO_NUM (a2
)];
3018 return SORTGT (pri2
, pri1
);
3020 /* If regs are equally good, sort by allocnos, so that the results of
3021 qsort leave nothing to chance. */
3022 return ALLOCNO_NUM (a1
) - ALLOCNO_NUM (a2
);
3025 /* Chaitin-Briggs coloring for allocnos in COLORING_ALLOCNO_BITMAP
3026 taking into account allocnos in CONSIDERATION_ALLOCNO_BITMAP. */
3028 color_allocnos (void)
3034 setup_profitable_hard_regs ();
3035 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap
, 0, i
, bi
)
3038 HARD_REG_SET conflict_hard_regs
;
3039 allocno_color_data_t data
;
3040 ira_pref_t pref
, next_pref
;
3042 a
= ira_allocnos
[i
];
3043 nr
= ALLOCNO_NUM_OBJECTS (a
);
3044 CLEAR_HARD_REG_SET (conflict_hard_regs
);
3045 for (l
= 0; l
< nr
; l
++)
3047 ira_object_t obj
= ALLOCNO_OBJECT (a
, l
);
3048 IOR_HARD_REG_SET (conflict_hard_regs
,
3049 OBJECT_CONFLICT_HARD_REGS (obj
));
3051 data
= ALLOCNO_COLOR_DATA (a
);
3052 for (pref
= ALLOCNO_PREFS (a
); pref
!= NULL
; pref
= next_pref
)
3054 next_pref
= pref
->next_pref
;
3055 if (! ira_hard_reg_in_set_p (pref
->hard_regno
,
3057 data
->profitable_hard_regs
))
3058 ira_remove_pref (pref
);
3061 if (flag_ira_algorithm
== IRA_ALGORITHM_PRIORITY
)
3064 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap
, 0, i
, bi
)
3066 a
= ira_allocnos
[i
];
3067 if (ALLOCNO_CLASS (a
) == NO_REGS
)
3069 ALLOCNO_HARD_REGNO (a
) = -1;
3070 ALLOCNO_ASSIGNED_P (a
) = true;
3071 ira_assert (ALLOCNO_UPDATED_HARD_REG_COSTS (a
) == NULL
);
3072 ira_assert (ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a
) == NULL
);
3073 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
3075 fprintf (ira_dump_file
, " Spill");
3076 ira_print_expanded_allocno (a
);
3077 fprintf (ira_dump_file
, "\n");
3081 sorted_allocnos
[n
++] = a
;
3085 setup_allocno_priorities (sorted_allocnos
, n
);
3086 qsort (sorted_allocnos
, n
, sizeof (ira_allocno_t
),
3087 allocno_priority_compare_func
);
3088 for (i
= 0; i
< n
; i
++)
3090 a
= sorted_allocnos
[i
];
3091 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
3093 fprintf (ira_dump_file
, " ");
3094 ira_print_expanded_allocno (a
);
3095 fprintf (ira_dump_file
, " -- ");
3097 if (assign_hard_reg (a
, false))
3099 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
3100 fprintf (ira_dump_file
, "assign hard reg %d\n",
3101 ALLOCNO_HARD_REGNO (a
));
3105 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
3106 fprintf (ira_dump_file
, "assign memory\n");
3113 form_allocno_hard_regs_nodes_forest ();
3114 if (internal_flag_ira_verbose
> 2 && ira_dump_file
!= NULL
)
3115 print_hard_regs_forest (ira_dump_file
);
3116 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap
, 0, i
, bi
)
3118 a
= ira_allocnos
[i
];
3119 if (ALLOCNO_CLASS (a
) != NO_REGS
&& ! empty_profitable_hard_regs (a
))
3121 ALLOCNO_COLOR_DATA (a
)->in_graph_p
= true;
3122 update_costs_from_prefs (a
);
3126 ALLOCNO_HARD_REGNO (a
) = -1;
3127 ALLOCNO_ASSIGNED_P (a
) = true;
3128 /* We don't need updated costs anymore. */
3129 ira_free_allocno_updated_costs (a
);
3130 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
3132 fprintf (ira_dump_file
, " Spill");
3133 ira_print_expanded_allocno (a
);
3134 fprintf (ira_dump_file
, "\n");
3138 /* Put the allocnos into the corresponding buckets. */
3139 colorable_allocno_bucket
= NULL
;
3140 uncolorable_allocno_bucket
= NULL
;
3141 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap
, 0, i
, bi
)
3143 a
= ira_allocnos
[i
];
3144 if (ALLOCNO_COLOR_DATA (a
)->in_graph_p
)
3145 put_allocno_into_bucket (a
);
3147 push_allocnos_to_stack ();
3148 pop_allocnos_from_stack ();
3149 finish_allocno_hard_regs_nodes_forest ();
3151 improve_allocation ();
3156 /* Output information about the loop given by its LOOP_TREE_NODE. */
3158 print_loop_title (ira_loop_tree_node_t loop_tree_node
)
3162 ira_loop_tree_node_t subloop_node
, dest_loop_node
;
3166 if (loop_tree_node
->parent
== NULL
)
3167 fprintf (ira_dump_file
,
3168 "\n Loop 0 (parent -1, header bb%d, depth 0)\n bbs:",
3172 ira_assert (current_loops
!= NULL
&& loop_tree_node
->loop
!= NULL
);
3173 fprintf (ira_dump_file
,
3174 "\n Loop %d (parent %d, header bb%d, depth %d)\n bbs:",
3175 loop_tree_node
->loop_num
, loop_tree_node
->parent
->loop_num
,
3176 loop_tree_node
->loop
->header
->index
,
3177 loop_depth (loop_tree_node
->loop
));
3179 for (subloop_node
= loop_tree_node
->children
;
3180 subloop_node
!= NULL
;
3181 subloop_node
= subloop_node
->next
)
3182 if (subloop_node
->bb
!= NULL
)
3184 fprintf (ira_dump_file
, " %d", subloop_node
->bb
->index
);
3185 FOR_EACH_EDGE (e
, ei
, subloop_node
->bb
->succs
)
3186 if (e
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
)
3187 && ((dest_loop_node
= IRA_BB_NODE (e
->dest
)->parent
)
3189 fprintf (ira_dump_file
, "(->%d:l%d)",
3190 e
->dest
->index
, dest_loop_node
->loop_num
);
3192 fprintf (ira_dump_file
, "\n all:");
3193 EXECUTE_IF_SET_IN_BITMAP (loop_tree_node
->all_allocnos
, 0, j
, bi
)
3194 fprintf (ira_dump_file
, " %dr%d", j
, ALLOCNO_REGNO (ira_allocnos
[j
]));
3195 fprintf (ira_dump_file
, "\n modified regnos:");
3196 EXECUTE_IF_SET_IN_BITMAP (loop_tree_node
->modified_regnos
, 0, j
, bi
)
3197 fprintf (ira_dump_file
, " %d", j
);
3198 fprintf (ira_dump_file
, "\n border:");
3199 EXECUTE_IF_SET_IN_BITMAP (loop_tree_node
->border_allocnos
, 0, j
, bi
)
3200 fprintf (ira_dump_file
, " %dr%d", j
, ALLOCNO_REGNO (ira_allocnos
[j
]));
3201 fprintf (ira_dump_file
, "\n Pressure:");
3202 for (j
= 0; (int) j
< ira_pressure_classes_num
; j
++)
3204 enum reg_class pclass
;
3206 pclass
= ira_pressure_classes
[j
];
3207 if (loop_tree_node
->reg_pressure
[pclass
] == 0)
3209 fprintf (ira_dump_file
, " %s=%d", reg_class_names
[pclass
],
3210 loop_tree_node
->reg_pressure
[pclass
]);
3212 fprintf (ira_dump_file
, "\n");
3215 /* Color the allocnos inside loop (in the extreme case it can be all
3216 of the function) given the corresponding LOOP_TREE_NODE. The
3217 function is called for each loop during top-down traverse of the
3220 color_pass (ira_loop_tree_node_t loop_tree_node
)
3222 int regno
, hard_regno
, index
= -1, n
;
3223 int cost
, exit_freq
, enter_freq
;
3227 enum reg_class rclass
, aclass
, pclass
;
3228 ira_allocno_t a
, subloop_allocno
;
3229 ira_loop_tree_node_t subloop_node
;
3231 ira_assert (loop_tree_node
->bb
== NULL
);
3232 if (internal_flag_ira_verbose
> 1 && ira_dump_file
!= NULL
)
3233 print_loop_title (loop_tree_node
);
3235 bitmap_copy (coloring_allocno_bitmap
, loop_tree_node
->all_allocnos
);
3236 bitmap_copy (consideration_allocno_bitmap
, coloring_allocno_bitmap
);
3238 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap
, 0, j
, bi
)
3240 a
= ira_allocnos
[j
];
3242 if (! ALLOCNO_ASSIGNED_P (a
))
3244 bitmap_clear_bit (coloring_allocno_bitmap
, ALLOCNO_NUM (a
));
3247 = (allocno_color_data_t
) ira_allocate (sizeof (struct allocno_color_data
)
3249 memset (allocno_color_data
, 0, sizeof (struct allocno_color_data
) * n
);
3250 curr_allocno_process
= 0;
3252 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap
, 0, j
, bi
)
3254 a
= ira_allocnos
[j
];
3255 ALLOCNO_ADD_DATA (a
) = allocno_color_data
+ n
;
3258 init_allocno_threads ();
3259 /* Color all mentioned allocnos including transparent ones. */
3261 /* Process caps. They are processed just once. */
3262 if (flag_ira_region
== IRA_REGION_MIXED
3263 || flag_ira_region
== IRA_REGION_ALL
)
3264 EXECUTE_IF_SET_IN_BITMAP (loop_tree_node
->all_allocnos
, 0, j
, bi
)
3266 a
= ira_allocnos
[j
];
3267 if (ALLOCNO_CAP_MEMBER (a
) == NULL
)
3269 /* Remove from processing in the next loop. */
3270 bitmap_clear_bit (consideration_allocno_bitmap
, j
);
3271 rclass
= ALLOCNO_CLASS (a
);
3272 pclass
= ira_pressure_class_translate
[rclass
];
3273 if (flag_ira_region
== IRA_REGION_MIXED
3274 && (loop_tree_node
->reg_pressure
[pclass
]
3275 <= ira_class_hard_regs_num
[pclass
]))
3277 mode
= ALLOCNO_MODE (a
);
3278 hard_regno
= ALLOCNO_HARD_REGNO (a
);
3279 if (hard_regno
>= 0)
3281 index
= ira_class_hard_reg_index
[rclass
][hard_regno
];
3282 ira_assert (index
>= 0);
3284 regno
= ALLOCNO_REGNO (a
);
3285 subloop_allocno
= ALLOCNO_CAP_MEMBER (a
);
3286 subloop_node
= ALLOCNO_LOOP_TREE_NODE (subloop_allocno
);
3287 ira_assert (!ALLOCNO_ASSIGNED_P (subloop_allocno
));
3288 ALLOCNO_HARD_REGNO (subloop_allocno
) = hard_regno
;
3289 ALLOCNO_ASSIGNED_P (subloop_allocno
) = true;
3290 if (hard_regno
>= 0)
3291 update_costs_from_copies (subloop_allocno
, true, true);
3292 /* We don't need updated costs anymore. */
3293 ira_free_allocno_updated_costs (subloop_allocno
);
3296 /* Update costs of the corresponding allocnos (not caps) in the
3298 for (subloop_node
= loop_tree_node
->subloops
;
3299 subloop_node
!= NULL
;
3300 subloop_node
= subloop_node
->subloop_next
)
3302 ira_assert (subloop_node
->bb
== NULL
);
3303 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap
, 0, j
, bi
)
3305 a
= ira_allocnos
[j
];
3306 ira_assert (ALLOCNO_CAP_MEMBER (a
) == NULL
);
3307 mode
= ALLOCNO_MODE (a
);
3308 rclass
= ALLOCNO_CLASS (a
);
3309 pclass
= ira_pressure_class_translate
[rclass
];
3310 hard_regno
= ALLOCNO_HARD_REGNO (a
);
3311 /* Use hard register class here. ??? */
3312 if (hard_regno
>= 0)
3314 index
= ira_class_hard_reg_index
[rclass
][hard_regno
];
3315 ira_assert (index
>= 0);
3317 regno
= ALLOCNO_REGNO (a
);
3318 /* ??? conflict costs */
3319 subloop_allocno
= subloop_node
->regno_allocno_map
[regno
];
3320 if (subloop_allocno
== NULL
3321 || ALLOCNO_CAP (subloop_allocno
) != NULL
)
3323 ira_assert (ALLOCNO_CLASS (subloop_allocno
) == rclass
);
3324 ira_assert (bitmap_bit_p (subloop_node
->all_allocnos
,
3325 ALLOCNO_NUM (subloop_allocno
)));
3326 if ((flag_ira_region
== IRA_REGION_MIXED
3327 && (loop_tree_node
->reg_pressure
[pclass
]
3328 <= ira_class_hard_regs_num
[pclass
]))
3329 || (pic_offset_table_rtx
!= NULL
3330 && regno
== (int) REGNO (pic_offset_table_rtx
))
3331 /* Avoid overlapped multi-registers. Moves between them
3332 might result in wrong code generation. */
3334 && ira_reg_class_max_nregs
[pclass
][mode
] > 1))
3336 if (! ALLOCNO_ASSIGNED_P (subloop_allocno
))
3338 ALLOCNO_HARD_REGNO (subloop_allocno
) = hard_regno
;
3339 ALLOCNO_ASSIGNED_P (subloop_allocno
) = true;
3340 if (hard_regno
>= 0)
3341 update_costs_from_copies (subloop_allocno
, true, true);
3342 /* We don't need updated costs anymore. */
3343 ira_free_allocno_updated_costs (subloop_allocno
);
3347 exit_freq
= ira_loop_edge_freq (subloop_node
, regno
, true);
3348 enter_freq
= ira_loop_edge_freq (subloop_node
, regno
, false);
3349 ira_assert (regno
< ira_reg_equiv_len
);
3350 if (ira_equiv_no_lvalue_p (regno
))
3352 if (! ALLOCNO_ASSIGNED_P (subloop_allocno
))
3354 ALLOCNO_HARD_REGNO (subloop_allocno
) = hard_regno
;
3355 ALLOCNO_ASSIGNED_P (subloop_allocno
) = true;
3356 if (hard_regno
>= 0)
3357 update_costs_from_copies (subloop_allocno
, true, true);
3358 /* We don't need updated costs anymore. */
3359 ira_free_allocno_updated_costs (subloop_allocno
);
3362 else if (hard_regno
< 0)
3364 ALLOCNO_UPDATED_MEMORY_COST (subloop_allocno
)
3365 -= ((ira_memory_move_cost
[mode
][rclass
][1] * enter_freq
)
3366 + (ira_memory_move_cost
[mode
][rclass
][0] * exit_freq
));
3370 aclass
= ALLOCNO_CLASS (subloop_allocno
);
3371 ira_init_register_move_cost_if_necessary (mode
);
3372 cost
= (ira_register_move_cost
[mode
][rclass
][rclass
]
3373 * (exit_freq
+ enter_freq
));
3374 ira_allocate_and_set_or_copy_costs
3375 (&ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno
), aclass
,
3376 ALLOCNO_UPDATED_CLASS_COST (subloop_allocno
),
3377 ALLOCNO_HARD_REG_COSTS (subloop_allocno
));
3378 ira_allocate_and_set_or_copy_costs
3379 (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (subloop_allocno
),
3380 aclass
, 0, ALLOCNO_CONFLICT_HARD_REG_COSTS (subloop_allocno
));
3381 ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno
)[index
] -= cost
;
3382 ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (subloop_allocno
)[index
]
3384 if (ALLOCNO_UPDATED_CLASS_COST (subloop_allocno
)
3385 > ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno
)[index
])
3386 ALLOCNO_UPDATED_CLASS_COST (subloop_allocno
)
3387 = ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno
)[index
];
3388 ALLOCNO_UPDATED_MEMORY_COST (subloop_allocno
)
3389 += (ira_memory_move_cost
[mode
][rclass
][0] * enter_freq
3390 + ira_memory_move_cost
[mode
][rclass
][1] * exit_freq
);
3394 ira_free (allocno_color_data
);
3395 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap
, 0, j
, bi
)
3397 a
= ira_allocnos
[j
];
3398 ALLOCNO_ADD_DATA (a
) = NULL
;
3402 /* Initialize the common data for coloring and calls functions to do
3403 Chaitin-Briggs and regional coloring. */
3407 coloring_allocno_bitmap
= ira_allocate_bitmap ();
3408 if (internal_flag_ira_verbose
> 0 && ira_dump_file
!= NULL
)
3409 fprintf (ira_dump_file
, "\n**** Allocnos coloring:\n\n");
3411 ira_traverse_loop_tree (false, ira_loop_tree_root
, color_pass
, NULL
);
3413 if (internal_flag_ira_verbose
> 1 && ira_dump_file
!= NULL
)
3414 ira_print_disposition (ira_dump_file
);
3416 ira_free_bitmap (coloring_allocno_bitmap
);
3421 /* Move spill/restore code, which are to be generated in ira-emit.c,
3422 to less frequent points (if it is profitable) by reassigning some
3423 allocnos (in loop with subloops containing in another loop) to
3424 memory which results in longer live-range where the corresponding
3425 pseudo-registers will be in memory. */
3427 move_spill_restore (void)
3429 int cost
, regno
, hard_regno
, hard_regno2
, index
;
3431 int enter_freq
, exit_freq
;
3433 enum reg_class rclass
;
3434 ira_allocno_t a
, parent_allocno
, subloop_allocno
;
3435 ira_loop_tree_node_t parent
, loop_node
, subloop_node
;
3436 ira_allocno_iterator ai
;
3441 if (internal_flag_ira_verbose
> 0 && ira_dump_file
!= NULL
)
3442 fprintf (ira_dump_file
, "New iteration of spill/restore move\n");
3443 FOR_EACH_ALLOCNO (a
, ai
)
3445 regno
= ALLOCNO_REGNO (a
);
3446 loop_node
= ALLOCNO_LOOP_TREE_NODE (a
);
3447 if (ALLOCNO_CAP_MEMBER (a
) != NULL
3448 || ALLOCNO_CAP (a
) != NULL
3449 || (hard_regno
= ALLOCNO_HARD_REGNO (a
)) < 0
3450 || loop_node
->children
== NULL
3451 /* don't do the optimization because it can create
3452 copies and the reload pass can spill the allocno set
3453 by copy although the allocno will not get memory
3455 || ira_equiv_no_lvalue_p (regno
)
3456 || !bitmap_bit_p (loop_node
->border_allocnos
, ALLOCNO_NUM (a
))
3457 /* Do not spill static chain pointer pseudo when
3458 non-local goto is used. */
3459 || non_spilled_static_chain_regno_p (regno
))
3461 mode
= ALLOCNO_MODE (a
);
3462 rclass
= ALLOCNO_CLASS (a
);
3463 index
= ira_class_hard_reg_index
[rclass
][hard_regno
];
3464 ira_assert (index
>= 0);
3465 cost
= (ALLOCNO_MEMORY_COST (a
)
3466 - (ALLOCNO_HARD_REG_COSTS (a
) == NULL
3467 ? ALLOCNO_CLASS_COST (a
)
3468 : ALLOCNO_HARD_REG_COSTS (a
)[index
]));
3469 ira_init_register_move_cost_if_necessary (mode
);
3470 for (subloop_node
= loop_node
->subloops
;
3471 subloop_node
!= NULL
;
3472 subloop_node
= subloop_node
->subloop_next
)
3474 ira_assert (subloop_node
->bb
== NULL
);
3475 subloop_allocno
= subloop_node
->regno_allocno_map
[regno
];
3476 if (subloop_allocno
== NULL
)
3478 ira_assert (rclass
== ALLOCNO_CLASS (subloop_allocno
));
3479 /* We have accumulated cost. To get the real cost of
3480 allocno usage in the loop we should subtract costs of
3481 the subloop allocnos. */
3482 cost
-= (ALLOCNO_MEMORY_COST (subloop_allocno
)
3483 - (ALLOCNO_HARD_REG_COSTS (subloop_allocno
) == NULL
3484 ? ALLOCNO_CLASS_COST (subloop_allocno
)
3485 : ALLOCNO_HARD_REG_COSTS (subloop_allocno
)[index
]));
3486 exit_freq
= ira_loop_edge_freq (subloop_node
, regno
, true);
3487 enter_freq
= ira_loop_edge_freq (subloop_node
, regno
, false);
3488 if ((hard_regno2
= ALLOCNO_HARD_REGNO (subloop_allocno
)) < 0)
3489 cost
-= (ira_memory_move_cost
[mode
][rclass
][0] * exit_freq
3490 + ira_memory_move_cost
[mode
][rclass
][1] * enter_freq
);
3494 += (ira_memory_move_cost
[mode
][rclass
][0] * exit_freq
3495 + ira_memory_move_cost
[mode
][rclass
][1] * enter_freq
);
3496 if (hard_regno2
!= hard_regno
)
3497 cost
-= (ira_register_move_cost
[mode
][rclass
][rclass
]
3498 * (exit_freq
+ enter_freq
));
3501 if ((parent
= loop_node
->parent
) != NULL
3502 && (parent_allocno
= parent
->regno_allocno_map
[regno
]) != NULL
)
3504 ira_assert (rclass
== ALLOCNO_CLASS (parent_allocno
));
3505 exit_freq
= ira_loop_edge_freq (loop_node
, regno
, true);
3506 enter_freq
= ira_loop_edge_freq (loop_node
, regno
, false);
3507 if ((hard_regno2
= ALLOCNO_HARD_REGNO (parent_allocno
)) < 0)
3508 cost
-= (ira_memory_move_cost
[mode
][rclass
][0] * exit_freq
3509 + ira_memory_move_cost
[mode
][rclass
][1] * enter_freq
);
3513 += (ira_memory_move_cost
[mode
][rclass
][1] * exit_freq
3514 + ira_memory_move_cost
[mode
][rclass
][0] * enter_freq
);
3515 if (hard_regno2
!= hard_regno
)
3516 cost
-= (ira_register_move_cost
[mode
][rclass
][rclass
]
3517 * (exit_freq
+ enter_freq
));
3522 ALLOCNO_HARD_REGNO (a
) = -1;
3523 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
3527 " Moving spill/restore for a%dr%d up from loop %d",
3528 ALLOCNO_NUM (a
), regno
, loop_node
->loop_num
);
3529 fprintf (ira_dump_file
, " - profit %d\n", -cost
);
3541 /* Update current hard reg costs and current conflict hard reg costs
3542 for allocno A. It is done by processing its copies containing
3543 other allocnos already assigned. */
3545 update_curr_costs (ira_allocno_t a
)
3547 int i
, hard_regno
, cost
;
3549 enum reg_class aclass
, rclass
;
3550 ira_allocno_t another_a
;
3551 ira_copy_t cp
, next_cp
;
3553 ira_free_allocno_updated_costs (a
);
3554 ira_assert (! ALLOCNO_ASSIGNED_P (a
));
3555 aclass
= ALLOCNO_CLASS (a
);
3556 if (aclass
== NO_REGS
)
3558 mode
= ALLOCNO_MODE (a
);
3559 ira_init_register_move_cost_if_necessary (mode
);
3560 for (cp
= ALLOCNO_COPIES (a
); cp
!= NULL
; cp
= next_cp
)
3564 next_cp
= cp
->next_first_allocno_copy
;
3565 another_a
= cp
->second
;
3567 else if (cp
->second
== a
)
3569 next_cp
= cp
->next_second_allocno_copy
;
3570 another_a
= cp
->first
;
3574 if (! ira_reg_classes_intersect_p
[aclass
][ALLOCNO_CLASS (another_a
)]
3575 || ! ALLOCNO_ASSIGNED_P (another_a
)
3576 || (hard_regno
= ALLOCNO_HARD_REGNO (another_a
)) < 0)
3578 rclass
= REGNO_REG_CLASS (hard_regno
);
3579 i
= ira_class_hard_reg_index
[aclass
][hard_regno
];
3582 cost
= (cp
->first
== a
3583 ? ira_register_move_cost
[mode
][rclass
][aclass
]
3584 : ira_register_move_cost
[mode
][aclass
][rclass
]);
3585 ira_allocate_and_set_or_copy_costs
3586 (&ALLOCNO_UPDATED_HARD_REG_COSTS (a
), aclass
, ALLOCNO_CLASS_COST (a
),
3587 ALLOCNO_HARD_REG_COSTS (a
));
3588 ira_allocate_and_set_or_copy_costs
3589 (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a
),
3590 aclass
, 0, ALLOCNO_CONFLICT_HARD_REG_COSTS (a
));
3591 ALLOCNO_UPDATED_HARD_REG_COSTS (a
)[i
] -= cp
->freq
* cost
;
3592 ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a
)[i
] -= cp
->freq
* cost
;
3596 /* Try to assign hard registers to the unassigned allocnos and
3597 allocnos conflicting with them or conflicting with allocnos whose
3598 regno >= START_REGNO. The function is called after ira_flattening,
3599 so more allocnos (including ones created in ira-emit.c) will have a
3600 chance to get a hard register. We use simple assignment algorithm
3601 based on priorities. */
3603 ira_reassign_conflict_allocnos (int start_regno
)
3605 int i
, allocnos_to_color_num
;
3607 enum reg_class aclass
;
3608 bitmap allocnos_to_color
;
3609 ira_allocno_iterator ai
;
3611 allocnos_to_color
= ira_allocate_bitmap ();
3612 allocnos_to_color_num
= 0;
3613 FOR_EACH_ALLOCNO (a
, ai
)
3615 int n
= ALLOCNO_NUM_OBJECTS (a
);
3617 if (! ALLOCNO_ASSIGNED_P (a
)
3618 && ! bitmap_bit_p (allocnos_to_color
, ALLOCNO_NUM (a
)))
3620 if (ALLOCNO_CLASS (a
) != NO_REGS
)
3621 sorted_allocnos
[allocnos_to_color_num
++] = a
;
3624 ALLOCNO_ASSIGNED_P (a
) = true;
3625 ALLOCNO_HARD_REGNO (a
) = -1;
3626 ira_assert (ALLOCNO_UPDATED_HARD_REG_COSTS (a
) == NULL
);
3627 ira_assert (ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a
) == NULL
);
3629 bitmap_set_bit (allocnos_to_color
, ALLOCNO_NUM (a
));
3631 if (ALLOCNO_REGNO (a
) < start_regno
3632 || (aclass
= ALLOCNO_CLASS (a
)) == NO_REGS
)
3634 for (i
= 0; i
< n
; i
++)
3636 ira_object_t obj
= ALLOCNO_OBJECT (a
, i
);
3637 ira_object_t conflict_obj
;
3638 ira_object_conflict_iterator oci
;
3640 FOR_EACH_OBJECT_CONFLICT (obj
, conflict_obj
, oci
)
3642 ira_allocno_t conflict_a
= OBJECT_ALLOCNO (conflict_obj
);
3644 ira_assert (ira_reg_classes_intersect_p
3645 [aclass
][ALLOCNO_CLASS (conflict_a
)]);
3646 if (!bitmap_set_bit (allocnos_to_color
, ALLOCNO_NUM (conflict_a
)))
3648 sorted_allocnos
[allocnos_to_color_num
++] = conflict_a
;
3652 ira_free_bitmap (allocnos_to_color
);
3653 if (allocnos_to_color_num
> 1)
3655 setup_allocno_priorities (sorted_allocnos
, allocnos_to_color_num
);
3656 qsort (sorted_allocnos
, allocnos_to_color_num
, sizeof (ira_allocno_t
),
3657 allocno_priority_compare_func
);
3659 for (i
= 0; i
< allocnos_to_color_num
; i
++)
3661 a
= sorted_allocnos
[i
];
3662 ALLOCNO_ASSIGNED_P (a
) = false;
3663 update_curr_costs (a
);
3665 for (i
= 0; i
< allocnos_to_color_num
; i
++)
3667 a
= sorted_allocnos
[i
];
3668 if (assign_hard_reg (a
, true))
3670 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
3673 " Secondary allocation: assign hard reg %d to reg %d\n",
3674 ALLOCNO_HARD_REGNO (a
), ALLOCNO_REGNO (a
));
3681 /* This page contains functions used to find conflicts using allocno
3684 #ifdef ENABLE_IRA_CHECKING
3686 /* Return TRUE if live ranges of pseudo-registers REGNO1 and REGNO2
3687 intersect. This should be used when there is only one region.
3688 Currently this is used during reload. */
3690 conflict_by_live_ranges_p (int regno1
, int regno2
)
3692 ira_allocno_t a1
, a2
;
3694 ira_assert (regno1
>= FIRST_PSEUDO_REGISTER
3695 && regno2
>= FIRST_PSEUDO_REGISTER
);
3696 /* Reg info calculated by dataflow infrastructure can be different
3697 from one calculated by regclass. */
3698 if ((a1
= ira_loop_tree_root
->regno_allocno_map
[regno1
]) == NULL
3699 || (a2
= ira_loop_tree_root
->regno_allocno_map
[regno2
]) == NULL
)
3701 return allocnos_conflict_by_live_ranges_p (a1
, a2
);
3708 /* This page contains code to coalesce memory stack slots used by
3709 spilled allocnos. This results in smaller stack frame, better data
3710 locality, and in smaller code for some architectures like
3711 x86/x86_64 where insn size depends on address displacement value.
3712 On the other hand, it can worsen insn scheduling after the RA but
3713 in practice it is less important than smaller stack frames. */
3715 /* TRUE if we coalesced some allocnos. In other words, if we got
3716 loops formed by members first_coalesced_allocno and
3717 next_coalesced_allocno containing more one allocno. */
3718 static bool allocno_coalesced_p
;
3720 /* Bitmap used to prevent a repeated allocno processing because of
3722 static bitmap processed_coalesced_allocno_bitmap
;
3725 typedef struct coalesce_data
*coalesce_data_t
;
3727 /* To decrease footprint of ira_allocno structure we store all data
3728 needed only for coalescing in the following structure. */
3729 struct coalesce_data
3731 /* Coalesced allocnos form a cyclic list. One allocno given by
3732 FIRST represents all coalesced allocnos. The
3733 list is chained by NEXT. */
3734 ira_allocno_t first
;
3739 /* Container for storing allocno data concerning coalescing. */
3740 static coalesce_data_t allocno_coalesce_data
;
3742 /* Macro to access the data concerning coalescing. */
3743 #define ALLOCNO_COALESCE_DATA(a) ((coalesce_data_t) ALLOCNO_ADD_DATA (a))
3745 /* Merge two sets of coalesced allocnos given correspondingly by
3746 allocnos A1 and A2 (more accurately merging A2 set into A1
3749 merge_allocnos (ira_allocno_t a1
, ira_allocno_t a2
)
3751 ira_allocno_t a
, first
, last
, next
;
3753 first
= ALLOCNO_COALESCE_DATA (a1
)->first
;
3754 a
= ALLOCNO_COALESCE_DATA (a2
)->first
;
3757 for (last
= a2
, a
= ALLOCNO_COALESCE_DATA (a2
)->next
;;
3758 a
= ALLOCNO_COALESCE_DATA (a
)->next
)
3760 ALLOCNO_COALESCE_DATA (a
)->first
= first
;
3765 next
= allocno_coalesce_data
[ALLOCNO_NUM (first
)].next
;
3766 allocno_coalesce_data
[ALLOCNO_NUM (first
)].next
= a2
;
3767 allocno_coalesce_data
[ALLOCNO_NUM (last
)].next
= next
;
3770 /* Return TRUE if there are conflicting allocnos from two sets of
3771 coalesced allocnos given correspondingly by allocnos A1 and A2. We
3772 use live ranges to find conflicts because conflicts are represented
3773 only for allocnos of the same allocno class and during the reload
3774 pass we coalesce allocnos for sharing stack memory slots. */
3776 coalesced_allocno_conflict_p (ira_allocno_t a1
, ira_allocno_t a2
)
3778 ira_allocno_t a
, conflict_a
;
3780 if (allocno_coalesced_p
)
3782 bitmap_clear (processed_coalesced_allocno_bitmap
);
3783 for (a
= ALLOCNO_COALESCE_DATA (a1
)->next
;;
3784 a
= ALLOCNO_COALESCE_DATA (a
)->next
)
3786 bitmap_set_bit (processed_coalesced_allocno_bitmap
, ALLOCNO_NUM (a
));
3791 for (a
= ALLOCNO_COALESCE_DATA (a2
)->next
;;
3792 a
= ALLOCNO_COALESCE_DATA (a
)->next
)
3794 for (conflict_a
= ALLOCNO_COALESCE_DATA (a1
)->next
;;
3795 conflict_a
= ALLOCNO_COALESCE_DATA (conflict_a
)->next
)
3797 if (allocnos_conflict_by_live_ranges_p (a
, conflict_a
))
3799 if (conflict_a
== a1
)
3808 /* The major function for aggressive allocno coalescing. We coalesce
3809 only spilled allocnos. If some allocnos have been coalesced, we
3810 set up flag allocno_coalesced_p. */
3812 coalesce_allocnos (void)
3815 ira_copy_t cp
, next_cp
;
3817 int i
, n
, cp_num
, regno
;
3821 /* Collect copies. */
3822 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap
, 0, j
, bi
)
3824 a
= ira_allocnos
[j
];
3825 regno
= ALLOCNO_REGNO (a
);
3826 if (! ALLOCNO_ASSIGNED_P (a
) || ALLOCNO_HARD_REGNO (a
) >= 0
3827 || ira_equiv_no_lvalue_p (regno
))
3829 for (cp
= ALLOCNO_COPIES (a
); cp
!= NULL
; cp
= next_cp
)
3833 next_cp
= cp
->next_first_allocno_copy
;
3834 regno
= ALLOCNO_REGNO (cp
->second
);
3835 /* For priority coloring we coalesce allocnos only with
3836 the same allocno class not with intersected allocno
3837 classes as it were possible. It is done for
3839 if ((cp
->insn
!= NULL
|| cp
->constraint_p
)
3840 && ALLOCNO_ASSIGNED_P (cp
->second
)
3841 && ALLOCNO_HARD_REGNO (cp
->second
) < 0
3842 && ! ira_equiv_no_lvalue_p (regno
))
3843 sorted_copies
[cp_num
++] = cp
;
3845 else if (cp
->second
== a
)
3846 next_cp
= cp
->next_second_allocno_copy
;
3851 qsort (sorted_copies
, cp_num
, sizeof (ira_copy_t
), copy_freq_compare_func
);
3852 /* Coalesced copies, most frequently executed first. */
3853 for (; cp_num
!= 0;)
3855 for (i
= 0; i
< cp_num
; i
++)
3857 cp
= sorted_copies
[i
];
3858 if (! coalesced_allocno_conflict_p (cp
->first
, cp
->second
))
3860 allocno_coalesced_p
= true;
3861 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
3864 " Coalescing copy %d:a%dr%d-a%dr%d (freq=%d)\n",
3865 cp
->num
, ALLOCNO_NUM (cp
->first
), ALLOCNO_REGNO (cp
->first
),
3866 ALLOCNO_NUM (cp
->second
), ALLOCNO_REGNO (cp
->second
),
3868 merge_allocnos (cp
->first
, cp
->second
);
3873 /* Collect the rest of copies. */
3874 for (n
= 0; i
< cp_num
; i
++)
3876 cp
= sorted_copies
[i
];
3877 if (allocno_coalesce_data
[ALLOCNO_NUM (cp
->first
)].first
3878 != allocno_coalesce_data
[ALLOCNO_NUM (cp
->second
)].first
)
3879 sorted_copies
[n
++] = cp
;
3885 /* Usage cost and order number of coalesced allocno set to which
3886 given pseudo register belongs to. */
3887 static int *regno_coalesced_allocno_cost
;
3888 static int *regno_coalesced_allocno_num
;
3890 /* Sort pseudos according frequencies of coalesced allocno sets they
3891 belong to (putting most frequently ones first), and according to
3892 coalesced allocno set order numbers. */
3894 coalesced_pseudo_reg_freq_compare (const void *v1p
, const void *v2p
)
3896 const int regno1
= *(const int *) v1p
;
3897 const int regno2
= *(const int *) v2p
;
3900 if ((diff
= (regno_coalesced_allocno_cost
[regno2
]
3901 - regno_coalesced_allocno_cost
[regno1
])) != 0)
3903 if ((diff
= (regno_coalesced_allocno_num
[regno1
]
3904 - regno_coalesced_allocno_num
[regno2
])) != 0)
3906 return regno1
- regno2
;
3909 /* Widest width in which each pseudo reg is referred to (via subreg).
3910 It is used for sorting pseudo registers. */
3911 static machine_mode
*regno_max_ref_mode
;
3913 /* Sort pseudos according their slot numbers (putting ones with
3914 smaller numbers first, or last when the frame pointer is not
3917 coalesced_pseudo_reg_slot_compare (const void *v1p
, const void *v2p
)
3919 const int regno1
= *(const int *) v1p
;
3920 const int regno2
= *(const int *) v2p
;
3921 ira_allocno_t a1
= ira_regno_allocno_map
[regno1
];
3922 ira_allocno_t a2
= ira_regno_allocno_map
[regno2
];
3923 int diff
, slot_num1
, slot_num2
;
3924 machine_mode mode1
, mode2
;
3926 if (a1
== NULL
|| ALLOCNO_HARD_REGNO (a1
) >= 0)
3928 if (a2
== NULL
|| ALLOCNO_HARD_REGNO (a2
) >= 0)
3929 return regno1
- regno2
;
3932 else if (a2
== NULL
|| ALLOCNO_HARD_REGNO (a2
) >= 0)
3934 slot_num1
= -ALLOCNO_HARD_REGNO (a1
);
3935 slot_num2
= -ALLOCNO_HARD_REGNO (a2
);
3936 if ((diff
= slot_num1
- slot_num2
) != 0)
3937 return (frame_pointer_needed
3938 || (!FRAME_GROWS_DOWNWARD
) == STACK_GROWS_DOWNWARD
? diff
: -diff
);
3939 mode1
= wider_subreg_mode (PSEUDO_REGNO_MODE (regno1
),
3940 regno_max_ref_mode
[regno1
]);
3941 mode2
= wider_subreg_mode (PSEUDO_REGNO_MODE (regno2
),
3942 regno_max_ref_mode
[regno2
]);
3943 if ((diff
= GET_MODE_SIZE (mode2
) - GET_MODE_SIZE (mode1
)) != 0)
3945 return regno1
- regno2
;
3948 /* Setup REGNO_COALESCED_ALLOCNO_COST and REGNO_COALESCED_ALLOCNO_NUM
3949 for coalesced allocno sets containing allocnos with their regnos
3950 given in array PSEUDO_REGNOS of length N. */
3952 setup_coalesced_allocno_costs_and_nums (int *pseudo_regnos
, int n
)
3954 int i
, num
, regno
, cost
;
3955 ira_allocno_t allocno
, a
;
3957 for (num
= i
= 0; i
< n
; i
++)
3959 regno
= pseudo_regnos
[i
];
3960 allocno
= ira_regno_allocno_map
[regno
];
3961 if (allocno
== NULL
)
3963 regno_coalesced_allocno_cost
[regno
] = 0;
3964 regno_coalesced_allocno_num
[regno
] = ++num
;
3967 if (ALLOCNO_COALESCE_DATA (allocno
)->first
!= allocno
)
3970 for (cost
= 0, a
= ALLOCNO_COALESCE_DATA (allocno
)->next
;;
3971 a
= ALLOCNO_COALESCE_DATA (a
)->next
)
3973 cost
+= ALLOCNO_FREQ (a
);
3977 for (a
= ALLOCNO_COALESCE_DATA (allocno
)->next
;;
3978 a
= ALLOCNO_COALESCE_DATA (a
)->next
)
3980 regno_coalesced_allocno_num
[ALLOCNO_REGNO (a
)] = num
;
3981 regno_coalesced_allocno_cost
[ALLOCNO_REGNO (a
)] = cost
;
3988 /* Collect spilled allocnos representing coalesced allocno sets (the
3989 first coalesced allocno). The collected allocnos are returned
3990 through array SPILLED_COALESCED_ALLOCNOS. The function returns the
3991 number of the collected allocnos. The allocnos are given by their
3992 regnos in array PSEUDO_REGNOS of length N. */
3994 collect_spilled_coalesced_allocnos (int *pseudo_regnos
, int n
,
3995 ira_allocno_t
*spilled_coalesced_allocnos
)
3998 ira_allocno_t allocno
;
4000 for (num
= i
= 0; i
< n
; i
++)
4002 regno
= pseudo_regnos
[i
];
4003 allocno
= ira_regno_allocno_map
[regno
];
4004 if (allocno
== NULL
|| ALLOCNO_HARD_REGNO (allocno
) >= 0
4005 || ALLOCNO_COALESCE_DATA (allocno
)->first
!= allocno
)
4007 spilled_coalesced_allocnos
[num
++] = allocno
;
4012 /* Array of live ranges of size IRA_ALLOCNOS_NUM. Live range for
4013 given slot contains live ranges of coalesced allocnos assigned to
4015 static live_range_t
*slot_coalesced_allocnos_live_ranges
;
4017 /* Return TRUE if coalesced allocnos represented by ALLOCNO has live
4018 ranges intersected with live ranges of coalesced allocnos assigned
4019 to slot with number N. */
4021 slot_coalesced_allocno_live_ranges_intersect_p (ira_allocno_t allocno
, int n
)
4025 for (a
= ALLOCNO_COALESCE_DATA (allocno
)->next
;;
4026 a
= ALLOCNO_COALESCE_DATA (a
)->next
)
4029 int nr
= ALLOCNO_NUM_OBJECTS (a
);
4031 for (i
= 0; i
< nr
; i
++)
4033 ira_object_t obj
= ALLOCNO_OBJECT (a
, i
);
4035 if (ira_live_ranges_intersect_p
4036 (slot_coalesced_allocnos_live_ranges
[n
],
4037 OBJECT_LIVE_RANGES (obj
)))
4046 /* Update live ranges of slot to which coalesced allocnos represented
4047 by ALLOCNO were assigned. */
4049 setup_slot_coalesced_allocno_live_ranges (ira_allocno_t allocno
)
4055 n
= ALLOCNO_COALESCE_DATA (allocno
)->temp
;
4056 for (a
= ALLOCNO_COALESCE_DATA (allocno
)->next
;;
4057 a
= ALLOCNO_COALESCE_DATA (a
)->next
)
4059 int nr
= ALLOCNO_NUM_OBJECTS (a
);
4060 for (i
= 0; i
< nr
; i
++)
4062 ira_object_t obj
= ALLOCNO_OBJECT (a
, i
);
4064 r
= ira_copy_live_range_list (OBJECT_LIVE_RANGES (obj
));
4065 slot_coalesced_allocnos_live_ranges
[n
]
4066 = ira_merge_live_ranges
4067 (slot_coalesced_allocnos_live_ranges
[n
], r
);
4074 /* We have coalesced allocnos involving in copies. Coalesce allocnos
4075 further in order to share the same memory stack slot. Allocnos
4076 representing sets of allocnos coalesced before the call are given
4077 in array SPILLED_COALESCED_ALLOCNOS of length NUM. Return TRUE if
4078 some allocnos were coalesced in the function. */
4080 coalesce_spill_slots (ira_allocno_t
*spilled_coalesced_allocnos
, int num
)
4082 int i
, j
, n
, last_coalesced_allocno_num
;
4083 ira_allocno_t allocno
, a
;
4084 bool merged_p
= false;
4085 bitmap set_jump_crosses
= regstat_get_setjmp_crosses ();
4087 slot_coalesced_allocnos_live_ranges
4088 = (live_range_t
*) ira_allocate (sizeof (live_range_t
) * ira_allocnos_num
);
4089 memset (slot_coalesced_allocnos_live_ranges
, 0,
4090 sizeof (live_range_t
) * ira_allocnos_num
);
4091 last_coalesced_allocno_num
= 0;
4092 /* Coalesce non-conflicting spilled allocnos preferring most
4094 for (i
= 0; i
< num
; i
++)
4096 allocno
= spilled_coalesced_allocnos
[i
];
4097 if (ALLOCNO_COALESCE_DATA (allocno
)->first
!= allocno
4098 || bitmap_bit_p (set_jump_crosses
, ALLOCNO_REGNO (allocno
))
4099 || ira_equiv_no_lvalue_p (ALLOCNO_REGNO (allocno
)))
4101 for (j
= 0; j
< i
; j
++)
4103 a
= spilled_coalesced_allocnos
[j
];
4104 n
= ALLOCNO_COALESCE_DATA (a
)->temp
;
4105 if (ALLOCNO_COALESCE_DATA (a
)->first
== a
4106 && ! bitmap_bit_p (set_jump_crosses
, ALLOCNO_REGNO (a
))
4107 && ! ira_equiv_no_lvalue_p (ALLOCNO_REGNO (a
))
4108 && ! slot_coalesced_allocno_live_ranges_intersect_p (allocno
, n
))
4113 /* No coalescing: set up number for coalesced allocnos
4114 represented by ALLOCNO. */
4115 ALLOCNO_COALESCE_DATA (allocno
)->temp
= last_coalesced_allocno_num
++;
4116 setup_slot_coalesced_allocno_live_ranges (allocno
);
4120 allocno_coalesced_p
= true;
4122 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
4123 fprintf (ira_dump_file
,
4124 " Coalescing spilled allocnos a%dr%d->a%dr%d\n",
4125 ALLOCNO_NUM (allocno
), ALLOCNO_REGNO (allocno
),
4126 ALLOCNO_NUM (a
), ALLOCNO_REGNO (a
));
4127 ALLOCNO_COALESCE_DATA (allocno
)->temp
4128 = ALLOCNO_COALESCE_DATA (a
)->temp
;
4129 setup_slot_coalesced_allocno_live_ranges (allocno
);
4130 merge_allocnos (a
, allocno
);
4131 ira_assert (ALLOCNO_COALESCE_DATA (a
)->first
== a
);
4134 for (i
= 0; i
< ira_allocnos_num
; i
++)
4135 ira_finish_live_range_list (slot_coalesced_allocnos_live_ranges
[i
]);
4136 ira_free (slot_coalesced_allocnos_live_ranges
);
4140 /* Sort pseudo-register numbers in array PSEUDO_REGNOS of length N for
4141 subsequent assigning stack slots to them in the reload pass. To do
4142 this we coalesce spilled allocnos first to decrease the number of
4143 memory-memory move insns. This function is called by the
4146 ira_sort_regnos_for_alter_reg (int *pseudo_regnos
, int n
,
4147 machine_mode
*reg_max_ref_mode
)
4149 int max_regno
= max_reg_num ();
4150 int i
, regno
, num
, slot_num
;
4151 ira_allocno_t allocno
, a
;
4152 ira_allocno_iterator ai
;
4153 ira_allocno_t
*spilled_coalesced_allocnos
;
4155 ira_assert (! ira_use_lra_p
);
4157 /* Set up allocnos can be coalesced. */
4158 coloring_allocno_bitmap
= ira_allocate_bitmap ();
4159 for (i
= 0; i
< n
; i
++)
4161 regno
= pseudo_regnos
[i
];
4162 allocno
= ira_regno_allocno_map
[regno
];
4163 if (allocno
!= NULL
)
4164 bitmap_set_bit (coloring_allocno_bitmap
, ALLOCNO_NUM (allocno
));
4166 allocno_coalesced_p
= false;
4167 processed_coalesced_allocno_bitmap
= ira_allocate_bitmap ();
4168 allocno_coalesce_data
4169 = (coalesce_data_t
) ira_allocate (sizeof (struct coalesce_data
)
4170 * ira_allocnos_num
);
4171 /* Initialize coalesce data for allocnos. */
4172 FOR_EACH_ALLOCNO (a
, ai
)
4174 ALLOCNO_ADD_DATA (a
) = allocno_coalesce_data
+ ALLOCNO_NUM (a
);
4175 ALLOCNO_COALESCE_DATA (a
)->first
= a
;
4176 ALLOCNO_COALESCE_DATA (a
)->next
= a
;
4178 coalesce_allocnos ();
4179 ira_free_bitmap (coloring_allocno_bitmap
);
4180 regno_coalesced_allocno_cost
4181 = (int *) ira_allocate (max_regno
* sizeof (int));
4182 regno_coalesced_allocno_num
4183 = (int *) ira_allocate (max_regno
* sizeof (int));
4184 memset (regno_coalesced_allocno_num
, 0, max_regno
* sizeof (int));
4185 setup_coalesced_allocno_costs_and_nums (pseudo_regnos
, n
);
4186 /* Sort regnos according frequencies of the corresponding coalesced
4188 qsort (pseudo_regnos
, n
, sizeof (int), coalesced_pseudo_reg_freq_compare
);
4189 spilled_coalesced_allocnos
4190 = (ira_allocno_t
*) ira_allocate (ira_allocnos_num
4191 * sizeof (ira_allocno_t
));
4192 /* Collect allocnos representing the spilled coalesced allocno
4194 num
= collect_spilled_coalesced_allocnos (pseudo_regnos
, n
,
4195 spilled_coalesced_allocnos
);
4196 if (flag_ira_share_spill_slots
4197 && coalesce_spill_slots (spilled_coalesced_allocnos
, num
))
4199 setup_coalesced_allocno_costs_and_nums (pseudo_regnos
, n
);
4200 qsort (pseudo_regnos
, n
, sizeof (int),
4201 coalesced_pseudo_reg_freq_compare
);
4202 num
= collect_spilled_coalesced_allocnos (pseudo_regnos
, n
,
4203 spilled_coalesced_allocnos
);
4205 ira_free_bitmap (processed_coalesced_allocno_bitmap
);
4206 allocno_coalesced_p
= false;
4207 /* Assign stack slot numbers to spilled allocno sets, use smaller
4208 numbers for most frequently used coalesced allocnos. -1 is
4209 reserved for dynamic search of stack slots for pseudos spilled by
4212 for (i
= 0; i
< num
; i
++)
4214 allocno
= spilled_coalesced_allocnos
[i
];
4215 if (ALLOCNO_COALESCE_DATA (allocno
)->first
!= allocno
4216 || ALLOCNO_HARD_REGNO (allocno
) >= 0
4217 || ira_equiv_no_lvalue_p (ALLOCNO_REGNO (allocno
)))
4219 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
4220 fprintf (ira_dump_file
, " Slot %d (freq,size):", slot_num
);
4222 for (a
= ALLOCNO_COALESCE_DATA (allocno
)->next
;;
4223 a
= ALLOCNO_COALESCE_DATA (a
)->next
)
4225 ira_assert (ALLOCNO_HARD_REGNO (a
) < 0);
4226 ALLOCNO_HARD_REGNO (a
) = -slot_num
;
4227 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
4229 machine_mode mode
= wider_subreg_mode
4230 (PSEUDO_REGNO_MODE (ALLOCNO_REGNO (a
)),
4231 reg_max_ref_mode
[ALLOCNO_REGNO (a
)]);
4232 fprintf (ira_dump_file
, " a%dr%d(%d,%d)",
4233 ALLOCNO_NUM (a
), ALLOCNO_REGNO (a
), ALLOCNO_FREQ (a
),
4234 GET_MODE_SIZE (mode
));
4240 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
4241 fprintf (ira_dump_file
, "\n");
4243 ira_spilled_reg_stack_slots_num
= slot_num
- 1;
4244 ira_free (spilled_coalesced_allocnos
);
4245 /* Sort regnos according the slot numbers. */
4246 regno_max_ref_mode
= reg_max_ref_mode
;
4247 qsort (pseudo_regnos
, n
, sizeof (int), coalesced_pseudo_reg_slot_compare
);
4248 FOR_EACH_ALLOCNO (a
, ai
)
4249 ALLOCNO_ADD_DATA (a
) = NULL
;
4250 ira_free (allocno_coalesce_data
);
4251 ira_free (regno_coalesced_allocno_num
);
4252 ira_free (regno_coalesced_allocno_cost
);
4257 /* This page contains code used by the reload pass to improve the
4260 /* The function is called from reload to mark changes in the
4261 allocation of REGNO made by the reload. Remember that reg_renumber
4262 reflects the change result. */
4264 ira_mark_allocation_change (int regno
)
4266 ira_allocno_t a
= ira_regno_allocno_map
[regno
];
4267 int old_hard_regno
, hard_regno
, cost
;
4268 enum reg_class aclass
= ALLOCNO_CLASS (a
);
4270 ira_assert (a
!= NULL
);
4271 hard_regno
= reg_renumber
[regno
];
4272 if ((old_hard_regno
= ALLOCNO_HARD_REGNO (a
)) == hard_regno
)
4274 if (old_hard_regno
< 0)
4275 cost
= -ALLOCNO_MEMORY_COST (a
);
4278 ira_assert (ira_class_hard_reg_index
[aclass
][old_hard_regno
] >= 0);
4279 cost
= -(ALLOCNO_HARD_REG_COSTS (a
) == NULL
4280 ? ALLOCNO_CLASS_COST (a
)
4281 : ALLOCNO_HARD_REG_COSTS (a
)
4282 [ira_class_hard_reg_index
[aclass
][old_hard_regno
]]);
4283 update_costs_from_copies (a
, false, false);
4285 ira_overall_cost
-= cost
;
4286 ALLOCNO_HARD_REGNO (a
) = hard_regno
;
4289 ALLOCNO_HARD_REGNO (a
) = -1;
4290 cost
+= ALLOCNO_MEMORY_COST (a
);
4292 else if (ira_class_hard_reg_index
[aclass
][hard_regno
] >= 0)
4294 cost
+= (ALLOCNO_HARD_REG_COSTS (a
) == NULL
4295 ? ALLOCNO_CLASS_COST (a
)
4296 : ALLOCNO_HARD_REG_COSTS (a
)
4297 [ira_class_hard_reg_index
[aclass
][hard_regno
]]);
4298 update_costs_from_copies (a
, true, false);
4301 /* Reload changed class of the allocno. */
4303 ira_overall_cost
+= cost
;
4306 /* This function is called when reload deletes memory-memory move. In
4307 this case we marks that the allocation of the corresponding
4308 allocnos should be not changed in future. Otherwise we risk to get
4311 ira_mark_memory_move_deletion (int dst_regno
, int src_regno
)
4313 ira_allocno_t dst
= ira_regno_allocno_map
[dst_regno
];
4314 ira_allocno_t src
= ira_regno_allocno_map
[src_regno
];
4316 ira_assert (dst
!= NULL
&& src
!= NULL
4317 && ALLOCNO_HARD_REGNO (dst
) < 0
4318 && ALLOCNO_HARD_REGNO (src
) < 0);
4319 ALLOCNO_DONT_REASSIGN_P (dst
) = true;
4320 ALLOCNO_DONT_REASSIGN_P (src
) = true;
4323 /* Try to assign a hard register (except for FORBIDDEN_REGS) to
4324 allocno A and return TRUE in the case of success. */
4326 allocno_reload_assign (ira_allocno_t a
, HARD_REG_SET forbidden_regs
)
4329 enum reg_class aclass
;
4330 int regno
= ALLOCNO_REGNO (a
);
4331 HARD_REG_SET saved
[2];
4334 n
= ALLOCNO_NUM_OBJECTS (a
);
4335 for (i
= 0; i
< n
; i
++)
4337 ira_object_t obj
= ALLOCNO_OBJECT (a
, i
);
4338 COPY_HARD_REG_SET (saved
[i
], OBJECT_TOTAL_CONFLICT_HARD_REGS (obj
));
4339 IOR_HARD_REG_SET (OBJECT_TOTAL_CONFLICT_HARD_REGS (obj
), forbidden_regs
);
4340 if (! flag_caller_saves
&& ALLOCNO_CALLS_CROSSED_NUM (a
) != 0)
4341 IOR_HARD_REG_SET (OBJECT_TOTAL_CONFLICT_HARD_REGS (obj
),
4344 ALLOCNO_ASSIGNED_P (a
) = false;
4345 aclass
= ALLOCNO_CLASS (a
);
4346 update_curr_costs (a
);
4347 assign_hard_reg (a
, true);
4348 hard_regno
= ALLOCNO_HARD_REGNO (a
);
4349 reg_renumber
[regno
] = hard_regno
;
4351 ALLOCNO_HARD_REGNO (a
) = -1;
4354 ira_assert (ira_class_hard_reg_index
[aclass
][hard_regno
] >= 0);
4356 -= (ALLOCNO_MEMORY_COST (a
)
4357 - (ALLOCNO_HARD_REG_COSTS (a
) == NULL
4358 ? ALLOCNO_CLASS_COST (a
)
4359 : ALLOCNO_HARD_REG_COSTS (a
)[ira_class_hard_reg_index
4360 [aclass
][hard_regno
]]));
4361 if (ALLOCNO_CALLS_CROSSED_NUM (a
) != 0
4362 && ira_hard_reg_set_intersection_p (hard_regno
, ALLOCNO_MODE (a
),
4365 ira_assert (flag_caller_saves
);
4366 caller_save_needed
= 1;
4370 /* If we found a hard register, modify the RTL for the pseudo
4371 register to show the hard register, and mark the pseudo register
4373 if (reg_renumber
[regno
] >= 0)
4375 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
4376 fprintf (ira_dump_file
, ": reassign to %d\n", reg_renumber
[regno
]);
4377 SET_REGNO (regno_reg_rtx
[regno
], reg_renumber
[regno
]);
4378 mark_home_live (regno
);
4380 else if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
4381 fprintf (ira_dump_file
, "\n");
4382 for (i
= 0; i
< n
; i
++)
4384 ira_object_t obj
= ALLOCNO_OBJECT (a
, i
);
4385 COPY_HARD_REG_SET (OBJECT_TOTAL_CONFLICT_HARD_REGS (obj
), saved
[i
]);
4387 return reg_renumber
[regno
] >= 0;
4390 /* Sort pseudos according their usage frequencies (putting most
4391 frequently ones first). */
4393 pseudo_reg_compare (const void *v1p
, const void *v2p
)
4395 int regno1
= *(const int *) v1p
;
4396 int regno2
= *(const int *) v2p
;
4399 if ((diff
= REG_FREQ (regno2
) - REG_FREQ (regno1
)) != 0)
4401 return regno1
- regno2
;
4404 /* Try to allocate hard registers to SPILLED_PSEUDO_REGS (there are
4405 NUM of them) or spilled pseudos conflicting with pseudos in
4406 SPILLED_PSEUDO_REGS. Return TRUE and update SPILLED, if the
4407 allocation has been changed. The function doesn't use
4408 BAD_SPILL_REGS and hard registers in PSEUDO_FORBIDDEN_REGS and
4409 PSEUDO_PREVIOUS_REGS for the corresponding pseudos. The function
4410 is called by the reload pass at the end of each reload
4413 ira_reassign_pseudos (int *spilled_pseudo_regs
, int num
,
4414 HARD_REG_SET bad_spill_regs
,
4415 HARD_REG_SET
*pseudo_forbidden_regs
,
4416 HARD_REG_SET
*pseudo_previous_regs
,
4422 HARD_REG_SET forbidden_regs
;
4423 bitmap temp
= BITMAP_ALLOC (NULL
);
4425 /* Add pseudos which conflict with pseudos already in
4426 SPILLED_PSEUDO_REGS to SPILLED_PSEUDO_REGS. This is preferable
4427 to allocating in two steps as some of the conflicts might have
4428 a higher priority than the pseudos passed in SPILLED_PSEUDO_REGS. */
4429 for (i
= 0; i
< num
; i
++)
4430 bitmap_set_bit (temp
, spilled_pseudo_regs
[i
]);
4432 for (i
= 0, n
= num
; i
< n
; i
++)
4435 int regno
= spilled_pseudo_regs
[i
];
4436 bitmap_set_bit (temp
, regno
);
4438 a
= ira_regno_allocno_map
[regno
];
4439 nr
= ALLOCNO_NUM_OBJECTS (a
);
4440 for (j
= 0; j
< nr
; j
++)
4442 ira_object_t conflict_obj
;
4443 ira_object_t obj
= ALLOCNO_OBJECT (a
, j
);
4444 ira_object_conflict_iterator oci
;
4446 FOR_EACH_OBJECT_CONFLICT (obj
, conflict_obj
, oci
)
4448 ira_allocno_t conflict_a
= OBJECT_ALLOCNO (conflict_obj
);
4449 if (ALLOCNO_HARD_REGNO (conflict_a
) < 0
4450 && ! ALLOCNO_DONT_REASSIGN_P (conflict_a
)
4451 && bitmap_set_bit (temp
, ALLOCNO_REGNO (conflict_a
)))
4453 spilled_pseudo_regs
[num
++] = ALLOCNO_REGNO (conflict_a
);
4454 /* ?!? This seems wrong. */
4455 bitmap_set_bit (consideration_allocno_bitmap
,
4456 ALLOCNO_NUM (conflict_a
));
4463 qsort (spilled_pseudo_regs
, num
, sizeof (int), pseudo_reg_compare
);
4465 /* Try to assign hard registers to pseudos from
4466 SPILLED_PSEUDO_REGS. */
4467 for (i
= 0; i
< num
; i
++)
4469 regno
= spilled_pseudo_regs
[i
];
4470 COPY_HARD_REG_SET (forbidden_regs
, bad_spill_regs
);
4471 IOR_HARD_REG_SET (forbidden_regs
, pseudo_forbidden_regs
[regno
]);
4472 IOR_HARD_REG_SET (forbidden_regs
, pseudo_previous_regs
[regno
]);
4473 gcc_assert (reg_renumber
[regno
] < 0);
4474 a
= ira_regno_allocno_map
[regno
];
4475 ira_mark_allocation_change (regno
);
4476 ira_assert (reg_renumber
[regno
] < 0);
4477 if (internal_flag_ira_verbose
> 3 && ira_dump_file
!= NULL
)
4478 fprintf (ira_dump_file
,
4479 " Try Assign %d(a%d), cost=%d", regno
, ALLOCNO_NUM (a
),
4480 ALLOCNO_MEMORY_COST (a
)
4481 - ALLOCNO_CLASS_COST (a
));
4482 allocno_reload_assign (a
, forbidden_regs
);
4483 if (reg_renumber
[regno
] >= 0)
4485 CLEAR_REGNO_REG_SET (spilled
, regno
);
4493 /* The function is called by reload and returns already allocated
4494 stack slot (if any) for REGNO with given INHERENT_SIZE and
4495 TOTAL_SIZE. In the case of failure to find a slot which can be
4496 used for REGNO, the function returns NULL. */
4498 ira_reuse_stack_slot (int regno
, unsigned int inherent_size
,
4499 unsigned int total_size
)
4502 int slot_num
, best_slot_num
;
4503 int cost
, best_cost
;
4504 ira_copy_t cp
, next_cp
;
4505 ira_allocno_t another_allocno
, allocno
= ira_regno_allocno_map
[regno
];
4508 struct ira_spilled_reg_stack_slot
*slot
= NULL
;
4510 ira_assert (! ira_use_lra_p
);
4512 ira_assert (inherent_size
== PSEUDO_REGNO_BYTES (regno
)
4513 && inherent_size
<= total_size
4514 && ALLOCNO_HARD_REGNO (allocno
) < 0);
4515 if (! flag_ira_share_spill_slots
)
4517 slot_num
= -ALLOCNO_HARD_REGNO (allocno
) - 2;
4520 slot
= &ira_spilled_reg_stack_slots
[slot_num
];
4525 best_cost
= best_slot_num
= -1;
4527 /* It means that the pseudo was spilled in the reload pass, try
4530 slot_num
< ira_spilled_reg_stack_slots_num
;
4533 slot
= &ira_spilled_reg_stack_slots
[slot_num
];
4534 if (slot
->mem
== NULL_RTX
)
4536 if (slot
->width
< total_size
4537 || GET_MODE_SIZE (GET_MODE (slot
->mem
)) < inherent_size
)
4540 EXECUTE_IF_SET_IN_BITMAP (&slot
->spilled_regs
,
4541 FIRST_PSEUDO_REGISTER
, i
, bi
)
4543 another_allocno
= ira_regno_allocno_map
[i
];
4544 if (allocnos_conflict_by_live_ranges_p (allocno
,
4548 for (cost
= 0, cp
= ALLOCNO_COPIES (allocno
);
4552 if (cp
->first
== allocno
)
4554 next_cp
= cp
->next_first_allocno_copy
;
4555 another_allocno
= cp
->second
;
4557 else if (cp
->second
== allocno
)
4559 next_cp
= cp
->next_second_allocno_copy
;
4560 another_allocno
= cp
->first
;
4564 if (cp
->insn
== NULL_RTX
)
4566 if (bitmap_bit_p (&slot
->spilled_regs
,
4567 ALLOCNO_REGNO (another_allocno
)))
4570 if (cost
> best_cost
)
4573 best_slot_num
= slot_num
;
4580 slot_num
= best_slot_num
;
4581 slot
= &ira_spilled_reg_stack_slots
[slot_num
];
4582 SET_REGNO_REG_SET (&slot
->spilled_regs
, regno
);
4584 ALLOCNO_HARD_REGNO (allocno
) = -slot_num
- 2;
4589 ira_assert (slot
->width
>= total_size
);
4590 #ifdef ENABLE_IRA_CHECKING
4591 EXECUTE_IF_SET_IN_BITMAP (&slot
->spilled_regs
,
4592 FIRST_PSEUDO_REGISTER
, i
, bi
)
4594 ira_assert (! conflict_by_live_ranges_p (regno
, i
));
4597 SET_REGNO_REG_SET (&slot
->spilled_regs
, regno
);
4598 if (internal_flag_ira_verbose
> 3 && ira_dump_file
)
4600 fprintf (ira_dump_file
, " Assigning %d(freq=%d) slot %d of",
4601 regno
, REG_FREQ (regno
), slot_num
);
4602 EXECUTE_IF_SET_IN_BITMAP (&slot
->spilled_regs
,
4603 FIRST_PSEUDO_REGISTER
, i
, bi
)
4605 if ((unsigned) regno
!= i
)
4606 fprintf (ira_dump_file
, " %d", i
);
4608 fprintf (ira_dump_file
, "\n");
4614 /* This is called by reload every time a new stack slot X with
4615 TOTAL_SIZE was allocated for REGNO. We store this info for
4616 subsequent ira_reuse_stack_slot calls. */
4618 ira_mark_new_stack_slot (rtx x
, int regno
, unsigned int total_size
)
4620 struct ira_spilled_reg_stack_slot
*slot
;
4622 ira_allocno_t allocno
;
4624 ira_assert (! ira_use_lra_p
);
4626 ira_assert (PSEUDO_REGNO_BYTES (regno
) <= total_size
);
4627 allocno
= ira_regno_allocno_map
[regno
];
4628 slot_num
= -ALLOCNO_HARD_REGNO (allocno
) - 2;
4631 slot_num
= ira_spilled_reg_stack_slots_num
++;
4632 ALLOCNO_HARD_REGNO (allocno
) = -slot_num
- 2;
4634 slot
= &ira_spilled_reg_stack_slots
[slot_num
];
4635 INIT_REG_SET (&slot
->spilled_regs
);
4636 SET_REGNO_REG_SET (&slot
->spilled_regs
, regno
);
4638 slot
->width
= total_size
;
4639 if (internal_flag_ira_verbose
> 3 && ira_dump_file
)
4640 fprintf (ira_dump_file
, " Assigning %d(freq=%d) a new slot %d\n",
4641 regno
, REG_FREQ (regno
), slot_num
);
4645 /* Return spill cost for pseudo-registers whose numbers are in array
4646 REGNOS (with a negative number as an end marker) for reload with
4647 given IN and OUT for INSN. Return also number points (through
4648 EXCESS_PRESSURE_LIVE_LENGTH) where the pseudo-register lives and
4649 the register pressure is high, number of references of the
4650 pseudo-registers (through NREFS), number of callee-clobbered
4651 hard-registers occupied by the pseudo-registers (through
4652 CALL_USED_COUNT), and the first hard regno occupied by the
4653 pseudo-registers (through FIRST_HARD_REGNO). */
4655 calculate_spill_cost (int *regnos
, rtx in
, rtx out
, rtx_insn
*insn
,
4656 int *excess_pressure_live_length
,
4657 int *nrefs
, int *call_used_count
, int *first_hard_regno
)
4659 int i
, cost
, regno
, hard_regno
, j
, count
, saved_cost
, nregs
;
4665 for (length
= count
= cost
= i
= 0;; i
++)
4670 *nrefs
+= REG_N_REFS (regno
);
4671 hard_regno
= reg_renumber
[regno
];
4672 ira_assert (hard_regno
>= 0);
4673 a
= ira_regno_allocno_map
[regno
];
4674 length
+= ALLOCNO_EXCESS_PRESSURE_POINTS_NUM (a
) / ALLOCNO_NUM_OBJECTS (a
);
4675 cost
+= ALLOCNO_MEMORY_COST (a
) - ALLOCNO_CLASS_COST (a
);
4676 nregs
= hard_regno_nregs (hard_regno
, ALLOCNO_MODE (a
));
4677 for (j
= 0; j
< nregs
; j
++)
4678 if (! TEST_HARD_REG_BIT (call_used_reg_set
, hard_regno
+ j
))
4682 in_p
= in
&& REG_P (in
) && (int) REGNO (in
) == hard_regno
;
4683 out_p
= out
&& REG_P (out
) && (int) REGNO (out
) == hard_regno
;
4685 && find_regno_note (insn
, REG_DEAD
, hard_regno
) != NULL_RTX
)
4689 saved_cost
+= ira_memory_move_cost
4690 [ALLOCNO_MODE (a
)][ALLOCNO_CLASS (a
)][1];
4693 += ira_memory_move_cost
4694 [ALLOCNO_MODE (a
)][ALLOCNO_CLASS (a
)][0];
4695 cost
-= REG_FREQ_FROM_BB (BLOCK_FOR_INSN (insn
)) * saved_cost
;
4698 *excess_pressure_live_length
= length
;
4699 *call_used_count
= count
;
4703 hard_regno
= reg_renumber
[regnos
[0]];
4705 *first_hard_regno
= hard_regno
;
4709 /* Return TRUE if spilling pseudo-registers whose numbers are in array
4710 REGNOS is better than spilling pseudo-registers with numbers in
4711 OTHER_REGNOS for reload with given IN and OUT for INSN. The
4712 function used by the reload pass to make better register spilling
4715 ira_better_spill_reload_regno_p (int *regnos
, int *other_regnos
,
4716 rtx in
, rtx out
, rtx_insn
*insn
)
4718 int cost
, other_cost
;
4719 int length
, other_length
;
4720 int nrefs
, other_nrefs
;
4721 int call_used_count
, other_call_used_count
;
4722 int hard_regno
, other_hard_regno
;
4724 cost
= calculate_spill_cost (regnos
, in
, out
, insn
,
4725 &length
, &nrefs
, &call_used_count
, &hard_regno
);
4726 other_cost
= calculate_spill_cost (other_regnos
, in
, out
, insn
,
4727 &other_length
, &other_nrefs
,
4728 &other_call_used_count
,
4730 if (nrefs
== 0 && other_nrefs
!= 0)
4732 if (nrefs
!= 0 && other_nrefs
== 0)
4734 if (cost
!= other_cost
)
4735 return cost
< other_cost
;
4736 if (length
!= other_length
)
4737 return length
> other_length
;
4738 #ifdef REG_ALLOC_ORDER
4739 if (hard_regno
>= 0 && other_hard_regno
>= 0)
4740 return (inv_reg_alloc_order
[hard_regno
]
4741 < inv_reg_alloc_order
[other_hard_regno
]);
4743 if (call_used_count
!= other_call_used_count
)
4744 return call_used_count
> other_call_used_count
;
4751 /* Allocate and initialize data necessary for assign_hard_reg. */
4753 ira_initiate_assign (void)
4756 = (ira_allocno_t
*) ira_allocate (sizeof (ira_allocno_t
)
4757 * ira_allocnos_num
);
4758 consideration_allocno_bitmap
= ira_allocate_bitmap ();
4759 initiate_cost_update ();
4760 allocno_priorities
= (int *) ira_allocate (sizeof (int) * ira_allocnos_num
);
4761 sorted_copies
= (ira_copy_t
*) ira_allocate (ira_copies_num
4762 * sizeof (ira_copy_t
));
4765 /* Deallocate data used by assign_hard_reg. */
4767 ira_finish_assign (void)
4769 ira_free (sorted_allocnos
);
4770 ira_free_bitmap (consideration_allocno_bitmap
);
4771 finish_cost_update ();
4772 ira_free (allocno_priorities
);
4773 ira_free (sorted_copies
);
4778 /* Entry function doing color-based register allocation. */
4782 allocno_stack_vec
.create (ira_allocnos_num
);
4783 memset (allocated_hardreg_p
, 0, sizeof (allocated_hardreg_p
));
4784 ira_initiate_assign ();
4786 ira_finish_assign ();
4787 allocno_stack_vec
.release ();
4788 move_spill_restore ();
4793 /* This page contains a simple register allocator without usage of
4794 allocno conflicts. This is used for fast allocation for -O0. */
4796 /* Do register allocation by not using allocno conflicts. It uses
4797 only allocno live ranges. The algorithm is close to Chow's
4798 priority coloring. */
4800 fast_allocation (void)
4802 int i
, j
, k
, num
, class_size
, hard_regno
;
4804 bool no_stack_reg_p
;
4806 enum reg_class aclass
;
4809 ira_allocno_iterator ai
;
4811 HARD_REG_SET conflict_hard_regs
, *used_hard_regs
;
4813 sorted_allocnos
= (ira_allocno_t
*) ira_allocate (sizeof (ira_allocno_t
)
4814 * ira_allocnos_num
);
4816 FOR_EACH_ALLOCNO (a
, ai
)
4817 sorted_allocnos
[num
++] = a
;
4818 allocno_priorities
= (int *) ira_allocate (sizeof (int) * ira_allocnos_num
);
4819 setup_allocno_priorities (sorted_allocnos
, num
);
4820 used_hard_regs
= (HARD_REG_SET
*) ira_allocate (sizeof (HARD_REG_SET
)
4822 for (i
= 0; i
< ira_max_point
; i
++)
4823 CLEAR_HARD_REG_SET (used_hard_regs
[i
]);
4824 qsort (sorted_allocnos
, num
, sizeof (ira_allocno_t
),
4825 allocno_priority_compare_func
);
4826 for (i
= 0; i
< num
; i
++)
4830 a
= sorted_allocnos
[i
];
4831 nr
= ALLOCNO_NUM_OBJECTS (a
);
4832 CLEAR_HARD_REG_SET (conflict_hard_regs
);
4833 for (l
= 0; l
< nr
; l
++)
4835 ira_object_t obj
= ALLOCNO_OBJECT (a
, l
);
4836 IOR_HARD_REG_SET (conflict_hard_regs
,
4837 OBJECT_CONFLICT_HARD_REGS (obj
));
4838 for (r
= OBJECT_LIVE_RANGES (obj
); r
!= NULL
; r
= r
->next
)
4839 for (j
= r
->start
; j
<= r
->finish
; j
++)
4840 IOR_HARD_REG_SET (conflict_hard_regs
, used_hard_regs
[j
]);
4842 aclass
= ALLOCNO_CLASS (a
);
4843 ALLOCNO_ASSIGNED_P (a
) = true;
4844 ALLOCNO_HARD_REGNO (a
) = -1;
4845 if (hard_reg_set_subset_p (reg_class_contents
[aclass
],
4846 conflict_hard_regs
))
4848 mode
= ALLOCNO_MODE (a
);
4850 no_stack_reg_p
= ALLOCNO_NO_STACK_REG_P (a
);
4852 class_size
= ira_class_hard_regs_num
[aclass
];
4853 for (j
= 0; j
< class_size
; j
++)
4855 hard_regno
= ira_class_hard_regs
[aclass
][j
];
4857 if (no_stack_reg_p
&& FIRST_STACK_REG
<= hard_regno
4858 && hard_regno
<= LAST_STACK_REG
)
4861 if (ira_hard_reg_set_intersection_p (hard_regno
, mode
, conflict_hard_regs
)
4862 || (TEST_HARD_REG_BIT
4863 (ira_prohibited_class_mode_regs
[aclass
][mode
], hard_regno
)))
4865 ALLOCNO_HARD_REGNO (a
) = hard_regno
;
4866 for (l
= 0; l
< nr
; l
++)
4868 ira_object_t obj
= ALLOCNO_OBJECT (a
, l
);
4869 for (r
= OBJECT_LIVE_RANGES (obj
); r
!= NULL
; r
= r
->next
)
4870 for (k
= r
->start
; k
<= r
->finish
; k
++)
4871 IOR_HARD_REG_SET (used_hard_regs
[k
],
4872 ira_reg_mode_hard_regset
[hard_regno
][mode
]);
4877 ira_free (sorted_allocnos
);
4878 ira_free (used_hard_regs
);
4879 ira_free (allocno_priorities
);
4880 if (internal_flag_ira_verbose
> 1 && ira_dump_file
!= NULL
)
4881 ira_print_disposition (ira_dump_file
);
4886 /* Entry function doing coloring. */
4891 ira_allocno_iterator ai
;
4893 /* Setup updated costs. */
4894 FOR_EACH_ALLOCNO (a
, ai
)
4896 ALLOCNO_UPDATED_MEMORY_COST (a
) = ALLOCNO_MEMORY_COST (a
);
4897 ALLOCNO_UPDATED_CLASS_COST (a
) = ALLOCNO_CLASS_COST (a
);
4899 if (ira_conflicts_p
)