tree-ssa-phiprop.c: use gimple_phi
[official-gcc.git] / gcc / ira-color.c
blobdffe40a15502b3633206bc0c3906333046c8f22e
1 /* IRA allocation based on graph coloring.
2 Copyright (C) 2006-2014 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
10 version.
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
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "tm.h"
25 #include "rtl.h"
26 #include "tm_p.h"
27 #include "target.h"
28 #include "regs.h"
29 #include "flags.h"
30 #include "sbitmap.h"
31 #include "bitmap.h"
32 #include "hash-table.h"
33 #include "hard-reg-set.h"
34 #include "basic-block.h"
35 #include "expr.h"
36 #include "diagnostic-core.h"
37 #include "reload.h"
38 #include "params.h"
39 #include "df.h"
40 #include "ira-int.h"
42 typedef struct allocno_hard_regs *allocno_hard_regs_t;
44 /* The structure contains information about hard registers can be
45 assigned to allocnos. Usually it is allocno profitable hard
46 registers but in some cases this set can be a bit different. Major
47 reason of the difference is a requirement to use hard register sets
48 that form a tree or a forest (set of trees), i.e. hard register set
49 of a node should contain hard register sets of its subnodes. */
50 struct allocno_hard_regs
52 /* Hard registers can be assigned to an allocno. */
53 HARD_REG_SET set;
54 /* Overall (spilling) cost of all allocnos with given register
55 set. */
56 int64_t cost;
59 typedef struct allocno_hard_regs_node *allocno_hard_regs_node_t;
61 /* A node representing allocno hard registers. Such nodes form a
62 forest (set of trees). Each subnode of given node in the forest
63 refers for hard register set (usually allocno profitable hard
64 register set) which is a subset of one referred from given
65 node. */
66 struct allocno_hard_regs_node
68 /* Set up number of the node in preorder traversing of the forest. */
69 int preorder_num;
70 /* Used for different calculation like finding conflict size of an
71 allocno. */
72 int check;
73 /* Used for calculation of conflict size of an allocno. The
74 conflict size of the allocno is maximal number of given allocno
75 hard registers needed for allocation of the conflicting allocnos.
76 Given allocno is trivially colored if this number plus the number
77 of hard registers needed for given allocno is not greater than
78 the number of given allocno hard register set. */
79 int conflict_size;
80 /* The number of hard registers given by member hard_regs. */
81 int hard_regs_num;
82 /* The following member is used to form the final forest. */
83 bool used_p;
84 /* Pointer to the corresponding profitable hard registers. */
85 allocno_hard_regs_t hard_regs;
86 /* Parent, first subnode, previous and next node with the same
87 parent in the forest. */
88 allocno_hard_regs_node_t parent, first, prev, next;
91 /* Info about changing hard reg costs of an allocno. */
92 struct update_cost_record
94 /* Hard regno for which we changed the cost. */
95 int hard_regno;
96 /* Divisor used when we changed the cost of HARD_REGNO. */
97 int divisor;
98 /* Next record for given allocno. */
99 struct update_cost_record *next;
102 /* To decrease footprint of ira_allocno structure we store all data
103 needed only for coloring in the following structure. */
104 struct allocno_color_data
106 /* TRUE value means that the allocno was not removed yet from the
107 conflicting graph during colouring. */
108 unsigned int in_graph_p : 1;
109 /* TRUE if it is put on the stack to make other allocnos
110 colorable. */
111 unsigned int may_be_spilled_p : 1;
112 /* TRUE if the allocno is trivially colorable. */
113 unsigned int colorable_p : 1;
114 /* Number of hard registers of the allocno class really
115 available for the allocno allocation. It is number of the
116 profitable hard regs. */
117 int available_regs_num;
118 /* Allocnos in a bucket (used in coloring) chained by the following
119 two members. */
120 ira_allocno_t next_bucket_allocno;
121 ira_allocno_t prev_bucket_allocno;
122 /* Used for temporary purposes. */
123 int temp;
124 /* Used to exclude repeated processing. */
125 int last_process;
126 /* Profitable hard regs available for this pseudo allocation. It
127 means that the set excludes unavailable hard regs and hard regs
128 conflicting with given pseudo. They should be of the allocno
129 class. */
130 HARD_REG_SET profitable_hard_regs;
131 /* The allocno hard registers node. */
132 allocno_hard_regs_node_t hard_regs_node;
133 /* Array of structures allocno_hard_regs_subnode representing
134 given allocno hard registers node (the 1st element in the array)
135 and all its subnodes in the tree (forest) of allocno hard
136 register nodes (see comments above). */
137 int hard_regs_subnodes_start;
138 /* The length of the previous array. */
139 int hard_regs_subnodes_num;
140 /* Records about updating allocno hard reg costs from copies. If
141 the allocno did not get expected hard register, these records are
142 used to restore original hard reg costs of allocnos connected to
143 this allocno by copies. */
144 struct update_cost_record *update_cost_records;
145 /* Threads. We collect allocnos connected by copies into threads
146 and try to assign hard regs to allocnos by threads. */
147 /* Allocno representing all thread. */
148 ira_allocno_t first_thread_allocno;
149 /* Allocnos in thread forms a cycle list through the following
150 member. */
151 ira_allocno_t next_thread_allocno;
152 /* All thread frequency. Defined only for first thread allocno. */
153 int thread_freq;
156 /* See above. */
157 typedef struct allocno_color_data *allocno_color_data_t;
159 /* Container for storing allocno data concerning coloring. */
160 static allocno_color_data_t allocno_color_data;
162 /* Macro to access the data concerning coloring. */
163 #define ALLOCNO_COLOR_DATA(a) ((allocno_color_data_t) ALLOCNO_ADD_DATA (a))
165 /* Used for finding allocno colorability to exclude repeated allocno
166 processing and for updating preferencing to exclude repeated
167 allocno processing during assignment. */
168 static int curr_allocno_process;
170 /* This file contains code for regional graph coloring, spill/restore
171 code placement optimization, and code helping the reload pass to do
172 a better job. */
174 /* Bitmap of allocnos which should be colored. */
175 static bitmap coloring_allocno_bitmap;
177 /* Bitmap of allocnos which should be taken into account during
178 coloring. In general case it contains allocnos from
179 coloring_allocno_bitmap plus other already colored conflicting
180 allocnos. */
181 static bitmap consideration_allocno_bitmap;
183 /* All allocnos sorted according their priorities. */
184 static ira_allocno_t *sorted_allocnos;
186 /* Vec representing the stack of allocnos used during coloring. */
187 static vec<ira_allocno_t> allocno_stack_vec;
189 /* Helper for qsort comparison callbacks - return a positive integer if
190 X > Y, or a negative value otherwise. Use a conditional expression
191 instead of a difference computation to insulate from possible overflow
192 issues, e.g. X - Y < 0 for some X > 0 and Y < 0. */
193 #define SORTGT(x,y) (((x) > (y)) ? 1 : -1)
197 /* Definition of vector of allocno hard registers. */
199 /* Vector of unique allocno hard registers. */
200 static vec<allocno_hard_regs_t> allocno_hard_regs_vec;
202 struct allocno_hard_regs_hasher : typed_noop_remove <allocno_hard_regs>
204 typedef allocno_hard_regs value_type;
205 typedef allocno_hard_regs compare_type;
206 static inline hashval_t hash (const value_type *);
207 static inline bool equal (const value_type *, const compare_type *);
210 /* Returns hash value for allocno hard registers V. */
211 inline hashval_t
212 allocno_hard_regs_hasher::hash (const value_type *hv)
214 return iterative_hash (&hv->set, sizeof (HARD_REG_SET), 0);
217 /* Compares allocno hard registers V1 and V2. */
218 inline bool
219 allocno_hard_regs_hasher::equal (const value_type *hv1, const compare_type *hv2)
221 return hard_reg_set_equal_p (hv1->set, hv2->set);
224 /* Hash table of unique allocno hard registers. */
225 static hash_table<allocno_hard_regs_hasher> *allocno_hard_regs_htab;
227 /* Return allocno hard registers in the hash table equal to HV. */
228 static allocno_hard_regs_t
229 find_hard_regs (allocno_hard_regs_t hv)
231 return allocno_hard_regs_htab->find (hv);
234 /* Insert allocno hard registers HV in the hash table (if it is not
235 there yet) and return the value which in the table. */
236 static allocno_hard_regs_t
237 insert_hard_regs (allocno_hard_regs_t hv)
239 allocno_hard_regs **slot = allocno_hard_regs_htab->find_slot (hv, INSERT);
241 if (*slot == NULL)
242 *slot = hv;
243 return *slot;
246 /* Initialize data concerning allocno hard registers. */
247 static void
248 init_allocno_hard_regs (void)
250 allocno_hard_regs_vec.create (200);
251 allocno_hard_regs_htab
252 = new hash_table<allocno_hard_regs_hasher> (200);
255 /* Add (or update info about) allocno hard registers with SET and
256 COST. */
257 static allocno_hard_regs_t
258 add_allocno_hard_regs (HARD_REG_SET set, int64_t cost)
260 struct allocno_hard_regs temp;
261 allocno_hard_regs_t hv;
263 gcc_assert (! hard_reg_set_empty_p (set));
264 COPY_HARD_REG_SET (temp.set, set);
265 if ((hv = find_hard_regs (&temp)) != NULL)
266 hv->cost += cost;
267 else
269 hv = ((struct allocno_hard_regs *)
270 ira_allocate (sizeof (struct allocno_hard_regs)));
271 COPY_HARD_REG_SET (hv->set, set);
272 hv->cost = cost;
273 allocno_hard_regs_vec.safe_push (hv);
274 insert_hard_regs (hv);
276 return hv;
279 /* Finalize data concerning allocno hard registers. */
280 static void
281 finish_allocno_hard_regs (void)
283 int i;
284 allocno_hard_regs_t hv;
286 for (i = 0;
287 allocno_hard_regs_vec.iterate (i, &hv);
288 i++)
289 ira_free (hv);
290 delete allocno_hard_regs_htab;
291 allocno_hard_regs_htab = NULL;
292 allocno_hard_regs_vec.release ();
295 /* Sort hard regs according to their frequency of usage. */
296 static int
297 allocno_hard_regs_compare (const void *v1p, const void *v2p)
299 allocno_hard_regs_t hv1 = *(const allocno_hard_regs_t *) v1p;
300 allocno_hard_regs_t hv2 = *(const allocno_hard_regs_t *) v2p;
302 if (hv2->cost > hv1->cost)
303 return 1;
304 else if (hv2->cost < hv1->cost)
305 return -1;
306 else
307 return 0;
312 /* Used for finding a common ancestor of two allocno hard registers
313 nodes in the forest. We use the current value of
314 'node_check_tick' to mark all nodes from one node to the top and
315 then walking up from another node until we find a marked node.
317 It is also used to figure out allocno colorability as a mark that
318 we already reset value of member 'conflict_size' for the forest
319 node corresponding to the processed allocno. */
320 static int node_check_tick;
322 /* Roots of the forest containing hard register sets can be assigned
323 to allocnos. */
324 static allocno_hard_regs_node_t hard_regs_roots;
326 /* Definition of vector of allocno hard register nodes. */
328 /* Vector used to create the forest. */
329 static vec<allocno_hard_regs_node_t> hard_regs_node_vec;
331 /* Create and return allocno hard registers node containing allocno
332 hard registers HV. */
333 static allocno_hard_regs_node_t
334 create_new_allocno_hard_regs_node (allocno_hard_regs_t hv)
336 allocno_hard_regs_node_t new_node;
338 new_node = ((struct allocno_hard_regs_node *)
339 ira_allocate (sizeof (struct allocno_hard_regs_node)));
340 new_node->check = 0;
341 new_node->hard_regs = hv;
342 new_node->hard_regs_num = hard_reg_set_size (hv->set);
343 new_node->first = NULL;
344 new_node->used_p = false;
345 return new_node;
348 /* Add allocno hard registers node NEW_NODE to the forest on its level
349 given by ROOTS. */
350 static void
351 add_new_allocno_hard_regs_node_to_forest (allocno_hard_regs_node_t *roots,
352 allocno_hard_regs_node_t new_node)
354 new_node->next = *roots;
355 if (new_node->next != NULL)
356 new_node->next->prev = new_node;
357 new_node->prev = NULL;
358 *roots = new_node;
361 /* Add allocno hard registers HV (or its best approximation if it is
362 not possible) to the forest on its level given by ROOTS. */
363 static void
364 add_allocno_hard_regs_to_forest (allocno_hard_regs_node_t *roots,
365 allocno_hard_regs_t hv)
367 unsigned int i, start;
368 allocno_hard_regs_node_t node, prev, new_node;
369 HARD_REG_SET temp_set;
370 allocno_hard_regs_t hv2;
372 start = hard_regs_node_vec.length ();
373 for (node = *roots; node != NULL; node = node->next)
375 if (hard_reg_set_equal_p (hv->set, node->hard_regs->set))
376 return;
377 if (hard_reg_set_subset_p (hv->set, node->hard_regs->set))
379 add_allocno_hard_regs_to_forest (&node->first, hv);
380 return;
382 if (hard_reg_set_subset_p (node->hard_regs->set, hv->set))
383 hard_regs_node_vec.safe_push (node);
384 else if (hard_reg_set_intersect_p (hv->set, node->hard_regs->set))
386 COPY_HARD_REG_SET (temp_set, hv->set);
387 AND_HARD_REG_SET (temp_set, node->hard_regs->set);
388 hv2 = add_allocno_hard_regs (temp_set, hv->cost);
389 add_allocno_hard_regs_to_forest (&node->first, hv2);
392 if (hard_regs_node_vec.length ()
393 > start + 1)
395 /* Create a new node which contains nodes in hard_regs_node_vec. */
396 CLEAR_HARD_REG_SET (temp_set);
397 for (i = start;
398 i < hard_regs_node_vec.length ();
399 i++)
401 node = hard_regs_node_vec[i];
402 IOR_HARD_REG_SET (temp_set, node->hard_regs->set);
404 hv = add_allocno_hard_regs (temp_set, hv->cost);
405 new_node = create_new_allocno_hard_regs_node (hv);
406 prev = NULL;
407 for (i = start;
408 i < hard_regs_node_vec.length ();
409 i++)
411 node = hard_regs_node_vec[i];
412 if (node->prev == NULL)
413 *roots = node->next;
414 else
415 node->prev->next = node->next;
416 if (node->next != NULL)
417 node->next->prev = node->prev;
418 if (prev == NULL)
419 new_node->first = node;
420 else
421 prev->next = node;
422 node->prev = prev;
423 node->next = NULL;
424 prev = node;
426 add_new_allocno_hard_regs_node_to_forest (roots, new_node);
428 hard_regs_node_vec.truncate (start);
431 /* Add allocno hard registers nodes starting with the forest level
432 given by FIRST which contains biggest set inside SET. */
433 static void
434 collect_allocno_hard_regs_cover (allocno_hard_regs_node_t first,
435 HARD_REG_SET set)
437 allocno_hard_regs_node_t node;
439 ira_assert (first != NULL);
440 for (node = first; node != NULL; node = node->next)
441 if (hard_reg_set_subset_p (node->hard_regs->set, set))
442 hard_regs_node_vec.safe_push (node);
443 else if (hard_reg_set_intersect_p (set, node->hard_regs->set))
444 collect_allocno_hard_regs_cover (node->first, set);
447 /* Set up field parent as PARENT in all allocno hard registers nodes
448 in forest given by FIRST. */
449 static void
450 setup_allocno_hard_regs_nodes_parent (allocno_hard_regs_node_t first,
451 allocno_hard_regs_node_t parent)
453 allocno_hard_regs_node_t node;
455 for (node = first; node != NULL; node = node->next)
457 node->parent = parent;
458 setup_allocno_hard_regs_nodes_parent (node->first, node);
462 /* Return allocno hard registers node which is a first common ancestor
463 node of FIRST and SECOND in the forest. */
464 static allocno_hard_regs_node_t
465 first_common_ancestor_node (allocno_hard_regs_node_t first,
466 allocno_hard_regs_node_t second)
468 allocno_hard_regs_node_t node;
470 node_check_tick++;
471 for (node = first; node != NULL; node = node->parent)
472 node->check = node_check_tick;
473 for (node = second; node != NULL; node = node->parent)
474 if (node->check == node_check_tick)
475 return node;
476 return first_common_ancestor_node (second, first);
479 /* Print hard reg set SET to F. */
480 static void
481 print_hard_reg_set (FILE *f, HARD_REG_SET set, bool new_line_p)
483 int i, start;
485 for (start = -1, i = 0; i < FIRST_PSEUDO_REGISTER; i++)
487 if (TEST_HARD_REG_BIT (set, i))
489 if (i == 0 || ! TEST_HARD_REG_BIT (set, i - 1))
490 start = i;
492 if (start >= 0
493 && (i == FIRST_PSEUDO_REGISTER - 1 || ! TEST_HARD_REG_BIT (set, i)))
495 if (start == i - 1)
496 fprintf (f, " %d", start);
497 else if (start == i - 2)
498 fprintf (f, " %d %d", start, start + 1);
499 else
500 fprintf (f, " %d-%d", start, i - 1);
501 start = -1;
504 if (new_line_p)
505 fprintf (f, "\n");
508 /* Print allocno hard register subforest given by ROOTS and its LEVEL
509 to F. */
510 static void
511 print_hard_regs_subforest (FILE *f, allocno_hard_regs_node_t roots,
512 int level)
514 int i;
515 allocno_hard_regs_node_t node;
517 for (node = roots; node != NULL; node = node->next)
519 fprintf (f, " ");
520 for (i = 0; i < level * 2; i++)
521 fprintf (f, " ");
522 fprintf (f, "%d:(", node->preorder_num);
523 print_hard_reg_set (f, node->hard_regs->set, false);
524 fprintf (f, ")@%"PRId64"\n", node->hard_regs->cost);
525 print_hard_regs_subforest (f, node->first, level + 1);
529 /* Print the allocno hard register forest to F. */
530 static void
531 print_hard_regs_forest (FILE *f)
533 fprintf (f, " Hard reg set forest:\n");
534 print_hard_regs_subforest (f, hard_regs_roots, 1);
537 /* Print the allocno hard register forest to stderr. */
538 void
539 ira_debug_hard_regs_forest (void)
541 print_hard_regs_forest (stderr);
544 /* Remove unused allocno hard registers nodes from forest given by its
545 *ROOTS. */
546 static void
547 remove_unused_allocno_hard_regs_nodes (allocno_hard_regs_node_t *roots)
549 allocno_hard_regs_node_t node, prev, next, last;
551 for (prev = NULL, node = *roots; node != NULL; node = next)
553 next = node->next;
554 if (node->used_p)
556 remove_unused_allocno_hard_regs_nodes (&node->first);
557 prev = node;
559 else
561 for (last = node->first;
562 last != NULL && last->next != NULL;
563 last = last->next)
565 if (last != NULL)
567 if (prev == NULL)
568 *roots = node->first;
569 else
570 prev->next = node->first;
571 if (next != NULL)
572 next->prev = last;
573 last->next = next;
574 next = node->first;
576 else
578 if (prev == NULL)
579 *roots = next;
580 else
581 prev->next = next;
582 if (next != NULL)
583 next->prev = prev;
585 ira_free (node);
590 /* Set up fields preorder_num starting with START_NUM in all allocno
591 hard registers nodes in forest given by FIRST. Return biggest set
592 PREORDER_NUM increased by 1. */
593 static int
594 enumerate_allocno_hard_regs_nodes (allocno_hard_regs_node_t first,
595 allocno_hard_regs_node_t parent,
596 int start_num)
598 allocno_hard_regs_node_t node;
600 for (node = first; node != NULL; node = node->next)
602 node->preorder_num = start_num++;
603 node->parent = parent;
604 start_num = enumerate_allocno_hard_regs_nodes (node->first, node,
605 start_num);
607 return start_num;
610 /* Number of allocno hard registers nodes in the forest. */
611 static int allocno_hard_regs_nodes_num;
613 /* Table preorder number of allocno hard registers node in the forest
614 -> the allocno hard registers node. */
615 static allocno_hard_regs_node_t *allocno_hard_regs_nodes;
617 /* See below. */
618 typedef struct allocno_hard_regs_subnode *allocno_hard_regs_subnode_t;
620 /* The structure is used to describes all subnodes (not only immediate
621 ones) in the mentioned above tree for given allocno hard register
622 node. The usage of such data accelerates calculation of
623 colorability of given allocno. */
624 struct allocno_hard_regs_subnode
626 /* The conflict size of conflicting allocnos whose hard register
627 sets are equal sets (plus supersets if given node is given
628 allocno hard registers node) of one in the given node. */
629 int left_conflict_size;
630 /* The summary conflict size of conflicting allocnos whose hard
631 register sets are strict subsets of one in the given node.
632 Overall conflict size is
633 left_conflict_subnodes_size
634 + MIN (max_node_impact - left_conflict_subnodes_size,
635 left_conflict_size)
637 short left_conflict_subnodes_size;
638 short max_node_impact;
641 /* Container for hard regs subnodes of all allocnos. */
642 static allocno_hard_regs_subnode_t allocno_hard_regs_subnodes;
644 /* Table (preorder number of allocno hard registers node in the
645 forest, preorder number of allocno hard registers subnode) -> index
646 of the subnode relative to the node. -1 if it is not a
647 subnode. */
648 static int *allocno_hard_regs_subnode_index;
650 /* Setup arrays ALLOCNO_HARD_REGS_NODES and
651 ALLOCNO_HARD_REGS_SUBNODE_INDEX. */
652 static void
653 setup_allocno_hard_regs_subnode_index (allocno_hard_regs_node_t first)
655 allocno_hard_regs_node_t node, parent;
656 int index;
658 for (node = first; node != NULL; node = node->next)
660 allocno_hard_regs_nodes[node->preorder_num] = node;
661 for (parent = node; parent != NULL; parent = parent->parent)
663 index = parent->preorder_num * allocno_hard_regs_nodes_num;
664 allocno_hard_regs_subnode_index[index + node->preorder_num]
665 = node->preorder_num - parent->preorder_num;
667 setup_allocno_hard_regs_subnode_index (node->first);
671 /* Count all allocno hard registers nodes in tree ROOT. */
672 static int
673 get_allocno_hard_regs_subnodes_num (allocno_hard_regs_node_t root)
675 int len = 1;
677 for (root = root->first; root != NULL; root = root->next)
678 len += get_allocno_hard_regs_subnodes_num (root);
679 return len;
682 /* Build the forest of allocno hard registers nodes and assign each
683 allocno a node from the forest. */
684 static void
685 form_allocno_hard_regs_nodes_forest (void)
687 unsigned int i, j, size, len;
688 int start;
689 ira_allocno_t a;
690 allocno_hard_regs_t hv;
691 bitmap_iterator bi;
692 HARD_REG_SET temp;
693 allocno_hard_regs_node_t node, allocno_hard_regs_node;
694 allocno_color_data_t allocno_data;
696 node_check_tick = 0;
697 init_allocno_hard_regs ();
698 hard_regs_roots = NULL;
699 hard_regs_node_vec.create (100);
700 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
701 if (! TEST_HARD_REG_BIT (ira_no_alloc_regs, i))
703 CLEAR_HARD_REG_SET (temp);
704 SET_HARD_REG_BIT (temp, i);
705 hv = add_allocno_hard_regs (temp, 0);
706 node = create_new_allocno_hard_regs_node (hv);
707 add_new_allocno_hard_regs_node_to_forest (&hard_regs_roots, node);
709 start = allocno_hard_regs_vec.length ();
710 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
712 a = ira_allocnos[i];
713 allocno_data = ALLOCNO_COLOR_DATA (a);
715 if (hard_reg_set_empty_p (allocno_data->profitable_hard_regs))
716 continue;
717 hv = (add_allocno_hard_regs
718 (allocno_data->profitable_hard_regs,
719 ALLOCNO_MEMORY_COST (a) - ALLOCNO_CLASS_COST (a)));
721 SET_HARD_REG_SET (temp);
722 AND_COMPL_HARD_REG_SET (temp, ira_no_alloc_regs);
723 add_allocno_hard_regs (temp, 0);
724 qsort (allocno_hard_regs_vec.address () + start,
725 allocno_hard_regs_vec.length () - start,
726 sizeof (allocno_hard_regs_t), allocno_hard_regs_compare);
727 for (i = start;
728 allocno_hard_regs_vec.iterate (i, &hv);
729 i++)
731 add_allocno_hard_regs_to_forest (&hard_regs_roots, hv);
732 ira_assert (hard_regs_node_vec.length () == 0);
734 /* We need to set up parent fields for right work of
735 first_common_ancestor_node. */
736 setup_allocno_hard_regs_nodes_parent (hard_regs_roots, NULL);
737 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
739 a = ira_allocnos[i];
740 allocno_data = ALLOCNO_COLOR_DATA (a);
741 if (hard_reg_set_empty_p (allocno_data->profitable_hard_regs))
742 continue;
743 hard_regs_node_vec.truncate (0);
744 collect_allocno_hard_regs_cover (hard_regs_roots,
745 allocno_data->profitable_hard_regs);
746 allocno_hard_regs_node = NULL;
747 for (j = 0; hard_regs_node_vec.iterate (j, &node); j++)
748 allocno_hard_regs_node
749 = (j == 0
750 ? node
751 : first_common_ancestor_node (node, allocno_hard_regs_node));
752 /* That is a temporary storage. */
753 allocno_hard_regs_node->used_p = true;
754 allocno_data->hard_regs_node = allocno_hard_regs_node;
756 ira_assert (hard_regs_roots->next == NULL);
757 hard_regs_roots->used_p = true;
758 remove_unused_allocno_hard_regs_nodes (&hard_regs_roots);
759 allocno_hard_regs_nodes_num
760 = enumerate_allocno_hard_regs_nodes (hard_regs_roots, NULL, 0);
761 allocno_hard_regs_nodes
762 = ((allocno_hard_regs_node_t *)
763 ira_allocate (allocno_hard_regs_nodes_num
764 * sizeof (allocno_hard_regs_node_t)));
765 size = allocno_hard_regs_nodes_num * allocno_hard_regs_nodes_num;
766 allocno_hard_regs_subnode_index
767 = (int *) ira_allocate (size * sizeof (int));
768 for (i = 0; i < size; i++)
769 allocno_hard_regs_subnode_index[i] = -1;
770 setup_allocno_hard_regs_subnode_index (hard_regs_roots);
771 start = 0;
772 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
774 a = ira_allocnos[i];
775 allocno_data = ALLOCNO_COLOR_DATA (a);
776 if (hard_reg_set_empty_p (allocno_data->profitable_hard_regs))
777 continue;
778 len = get_allocno_hard_regs_subnodes_num (allocno_data->hard_regs_node);
779 allocno_data->hard_regs_subnodes_start = start;
780 allocno_data->hard_regs_subnodes_num = len;
781 start += len;
783 allocno_hard_regs_subnodes
784 = ((allocno_hard_regs_subnode_t)
785 ira_allocate (sizeof (struct allocno_hard_regs_subnode) * start));
786 hard_regs_node_vec.release ();
789 /* Free tree of allocno hard registers nodes given by its ROOT. */
790 static void
791 finish_allocno_hard_regs_nodes_tree (allocno_hard_regs_node_t root)
793 allocno_hard_regs_node_t child, next;
795 for (child = root->first; child != NULL; child = next)
797 next = child->next;
798 finish_allocno_hard_regs_nodes_tree (child);
800 ira_free (root);
803 /* Finish work with the forest of allocno hard registers nodes. */
804 static void
805 finish_allocno_hard_regs_nodes_forest (void)
807 allocno_hard_regs_node_t node, next;
809 ira_free (allocno_hard_regs_subnodes);
810 for (node = hard_regs_roots; node != NULL; node = next)
812 next = node->next;
813 finish_allocno_hard_regs_nodes_tree (node);
815 ira_free (allocno_hard_regs_nodes);
816 ira_free (allocno_hard_regs_subnode_index);
817 finish_allocno_hard_regs ();
820 /* Set up left conflict sizes and left conflict subnodes sizes of hard
821 registers subnodes of allocno A. Return TRUE if allocno A is
822 trivially colorable. */
823 static bool
824 setup_left_conflict_sizes_p (ira_allocno_t a)
826 int i, k, nobj, start;
827 int conflict_size, left_conflict_subnodes_size, node_preorder_num;
828 allocno_color_data_t data;
829 HARD_REG_SET profitable_hard_regs;
830 allocno_hard_regs_subnode_t subnodes;
831 allocno_hard_regs_node_t node;
832 HARD_REG_SET node_set;
834 nobj = ALLOCNO_NUM_OBJECTS (a);
835 conflict_size = 0;
836 data = ALLOCNO_COLOR_DATA (a);
837 subnodes = allocno_hard_regs_subnodes + data->hard_regs_subnodes_start;
838 COPY_HARD_REG_SET (profitable_hard_regs, data->profitable_hard_regs);
839 node = data->hard_regs_node;
840 node_preorder_num = node->preorder_num;
841 COPY_HARD_REG_SET (node_set, node->hard_regs->set);
842 node_check_tick++;
843 for (k = 0; k < nobj; k++)
845 ira_object_t obj = ALLOCNO_OBJECT (a, k);
846 ira_object_t conflict_obj;
847 ira_object_conflict_iterator oci;
849 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
851 int size;
852 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
853 allocno_hard_regs_node_t conflict_node, temp_node;
854 HARD_REG_SET conflict_node_set;
855 allocno_color_data_t conflict_data;
857 conflict_data = ALLOCNO_COLOR_DATA (conflict_a);
858 if (! ALLOCNO_COLOR_DATA (conflict_a)->in_graph_p
859 || ! hard_reg_set_intersect_p (profitable_hard_regs,
860 conflict_data
861 ->profitable_hard_regs))
862 continue;
863 conflict_node = conflict_data->hard_regs_node;
864 COPY_HARD_REG_SET (conflict_node_set, conflict_node->hard_regs->set);
865 if (hard_reg_set_subset_p (node_set, conflict_node_set))
866 temp_node = node;
867 else
869 ira_assert (hard_reg_set_subset_p (conflict_node_set, node_set));
870 temp_node = conflict_node;
872 if (temp_node->check != node_check_tick)
874 temp_node->check = node_check_tick;
875 temp_node->conflict_size = 0;
877 size = (ira_reg_class_max_nregs
878 [ALLOCNO_CLASS (conflict_a)][ALLOCNO_MODE (conflict_a)]);
879 if (ALLOCNO_NUM_OBJECTS (conflict_a) > 1)
880 /* We will deal with the subwords individually. */
881 size = 1;
882 temp_node->conflict_size += size;
885 for (i = 0; i < data->hard_regs_subnodes_num; i++)
887 allocno_hard_regs_node_t temp_node;
889 temp_node = allocno_hard_regs_nodes[i + node_preorder_num];
890 ira_assert (temp_node->preorder_num == i + node_preorder_num);
891 subnodes[i].left_conflict_size = (temp_node->check != node_check_tick
892 ? 0 : temp_node->conflict_size);
893 if (hard_reg_set_subset_p (temp_node->hard_regs->set,
894 profitable_hard_regs))
895 subnodes[i].max_node_impact = temp_node->hard_regs_num;
896 else
898 HARD_REG_SET temp_set;
899 int j, n, hard_regno;
900 enum reg_class aclass;
902 COPY_HARD_REG_SET (temp_set, temp_node->hard_regs->set);
903 AND_HARD_REG_SET (temp_set, profitable_hard_regs);
904 aclass = ALLOCNO_CLASS (a);
905 for (n = 0, j = ira_class_hard_regs_num[aclass] - 1; j >= 0; j--)
907 hard_regno = ira_class_hard_regs[aclass][j];
908 if (TEST_HARD_REG_BIT (temp_set, hard_regno))
909 n++;
911 subnodes[i].max_node_impact = n;
913 subnodes[i].left_conflict_subnodes_size = 0;
915 start = node_preorder_num * allocno_hard_regs_nodes_num;
916 for (i = data->hard_regs_subnodes_num - 1; i >= 0; i--)
918 int size, parent_i;
919 allocno_hard_regs_node_t parent;
921 size = (subnodes[i].left_conflict_subnodes_size
922 + MIN (subnodes[i].max_node_impact
923 - subnodes[i].left_conflict_subnodes_size,
924 subnodes[i].left_conflict_size));
925 parent = allocno_hard_regs_nodes[i + node_preorder_num]->parent;
926 if (parent == NULL)
927 continue;
928 parent_i
929 = allocno_hard_regs_subnode_index[start + parent->preorder_num];
930 if (parent_i < 0)
931 continue;
932 subnodes[parent_i].left_conflict_subnodes_size += size;
934 left_conflict_subnodes_size = subnodes[0].left_conflict_subnodes_size;
935 conflict_size
936 += (left_conflict_subnodes_size
937 + MIN (subnodes[0].max_node_impact - left_conflict_subnodes_size,
938 subnodes[0].left_conflict_size));
939 conflict_size += ira_reg_class_max_nregs[ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)];
940 data->colorable_p = conflict_size <= data->available_regs_num;
941 return data->colorable_p;
944 /* Update left conflict sizes of hard registers subnodes of allocno A
945 after removing allocno REMOVED_A with SIZE from the conflict graph.
946 Return TRUE if A is trivially colorable. */
947 static bool
948 update_left_conflict_sizes_p (ira_allocno_t a,
949 ira_allocno_t removed_a, int size)
951 int i, conflict_size, before_conflict_size, diff, start;
952 int node_preorder_num, parent_i;
953 allocno_hard_regs_node_t node, removed_node, parent;
954 allocno_hard_regs_subnode_t subnodes;
955 allocno_color_data_t data = ALLOCNO_COLOR_DATA (a);
957 ira_assert (! data->colorable_p);
958 node = data->hard_regs_node;
959 node_preorder_num = node->preorder_num;
960 removed_node = ALLOCNO_COLOR_DATA (removed_a)->hard_regs_node;
961 ira_assert (hard_reg_set_subset_p (removed_node->hard_regs->set,
962 node->hard_regs->set)
963 || hard_reg_set_subset_p (node->hard_regs->set,
964 removed_node->hard_regs->set));
965 start = node_preorder_num * allocno_hard_regs_nodes_num;
966 i = allocno_hard_regs_subnode_index[start + removed_node->preorder_num];
967 if (i < 0)
968 i = 0;
969 subnodes = allocno_hard_regs_subnodes + data->hard_regs_subnodes_start;
970 before_conflict_size
971 = (subnodes[i].left_conflict_subnodes_size
972 + MIN (subnodes[i].max_node_impact
973 - subnodes[i].left_conflict_subnodes_size,
974 subnodes[i].left_conflict_size));
975 subnodes[i].left_conflict_size -= size;
976 for (;;)
978 conflict_size
979 = (subnodes[i].left_conflict_subnodes_size
980 + MIN (subnodes[i].max_node_impact
981 - subnodes[i].left_conflict_subnodes_size,
982 subnodes[i].left_conflict_size));
983 if ((diff = before_conflict_size - conflict_size) == 0)
984 break;
985 ira_assert (conflict_size < before_conflict_size);
986 parent = allocno_hard_regs_nodes[i + node_preorder_num]->parent;
987 if (parent == NULL)
988 break;
989 parent_i
990 = allocno_hard_regs_subnode_index[start + parent->preorder_num];
991 if (parent_i < 0)
992 break;
993 i = parent_i;
994 before_conflict_size
995 = (subnodes[i].left_conflict_subnodes_size
996 + MIN (subnodes[i].max_node_impact
997 - subnodes[i].left_conflict_subnodes_size,
998 subnodes[i].left_conflict_size));
999 subnodes[i].left_conflict_subnodes_size -= diff;
1001 if (i != 0
1002 || (conflict_size
1003 + ira_reg_class_max_nregs[ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)]
1004 > data->available_regs_num))
1005 return false;
1006 data->colorable_p = true;
1007 return true;
1010 /* Return true if allocno A has empty profitable hard regs. */
1011 static bool
1012 empty_profitable_hard_regs (ira_allocno_t a)
1014 allocno_color_data_t data = ALLOCNO_COLOR_DATA (a);
1016 return hard_reg_set_empty_p (data->profitable_hard_regs);
1019 /* Set up profitable hard registers for each allocno being
1020 colored. */
1021 static void
1022 setup_profitable_hard_regs (void)
1024 unsigned int i;
1025 int j, k, nobj, hard_regno, nregs, class_size;
1026 ira_allocno_t a;
1027 bitmap_iterator bi;
1028 enum reg_class aclass;
1029 enum machine_mode mode;
1030 allocno_color_data_t data;
1032 /* Initial set up from allocno classes and explicitly conflicting
1033 hard regs. */
1034 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
1036 a = ira_allocnos[i];
1037 if ((aclass = ALLOCNO_CLASS (a)) == NO_REGS)
1038 continue;
1039 data = ALLOCNO_COLOR_DATA (a);
1040 if (ALLOCNO_UPDATED_HARD_REG_COSTS (a) == NULL
1041 && ALLOCNO_CLASS_COST (a) > ALLOCNO_MEMORY_COST (a))
1042 CLEAR_HARD_REG_SET (data->profitable_hard_regs);
1043 else
1045 mode = ALLOCNO_MODE (a);
1046 COPY_HARD_REG_SET (data->profitable_hard_regs,
1047 ira_useful_class_mode_regs[aclass][mode]);
1048 nobj = ALLOCNO_NUM_OBJECTS (a);
1049 for (k = 0; k < nobj; k++)
1051 ira_object_t obj = ALLOCNO_OBJECT (a, k);
1053 AND_COMPL_HARD_REG_SET (data->profitable_hard_regs,
1054 OBJECT_TOTAL_CONFLICT_HARD_REGS (obj));
1058 /* Exclude hard regs already assigned for conflicting objects. */
1059 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, i, bi)
1061 a = ira_allocnos[i];
1062 if ((aclass = ALLOCNO_CLASS (a)) == NO_REGS
1063 || ! ALLOCNO_ASSIGNED_P (a)
1064 || (hard_regno = ALLOCNO_HARD_REGNO (a)) < 0)
1065 continue;
1066 mode = ALLOCNO_MODE (a);
1067 nregs = hard_regno_nregs[hard_regno][mode];
1068 nobj = ALLOCNO_NUM_OBJECTS (a);
1069 for (k = 0; k < nobj; k++)
1071 ira_object_t obj = ALLOCNO_OBJECT (a, k);
1072 ira_object_t conflict_obj;
1073 ira_object_conflict_iterator oci;
1075 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
1077 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
1079 /* We can process the conflict allocno repeatedly with
1080 the same result. */
1081 if (nregs == nobj && nregs > 1)
1083 int num = OBJECT_SUBWORD (conflict_obj);
1085 if (REG_WORDS_BIG_ENDIAN)
1086 CLEAR_HARD_REG_BIT
1087 (ALLOCNO_COLOR_DATA (conflict_a)->profitable_hard_regs,
1088 hard_regno + nobj - num - 1);
1089 else
1090 CLEAR_HARD_REG_BIT
1091 (ALLOCNO_COLOR_DATA (conflict_a)->profitable_hard_regs,
1092 hard_regno + num);
1094 else
1095 AND_COMPL_HARD_REG_SET
1096 (ALLOCNO_COLOR_DATA (conflict_a)->profitable_hard_regs,
1097 ira_reg_mode_hard_regset[hard_regno][mode]);
1101 /* Exclude too costly hard regs. */
1102 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
1104 int min_cost = INT_MAX;
1105 int *costs;
1107 a = ira_allocnos[i];
1108 if ((aclass = ALLOCNO_CLASS (a)) == NO_REGS
1109 || empty_profitable_hard_regs (a))
1110 continue;
1111 data = ALLOCNO_COLOR_DATA (a);
1112 mode = ALLOCNO_MODE (a);
1113 if ((costs = ALLOCNO_UPDATED_HARD_REG_COSTS (a)) != NULL
1114 || (costs = ALLOCNO_HARD_REG_COSTS (a)) != NULL)
1116 class_size = ira_class_hard_regs_num[aclass];
1117 for (j = 0; j < class_size; j++)
1119 hard_regno = ira_class_hard_regs[aclass][j];
1120 if (! TEST_HARD_REG_BIT (data->profitable_hard_regs,
1121 hard_regno))
1122 continue;
1123 if (ALLOCNO_UPDATED_MEMORY_COST (a) < costs[j])
1124 CLEAR_HARD_REG_BIT (data->profitable_hard_regs,
1125 hard_regno);
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 CLEAR_HARD_REG_SET (data->profitable_hard_regs);
1133 if (ALLOCNO_UPDATED_CLASS_COST (a) > min_cost)
1134 ALLOCNO_UPDATED_CLASS_COST (a) = min_cost;
1140 /* This page contains functions used to choose hard registers for
1141 allocnos. */
1143 /* Pool for update cost records. */
1144 static alloc_pool update_cost_record_pool;
1146 /* Initiate update cost records. */
1147 static void
1148 init_update_cost_records (void)
1150 update_cost_record_pool
1151 = create_alloc_pool ("update cost records",
1152 sizeof (struct update_cost_record), 100);
1155 /* Return new update cost record with given params. */
1156 static struct update_cost_record *
1157 get_update_cost_record (int hard_regno, int divisor,
1158 struct update_cost_record *next)
1160 struct update_cost_record *record;
1162 record = (struct update_cost_record *) pool_alloc (update_cost_record_pool);
1163 record->hard_regno = hard_regno;
1164 record->divisor = divisor;
1165 record->next = next;
1166 return record;
1169 /* Free memory for all records in LIST. */
1170 static void
1171 free_update_cost_record_list (struct update_cost_record *list)
1173 struct update_cost_record *next;
1175 while (list != NULL)
1177 next = list->next;
1178 pool_free (update_cost_record_pool, list);
1179 list = next;
1183 /* Free memory allocated for all update cost records. */
1184 static void
1185 finish_update_cost_records (void)
1187 free_alloc_pool (update_cost_record_pool);
1190 /* Array whose element value is TRUE if the corresponding hard
1191 register was already allocated for an allocno. */
1192 static bool allocated_hardreg_p[FIRST_PSEUDO_REGISTER];
1194 /* Describes one element in a queue of allocnos whose costs need to be
1195 updated. Each allocno in the queue is known to have an allocno
1196 class. */
1197 struct update_cost_queue_elem
1199 /* This element is in the queue iff CHECK == update_cost_check. */
1200 int check;
1202 /* COST_HOP_DIVISOR**N, where N is the length of the shortest path
1203 connecting this allocno to the one being allocated. */
1204 int divisor;
1206 /* Allocno from which we are chaning costs of connected allocnos.
1207 It is used not go back in graph of allocnos connected by
1208 copies. */
1209 ira_allocno_t from;
1211 /* The next allocno in the queue, or null if this is the last element. */
1212 ira_allocno_t next;
1215 /* The first element in a queue of allocnos whose copy costs need to be
1216 updated. Null if the queue is empty. */
1217 static ira_allocno_t update_cost_queue;
1219 /* The last element in the queue described by update_cost_queue.
1220 Not valid if update_cost_queue is null. */
1221 static struct update_cost_queue_elem *update_cost_queue_tail;
1223 /* A pool of elements in the queue described by update_cost_queue.
1224 Elements are indexed by ALLOCNO_NUM. */
1225 static struct update_cost_queue_elem *update_cost_queue_elems;
1227 /* The current value of update_costs_from_copies call count. */
1228 static int update_cost_check;
1230 /* Allocate and initialize data necessary for function
1231 update_costs_from_copies. */
1232 static void
1233 initiate_cost_update (void)
1235 size_t size;
1237 size = ira_allocnos_num * sizeof (struct update_cost_queue_elem);
1238 update_cost_queue_elems
1239 = (struct update_cost_queue_elem *) ira_allocate (size);
1240 memset (update_cost_queue_elems, 0, size);
1241 update_cost_check = 0;
1242 init_update_cost_records ();
1245 /* Deallocate data used by function update_costs_from_copies. */
1246 static void
1247 finish_cost_update (void)
1249 ira_free (update_cost_queue_elems);
1250 finish_update_cost_records ();
1253 /* When we traverse allocnos to update hard register costs, the cost
1254 divisor will be multiplied by the following macro value for each
1255 hop from given allocno to directly connected allocnos. */
1256 #define COST_HOP_DIVISOR 4
1258 /* Start a new cost-updating pass. */
1259 static void
1260 start_update_cost (void)
1262 update_cost_check++;
1263 update_cost_queue = NULL;
1266 /* Add (ALLOCNO, FROM, DIVISOR) to the end of update_cost_queue, unless
1267 ALLOCNO is already in the queue, or has NO_REGS class. */
1268 static inline void
1269 queue_update_cost (ira_allocno_t allocno, ira_allocno_t from, int divisor)
1271 struct update_cost_queue_elem *elem;
1273 elem = &update_cost_queue_elems[ALLOCNO_NUM (allocno)];
1274 if (elem->check != update_cost_check
1275 && ALLOCNO_CLASS (allocno) != NO_REGS)
1277 elem->check = update_cost_check;
1278 elem->from = from;
1279 elem->divisor = divisor;
1280 elem->next = NULL;
1281 if (update_cost_queue == NULL)
1282 update_cost_queue = allocno;
1283 else
1284 update_cost_queue_tail->next = allocno;
1285 update_cost_queue_tail = elem;
1289 /* Try to remove the first element from update_cost_queue. Return
1290 false if the queue was empty, otherwise make (*ALLOCNO, *FROM,
1291 *DIVISOR) describe the removed element. */
1292 static inline bool
1293 get_next_update_cost (ira_allocno_t *allocno, ira_allocno_t *from, int *divisor)
1295 struct update_cost_queue_elem *elem;
1297 if (update_cost_queue == NULL)
1298 return false;
1300 *allocno = update_cost_queue;
1301 elem = &update_cost_queue_elems[ALLOCNO_NUM (*allocno)];
1302 *from = elem->from;
1303 *divisor = elem->divisor;
1304 update_cost_queue = elem->next;
1305 return true;
1308 /* Increase costs of HARD_REGNO by UPDATE_COST for ALLOCNO. Return
1309 true if we really modified the cost. */
1310 static bool
1311 update_allocno_cost (ira_allocno_t allocno, int hard_regno, int update_cost)
1313 int i;
1314 enum reg_class aclass = ALLOCNO_CLASS (allocno);
1316 i = ira_class_hard_reg_index[aclass][hard_regno];
1317 if (i < 0)
1318 return false;
1319 ira_allocate_and_set_or_copy_costs
1320 (&ALLOCNO_UPDATED_HARD_REG_COSTS (allocno), aclass,
1321 ALLOCNO_UPDATED_CLASS_COST (allocno),
1322 ALLOCNO_HARD_REG_COSTS (allocno));
1323 ira_allocate_and_set_or_copy_costs
1324 (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (allocno),
1325 aclass, 0, ALLOCNO_CONFLICT_HARD_REG_COSTS (allocno));
1326 ALLOCNO_UPDATED_HARD_REG_COSTS (allocno)[i] += update_cost;
1327 ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (allocno)[i] += update_cost;
1328 return true;
1331 /* Update (decrease if DECR_P) HARD_REGNO cost of allocnos connected
1332 by copies to ALLOCNO to increase chances to remove some copies as
1333 the result of subsequent assignment. Record cost updates if
1334 RECORD_P is true. */
1335 static void
1336 update_costs_from_allocno (ira_allocno_t allocno, int hard_regno,
1337 int divisor, bool decr_p, bool record_p)
1339 int cost, update_cost;
1340 enum machine_mode mode;
1341 enum reg_class rclass, aclass;
1342 ira_allocno_t another_allocno, from = NULL;
1343 ira_copy_t cp, next_cp;
1345 rclass = REGNO_REG_CLASS (hard_regno);
1348 mode = ALLOCNO_MODE (allocno);
1349 ira_init_register_move_cost_if_necessary (mode);
1350 for (cp = ALLOCNO_COPIES (allocno); cp != NULL; cp = next_cp)
1352 if (cp->first == allocno)
1354 next_cp = cp->next_first_allocno_copy;
1355 another_allocno = cp->second;
1357 else if (cp->second == allocno)
1359 next_cp = cp->next_second_allocno_copy;
1360 another_allocno = cp->first;
1362 else
1363 gcc_unreachable ();
1365 if (another_allocno == from)
1366 continue;
1368 aclass = ALLOCNO_CLASS (another_allocno);
1369 if (! TEST_HARD_REG_BIT (reg_class_contents[aclass],
1370 hard_regno)
1371 || ALLOCNO_ASSIGNED_P (another_allocno))
1372 continue;
1374 cost = (cp->second == allocno
1375 ? ira_register_move_cost[mode][rclass][aclass]
1376 : ira_register_move_cost[mode][aclass][rclass]);
1377 if (decr_p)
1378 cost = -cost;
1380 update_cost = cp->freq * cost / divisor;
1381 if (update_cost == 0)
1382 continue;
1384 if (! update_allocno_cost (another_allocno, hard_regno, update_cost))
1385 continue;
1386 queue_update_cost (another_allocno, allocno, divisor * COST_HOP_DIVISOR);
1387 if (record_p && ALLOCNO_COLOR_DATA (another_allocno) != NULL)
1388 ALLOCNO_COLOR_DATA (another_allocno)->update_cost_records
1389 = get_update_cost_record (hard_regno, divisor,
1390 ALLOCNO_COLOR_DATA (another_allocno)
1391 ->update_cost_records);
1394 while (get_next_update_cost (&allocno, &from, &divisor));
1397 /* Decrease preferred ALLOCNO hard register costs and costs of
1398 allocnos connected to ALLOCNO through copy. */
1399 static void
1400 update_costs_from_prefs (ira_allocno_t allocno)
1402 ira_pref_t pref;
1404 start_update_cost ();
1405 for (pref = ALLOCNO_PREFS (allocno); pref != NULL; pref = pref->next_pref)
1406 update_costs_from_allocno (allocno, pref->hard_regno,
1407 COST_HOP_DIVISOR, true, true);
1410 /* Update (decrease if DECR_P) the cost of allocnos connected to
1411 ALLOCNO through copies to increase chances to remove some copies as
1412 the result of subsequent assignment. ALLOCNO was just assigned to
1413 a hard register. Record cost updates if RECORD_P is true. */
1414 static void
1415 update_costs_from_copies (ira_allocno_t allocno, bool decr_p, bool record_p)
1417 int hard_regno;
1419 hard_regno = ALLOCNO_HARD_REGNO (allocno);
1420 ira_assert (hard_regno >= 0 && ALLOCNO_CLASS (allocno) != NO_REGS);
1421 start_update_cost ();
1422 update_costs_from_allocno (allocno, hard_regno, 1, decr_p, record_p);
1425 /* Restore costs of allocnos connected to ALLOCNO by copies as it was
1426 before updating costs of these allocnos from given allocno. This
1427 is a wise thing to do as if given allocno did not get an expected
1428 hard reg, using smaller cost of the hard reg for allocnos connected
1429 by copies to given allocno becomes actually misleading. Free all
1430 update cost records for ALLOCNO as we don't need them anymore. */
1431 static void
1432 restore_costs_from_copies (ira_allocno_t allocno)
1434 struct update_cost_record *records, *curr;
1436 if (ALLOCNO_COLOR_DATA (allocno) == NULL)
1437 return;
1438 records = ALLOCNO_COLOR_DATA (allocno)->update_cost_records;
1439 start_update_cost ();
1440 for (curr = records; curr != NULL; curr = curr->next)
1441 update_costs_from_allocno (allocno, curr->hard_regno,
1442 curr->divisor, true, false);
1443 free_update_cost_record_list (records);
1444 ALLOCNO_COLOR_DATA (allocno)->update_cost_records = NULL;
1447 /* This function updates COSTS (decrease if DECR_P) for hard_registers
1448 of ACLASS by conflict costs of the unassigned allocnos
1449 connected by copies with allocnos in update_cost_queue. This
1450 update increases chances to remove some copies. */
1451 static void
1452 update_conflict_hard_regno_costs (int *costs, enum reg_class aclass,
1453 bool decr_p)
1455 int i, cost, class_size, freq, mult, div, divisor;
1456 int index, hard_regno;
1457 int *conflict_costs;
1458 bool cont_p;
1459 enum reg_class another_aclass;
1460 ira_allocno_t allocno, another_allocno, from;
1461 ira_copy_t cp, next_cp;
1463 while (get_next_update_cost (&allocno, &from, &divisor))
1464 for (cp = ALLOCNO_COPIES (allocno); cp != NULL; cp = next_cp)
1466 if (cp->first == allocno)
1468 next_cp = cp->next_first_allocno_copy;
1469 another_allocno = cp->second;
1471 else if (cp->second == allocno)
1473 next_cp = cp->next_second_allocno_copy;
1474 another_allocno = cp->first;
1476 else
1477 gcc_unreachable ();
1479 if (another_allocno == from)
1480 continue;
1482 another_aclass = ALLOCNO_CLASS (another_allocno);
1483 if (! ira_reg_classes_intersect_p[aclass][another_aclass]
1484 || ALLOCNO_ASSIGNED_P (another_allocno)
1485 || ALLOCNO_COLOR_DATA (another_allocno)->may_be_spilled_p)
1486 continue;
1487 class_size = ira_class_hard_regs_num[another_aclass];
1488 ira_allocate_and_copy_costs
1489 (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (another_allocno),
1490 another_aclass, ALLOCNO_CONFLICT_HARD_REG_COSTS (another_allocno));
1491 conflict_costs
1492 = ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (another_allocno);
1493 if (conflict_costs == NULL)
1494 cont_p = true;
1495 else
1497 mult = cp->freq;
1498 freq = ALLOCNO_FREQ (another_allocno);
1499 if (freq == 0)
1500 freq = 1;
1501 div = freq * divisor;
1502 cont_p = false;
1503 for (i = class_size - 1; i >= 0; i--)
1505 hard_regno = ira_class_hard_regs[another_aclass][i];
1506 ira_assert (hard_regno >= 0);
1507 index = ira_class_hard_reg_index[aclass][hard_regno];
1508 if (index < 0)
1509 continue;
1510 cost = (int) ((unsigned) conflict_costs [i] * mult) / div;
1511 if (cost == 0)
1512 continue;
1513 cont_p = true;
1514 if (decr_p)
1515 cost = -cost;
1516 costs[index] += cost;
1519 /* Probably 5 hops will be enough. */
1520 if (cont_p
1521 && divisor <= (COST_HOP_DIVISOR
1522 * COST_HOP_DIVISOR
1523 * COST_HOP_DIVISOR
1524 * COST_HOP_DIVISOR))
1525 queue_update_cost (another_allocno, allocno, divisor * COST_HOP_DIVISOR);
1529 /* Set up conflicting (through CONFLICT_REGS) for each object of
1530 allocno A and the start allocno profitable regs (through
1531 START_PROFITABLE_REGS). Remember that the start profitable regs
1532 exclude hard regs which can not hold value of mode of allocno A.
1533 This covers mostly cases when multi-register value should be
1534 aligned. */
1535 static inline void
1536 get_conflict_and_start_profitable_regs (ira_allocno_t a, bool retry_p,
1537 HARD_REG_SET *conflict_regs,
1538 HARD_REG_SET *start_profitable_regs)
1540 int i, nwords;
1541 ira_object_t obj;
1543 nwords = ALLOCNO_NUM_OBJECTS (a);
1544 for (i = 0; i < nwords; i++)
1546 obj = ALLOCNO_OBJECT (a, i);
1547 COPY_HARD_REG_SET (conflict_regs[i],
1548 OBJECT_TOTAL_CONFLICT_HARD_REGS (obj));
1550 if (retry_p)
1552 COPY_HARD_REG_SET (*start_profitable_regs,
1553 reg_class_contents[ALLOCNO_CLASS (a)]);
1554 AND_COMPL_HARD_REG_SET (*start_profitable_regs,
1555 ira_prohibited_class_mode_regs
1556 [ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)]);
1558 else
1559 COPY_HARD_REG_SET (*start_profitable_regs,
1560 ALLOCNO_COLOR_DATA (a)->profitable_hard_regs);
1563 /* Return true if HARD_REGNO is ok for assigning to allocno A with
1564 PROFITABLE_REGS and whose objects have CONFLICT_REGS. */
1565 static inline bool
1566 check_hard_reg_p (ira_allocno_t a, int hard_regno,
1567 HARD_REG_SET *conflict_regs, HARD_REG_SET profitable_regs)
1569 int j, nwords, nregs;
1570 enum reg_class aclass;
1571 enum machine_mode mode;
1573 aclass = ALLOCNO_CLASS (a);
1574 mode = ALLOCNO_MODE (a);
1575 if (TEST_HARD_REG_BIT (ira_prohibited_class_mode_regs[aclass][mode],
1576 hard_regno))
1577 return false;
1578 /* Checking only profitable hard regs. */
1579 if (! TEST_HARD_REG_BIT (profitable_regs, hard_regno))
1580 return false;
1581 nregs = hard_regno_nregs[hard_regno][mode];
1582 nwords = ALLOCNO_NUM_OBJECTS (a);
1583 for (j = 0; j < nregs; j++)
1585 int k;
1586 int set_to_test_start = 0, set_to_test_end = nwords;
1588 if (nregs == nwords)
1590 if (REG_WORDS_BIG_ENDIAN)
1591 set_to_test_start = nwords - j - 1;
1592 else
1593 set_to_test_start = j;
1594 set_to_test_end = set_to_test_start + 1;
1596 for (k = set_to_test_start; k < set_to_test_end; k++)
1597 if (TEST_HARD_REG_BIT (conflict_regs[k], hard_regno + j))
1598 break;
1599 if (k != set_to_test_end)
1600 break;
1602 return j == nregs;
1605 /* Return number of registers needed to be saved and restored at
1606 function prologue/epilogue if we allocate HARD_REGNO to hold value
1607 of MODE. */
1608 static int
1609 calculate_saved_nregs (int hard_regno, enum machine_mode mode)
1611 int i;
1612 int nregs = 0;
1614 ira_assert (hard_regno >= 0);
1615 for (i = hard_regno_nregs[hard_regno][mode] - 1; i >= 0; i--)
1616 if (!allocated_hardreg_p[hard_regno + i]
1617 && !TEST_HARD_REG_BIT (call_used_reg_set, hard_regno + i)
1618 && !LOCAL_REGNO (hard_regno + i))
1619 nregs++;
1620 return nregs;
1623 /* Choose a hard register for allocno A. If RETRY_P is TRUE, it means
1624 that the function called from function
1625 `ira_reassign_conflict_allocnos' and `allocno_reload_assign'. In
1626 this case some allocno data are not defined or updated and we
1627 should not touch these data. The function returns true if we
1628 managed to assign a hard register to the allocno.
1630 To assign a hard register, first of all we calculate all conflict
1631 hard registers which can come from conflicting allocnos with
1632 already assigned hard registers. After that we find first free
1633 hard register with the minimal cost. During hard register cost
1634 calculation we take conflict hard register costs into account to
1635 give a chance for conflicting allocnos to get a better hard
1636 register in the future.
1638 If the best hard register cost is bigger than cost of memory usage
1639 for the allocno, we don't assign a hard register to given allocno
1640 at all.
1642 If we assign a hard register to the allocno, we update costs of the
1643 hard register for allocnos connected by copies to improve a chance
1644 to coalesce insns represented by the copies when we assign hard
1645 registers to the allocnos connected by the copies. */
1646 static bool
1647 assign_hard_reg (ira_allocno_t a, bool retry_p)
1649 HARD_REG_SET conflicting_regs[2], profitable_hard_regs;
1650 int i, j, hard_regno, best_hard_regno, class_size;
1651 int cost, mem_cost, min_cost, full_cost, min_full_cost, nwords, word;
1652 int *a_costs;
1653 enum reg_class aclass;
1654 enum machine_mode mode;
1655 static int costs[FIRST_PSEUDO_REGISTER], full_costs[FIRST_PSEUDO_REGISTER];
1656 int saved_nregs;
1657 enum reg_class rclass;
1658 int add_cost;
1659 #ifdef STACK_REGS
1660 bool no_stack_reg_p;
1661 #endif
1663 ira_assert (! ALLOCNO_ASSIGNED_P (a));
1664 get_conflict_and_start_profitable_regs (a, retry_p,
1665 conflicting_regs,
1666 &profitable_hard_regs);
1667 aclass = ALLOCNO_CLASS (a);
1668 class_size = ira_class_hard_regs_num[aclass];
1669 best_hard_regno = -1;
1670 memset (full_costs, 0, sizeof (int) * class_size);
1671 mem_cost = 0;
1672 memset (costs, 0, sizeof (int) * class_size);
1673 memset (full_costs, 0, sizeof (int) * class_size);
1674 #ifdef STACK_REGS
1675 no_stack_reg_p = false;
1676 #endif
1677 if (! retry_p)
1678 start_update_cost ();
1679 mem_cost += ALLOCNO_UPDATED_MEMORY_COST (a);
1681 ira_allocate_and_copy_costs (&ALLOCNO_UPDATED_HARD_REG_COSTS (a),
1682 aclass, ALLOCNO_HARD_REG_COSTS (a));
1683 a_costs = ALLOCNO_UPDATED_HARD_REG_COSTS (a);
1684 #ifdef STACK_REGS
1685 no_stack_reg_p = no_stack_reg_p || ALLOCNO_TOTAL_NO_STACK_REG_P (a);
1686 #endif
1687 cost = ALLOCNO_UPDATED_CLASS_COST (a);
1688 for (i = 0; i < class_size; i++)
1689 if (a_costs != NULL)
1691 costs[i] += a_costs[i];
1692 full_costs[i] += a_costs[i];
1694 else
1696 costs[i] += cost;
1697 full_costs[i] += cost;
1699 nwords = ALLOCNO_NUM_OBJECTS (a);
1700 curr_allocno_process++;
1701 for (word = 0; word < nwords; word++)
1703 ira_object_t conflict_obj;
1704 ira_object_t obj = ALLOCNO_OBJECT (a, word);
1705 ira_object_conflict_iterator oci;
1707 /* Take preferences of conflicting allocnos into account. */
1708 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
1710 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
1711 enum reg_class conflict_aclass;
1712 allocno_color_data_t data = ALLOCNO_COLOR_DATA (conflict_a);
1714 /* Reload can give another class so we need to check all
1715 allocnos. */
1716 if (!retry_p
1717 && (!bitmap_bit_p (consideration_allocno_bitmap,
1718 ALLOCNO_NUM (conflict_a))
1719 || ((!ALLOCNO_ASSIGNED_P (conflict_a)
1720 || ALLOCNO_HARD_REGNO (conflict_a) < 0)
1721 && !(hard_reg_set_intersect_p
1722 (profitable_hard_regs,
1723 ALLOCNO_COLOR_DATA
1724 (conflict_a)->profitable_hard_regs)))))
1725 continue;
1726 conflict_aclass = ALLOCNO_CLASS (conflict_a);
1727 ira_assert (ira_reg_classes_intersect_p
1728 [aclass][conflict_aclass]);
1729 if (ALLOCNO_ASSIGNED_P (conflict_a))
1731 hard_regno = ALLOCNO_HARD_REGNO (conflict_a);
1732 if (hard_regno >= 0
1733 && (ira_hard_reg_set_intersection_p
1734 (hard_regno, ALLOCNO_MODE (conflict_a),
1735 reg_class_contents[aclass])))
1737 int n_objects = ALLOCNO_NUM_OBJECTS (conflict_a);
1738 int conflict_nregs;
1740 mode = ALLOCNO_MODE (conflict_a);
1741 conflict_nregs = hard_regno_nregs[hard_regno][mode];
1742 if (conflict_nregs == n_objects && conflict_nregs > 1)
1744 int num = OBJECT_SUBWORD (conflict_obj);
1746 if (REG_WORDS_BIG_ENDIAN)
1747 SET_HARD_REG_BIT (conflicting_regs[word],
1748 hard_regno + n_objects - num - 1);
1749 else
1750 SET_HARD_REG_BIT (conflicting_regs[word],
1751 hard_regno + num);
1753 else
1754 IOR_HARD_REG_SET
1755 (conflicting_regs[word],
1756 ira_reg_mode_hard_regset[hard_regno][mode]);
1757 if (hard_reg_set_subset_p (profitable_hard_regs,
1758 conflicting_regs[word]))
1759 goto fail;
1762 else if (! retry_p
1763 && ! ALLOCNO_COLOR_DATA (conflict_a)->may_be_spilled_p
1764 /* Don't process the conflict allocno twice. */
1765 && (ALLOCNO_COLOR_DATA (conflict_a)->last_process
1766 != curr_allocno_process))
1768 int k, *conflict_costs;
1770 ALLOCNO_COLOR_DATA (conflict_a)->last_process
1771 = curr_allocno_process;
1772 ira_allocate_and_copy_costs
1773 (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (conflict_a),
1774 conflict_aclass,
1775 ALLOCNO_CONFLICT_HARD_REG_COSTS (conflict_a));
1776 conflict_costs
1777 = ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (conflict_a);
1778 if (conflict_costs != NULL)
1779 for (j = class_size - 1; j >= 0; j--)
1781 hard_regno = ira_class_hard_regs[aclass][j];
1782 ira_assert (hard_regno >= 0);
1783 k = ira_class_hard_reg_index[conflict_aclass][hard_regno];
1784 if (k < 0
1785 /* If HARD_REGNO is not available for CONFLICT_A,
1786 the conflict would be ignored, since HARD_REGNO
1787 will never be assigned to CONFLICT_A. */
1788 || !TEST_HARD_REG_BIT (data->profitable_hard_regs,
1789 hard_regno))
1790 continue;
1791 full_costs[j] -= conflict_costs[k];
1793 queue_update_cost (conflict_a, NULL, COST_HOP_DIVISOR);
1798 if (! retry_p)
1799 /* Take into account preferences of allocnos connected by copies to
1800 the conflict allocnos. */
1801 update_conflict_hard_regno_costs (full_costs, aclass, true);
1803 /* Take preferences of allocnos connected by copies into
1804 account. */
1805 if (! retry_p)
1807 start_update_cost ();
1808 queue_update_cost (a, NULL, COST_HOP_DIVISOR);
1809 update_conflict_hard_regno_costs (full_costs, aclass, false);
1811 min_cost = min_full_cost = INT_MAX;
1812 /* We don't care about giving callee saved registers to allocnos no
1813 living through calls because call clobbered registers are
1814 allocated first (it is usual practice to put them first in
1815 REG_ALLOC_ORDER). */
1816 mode = ALLOCNO_MODE (a);
1817 for (i = 0; i < class_size; i++)
1819 hard_regno = ira_class_hard_regs[aclass][i];
1820 #ifdef STACK_REGS
1821 if (no_stack_reg_p
1822 && FIRST_STACK_REG <= hard_regno && hard_regno <= LAST_STACK_REG)
1823 continue;
1824 #endif
1825 if (! check_hard_reg_p (a, hard_regno,
1826 conflicting_regs, profitable_hard_regs))
1827 continue;
1828 cost = costs[i];
1829 full_cost = full_costs[i];
1830 if (!HONOR_REG_ALLOC_ORDER)
1832 if ((saved_nregs = calculate_saved_nregs (hard_regno, mode)) != 0)
1833 /* We need to save/restore the hard register in
1834 epilogue/prologue. Therefore we increase the cost. */
1836 rclass = REGNO_REG_CLASS (hard_regno);
1837 add_cost = ((ira_memory_move_cost[mode][rclass][0]
1838 + ira_memory_move_cost[mode][rclass][1])
1839 * saved_nregs / hard_regno_nregs[hard_regno][mode] - 1);
1840 cost += add_cost;
1841 full_cost += add_cost;
1844 if (min_cost > cost)
1845 min_cost = cost;
1846 if (min_full_cost > full_cost)
1848 min_full_cost = full_cost;
1849 best_hard_regno = hard_regno;
1850 ira_assert (hard_regno >= 0);
1853 if (min_full_cost > mem_cost)
1855 if (! retry_p && internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
1856 fprintf (ira_dump_file, "(memory is more profitable %d vs %d) ",
1857 mem_cost, min_full_cost);
1858 best_hard_regno = -1;
1860 fail:
1861 if (best_hard_regno >= 0)
1863 for (i = hard_regno_nregs[best_hard_regno][mode] - 1; i >= 0; i--)
1864 allocated_hardreg_p[best_hard_regno + i] = true;
1866 if (! retry_p)
1867 restore_costs_from_copies (a);
1868 ALLOCNO_HARD_REGNO (a) = best_hard_regno;
1869 ALLOCNO_ASSIGNED_P (a) = true;
1870 if (best_hard_regno >= 0)
1871 update_costs_from_copies (a, true, ! retry_p);
1872 ira_assert (ALLOCNO_CLASS (a) == aclass);
1873 /* We don't need updated costs anymore. */
1874 ira_free_allocno_updated_costs (a);
1875 return best_hard_regno >= 0;
1880 /* An array used to sort copies. */
1881 static ira_copy_t *sorted_copies;
1883 /* Return TRUE if live ranges of allocnos A1 and A2 intersect. It is
1884 used to find a conflict for new allocnos or allocnos with the
1885 different allocno classes. */
1886 static bool
1887 allocnos_conflict_by_live_ranges_p (ira_allocno_t a1, ira_allocno_t a2)
1889 rtx reg1, reg2;
1890 int i, j;
1891 int n1 = ALLOCNO_NUM_OBJECTS (a1);
1892 int n2 = ALLOCNO_NUM_OBJECTS (a2);
1894 if (a1 == a2)
1895 return false;
1896 reg1 = regno_reg_rtx[ALLOCNO_REGNO (a1)];
1897 reg2 = regno_reg_rtx[ALLOCNO_REGNO (a2)];
1898 if (reg1 != NULL && reg2 != NULL
1899 && ORIGINAL_REGNO (reg1) == ORIGINAL_REGNO (reg2))
1900 return false;
1902 for (i = 0; i < n1; i++)
1904 ira_object_t c1 = ALLOCNO_OBJECT (a1, i);
1906 for (j = 0; j < n2; j++)
1908 ira_object_t c2 = ALLOCNO_OBJECT (a2, j);
1910 if (ira_live_ranges_intersect_p (OBJECT_LIVE_RANGES (c1),
1911 OBJECT_LIVE_RANGES (c2)))
1912 return true;
1915 return false;
1918 /* The function is used to sort copies according to their execution
1919 frequencies. */
1920 static int
1921 copy_freq_compare_func (const void *v1p, const void *v2p)
1923 ira_copy_t cp1 = *(const ira_copy_t *) v1p, cp2 = *(const ira_copy_t *) v2p;
1924 int pri1, pri2;
1926 pri1 = cp1->freq;
1927 pri2 = cp2->freq;
1928 if (pri2 - pri1)
1929 return pri2 - pri1;
1931 /* If freqencies are equal, sort by copies, so that the results of
1932 qsort leave nothing to chance. */
1933 return cp1->num - cp2->num;
1938 /* Return true if any allocno from thread of A1 conflicts with any
1939 allocno from thread A2. */
1940 static bool
1941 allocno_thread_conflict_p (ira_allocno_t a1, ira_allocno_t a2)
1943 ira_allocno_t a, conflict_a;
1945 for (a = ALLOCNO_COLOR_DATA (a2)->next_thread_allocno;;
1946 a = ALLOCNO_COLOR_DATA (a)->next_thread_allocno)
1948 for (conflict_a = ALLOCNO_COLOR_DATA (a1)->next_thread_allocno;;
1949 conflict_a = ALLOCNO_COLOR_DATA (conflict_a)->next_thread_allocno)
1951 if (allocnos_conflict_by_live_ranges_p (a, conflict_a))
1952 return true;
1953 if (conflict_a == a1)
1954 break;
1956 if (a == a2)
1957 break;
1959 return false;
1962 /* Merge two threads given correspondingly by their first allocnos T1
1963 and T2 (more accurately merging T2 into T1). */
1964 static void
1965 merge_threads (ira_allocno_t t1, ira_allocno_t t2)
1967 ira_allocno_t a, next, last;
1969 gcc_assert (t1 != t2
1970 && ALLOCNO_COLOR_DATA (t1)->first_thread_allocno == t1
1971 && ALLOCNO_COLOR_DATA (t2)->first_thread_allocno == t2);
1972 for (last = t2, a = ALLOCNO_COLOR_DATA (t2)->next_thread_allocno;;
1973 a = ALLOCNO_COLOR_DATA (a)->next_thread_allocno)
1975 ALLOCNO_COLOR_DATA (a)->first_thread_allocno = t1;
1976 if (a == t2)
1977 break;
1978 last = a;
1980 next = ALLOCNO_COLOR_DATA (t1)->next_thread_allocno;
1981 ALLOCNO_COLOR_DATA (t1)->next_thread_allocno = t2;
1982 ALLOCNO_COLOR_DATA (last)->next_thread_allocno = next;
1983 ALLOCNO_COLOR_DATA (t1)->thread_freq += ALLOCNO_COLOR_DATA (t2)->thread_freq;
1986 /* Create threads by processing CP_NUM copies from sorted)ciopeis. We
1987 process the most expensive copies first. */
1988 static void
1989 form_threads_from_copies (int cp_num)
1991 ira_allocno_t a, thread1, thread2;
1992 ira_copy_t cp;
1993 int i, n;
1995 qsort (sorted_copies, cp_num, sizeof (ira_copy_t), copy_freq_compare_func);
1996 /* Form threads processing copies, most frequently executed
1997 first. */
1998 for (; cp_num != 0;)
2000 for (i = 0; i < cp_num; i++)
2002 cp = sorted_copies[i];
2003 thread1 = ALLOCNO_COLOR_DATA (cp->first)->first_thread_allocno;
2004 thread2 = ALLOCNO_COLOR_DATA (cp->second)->first_thread_allocno;
2005 if (thread1 == thread2)
2006 continue;
2007 if (! allocno_thread_conflict_p (thread1, thread2))
2009 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2010 fprintf
2011 (ira_dump_file,
2012 " Forming thread by copy %d:a%dr%d-a%dr%d (freq=%d):\n",
2013 cp->num, ALLOCNO_NUM (cp->first), ALLOCNO_REGNO (cp->first),
2014 ALLOCNO_NUM (cp->second), ALLOCNO_REGNO (cp->second),
2015 cp->freq);
2016 merge_threads (thread1, thread2);
2017 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2019 thread1 = ALLOCNO_COLOR_DATA (thread1)->first_thread_allocno;
2020 fprintf (ira_dump_file, " Result (freq=%d): a%dr%d(%d)",
2021 ALLOCNO_COLOR_DATA (thread1)->thread_freq,
2022 ALLOCNO_NUM (thread1), ALLOCNO_REGNO (thread1),
2023 ALLOCNO_FREQ (thread1));
2024 for (a = ALLOCNO_COLOR_DATA (thread1)->next_thread_allocno;
2025 a != thread1;
2026 a = ALLOCNO_COLOR_DATA (a)->next_thread_allocno)
2027 fprintf (ira_dump_file, " a%dr%d(%d)",
2028 ALLOCNO_NUM (a), ALLOCNO_REGNO (a),
2029 ALLOCNO_FREQ (a));
2030 fprintf (ira_dump_file, "\n");
2032 i++;
2033 break;
2036 /* Collect the rest of copies. */
2037 for (n = 0; i < cp_num; i++)
2039 cp = sorted_copies[i];
2040 if (ALLOCNO_COLOR_DATA (cp->first)->first_thread_allocno
2041 != ALLOCNO_COLOR_DATA (cp->second)->first_thread_allocno)
2042 sorted_copies[n++] = cp;
2044 cp_num = n;
2048 /* Create threads by processing copies of all alocnos from BUCKET. We
2049 process the most expensive copies first. */
2050 static void
2051 form_threads_from_bucket (ira_allocno_t bucket)
2053 ira_allocno_t a;
2054 ira_copy_t cp, next_cp;
2055 int cp_num = 0;
2057 for (a = bucket; a != NULL; a = ALLOCNO_COLOR_DATA (a)->next_bucket_allocno)
2059 for (cp = ALLOCNO_COPIES (a); cp != NULL; cp = next_cp)
2061 if (cp->first == a)
2063 next_cp = cp->next_first_allocno_copy;
2064 sorted_copies[cp_num++] = cp;
2066 else if (cp->second == a)
2067 next_cp = cp->next_second_allocno_copy;
2068 else
2069 gcc_unreachable ();
2072 form_threads_from_copies (cp_num);
2075 /* Create threads by processing copies of colorable allocno A. We
2076 process most expensive copies first. */
2077 static void
2078 form_threads_from_colorable_allocno (ira_allocno_t a)
2080 ira_allocno_t another_a;
2081 ira_copy_t cp, next_cp;
2082 int cp_num = 0;
2084 for (cp = ALLOCNO_COPIES (a); cp != NULL; cp = next_cp)
2086 if (cp->first == a)
2088 next_cp = cp->next_first_allocno_copy;
2089 another_a = cp->second;
2091 else if (cp->second == a)
2093 next_cp = cp->next_second_allocno_copy;
2094 another_a = cp->first;
2096 else
2097 gcc_unreachable ();
2098 if ((! ALLOCNO_COLOR_DATA (another_a)->in_graph_p
2099 && !ALLOCNO_COLOR_DATA (another_a)->may_be_spilled_p)
2100 || ALLOCNO_COLOR_DATA (another_a)->colorable_p)
2101 sorted_copies[cp_num++] = cp;
2103 form_threads_from_copies (cp_num);
2106 /* Form initial threads which contain only one allocno. */
2107 static void
2108 init_allocno_threads (void)
2110 ira_allocno_t a;
2111 unsigned int j;
2112 bitmap_iterator bi;
2114 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, j, bi)
2116 a = ira_allocnos[j];
2117 /* Set up initial thread data: */
2118 ALLOCNO_COLOR_DATA (a)->first_thread_allocno
2119 = ALLOCNO_COLOR_DATA (a)->next_thread_allocno = a;
2120 ALLOCNO_COLOR_DATA (a)->thread_freq = ALLOCNO_FREQ (a);
2126 /* This page contains the allocator based on the Chaitin-Briggs algorithm. */
2128 /* Bucket of allocnos that can colored currently without spilling. */
2129 static ira_allocno_t colorable_allocno_bucket;
2131 /* Bucket of allocnos that might be not colored currently without
2132 spilling. */
2133 static ira_allocno_t uncolorable_allocno_bucket;
2135 /* The current number of allocnos in the uncolorable_bucket. */
2136 static int uncolorable_allocnos_num;
2138 /* Return the current spill priority of allocno A. The less the
2139 number, the more preferable the allocno for spilling. */
2140 static inline int
2141 allocno_spill_priority (ira_allocno_t a)
2143 allocno_color_data_t data = ALLOCNO_COLOR_DATA (a);
2145 return (data->temp
2146 / (ALLOCNO_EXCESS_PRESSURE_POINTS_NUM (a)
2147 * ira_reg_class_max_nregs[ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)]
2148 + 1));
2151 /* Add allocno A to bucket *BUCKET_PTR. A should be not in a bucket
2152 before the call. */
2153 static void
2154 add_allocno_to_bucket (ira_allocno_t a, ira_allocno_t *bucket_ptr)
2156 ira_allocno_t first_a;
2157 allocno_color_data_t data;
2159 if (bucket_ptr == &uncolorable_allocno_bucket
2160 && ALLOCNO_CLASS (a) != NO_REGS)
2162 uncolorable_allocnos_num++;
2163 ira_assert (uncolorable_allocnos_num > 0);
2165 first_a = *bucket_ptr;
2166 data = ALLOCNO_COLOR_DATA (a);
2167 data->next_bucket_allocno = first_a;
2168 data->prev_bucket_allocno = NULL;
2169 if (first_a != NULL)
2170 ALLOCNO_COLOR_DATA (first_a)->prev_bucket_allocno = a;
2171 *bucket_ptr = a;
2174 /* Compare two allocnos to define which allocno should be pushed first
2175 into the coloring stack. If the return is a negative number, the
2176 allocno given by the first parameter will be pushed first. In this
2177 case such allocno has less priority than the second one and the
2178 hard register will be assigned to it after assignment to the second
2179 one. As the result of such assignment order, the second allocno
2180 has a better chance to get the best hard register. */
2181 static int
2182 bucket_allocno_compare_func (const void *v1p, const void *v2p)
2184 ira_allocno_t a1 = *(const ira_allocno_t *) v1p;
2185 ira_allocno_t a2 = *(const ira_allocno_t *) v2p;
2186 int diff, freq1, freq2, a1_num, a2_num;
2187 ira_allocno_t t1 = ALLOCNO_COLOR_DATA (a1)->first_thread_allocno;
2188 ira_allocno_t t2 = ALLOCNO_COLOR_DATA (a2)->first_thread_allocno;
2189 int cl1 = ALLOCNO_CLASS (a1), cl2 = ALLOCNO_CLASS (a2);
2191 freq1 = ALLOCNO_COLOR_DATA (t1)->thread_freq;
2192 freq2 = ALLOCNO_COLOR_DATA (t2)->thread_freq;
2193 if ((diff = freq1 - freq2) != 0)
2194 return diff;
2196 if ((diff = ALLOCNO_NUM (t2) - ALLOCNO_NUM (t1)) != 0)
2197 return diff;
2199 /* Push pseudos requiring less hard registers first. It means that
2200 we will assign pseudos requiring more hard registers first
2201 avoiding creation small holes in free hard register file into
2202 which the pseudos requiring more hard registers can not fit. */
2203 if ((diff = (ira_reg_class_max_nregs[cl1][ALLOCNO_MODE (a1)]
2204 - ira_reg_class_max_nregs[cl2][ALLOCNO_MODE (a2)])) != 0)
2205 return diff;
2207 freq1 = ALLOCNO_FREQ (a1);
2208 freq2 = ALLOCNO_FREQ (a2);
2209 if ((diff = freq1 - freq2) != 0)
2210 return diff;
2212 a1_num = ALLOCNO_COLOR_DATA (a1)->available_regs_num;
2213 a2_num = ALLOCNO_COLOR_DATA (a2)->available_regs_num;
2214 if ((diff = a2_num - a1_num) != 0)
2215 return diff;
2216 return ALLOCNO_NUM (a2) - ALLOCNO_NUM (a1);
2219 /* Sort bucket *BUCKET_PTR and return the result through
2220 BUCKET_PTR. */
2221 static void
2222 sort_bucket (ira_allocno_t *bucket_ptr,
2223 int (*compare_func) (const void *, const void *))
2225 ira_allocno_t a, head;
2226 int n;
2228 for (n = 0, a = *bucket_ptr;
2229 a != NULL;
2230 a = ALLOCNO_COLOR_DATA (a)->next_bucket_allocno)
2231 sorted_allocnos[n++] = a;
2232 if (n <= 1)
2233 return;
2234 qsort (sorted_allocnos, n, sizeof (ira_allocno_t), compare_func);
2235 head = NULL;
2236 for (n--; n >= 0; n--)
2238 a = sorted_allocnos[n];
2239 ALLOCNO_COLOR_DATA (a)->next_bucket_allocno = head;
2240 ALLOCNO_COLOR_DATA (a)->prev_bucket_allocno = NULL;
2241 if (head != NULL)
2242 ALLOCNO_COLOR_DATA (head)->prev_bucket_allocno = a;
2243 head = a;
2245 *bucket_ptr = head;
2248 /* Add ALLOCNO to colorable bucket maintaining the order according
2249 their priority. ALLOCNO should be not in a bucket before the
2250 call. */
2251 static void
2252 add_allocno_to_ordered_colorable_bucket (ira_allocno_t allocno)
2254 ira_allocno_t before, after;
2256 form_threads_from_colorable_allocno (allocno);
2257 for (before = colorable_allocno_bucket, after = NULL;
2258 before != NULL;
2259 after = before,
2260 before = ALLOCNO_COLOR_DATA (before)->next_bucket_allocno)
2261 if (bucket_allocno_compare_func (&allocno, &before) < 0)
2262 break;
2263 ALLOCNO_COLOR_DATA (allocno)->next_bucket_allocno = before;
2264 ALLOCNO_COLOR_DATA (allocno)->prev_bucket_allocno = after;
2265 if (after == NULL)
2266 colorable_allocno_bucket = allocno;
2267 else
2268 ALLOCNO_COLOR_DATA (after)->next_bucket_allocno = allocno;
2269 if (before != NULL)
2270 ALLOCNO_COLOR_DATA (before)->prev_bucket_allocno = allocno;
2273 /* Delete ALLOCNO from bucket *BUCKET_PTR. It should be there before
2274 the call. */
2275 static void
2276 delete_allocno_from_bucket (ira_allocno_t allocno, ira_allocno_t *bucket_ptr)
2278 ira_allocno_t prev_allocno, next_allocno;
2280 if (bucket_ptr == &uncolorable_allocno_bucket
2281 && ALLOCNO_CLASS (allocno) != NO_REGS)
2283 uncolorable_allocnos_num--;
2284 ira_assert (uncolorable_allocnos_num >= 0);
2286 prev_allocno = ALLOCNO_COLOR_DATA (allocno)->prev_bucket_allocno;
2287 next_allocno = ALLOCNO_COLOR_DATA (allocno)->next_bucket_allocno;
2288 if (prev_allocno != NULL)
2289 ALLOCNO_COLOR_DATA (prev_allocno)->next_bucket_allocno = next_allocno;
2290 else
2292 ira_assert (*bucket_ptr == allocno);
2293 *bucket_ptr = next_allocno;
2295 if (next_allocno != NULL)
2296 ALLOCNO_COLOR_DATA (next_allocno)->prev_bucket_allocno = prev_allocno;
2299 /* Put allocno A onto the coloring stack without removing it from its
2300 bucket. Pushing allocno to the coloring stack can result in moving
2301 conflicting allocnos from the uncolorable bucket to the colorable
2302 one. */
2303 static void
2304 push_allocno_to_stack (ira_allocno_t a)
2306 enum reg_class aclass;
2307 allocno_color_data_t data, conflict_data;
2308 int size, i, n = ALLOCNO_NUM_OBJECTS (a);
2310 data = ALLOCNO_COLOR_DATA (a);
2311 data->in_graph_p = false;
2312 allocno_stack_vec.safe_push (a);
2313 aclass = ALLOCNO_CLASS (a);
2314 if (aclass == NO_REGS)
2315 return;
2316 size = ira_reg_class_max_nregs[aclass][ALLOCNO_MODE (a)];
2317 if (n > 1)
2319 /* We will deal with the subwords individually. */
2320 gcc_assert (size == ALLOCNO_NUM_OBJECTS (a));
2321 size = 1;
2323 for (i = 0; i < n; i++)
2325 ira_object_t obj = ALLOCNO_OBJECT (a, i);
2326 ira_object_t conflict_obj;
2327 ira_object_conflict_iterator oci;
2329 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
2331 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
2333 conflict_data = ALLOCNO_COLOR_DATA (conflict_a);
2334 if (conflict_data->colorable_p
2335 || ! conflict_data->in_graph_p
2336 || ALLOCNO_ASSIGNED_P (conflict_a)
2337 || !(hard_reg_set_intersect_p
2338 (ALLOCNO_COLOR_DATA (a)->profitable_hard_regs,
2339 conflict_data->profitable_hard_regs)))
2340 continue;
2341 ira_assert (bitmap_bit_p (coloring_allocno_bitmap,
2342 ALLOCNO_NUM (conflict_a)));
2343 if (update_left_conflict_sizes_p (conflict_a, a, size))
2345 delete_allocno_from_bucket
2346 (conflict_a, &uncolorable_allocno_bucket);
2347 add_allocno_to_ordered_colorable_bucket (conflict_a);
2348 if (internal_flag_ira_verbose > 4 && ira_dump_file != NULL)
2350 fprintf (ira_dump_file, " Making");
2351 ira_print_expanded_allocno (conflict_a);
2352 fprintf (ira_dump_file, " colorable\n");
2360 /* Put ALLOCNO onto the coloring stack and remove it from its bucket.
2361 The allocno is in the colorable bucket if COLORABLE_P is TRUE. */
2362 static void
2363 remove_allocno_from_bucket_and_push (ira_allocno_t allocno, bool colorable_p)
2365 if (colorable_p)
2366 delete_allocno_from_bucket (allocno, &colorable_allocno_bucket);
2367 else
2368 delete_allocno_from_bucket (allocno, &uncolorable_allocno_bucket);
2369 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2371 fprintf (ira_dump_file, " Pushing");
2372 ira_print_expanded_allocno (allocno);
2373 if (colorable_p)
2374 fprintf (ira_dump_file, "(cost %d)\n",
2375 ALLOCNO_COLOR_DATA (allocno)->temp);
2376 else
2377 fprintf (ira_dump_file, "(potential spill: %spri=%d, cost=%d)\n",
2378 ALLOCNO_BAD_SPILL_P (allocno) ? "bad spill, " : "",
2379 allocno_spill_priority (allocno),
2380 ALLOCNO_COLOR_DATA (allocno)->temp);
2382 if (! colorable_p)
2383 ALLOCNO_COLOR_DATA (allocno)->may_be_spilled_p = true;
2384 push_allocno_to_stack (allocno);
2387 /* Put all allocnos from colorable bucket onto the coloring stack. */
2388 static void
2389 push_only_colorable (void)
2391 form_threads_from_bucket (colorable_allocno_bucket);
2392 sort_bucket (&colorable_allocno_bucket, bucket_allocno_compare_func);
2393 for (;colorable_allocno_bucket != NULL;)
2394 remove_allocno_from_bucket_and_push (colorable_allocno_bucket, true);
2397 /* Return the frequency of exit edges (if EXIT_P) or entry from/to the
2398 loop given by its LOOP_NODE. */
2400 ira_loop_edge_freq (ira_loop_tree_node_t loop_node, int regno, bool exit_p)
2402 int freq, i;
2403 edge_iterator ei;
2404 edge e;
2405 vec<edge> edges;
2407 ira_assert (current_loops != NULL && loop_node->loop != NULL
2408 && (regno < 0 || regno >= FIRST_PSEUDO_REGISTER));
2409 freq = 0;
2410 if (! exit_p)
2412 FOR_EACH_EDGE (e, ei, loop_node->loop->header->preds)
2413 if (e->src != loop_node->loop->latch
2414 && (regno < 0
2415 || (bitmap_bit_p (df_get_live_out (e->src), regno)
2416 && bitmap_bit_p (df_get_live_in (e->dest), regno))))
2417 freq += EDGE_FREQUENCY (e);
2419 else
2421 edges = get_loop_exit_edges (loop_node->loop);
2422 FOR_EACH_VEC_ELT (edges, i, e)
2423 if (regno < 0
2424 || (bitmap_bit_p (df_get_live_out (e->src), regno)
2425 && bitmap_bit_p (df_get_live_in (e->dest), regno)))
2426 freq += EDGE_FREQUENCY (e);
2427 edges.release ();
2430 return REG_FREQ_FROM_EDGE_FREQ (freq);
2433 /* Calculate and return the cost of putting allocno A into memory. */
2434 static int
2435 calculate_allocno_spill_cost (ira_allocno_t a)
2437 int regno, cost;
2438 enum machine_mode mode;
2439 enum reg_class rclass;
2440 ira_allocno_t parent_allocno;
2441 ira_loop_tree_node_t parent_node, loop_node;
2443 regno = ALLOCNO_REGNO (a);
2444 cost = ALLOCNO_UPDATED_MEMORY_COST (a) - ALLOCNO_UPDATED_CLASS_COST (a);
2445 if (ALLOCNO_CAP (a) != NULL)
2446 return cost;
2447 loop_node = ALLOCNO_LOOP_TREE_NODE (a);
2448 if ((parent_node = loop_node->parent) == NULL)
2449 return cost;
2450 if ((parent_allocno = parent_node->regno_allocno_map[regno]) == NULL)
2451 return cost;
2452 mode = ALLOCNO_MODE (a);
2453 rclass = ALLOCNO_CLASS (a);
2454 if (ALLOCNO_HARD_REGNO (parent_allocno) < 0)
2455 cost -= (ira_memory_move_cost[mode][rclass][0]
2456 * ira_loop_edge_freq (loop_node, regno, true)
2457 + ira_memory_move_cost[mode][rclass][1]
2458 * ira_loop_edge_freq (loop_node, regno, false));
2459 else
2461 ira_init_register_move_cost_if_necessary (mode);
2462 cost += ((ira_memory_move_cost[mode][rclass][1]
2463 * ira_loop_edge_freq (loop_node, regno, true)
2464 + ira_memory_move_cost[mode][rclass][0]
2465 * ira_loop_edge_freq (loop_node, regno, false))
2466 - (ira_register_move_cost[mode][rclass][rclass]
2467 * (ira_loop_edge_freq (loop_node, regno, false)
2468 + ira_loop_edge_freq (loop_node, regno, true))));
2470 return cost;
2473 /* Used for sorting allocnos for spilling. */
2474 static inline int
2475 allocno_spill_priority_compare (ira_allocno_t a1, ira_allocno_t a2)
2477 int pri1, pri2, diff;
2479 if (ALLOCNO_BAD_SPILL_P (a1) && ! ALLOCNO_BAD_SPILL_P (a2))
2480 return 1;
2481 if (ALLOCNO_BAD_SPILL_P (a2) && ! ALLOCNO_BAD_SPILL_P (a1))
2482 return -1;
2483 pri1 = allocno_spill_priority (a1);
2484 pri2 = allocno_spill_priority (a2);
2485 if ((diff = pri1 - pri2) != 0)
2486 return diff;
2487 if ((diff
2488 = ALLOCNO_COLOR_DATA (a1)->temp - ALLOCNO_COLOR_DATA (a2)->temp) != 0)
2489 return diff;
2490 return ALLOCNO_NUM (a1) - ALLOCNO_NUM (a2);
2493 /* Used for sorting allocnos for spilling. */
2494 static int
2495 allocno_spill_sort_compare (const void *v1p, const void *v2p)
2497 ira_allocno_t p1 = *(const ira_allocno_t *) v1p;
2498 ira_allocno_t p2 = *(const ira_allocno_t *) v2p;
2500 return allocno_spill_priority_compare (p1, p2);
2503 /* Push allocnos to the coloring stack. The order of allocnos in the
2504 stack defines the order for the subsequent coloring. */
2505 static void
2506 push_allocnos_to_stack (void)
2508 ira_allocno_t a;
2509 int cost;
2511 /* Calculate uncolorable allocno spill costs. */
2512 for (a = uncolorable_allocno_bucket;
2513 a != NULL;
2514 a = ALLOCNO_COLOR_DATA (a)->next_bucket_allocno)
2515 if (ALLOCNO_CLASS (a) != NO_REGS)
2517 cost = calculate_allocno_spill_cost (a);
2518 /* ??? Remove cost of copies between the coalesced
2519 allocnos. */
2520 ALLOCNO_COLOR_DATA (a)->temp = cost;
2522 sort_bucket (&uncolorable_allocno_bucket, allocno_spill_sort_compare);
2523 for (;;)
2525 push_only_colorable ();
2526 a = uncolorable_allocno_bucket;
2527 if (a == NULL)
2528 break;
2529 remove_allocno_from_bucket_and_push (a, false);
2531 ira_assert (colorable_allocno_bucket == NULL
2532 && uncolorable_allocno_bucket == NULL);
2533 ira_assert (uncolorable_allocnos_num == 0);
2536 /* Pop the coloring stack and assign hard registers to the popped
2537 allocnos. */
2538 static void
2539 pop_allocnos_from_stack (void)
2541 ira_allocno_t allocno;
2542 enum reg_class aclass;
2544 for (;allocno_stack_vec.length () != 0;)
2546 allocno = allocno_stack_vec.pop ();
2547 aclass = ALLOCNO_CLASS (allocno);
2548 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2550 fprintf (ira_dump_file, " Popping");
2551 ira_print_expanded_allocno (allocno);
2552 fprintf (ira_dump_file, " -- ");
2554 if (aclass == NO_REGS)
2556 ALLOCNO_HARD_REGNO (allocno) = -1;
2557 ALLOCNO_ASSIGNED_P (allocno) = true;
2558 ira_assert (ALLOCNO_UPDATED_HARD_REG_COSTS (allocno) == NULL);
2559 ira_assert
2560 (ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (allocno) == NULL);
2561 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2562 fprintf (ira_dump_file, "assign memory\n");
2564 else if (assign_hard_reg (allocno, false))
2566 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2567 fprintf (ira_dump_file, "assign reg %d\n",
2568 ALLOCNO_HARD_REGNO (allocno));
2570 else if (ALLOCNO_ASSIGNED_P (allocno))
2572 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2573 fprintf (ira_dump_file, "spill%s\n",
2574 ALLOCNO_COLOR_DATA (allocno)->may_be_spilled_p
2575 ? "" : "!");
2577 ALLOCNO_COLOR_DATA (allocno)->in_graph_p = true;
2581 /* Set up number of available hard registers for allocno A. */
2582 static void
2583 setup_allocno_available_regs_num (ira_allocno_t a)
2585 int i, n, hard_regno, hard_regs_num, nwords;
2586 enum reg_class aclass;
2587 allocno_color_data_t data;
2589 aclass = ALLOCNO_CLASS (a);
2590 data = ALLOCNO_COLOR_DATA (a);
2591 data->available_regs_num = 0;
2592 if (aclass == NO_REGS)
2593 return;
2594 hard_regs_num = ira_class_hard_regs_num[aclass];
2595 nwords = ALLOCNO_NUM_OBJECTS (a);
2596 for (n = 0, i = hard_regs_num - 1; i >= 0; i--)
2598 hard_regno = ira_class_hard_regs[aclass][i];
2599 /* Checking only profitable hard regs. */
2600 if (TEST_HARD_REG_BIT (data->profitable_hard_regs, hard_regno))
2601 n++;
2603 data->available_regs_num = n;
2604 if (internal_flag_ira_verbose <= 2 || ira_dump_file == NULL)
2605 return;
2606 fprintf
2607 (ira_dump_file,
2608 " Allocno a%dr%d of %s(%d) has %d avail. regs ",
2609 ALLOCNO_NUM (a), ALLOCNO_REGNO (a),
2610 reg_class_names[aclass], ira_class_hard_regs_num[aclass], n);
2611 print_hard_reg_set (ira_dump_file, data->profitable_hard_regs, false);
2612 fprintf (ira_dump_file, ", %snode: ",
2613 hard_reg_set_equal_p (data->profitable_hard_regs,
2614 data->hard_regs_node->hard_regs->set)
2615 ? "" : "^");
2616 print_hard_reg_set (ira_dump_file,
2617 data->hard_regs_node->hard_regs->set, false);
2618 for (i = 0; i < nwords; i++)
2620 ira_object_t obj = ALLOCNO_OBJECT (a, i);
2622 if (nwords != 1)
2624 if (i != 0)
2625 fprintf (ira_dump_file, ", ");
2626 fprintf (ira_dump_file, " obj %d", i);
2628 fprintf (ira_dump_file, " (confl regs = ");
2629 print_hard_reg_set (ira_dump_file, OBJECT_TOTAL_CONFLICT_HARD_REGS (obj),
2630 false);
2631 fprintf (ira_dump_file, ")");
2633 fprintf (ira_dump_file, "\n");
2636 /* Put ALLOCNO in a bucket corresponding to its number and size of its
2637 conflicting allocnos and hard registers. */
2638 static void
2639 put_allocno_into_bucket (ira_allocno_t allocno)
2641 ALLOCNO_COLOR_DATA (allocno)->in_graph_p = true;
2642 setup_allocno_available_regs_num (allocno);
2643 if (setup_left_conflict_sizes_p (allocno))
2644 add_allocno_to_bucket (allocno, &colorable_allocno_bucket);
2645 else
2646 add_allocno_to_bucket (allocno, &uncolorable_allocno_bucket);
2649 /* Map: allocno number -> allocno priority. */
2650 static int *allocno_priorities;
2652 /* Set up priorities for N allocnos in array
2653 CONSIDERATION_ALLOCNOS. */
2654 static void
2655 setup_allocno_priorities (ira_allocno_t *consideration_allocnos, int n)
2657 int i, length, nrefs, priority, max_priority, mult;
2658 ira_allocno_t a;
2660 max_priority = 0;
2661 for (i = 0; i < n; i++)
2663 a = consideration_allocnos[i];
2664 nrefs = ALLOCNO_NREFS (a);
2665 ira_assert (nrefs >= 0);
2666 mult = floor_log2 (ALLOCNO_NREFS (a)) + 1;
2667 ira_assert (mult >= 0);
2668 allocno_priorities[ALLOCNO_NUM (a)]
2669 = priority
2670 = (mult
2671 * (ALLOCNO_MEMORY_COST (a) - ALLOCNO_CLASS_COST (a))
2672 * ira_reg_class_max_nregs[ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)]);
2673 if (priority < 0)
2674 priority = -priority;
2675 if (max_priority < priority)
2676 max_priority = priority;
2678 mult = max_priority == 0 ? 1 : INT_MAX / max_priority;
2679 for (i = 0; i < n; i++)
2681 a = consideration_allocnos[i];
2682 length = ALLOCNO_EXCESS_PRESSURE_POINTS_NUM (a);
2683 if (ALLOCNO_NUM_OBJECTS (a) > 1)
2684 length /= ALLOCNO_NUM_OBJECTS (a);
2685 if (length <= 0)
2686 length = 1;
2687 allocno_priorities[ALLOCNO_NUM (a)]
2688 = allocno_priorities[ALLOCNO_NUM (a)] * mult / length;
2692 /* Sort allocnos according to the profit of usage of a hard register
2693 instead of memory for them. */
2694 static int
2695 allocno_cost_compare_func (const void *v1p, const void *v2p)
2697 ira_allocno_t p1 = *(const ira_allocno_t *) v1p;
2698 ira_allocno_t p2 = *(const ira_allocno_t *) v2p;
2699 int c1, c2;
2701 c1 = ALLOCNO_UPDATED_MEMORY_COST (p1) - ALLOCNO_UPDATED_CLASS_COST (p1);
2702 c2 = ALLOCNO_UPDATED_MEMORY_COST (p2) - ALLOCNO_UPDATED_CLASS_COST (p2);
2703 if (c1 - c2)
2704 return c1 - c2;
2706 /* If regs are equally good, sort by allocno numbers, so that the
2707 results of qsort leave nothing to chance. */
2708 return ALLOCNO_NUM (p1) - ALLOCNO_NUM (p2);
2711 /* We used Chaitin-Briggs coloring to assign as many pseudos as
2712 possible to hard registers. Let us try to improve allocation with
2713 cost point of view. This function improves the allocation by
2714 spilling some allocnos and assigning the freed hard registers to
2715 other allocnos if it decreases the overall allocation cost. */
2716 static void
2717 improve_allocation (void)
2719 unsigned int i;
2720 int j, k, n, hregno, conflict_hregno, base_cost, class_size, word, nwords;
2721 int check, spill_cost, min_cost, nregs, conflict_nregs, r, best;
2722 bool try_p;
2723 enum reg_class aclass;
2724 enum machine_mode mode;
2725 int *allocno_costs;
2726 int costs[FIRST_PSEUDO_REGISTER];
2727 HARD_REG_SET conflicting_regs[2], profitable_hard_regs;
2728 ira_allocno_t a;
2729 bitmap_iterator bi;
2731 /* Clear counts used to process conflicting allocnos only once for
2732 each allocno. */
2733 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
2734 ALLOCNO_COLOR_DATA (ira_allocnos[i])->temp = 0;
2735 check = n = 0;
2736 /* Process each allocno and try to assign a hard register to it by
2737 spilling some its conflicting allocnos. */
2738 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
2740 a = ira_allocnos[i];
2741 ALLOCNO_COLOR_DATA (a)->temp = 0;
2742 if (empty_profitable_hard_regs (a))
2743 continue;
2744 check++;
2745 aclass = ALLOCNO_CLASS (a);
2746 allocno_costs = ALLOCNO_UPDATED_HARD_REG_COSTS (a);
2747 if (allocno_costs == NULL)
2748 allocno_costs = ALLOCNO_HARD_REG_COSTS (a);
2749 if ((hregno = ALLOCNO_HARD_REGNO (a)) < 0)
2750 base_cost = ALLOCNO_UPDATED_MEMORY_COST (a);
2751 else if (allocno_costs == NULL)
2752 /* It means that assigning a hard register is not profitable
2753 (we don't waste memory for hard register costs in this
2754 case). */
2755 continue;
2756 else
2757 base_cost = allocno_costs[ira_class_hard_reg_index[aclass][hregno]];
2758 try_p = false;
2759 get_conflict_and_start_profitable_regs (a, false,
2760 conflicting_regs,
2761 &profitable_hard_regs);
2762 class_size = ira_class_hard_regs_num[aclass];
2763 /* Set up cost improvement for usage of each profitable hard
2764 register for allocno A. */
2765 for (j = 0; j < class_size; j++)
2767 hregno = ira_class_hard_regs[aclass][j];
2768 if (! check_hard_reg_p (a, hregno,
2769 conflicting_regs, profitable_hard_regs))
2770 continue;
2771 ira_assert (ira_class_hard_reg_index[aclass][hregno] == j);
2772 k = allocno_costs == NULL ? 0 : j;
2773 costs[hregno] = (allocno_costs == NULL
2774 ? ALLOCNO_UPDATED_CLASS_COST (a) : allocno_costs[k]);
2775 costs[hregno] -= base_cost;
2776 if (costs[hregno] < 0)
2777 try_p = true;
2779 if (! try_p)
2780 /* There is no chance to improve the allocation cost by
2781 assigning hard register to allocno A even without spilling
2782 conflicting allocnos. */
2783 continue;
2784 mode = ALLOCNO_MODE (a);
2785 nwords = ALLOCNO_NUM_OBJECTS (a);
2786 /* Process each allocno conflicting with A and update the cost
2787 improvement for profitable hard registers of A. To use a
2788 hard register for A we need to spill some conflicting
2789 allocnos and that creates penalty for the cost
2790 improvement. */
2791 for (word = 0; word < nwords; word++)
2793 ira_object_t conflict_obj;
2794 ira_object_t obj = ALLOCNO_OBJECT (a, word);
2795 ira_object_conflict_iterator oci;
2797 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
2799 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
2801 if (ALLOCNO_COLOR_DATA (conflict_a)->temp == check)
2802 /* We already processed this conflicting allocno
2803 because we processed earlier another object of the
2804 conflicting allocno. */
2805 continue;
2806 ALLOCNO_COLOR_DATA (conflict_a)->temp = check;
2807 if ((conflict_hregno = ALLOCNO_HARD_REGNO (conflict_a)) < 0)
2808 continue;
2809 spill_cost = ALLOCNO_UPDATED_MEMORY_COST (conflict_a);
2810 k = (ira_class_hard_reg_index
2811 [ALLOCNO_CLASS (conflict_a)][conflict_hregno]);
2812 ira_assert (k >= 0);
2813 if ((allocno_costs = ALLOCNO_UPDATED_HARD_REG_COSTS (conflict_a))
2814 != NULL)
2815 spill_cost -= allocno_costs[k];
2816 else if ((allocno_costs = ALLOCNO_HARD_REG_COSTS (conflict_a))
2817 != NULL)
2818 spill_cost -= allocno_costs[k];
2819 else
2820 spill_cost -= ALLOCNO_UPDATED_CLASS_COST (conflict_a);
2821 conflict_nregs
2822 = hard_regno_nregs[conflict_hregno][ALLOCNO_MODE (conflict_a)];
2823 for (r = conflict_hregno;
2824 r >= 0 && r + hard_regno_nregs[r][mode] > conflict_hregno;
2825 r--)
2826 if (check_hard_reg_p (a, r,
2827 conflicting_regs, profitable_hard_regs))
2828 costs[r] += spill_cost;
2829 for (r = conflict_hregno + 1;
2830 r < conflict_hregno + conflict_nregs;
2831 r++)
2832 if (check_hard_reg_p (a, r,
2833 conflicting_regs, profitable_hard_regs))
2834 costs[r] += spill_cost;
2837 min_cost = INT_MAX;
2838 best = -1;
2839 /* Now we choose hard register for A which results in highest
2840 allocation cost improvement. */
2841 for (j = 0; j < class_size; j++)
2843 hregno = ira_class_hard_regs[aclass][j];
2844 if (check_hard_reg_p (a, hregno,
2845 conflicting_regs, profitable_hard_regs)
2846 && min_cost > costs[hregno])
2848 best = hregno;
2849 min_cost = costs[hregno];
2852 if (min_cost >= 0)
2853 /* We are in a situation when assigning any hard register to A
2854 by spilling some conflicting allocnos does not improve the
2855 allocation cost. */
2856 continue;
2857 nregs = hard_regno_nregs[best][mode];
2858 /* Now spill conflicting allocnos which contain a hard register
2859 of A when we assign the best chosen hard register to it. */
2860 for (word = 0; word < nwords; word++)
2862 ira_object_t conflict_obj;
2863 ira_object_t obj = ALLOCNO_OBJECT (a, word);
2864 ira_object_conflict_iterator oci;
2866 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
2868 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
2870 if ((conflict_hregno = ALLOCNO_HARD_REGNO (conflict_a)) < 0)
2871 continue;
2872 conflict_nregs
2873 = hard_regno_nregs[conflict_hregno][ALLOCNO_MODE (conflict_a)];
2874 if (best + nregs <= conflict_hregno
2875 || conflict_hregno + conflict_nregs <= best)
2876 /* No intersection. */
2877 continue;
2878 ALLOCNO_HARD_REGNO (conflict_a) = -1;
2879 sorted_allocnos[n++] = conflict_a;
2880 if (internal_flag_ira_verbose > 2 && ira_dump_file != NULL)
2881 fprintf (ira_dump_file, "Spilling a%dr%d for a%dr%d\n",
2882 ALLOCNO_NUM (conflict_a), ALLOCNO_REGNO (conflict_a),
2883 ALLOCNO_NUM (a), ALLOCNO_REGNO (a));
2886 /* Assign the best chosen hard register to A. */
2887 ALLOCNO_HARD_REGNO (a) = best;
2888 if (internal_flag_ira_verbose > 2 && ira_dump_file != NULL)
2889 fprintf (ira_dump_file, "Assigning %d to a%dr%d\n",
2890 best, ALLOCNO_NUM (a), ALLOCNO_REGNO (a));
2892 if (n == 0)
2893 return;
2894 /* We spilled some allocnos to assign their hard registers to other
2895 allocnos. The spilled allocnos are now in array
2896 'sorted_allocnos'. There is still a possibility that some of the
2897 spilled allocnos can get hard registers. So let us try assign
2898 them hard registers again (just a reminder -- function
2899 'assign_hard_reg' assigns hard registers only if it is possible
2900 and profitable). We process the spilled allocnos with biggest
2901 benefit to get hard register first -- see function
2902 'allocno_cost_compare_func'. */
2903 qsort (sorted_allocnos, n, sizeof (ira_allocno_t),
2904 allocno_cost_compare_func);
2905 for (j = 0; j < n; j++)
2907 a = sorted_allocnos[j];
2908 ALLOCNO_ASSIGNED_P (a) = false;
2909 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2911 fprintf (ira_dump_file, " ");
2912 ira_print_expanded_allocno (a);
2913 fprintf (ira_dump_file, " -- ");
2915 if (assign_hard_reg (a, false))
2917 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2918 fprintf (ira_dump_file, "assign hard reg %d\n",
2919 ALLOCNO_HARD_REGNO (a));
2921 else
2923 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2924 fprintf (ira_dump_file, "assign memory\n");
2929 /* Sort allocnos according to their priorities. */
2930 static int
2931 allocno_priority_compare_func (const void *v1p, const void *v2p)
2933 ira_allocno_t a1 = *(const ira_allocno_t *) v1p;
2934 ira_allocno_t a2 = *(const ira_allocno_t *) v2p;
2935 int pri1, pri2;
2937 pri1 = allocno_priorities[ALLOCNO_NUM (a1)];
2938 pri2 = allocno_priorities[ALLOCNO_NUM (a2)];
2939 if (pri2 != pri1)
2940 return SORTGT (pri2, pri1);
2942 /* If regs are equally good, sort by allocnos, so that the results of
2943 qsort leave nothing to chance. */
2944 return ALLOCNO_NUM (a1) - ALLOCNO_NUM (a2);
2947 /* Chaitin-Briggs coloring for allocnos in COLORING_ALLOCNO_BITMAP
2948 taking into account allocnos in CONSIDERATION_ALLOCNO_BITMAP. */
2949 static void
2950 color_allocnos (void)
2952 unsigned int i, n;
2953 bitmap_iterator bi;
2954 ira_allocno_t a;
2956 setup_profitable_hard_regs ();
2957 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
2959 int l, nr;
2960 HARD_REG_SET conflict_hard_regs;
2961 allocno_color_data_t data;
2962 ira_pref_t pref, next_pref;
2964 a = ira_allocnos[i];
2965 nr = ALLOCNO_NUM_OBJECTS (a);
2966 CLEAR_HARD_REG_SET (conflict_hard_regs);
2967 for (l = 0; l < nr; l++)
2969 ira_object_t obj = ALLOCNO_OBJECT (a, l);
2970 IOR_HARD_REG_SET (conflict_hard_regs,
2971 OBJECT_CONFLICT_HARD_REGS (obj));
2973 data = ALLOCNO_COLOR_DATA (a);
2974 for (pref = ALLOCNO_PREFS (a); pref != NULL; pref = next_pref)
2976 next_pref = pref->next_pref;
2977 if (! ira_hard_reg_in_set_p (pref->hard_regno,
2978 ALLOCNO_MODE (a),
2979 data->profitable_hard_regs))
2980 ira_remove_pref (pref);
2983 if (flag_ira_algorithm == IRA_ALGORITHM_PRIORITY)
2985 n = 0;
2986 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
2988 a = ira_allocnos[i];
2989 if (ALLOCNO_CLASS (a) == NO_REGS)
2991 ALLOCNO_HARD_REGNO (a) = -1;
2992 ALLOCNO_ASSIGNED_P (a) = true;
2993 ira_assert (ALLOCNO_UPDATED_HARD_REG_COSTS (a) == NULL);
2994 ira_assert (ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a) == NULL);
2995 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
2997 fprintf (ira_dump_file, " Spill");
2998 ira_print_expanded_allocno (a);
2999 fprintf (ira_dump_file, "\n");
3001 continue;
3003 sorted_allocnos[n++] = a;
3005 if (n != 0)
3007 setup_allocno_priorities (sorted_allocnos, n);
3008 qsort (sorted_allocnos, n, sizeof (ira_allocno_t),
3009 allocno_priority_compare_func);
3010 for (i = 0; i < n; i++)
3012 a = sorted_allocnos[i];
3013 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3015 fprintf (ira_dump_file, " ");
3016 ira_print_expanded_allocno (a);
3017 fprintf (ira_dump_file, " -- ");
3019 if (assign_hard_reg (a, false))
3021 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3022 fprintf (ira_dump_file, "assign hard reg %d\n",
3023 ALLOCNO_HARD_REGNO (a));
3025 else
3027 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3028 fprintf (ira_dump_file, "assign memory\n");
3033 else
3035 form_allocno_hard_regs_nodes_forest ();
3036 if (internal_flag_ira_verbose > 2 && ira_dump_file != NULL)
3037 print_hard_regs_forest (ira_dump_file);
3038 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
3040 a = ira_allocnos[i];
3041 if (ALLOCNO_CLASS (a) != NO_REGS && ! empty_profitable_hard_regs (a))
3043 ALLOCNO_COLOR_DATA (a)->in_graph_p = true;
3044 update_costs_from_prefs (a);
3046 else
3048 ALLOCNO_HARD_REGNO (a) = -1;
3049 ALLOCNO_ASSIGNED_P (a) = true;
3050 /* We don't need updated costs anymore. */
3051 ira_free_allocno_updated_costs (a);
3052 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3054 fprintf (ira_dump_file, " Spill");
3055 ira_print_expanded_allocno (a);
3056 fprintf (ira_dump_file, "\n");
3060 /* Put the allocnos into the corresponding buckets. */
3061 colorable_allocno_bucket = NULL;
3062 uncolorable_allocno_bucket = NULL;
3063 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi)
3065 a = ira_allocnos[i];
3066 if (ALLOCNO_COLOR_DATA (a)->in_graph_p)
3067 put_allocno_into_bucket (a);
3069 push_allocnos_to_stack ();
3070 pop_allocnos_from_stack ();
3071 finish_allocno_hard_regs_nodes_forest ();
3073 improve_allocation ();
3078 /* Output information about the loop given by its LOOP_TREE_NODE. */
3079 static void
3080 print_loop_title (ira_loop_tree_node_t loop_tree_node)
3082 unsigned int j;
3083 bitmap_iterator bi;
3084 ira_loop_tree_node_t subloop_node, dest_loop_node;
3085 edge e;
3086 edge_iterator ei;
3088 if (loop_tree_node->parent == NULL)
3089 fprintf (ira_dump_file,
3090 "\n Loop 0 (parent -1, header bb%d, depth 0)\n bbs:",
3091 NUM_FIXED_BLOCKS);
3092 else
3094 ira_assert (current_loops != NULL && loop_tree_node->loop != NULL);
3095 fprintf (ira_dump_file,
3096 "\n Loop %d (parent %d, header bb%d, depth %d)\n bbs:",
3097 loop_tree_node->loop_num, loop_tree_node->parent->loop_num,
3098 loop_tree_node->loop->header->index,
3099 loop_depth (loop_tree_node->loop));
3101 for (subloop_node = loop_tree_node->children;
3102 subloop_node != NULL;
3103 subloop_node = subloop_node->next)
3104 if (subloop_node->bb != NULL)
3106 fprintf (ira_dump_file, " %d", subloop_node->bb->index);
3107 FOR_EACH_EDGE (e, ei, subloop_node->bb->succs)
3108 if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
3109 && ((dest_loop_node = IRA_BB_NODE (e->dest)->parent)
3110 != loop_tree_node))
3111 fprintf (ira_dump_file, "(->%d:l%d)",
3112 e->dest->index, dest_loop_node->loop_num);
3114 fprintf (ira_dump_file, "\n all:");
3115 EXECUTE_IF_SET_IN_BITMAP (loop_tree_node->all_allocnos, 0, j, bi)
3116 fprintf (ira_dump_file, " %dr%d", j, ALLOCNO_REGNO (ira_allocnos[j]));
3117 fprintf (ira_dump_file, "\n modified regnos:");
3118 EXECUTE_IF_SET_IN_BITMAP (loop_tree_node->modified_regnos, 0, j, bi)
3119 fprintf (ira_dump_file, " %d", j);
3120 fprintf (ira_dump_file, "\n border:");
3121 EXECUTE_IF_SET_IN_BITMAP (loop_tree_node->border_allocnos, 0, j, bi)
3122 fprintf (ira_dump_file, " %dr%d", j, ALLOCNO_REGNO (ira_allocnos[j]));
3123 fprintf (ira_dump_file, "\n Pressure:");
3124 for (j = 0; (int) j < ira_pressure_classes_num; j++)
3126 enum reg_class pclass;
3128 pclass = ira_pressure_classes[j];
3129 if (loop_tree_node->reg_pressure[pclass] == 0)
3130 continue;
3131 fprintf (ira_dump_file, " %s=%d", reg_class_names[pclass],
3132 loop_tree_node->reg_pressure[pclass]);
3134 fprintf (ira_dump_file, "\n");
3137 /* Color the allocnos inside loop (in the extreme case it can be all
3138 of the function) given the corresponding LOOP_TREE_NODE. The
3139 function is called for each loop during top-down traverse of the
3140 loop tree. */
3141 static void
3142 color_pass (ira_loop_tree_node_t loop_tree_node)
3144 int regno, hard_regno, index = -1, n;
3145 int cost, exit_freq, enter_freq;
3146 unsigned int j;
3147 bitmap_iterator bi;
3148 enum machine_mode mode;
3149 enum reg_class rclass, aclass, pclass;
3150 ira_allocno_t a, subloop_allocno;
3151 ira_loop_tree_node_t subloop_node;
3153 ira_assert (loop_tree_node->bb == NULL);
3154 if (internal_flag_ira_verbose > 1 && ira_dump_file != NULL)
3155 print_loop_title (loop_tree_node);
3157 bitmap_copy (coloring_allocno_bitmap, loop_tree_node->all_allocnos);
3158 bitmap_copy (consideration_allocno_bitmap, coloring_allocno_bitmap);
3159 n = 0;
3160 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, j, bi)
3162 a = ira_allocnos[j];
3163 n++;
3164 if (! ALLOCNO_ASSIGNED_P (a))
3165 continue;
3166 bitmap_clear_bit (coloring_allocno_bitmap, ALLOCNO_NUM (a));
3168 allocno_color_data
3169 = (allocno_color_data_t) ira_allocate (sizeof (struct allocno_color_data)
3170 * n);
3171 memset (allocno_color_data, 0, sizeof (struct allocno_color_data) * n);
3172 curr_allocno_process = 0;
3173 n = 0;
3174 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, j, bi)
3176 a = ira_allocnos[j];
3177 ALLOCNO_ADD_DATA (a) = allocno_color_data + n;
3178 n++;
3180 init_allocno_threads ();
3181 /* Color all mentioned allocnos including transparent ones. */
3182 color_allocnos ();
3183 /* Process caps. They are processed just once. */
3184 if (flag_ira_region == IRA_REGION_MIXED
3185 || flag_ira_region == IRA_REGION_ALL)
3186 EXECUTE_IF_SET_IN_BITMAP (loop_tree_node->all_allocnos, 0, j, bi)
3188 a = ira_allocnos[j];
3189 if (ALLOCNO_CAP_MEMBER (a) == NULL)
3190 continue;
3191 /* Remove from processing in the next loop. */
3192 bitmap_clear_bit (consideration_allocno_bitmap, j);
3193 rclass = ALLOCNO_CLASS (a);
3194 pclass = ira_pressure_class_translate[rclass];
3195 if (flag_ira_region == IRA_REGION_MIXED
3196 && (loop_tree_node->reg_pressure[pclass]
3197 <= ira_class_hard_regs_num[pclass]))
3199 mode = ALLOCNO_MODE (a);
3200 hard_regno = ALLOCNO_HARD_REGNO (a);
3201 if (hard_regno >= 0)
3203 index = ira_class_hard_reg_index[rclass][hard_regno];
3204 ira_assert (index >= 0);
3206 regno = ALLOCNO_REGNO (a);
3207 subloop_allocno = ALLOCNO_CAP_MEMBER (a);
3208 subloop_node = ALLOCNO_LOOP_TREE_NODE (subloop_allocno);
3209 ira_assert (!ALLOCNO_ASSIGNED_P (subloop_allocno));
3210 ALLOCNO_HARD_REGNO (subloop_allocno) = hard_regno;
3211 ALLOCNO_ASSIGNED_P (subloop_allocno) = true;
3212 if (hard_regno >= 0)
3213 update_costs_from_copies (subloop_allocno, true, true);
3214 /* We don't need updated costs anymore. */
3215 ira_free_allocno_updated_costs (subloop_allocno);
3218 /* Update costs of the corresponding allocnos (not caps) in the
3219 subloops. */
3220 for (subloop_node = loop_tree_node->subloops;
3221 subloop_node != NULL;
3222 subloop_node = subloop_node->subloop_next)
3224 ira_assert (subloop_node->bb == NULL);
3225 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, j, bi)
3227 a = ira_allocnos[j];
3228 ira_assert (ALLOCNO_CAP_MEMBER (a) == NULL);
3229 mode = ALLOCNO_MODE (a);
3230 rclass = ALLOCNO_CLASS (a);
3231 pclass = ira_pressure_class_translate[rclass];
3232 hard_regno = ALLOCNO_HARD_REGNO (a);
3233 /* Use hard register class here. ??? */
3234 if (hard_regno >= 0)
3236 index = ira_class_hard_reg_index[rclass][hard_regno];
3237 ira_assert (index >= 0);
3239 regno = ALLOCNO_REGNO (a);
3240 /* ??? conflict costs */
3241 subloop_allocno = subloop_node->regno_allocno_map[regno];
3242 if (subloop_allocno == NULL
3243 || ALLOCNO_CAP (subloop_allocno) != NULL)
3244 continue;
3245 ira_assert (ALLOCNO_CLASS (subloop_allocno) == rclass);
3246 ira_assert (bitmap_bit_p (subloop_node->all_allocnos,
3247 ALLOCNO_NUM (subloop_allocno)));
3248 if ((flag_ira_region == IRA_REGION_MIXED
3249 && (loop_tree_node->reg_pressure[pclass]
3250 <= ira_class_hard_regs_num[pclass]))
3251 || (pic_offset_table_rtx != NULL
3252 && regno == (int) REGNO (pic_offset_table_rtx)))
3254 if (! ALLOCNO_ASSIGNED_P (subloop_allocno))
3256 ALLOCNO_HARD_REGNO (subloop_allocno) = hard_regno;
3257 ALLOCNO_ASSIGNED_P (subloop_allocno) = true;
3258 if (hard_regno >= 0)
3259 update_costs_from_copies (subloop_allocno, true, true);
3260 /* We don't need updated costs anymore. */
3261 ira_free_allocno_updated_costs (subloop_allocno);
3263 continue;
3265 exit_freq = ira_loop_edge_freq (subloop_node, regno, true);
3266 enter_freq = ira_loop_edge_freq (subloop_node, regno, false);
3267 ira_assert (regno < ira_reg_equiv_len);
3268 if (ira_equiv_no_lvalue_p (regno))
3270 if (! ALLOCNO_ASSIGNED_P (subloop_allocno))
3272 ALLOCNO_HARD_REGNO (subloop_allocno) = hard_regno;
3273 ALLOCNO_ASSIGNED_P (subloop_allocno) = true;
3274 if (hard_regno >= 0)
3275 update_costs_from_copies (subloop_allocno, true, true);
3276 /* We don't need updated costs anymore. */
3277 ira_free_allocno_updated_costs (subloop_allocno);
3280 else if (hard_regno < 0)
3282 ALLOCNO_UPDATED_MEMORY_COST (subloop_allocno)
3283 -= ((ira_memory_move_cost[mode][rclass][1] * enter_freq)
3284 + (ira_memory_move_cost[mode][rclass][0] * exit_freq));
3286 else
3288 aclass = ALLOCNO_CLASS (subloop_allocno);
3289 ira_init_register_move_cost_if_necessary (mode);
3290 cost = (ira_register_move_cost[mode][rclass][rclass]
3291 * (exit_freq + enter_freq));
3292 ira_allocate_and_set_or_copy_costs
3293 (&ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno), aclass,
3294 ALLOCNO_UPDATED_CLASS_COST (subloop_allocno),
3295 ALLOCNO_HARD_REG_COSTS (subloop_allocno));
3296 ira_allocate_and_set_or_copy_costs
3297 (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (subloop_allocno),
3298 aclass, 0, ALLOCNO_CONFLICT_HARD_REG_COSTS (subloop_allocno));
3299 ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno)[index] -= cost;
3300 ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (subloop_allocno)[index]
3301 -= cost;
3302 if (ALLOCNO_UPDATED_CLASS_COST (subloop_allocno)
3303 > ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno)[index])
3304 ALLOCNO_UPDATED_CLASS_COST (subloop_allocno)
3305 = ALLOCNO_UPDATED_HARD_REG_COSTS (subloop_allocno)[index];
3306 ALLOCNO_UPDATED_MEMORY_COST (subloop_allocno)
3307 += (ira_memory_move_cost[mode][rclass][0] * enter_freq
3308 + ira_memory_move_cost[mode][rclass][1] * exit_freq);
3312 ira_free (allocno_color_data);
3313 EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, j, bi)
3315 a = ira_allocnos[j];
3316 ALLOCNO_ADD_DATA (a) = NULL;
3320 /* Initialize the common data for coloring and calls functions to do
3321 Chaitin-Briggs and regional coloring. */
3322 static void
3323 do_coloring (void)
3325 coloring_allocno_bitmap = ira_allocate_bitmap ();
3326 if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL)
3327 fprintf (ira_dump_file, "\n**** Allocnos coloring:\n\n");
3329 ira_traverse_loop_tree (false, ira_loop_tree_root, color_pass, NULL);
3331 if (internal_flag_ira_verbose > 1 && ira_dump_file != NULL)
3332 ira_print_disposition (ira_dump_file);
3334 ira_free_bitmap (coloring_allocno_bitmap);
3339 /* Move spill/restore code, which are to be generated in ira-emit.c,
3340 to less frequent points (if it is profitable) by reassigning some
3341 allocnos (in loop with subloops containing in another loop) to
3342 memory which results in longer live-range where the corresponding
3343 pseudo-registers will be in memory. */
3344 static void
3345 move_spill_restore (void)
3347 int cost, regno, hard_regno, hard_regno2, index;
3348 bool changed_p;
3349 int enter_freq, exit_freq;
3350 enum machine_mode mode;
3351 enum reg_class rclass;
3352 ira_allocno_t a, parent_allocno, subloop_allocno;
3353 ira_loop_tree_node_t parent, loop_node, subloop_node;
3354 ira_allocno_iterator ai;
3356 for (;;)
3358 changed_p = false;
3359 if (internal_flag_ira_verbose > 0 && ira_dump_file != NULL)
3360 fprintf (ira_dump_file, "New iteration of spill/restore move\n");
3361 FOR_EACH_ALLOCNO (a, ai)
3363 regno = ALLOCNO_REGNO (a);
3364 loop_node = ALLOCNO_LOOP_TREE_NODE (a);
3365 if (ALLOCNO_CAP_MEMBER (a) != NULL
3366 || ALLOCNO_CAP (a) != NULL
3367 || (hard_regno = ALLOCNO_HARD_REGNO (a)) < 0
3368 || loop_node->children == NULL
3369 /* don't do the optimization because it can create
3370 copies and the reload pass can spill the allocno set
3371 by copy although the allocno will not get memory
3372 slot. */
3373 || ira_equiv_no_lvalue_p (regno)
3374 || !bitmap_bit_p (loop_node->border_allocnos, ALLOCNO_NUM (a)))
3375 continue;
3376 mode = ALLOCNO_MODE (a);
3377 rclass = ALLOCNO_CLASS (a);
3378 index = ira_class_hard_reg_index[rclass][hard_regno];
3379 ira_assert (index >= 0);
3380 cost = (ALLOCNO_MEMORY_COST (a)
3381 - (ALLOCNO_HARD_REG_COSTS (a) == NULL
3382 ? ALLOCNO_CLASS_COST (a)
3383 : ALLOCNO_HARD_REG_COSTS (a)[index]));
3384 ira_init_register_move_cost_if_necessary (mode);
3385 for (subloop_node = loop_node->subloops;
3386 subloop_node != NULL;
3387 subloop_node = subloop_node->subloop_next)
3389 ira_assert (subloop_node->bb == NULL);
3390 subloop_allocno = subloop_node->regno_allocno_map[regno];
3391 if (subloop_allocno == NULL)
3392 continue;
3393 ira_assert (rclass == ALLOCNO_CLASS (subloop_allocno));
3394 /* We have accumulated cost. To get the real cost of
3395 allocno usage in the loop we should subtract costs of
3396 the subloop allocnos. */
3397 cost -= (ALLOCNO_MEMORY_COST (subloop_allocno)
3398 - (ALLOCNO_HARD_REG_COSTS (subloop_allocno) == NULL
3399 ? ALLOCNO_CLASS_COST (subloop_allocno)
3400 : ALLOCNO_HARD_REG_COSTS (subloop_allocno)[index]));
3401 exit_freq = ira_loop_edge_freq (subloop_node, regno, true);
3402 enter_freq = ira_loop_edge_freq (subloop_node, regno, false);
3403 if ((hard_regno2 = ALLOCNO_HARD_REGNO (subloop_allocno)) < 0)
3404 cost -= (ira_memory_move_cost[mode][rclass][0] * exit_freq
3405 + ira_memory_move_cost[mode][rclass][1] * enter_freq);
3406 else
3408 cost
3409 += (ira_memory_move_cost[mode][rclass][0] * exit_freq
3410 + ira_memory_move_cost[mode][rclass][1] * enter_freq);
3411 if (hard_regno2 != hard_regno)
3412 cost -= (ira_register_move_cost[mode][rclass][rclass]
3413 * (exit_freq + enter_freq));
3416 if ((parent = loop_node->parent) != NULL
3417 && (parent_allocno = parent->regno_allocno_map[regno]) != NULL)
3419 ira_assert (rclass == ALLOCNO_CLASS (parent_allocno));
3420 exit_freq = ira_loop_edge_freq (loop_node, regno, true);
3421 enter_freq = ira_loop_edge_freq (loop_node, regno, false);
3422 if ((hard_regno2 = ALLOCNO_HARD_REGNO (parent_allocno)) < 0)
3423 cost -= (ira_memory_move_cost[mode][rclass][0] * exit_freq
3424 + ira_memory_move_cost[mode][rclass][1] * enter_freq);
3425 else
3427 cost
3428 += (ira_memory_move_cost[mode][rclass][1] * exit_freq
3429 + ira_memory_move_cost[mode][rclass][0] * enter_freq);
3430 if (hard_regno2 != hard_regno)
3431 cost -= (ira_register_move_cost[mode][rclass][rclass]
3432 * (exit_freq + enter_freq));
3435 if (cost < 0)
3437 ALLOCNO_HARD_REGNO (a) = -1;
3438 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3440 fprintf
3441 (ira_dump_file,
3442 " Moving spill/restore for a%dr%d up from loop %d",
3443 ALLOCNO_NUM (a), regno, loop_node->loop_num);
3444 fprintf (ira_dump_file, " - profit %d\n", -cost);
3446 changed_p = true;
3449 if (! changed_p)
3450 break;
3456 /* Update current hard reg costs and current conflict hard reg costs
3457 for allocno A. It is done by processing its copies containing
3458 other allocnos already assigned. */
3459 static void
3460 update_curr_costs (ira_allocno_t a)
3462 int i, hard_regno, cost;
3463 enum machine_mode mode;
3464 enum reg_class aclass, rclass;
3465 ira_allocno_t another_a;
3466 ira_copy_t cp, next_cp;
3468 ira_free_allocno_updated_costs (a);
3469 ira_assert (! ALLOCNO_ASSIGNED_P (a));
3470 aclass = ALLOCNO_CLASS (a);
3471 if (aclass == NO_REGS)
3472 return;
3473 mode = ALLOCNO_MODE (a);
3474 ira_init_register_move_cost_if_necessary (mode);
3475 for (cp = ALLOCNO_COPIES (a); cp != NULL; cp = next_cp)
3477 if (cp->first == a)
3479 next_cp = cp->next_first_allocno_copy;
3480 another_a = cp->second;
3482 else if (cp->second == a)
3484 next_cp = cp->next_second_allocno_copy;
3485 another_a = cp->first;
3487 else
3488 gcc_unreachable ();
3489 if (! ira_reg_classes_intersect_p[aclass][ALLOCNO_CLASS (another_a)]
3490 || ! ALLOCNO_ASSIGNED_P (another_a)
3491 || (hard_regno = ALLOCNO_HARD_REGNO (another_a)) < 0)
3492 continue;
3493 rclass = REGNO_REG_CLASS (hard_regno);
3494 i = ira_class_hard_reg_index[aclass][hard_regno];
3495 if (i < 0)
3496 continue;
3497 cost = (cp->first == a
3498 ? ira_register_move_cost[mode][rclass][aclass]
3499 : ira_register_move_cost[mode][aclass][rclass]);
3500 ira_allocate_and_set_or_copy_costs
3501 (&ALLOCNO_UPDATED_HARD_REG_COSTS (a), aclass, ALLOCNO_CLASS_COST (a),
3502 ALLOCNO_HARD_REG_COSTS (a));
3503 ira_allocate_and_set_or_copy_costs
3504 (&ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a),
3505 aclass, 0, ALLOCNO_CONFLICT_HARD_REG_COSTS (a));
3506 ALLOCNO_UPDATED_HARD_REG_COSTS (a)[i] -= cp->freq * cost;
3507 ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a)[i] -= cp->freq * cost;
3511 /* Try to assign hard registers to the unassigned allocnos and
3512 allocnos conflicting with them or conflicting with allocnos whose
3513 regno >= START_REGNO. The function is called after ira_flattening,
3514 so more allocnos (including ones created in ira-emit.c) will have a
3515 chance to get a hard register. We use simple assignment algorithm
3516 based on priorities. */
3517 void
3518 ira_reassign_conflict_allocnos (int start_regno)
3520 int i, allocnos_to_color_num;
3521 ira_allocno_t a;
3522 enum reg_class aclass;
3523 bitmap allocnos_to_color;
3524 ira_allocno_iterator ai;
3526 allocnos_to_color = ira_allocate_bitmap ();
3527 allocnos_to_color_num = 0;
3528 FOR_EACH_ALLOCNO (a, ai)
3530 int n = ALLOCNO_NUM_OBJECTS (a);
3532 if (! ALLOCNO_ASSIGNED_P (a)
3533 && ! bitmap_bit_p (allocnos_to_color, ALLOCNO_NUM (a)))
3535 if (ALLOCNO_CLASS (a) != NO_REGS)
3536 sorted_allocnos[allocnos_to_color_num++] = a;
3537 else
3539 ALLOCNO_ASSIGNED_P (a) = true;
3540 ALLOCNO_HARD_REGNO (a) = -1;
3541 ira_assert (ALLOCNO_UPDATED_HARD_REG_COSTS (a) == NULL);
3542 ira_assert (ALLOCNO_UPDATED_CONFLICT_HARD_REG_COSTS (a) == NULL);
3544 bitmap_set_bit (allocnos_to_color, ALLOCNO_NUM (a));
3546 if (ALLOCNO_REGNO (a) < start_regno
3547 || (aclass = ALLOCNO_CLASS (a)) == NO_REGS)
3548 continue;
3549 for (i = 0; i < n; i++)
3551 ira_object_t obj = ALLOCNO_OBJECT (a, i);
3552 ira_object_t conflict_obj;
3553 ira_object_conflict_iterator oci;
3555 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
3557 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
3559 ira_assert (ira_reg_classes_intersect_p
3560 [aclass][ALLOCNO_CLASS (conflict_a)]);
3561 if (!bitmap_set_bit (allocnos_to_color, ALLOCNO_NUM (conflict_a)))
3562 continue;
3563 sorted_allocnos[allocnos_to_color_num++] = conflict_a;
3567 ira_free_bitmap (allocnos_to_color);
3568 if (allocnos_to_color_num > 1)
3570 setup_allocno_priorities (sorted_allocnos, allocnos_to_color_num);
3571 qsort (sorted_allocnos, allocnos_to_color_num, sizeof (ira_allocno_t),
3572 allocno_priority_compare_func);
3574 for (i = 0; i < allocnos_to_color_num; i++)
3576 a = sorted_allocnos[i];
3577 ALLOCNO_ASSIGNED_P (a) = false;
3578 update_curr_costs (a);
3580 for (i = 0; i < allocnos_to_color_num; i++)
3582 a = sorted_allocnos[i];
3583 if (assign_hard_reg (a, true))
3585 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3586 fprintf
3587 (ira_dump_file,
3588 " Secondary allocation: assign hard reg %d to reg %d\n",
3589 ALLOCNO_HARD_REGNO (a), ALLOCNO_REGNO (a));
3596 /* This page contains functions used to find conflicts using allocno
3597 live ranges. */
3599 #ifdef ENABLE_IRA_CHECKING
3601 /* Return TRUE if live ranges of pseudo-registers REGNO1 and REGNO2
3602 intersect. This should be used when there is only one region.
3603 Currently this is used during reload. */
3604 static bool
3605 conflict_by_live_ranges_p (int regno1, int regno2)
3607 ira_allocno_t a1, a2;
3609 ira_assert (regno1 >= FIRST_PSEUDO_REGISTER
3610 && regno2 >= FIRST_PSEUDO_REGISTER);
3611 /* Reg info caclulated by dataflow infrastructure can be different
3612 from one calculated by regclass. */
3613 if ((a1 = ira_loop_tree_root->regno_allocno_map[regno1]) == NULL
3614 || (a2 = ira_loop_tree_root->regno_allocno_map[regno2]) == NULL)
3615 return false;
3616 return allocnos_conflict_by_live_ranges_p (a1, a2);
3619 #endif
3623 /* This page contains code to coalesce memory stack slots used by
3624 spilled allocnos. This results in smaller stack frame, better data
3625 locality, and in smaller code for some architectures like
3626 x86/x86_64 where insn size depends on address displacement value.
3627 On the other hand, it can worsen insn scheduling after the RA but
3628 in practice it is less important than smaller stack frames. */
3630 /* TRUE if we coalesced some allocnos. In other words, if we got
3631 loops formed by members first_coalesced_allocno and
3632 next_coalesced_allocno containing more one allocno. */
3633 static bool allocno_coalesced_p;
3635 /* Bitmap used to prevent a repeated allocno processing because of
3636 coalescing. */
3637 static bitmap processed_coalesced_allocno_bitmap;
3639 /* See below. */
3640 typedef struct coalesce_data *coalesce_data_t;
3642 /* To decrease footprint of ira_allocno structure we store all data
3643 needed only for coalescing in the following structure. */
3644 struct coalesce_data
3646 /* Coalesced allocnos form a cyclic list. One allocno given by
3647 FIRST represents all coalesced allocnos. The
3648 list is chained by NEXT. */
3649 ira_allocno_t first;
3650 ira_allocno_t next;
3651 int temp;
3654 /* Container for storing allocno data concerning coalescing. */
3655 static coalesce_data_t allocno_coalesce_data;
3657 /* Macro to access the data concerning coalescing. */
3658 #define ALLOCNO_COALESCE_DATA(a) ((coalesce_data_t) ALLOCNO_ADD_DATA (a))
3660 /* Merge two sets of coalesced allocnos given correspondingly by
3661 allocnos A1 and A2 (more accurately merging A2 set into A1
3662 set). */
3663 static void
3664 merge_allocnos (ira_allocno_t a1, ira_allocno_t a2)
3666 ira_allocno_t a, first, last, next;
3668 first = ALLOCNO_COALESCE_DATA (a1)->first;
3669 a = ALLOCNO_COALESCE_DATA (a2)->first;
3670 if (first == a)
3671 return;
3672 for (last = a2, a = ALLOCNO_COALESCE_DATA (a2)->next;;
3673 a = ALLOCNO_COALESCE_DATA (a)->next)
3675 ALLOCNO_COALESCE_DATA (a)->first = first;
3676 if (a == a2)
3677 break;
3678 last = a;
3680 next = allocno_coalesce_data[ALLOCNO_NUM (first)].next;
3681 allocno_coalesce_data[ALLOCNO_NUM (first)].next = a2;
3682 allocno_coalesce_data[ALLOCNO_NUM (last)].next = next;
3685 /* Return TRUE if there are conflicting allocnos from two sets of
3686 coalesced allocnos given correspondingly by allocnos A1 and A2. We
3687 use live ranges to find conflicts because conflicts are represented
3688 only for allocnos of the same allocno class and during the reload
3689 pass we coalesce allocnos for sharing stack memory slots. */
3690 static bool
3691 coalesced_allocno_conflict_p (ira_allocno_t a1, ira_allocno_t a2)
3693 ira_allocno_t a, conflict_a;
3695 if (allocno_coalesced_p)
3697 bitmap_clear (processed_coalesced_allocno_bitmap);
3698 for (a = ALLOCNO_COALESCE_DATA (a1)->next;;
3699 a = ALLOCNO_COALESCE_DATA (a)->next)
3701 bitmap_set_bit (processed_coalesced_allocno_bitmap, ALLOCNO_NUM (a));
3702 if (a == a1)
3703 break;
3706 for (a = ALLOCNO_COALESCE_DATA (a2)->next;;
3707 a = ALLOCNO_COALESCE_DATA (a)->next)
3709 for (conflict_a = ALLOCNO_COALESCE_DATA (a1)->next;;
3710 conflict_a = ALLOCNO_COALESCE_DATA (conflict_a)->next)
3712 if (allocnos_conflict_by_live_ranges_p (a, conflict_a))
3713 return true;
3714 if (conflict_a == a1)
3715 break;
3717 if (a == a2)
3718 break;
3720 return false;
3723 /* The major function for aggressive allocno coalescing. We coalesce
3724 only spilled allocnos. If some allocnos have been coalesced, we
3725 set up flag allocno_coalesced_p. */
3726 static void
3727 coalesce_allocnos (void)
3729 ira_allocno_t a;
3730 ira_copy_t cp, next_cp;
3731 unsigned int j;
3732 int i, n, cp_num, regno;
3733 bitmap_iterator bi;
3735 cp_num = 0;
3736 /* Collect copies. */
3737 EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, j, bi)
3739 a = ira_allocnos[j];
3740 regno = ALLOCNO_REGNO (a);
3741 if (! ALLOCNO_ASSIGNED_P (a) || ALLOCNO_HARD_REGNO (a) >= 0
3742 || ira_equiv_no_lvalue_p (regno))
3743 continue;
3744 for (cp = ALLOCNO_COPIES (a); cp != NULL; cp = next_cp)
3746 if (cp->first == a)
3748 next_cp = cp->next_first_allocno_copy;
3749 regno = ALLOCNO_REGNO (cp->second);
3750 /* For priority coloring we coalesce allocnos only with
3751 the same allocno class not with intersected allocno
3752 classes as it were possible. It is done for
3753 simplicity. */
3754 if ((cp->insn != NULL || cp->constraint_p)
3755 && ALLOCNO_ASSIGNED_P (cp->second)
3756 && ALLOCNO_HARD_REGNO (cp->second) < 0
3757 && ! ira_equiv_no_lvalue_p (regno))
3758 sorted_copies[cp_num++] = cp;
3760 else if (cp->second == a)
3761 next_cp = cp->next_second_allocno_copy;
3762 else
3763 gcc_unreachable ();
3766 qsort (sorted_copies, cp_num, sizeof (ira_copy_t), copy_freq_compare_func);
3767 /* Coalesced copies, most frequently executed first. */
3768 for (; cp_num != 0;)
3770 for (i = 0; i < cp_num; i++)
3772 cp = sorted_copies[i];
3773 if (! coalesced_allocno_conflict_p (cp->first, cp->second))
3775 allocno_coalesced_p = true;
3776 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
3777 fprintf
3778 (ira_dump_file,
3779 " Coalescing copy %d:a%dr%d-a%dr%d (freq=%d)\n",
3780 cp->num, ALLOCNO_NUM (cp->first), ALLOCNO_REGNO (cp->first),
3781 ALLOCNO_NUM (cp->second), ALLOCNO_REGNO (cp->second),
3782 cp->freq);
3783 merge_allocnos (cp->first, cp->second);
3784 i++;
3785 break;
3788 /* Collect the rest of copies. */
3789 for (n = 0; i < cp_num; i++)
3791 cp = sorted_copies[i];
3792 if (allocno_coalesce_data[ALLOCNO_NUM (cp->first)].first
3793 != allocno_coalesce_data[ALLOCNO_NUM (cp->second)].first)
3794 sorted_copies[n++] = cp;
3796 cp_num = n;
3800 /* Usage cost and order number of coalesced allocno set to which
3801 given pseudo register belongs to. */
3802 static int *regno_coalesced_allocno_cost;
3803 static int *regno_coalesced_allocno_num;
3805 /* Sort pseudos according frequencies of coalesced allocno sets they
3806 belong to (putting most frequently ones first), and according to
3807 coalesced allocno set order numbers. */
3808 static int
3809 coalesced_pseudo_reg_freq_compare (const void *v1p, const void *v2p)
3811 const int regno1 = *(const int *) v1p;
3812 const int regno2 = *(const int *) v2p;
3813 int diff;
3815 if ((diff = (regno_coalesced_allocno_cost[regno2]
3816 - regno_coalesced_allocno_cost[regno1])) != 0)
3817 return diff;
3818 if ((diff = (regno_coalesced_allocno_num[regno1]
3819 - regno_coalesced_allocno_num[regno2])) != 0)
3820 return diff;
3821 return regno1 - regno2;
3824 /* Widest width in which each pseudo reg is referred to (via subreg).
3825 It is used for sorting pseudo registers. */
3826 static unsigned int *regno_max_ref_width;
3828 /* Redefine STACK_GROWS_DOWNWARD in terms of 0 or 1. */
3829 #ifdef STACK_GROWS_DOWNWARD
3830 # undef STACK_GROWS_DOWNWARD
3831 # define STACK_GROWS_DOWNWARD 1
3832 #else
3833 # define STACK_GROWS_DOWNWARD 0
3834 #endif
3836 /* Sort pseudos according their slot numbers (putting ones with
3837 smaller numbers first, or last when the frame pointer is not
3838 needed). */
3839 static int
3840 coalesced_pseudo_reg_slot_compare (const void *v1p, const void *v2p)
3842 const int regno1 = *(const int *) v1p;
3843 const int regno2 = *(const int *) v2p;
3844 ira_allocno_t a1 = ira_regno_allocno_map[regno1];
3845 ira_allocno_t a2 = ira_regno_allocno_map[regno2];
3846 int diff, slot_num1, slot_num2;
3847 int total_size1, total_size2;
3849 if (a1 == NULL || ALLOCNO_HARD_REGNO (a1) >= 0)
3851 if (a2 == NULL || ALLOCNO_HARD_REGNO (a2) >= 0)
3852 return regno1 - regno2;
3853 return 1;
3855 else if (a2 == NULL || ALLOCNO_HARD_REGNO (a2) >= 0)
3856 return -1;
3857 slot_num1 = -ALLOCNO_HARD_REGNO (a1);
3858 slot_num2 = -ALLOCNO_HARD_REGNO (a2);
3859 if ((diff = slot_num1 - slot_num2) != 0)
3860 return (frame_pointer_needed
3861 || (!FRAME_GROWS_DOWNWARD) == STACK_GROWS_DOWNWARD ? diff : -diff);
3862 total_size1 = MAX (PSEUDO_REGNO_BYTES (regno1),
3863 regno_max_ref_width[regno1]);
3864 total_size2 = MAX (PSEUDO_REGNO_BYTES (regno2),
3865 regno_max_ref_width[regno2]);
3866 if ((diff = total_size2 - total_size1) != 0)
3867 return diff;
3868 return regno1 - regno2;
3871 /* Setup REGNO_COALESCED_ALLOCNO_COST and REGNO_COALESCED_ALLOCNO_NUM
3872 for coalesced allocno sets containing allocnos with their regnos
3873 given in array PSEUDO_REGNOS of length N. */
3874 static void
3875 setup_coalesced_allocno_costs_and_nums (int *pseudo_regnos, int n)
3877 int i, num, regno, cost;
3878 ira_allocno_t allocno, a;
3880 for (num = i = 0; i < n; i++)
3882 regno = pseudo_regnos[i];
3883 allocno = ira_regno_allocno_map[regno];
3884 if (allocno == NULL)
3886 regno_coalesced_allocno_cost[regno] = 0;
3887 regno_coalesced_allocno_num[regno] = ++num;
3888 continue;
3890 if (ALLOCNO_COALESCE_DATA (allocno)->first != allocno)
3891 continue;
3892 num++;
3893 for (cost = 0, a = ALLOCNO_COALESCE_DATA (allocno)->next;;
3894 a = ALLOCNO_COALESCE_DATA (a)->next)
3896 cost += ALLOCNO_FREQ (a);
3897 if (a == allocno)
3898 break;
3900 for (a = ALLOCNO_COALESCE_DATA (allocno)->next;;
3901 a = ALLOCNO_COALESCE_DATA (a)->next)
3903 regno_coalesced_allocno_num[ALLOCNO_REGNO (a)] = num;
3904 regno_coalesced_allocno_cost[ALLOCNO_REGNO (a)] = cost;
3905 if (a == allocno)
3906 break;
3911 /* Collect spilled allocnos representing coalesced allocno sets (the
3912 first coalesced allocno). The collected allocnos are returned
3913 through array SPILLED_COALESCED_ALLOCNOS. The function returns the
3914 number of the collected allocnos. The allocnos are given by their
3915 regnos in array PSEUDO_REGNOS of length N. */
3916 static int
3917 collect_spilled_coalesced_allocnos (int *pseudo_regnos, int n,
3918 ira_allocno_t *spilled_coalesced_allocnos)
3920 int i, num, regno;
3921 ira_allocno_t allocno;
3923 for (num = i = 0; i < n; i++)
3925 regno = pseudo_regnos[i];
3926 allocno = ira_regno_allocno_map[regno];
3927 if (allocno == NULL || ALLOCNO_HARD_REGNO (allocno) >= 0
3928 || ALLOCNO_COALESCE_DATA (allocno)->first != allocno)
3929 continue;
3930 spilled_coalesced_allocnos[num++] = allocno;
3932 return num;
3935 /* Array of live ranges of size IRA_ALLOCNOS_NUM. Live range for
3936 given slot contains live ranges of coalesced allocnos assigned to
3937 given slot. */
3938 static live_range_t *slot_coalesced_allocnos_live_ranges;
3940 /* Return TRUE if coalesced allocnos represented by ALLOCNO has live
3941 ranges intersected with live ranges of coalesced allocnos assigned
3942 to slot with number N. */
3943 static bool
3944 slot_coalesced_allocno_live_ranges_intersect_p (ira_allocno_t allocno, int n)
3946 ira_allocno_t a;
3948 for (a = ALLOCNO_COALESCE_DATA (allocno)->next;;
3949 a = ALLOCNO_COALESCE_DATA (a)->next)
3951 int i;
3952 int nr = ALLOCNO_NUM_OBJECTS (a);
3954 for (i = 0; i < nr; i++)
3956 ira_object_t obj = ALLOCNO_OBJECT (a, i);
3958 if (ira_live_ranges_intersect_p
3959 (slot_coalesced_allocnos_live_ranges[n],
3960 OBJECT_LIVE_RANGES (obj)))
3961 return true;
3963 if (a == allocno)
3964 break;
3966 return false;
3969 /* Update live ranges of slot to which coalesced allocnos represented
3970 by ALLOCNO were assigned. */
3971 static void
3972 setup_slot_coalesced_allocno_live_ranges (ira_allocno_t allocno)
3974 int i, n;
3975 ira_allocno_t a;
3976 live_range_t r;
3978 n = ALLOCNO_COALESCE_DATA (allocno)->temp;
3979 for (a = ALLOCNO_COALESCE_DATA (allocno)->next;;
3980 a = ALLOCNO_COALESCE_DATA (a)->next)
3982 int nr = ALLOCNO_NUM_OBJECTS (a);
3983 for (i = 0; i < nr; i++)
3985 ira_object_t obj = ALLOCNO_OBJECT (a, i);
3987 r = ira_copy_live_range_list (OBJECT_LIVE_RANGES (obj));
3988 slot_coalesced_allocnos_live_ranges[n]
3989 = ira_merge_live_ranges
3990 (slot_coalesced_allocnos_live_ranges[n], r);
3992 if (a == allocno)
3993 break;
3997 /* We have coalesced allocnos involving in copies. Coalesce allocnos
3998 further in order to share the same memory stack slot. Allocnos
3999 representing sets of allocnos coalesced before the call are given
4000 in array SPILLED_COALESCED_ALLOCNOS of length NUM. Return TRUE if
4001 some allocnos were coalesced in the function. */
4002 static bool
4003 coalesce_spill_slots (ira_allocno_t *spilled_coalesced_allocnos, int num)
4005 int i, j, n, last_coalesced_allocno_num;
4006 ira_allocno_t allocno, a;
4007 bool merged_p = false;
4008 bitmap set_jump_crosses = regstat_get_setjmp_crosses ();
4010 slot_coalesced_allocnos_live_ranges
4011 = (live_range_t *) ira_allocate (sizeof (live_range_t) * ira_allocnos_num);
4012 memset (slot_coalesced_allocnos_live_ranges, 0,
4013 sizeof (live_range_t) * ira_allocnos_num);
4014 last_coalesced_allocno_num = 0;
4015 /* Coalesce non-conflicting spilled allocnos preferring most
4016 frequently used. */
4017 for (i = 0; i < num; i++)
4019 allocno = spilled_coalesced_allocnos[i];
4020 if (ALLOCNO_COALESCE_DATA (allocno)->first != allocno
4021 || bitmap_bit_p (set_jump_crosses, ALLOCNO_REGNO (allocno))
4022 || ira_equiv_no_lvalue_p (ALLOCNO_REGNO (allocno)))
4023 continue;
4024 for (j = 0; j < i; j++)
4026 a = spilled_coalesced_allocnos[j];
4027 n = ALLOCNO_COALESCE_DATA (a)->temp;
4028 if (ALLOCNO_COALESCE_DATA (a)->first == a
4029 && ! bitmap_bit_p (set_jump_crosses, ALLOCNO_REGNO (a))
4030 && ! ira_equiv_no_lvalue_p (ALLOCNO_REGNO (a))
4031 && ! slot_coalesced_allocno_live_ranges_intersect_p (allocno, n))
4032 break;
4034 if (j >= i)
4036 /* No coalescing: set up number for coalesced allocnos
4037 represented by ALLOCNO. */
4038 ALLOCNO_COALESCE_DATA (allocno)->temp = last_coalesced_allocno_num++;
4039 setup_slot_coalesced_allocno_live_ranges (allocno);
4041 else
4043 allocno_coalesced_p = true;
4044 merged_p = true;
4045 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4046 fprintf (ira_dump_file,
4047 " Coalescing spilled allocnos a%dr%d->a%dr%d\n",
4048 ALLOCNO_NUM (allocno), ALLOCNO_REGNO (allocno),
4049 ALLOCNO_NUM (a), ALLOCNO_REGNO (a));
4050 ALLOCNO_COALESCE_DATA (allocno)->temp
4051 = ALLOCNO_COALESCE_DATA (a)->temp;
4052 setup_slot_coalesced_allocno_live_ranges (allocno);
4053 merge_allocnos (a, allocno);
4054 ira_assert (ALLOCNO_COALESCE_DATA (a)->first == a);
4057 for (i = 0; i < ira_allocnos_num; i++)
4058 ira_finish_live_range_list (slot_coalesced_allocnos_live_ranges[i]);
4059 ira_free (slot_coalesced_allocnos_live_ranges);
4060 return merged_p;
4063 /* Sort pseudo-register numbers in array PSEUDO_REGNOS of length N for
4064 subsequent assigning stack slots to them in the reload pass. To do
4065 this we coalesce spilled allocnos first to decrease the number of
4066 memory-memory move insns. This function is called by the
4067 reload. */
4068 void
4069 ira_sort_regnos_for_alter_reg (int *pseudo_regnos, int n,
4070 unsigned int *reg_max_ref_width)
4072 int max_regno = max_reg_num ();
4073 int i, regno, num, slot_num;
4074 ira_allocno_t allocno, a;
4075 ira_allocno_iterator ai;
4076 ira_allocno_t *spilled_coalesced_allocnos;
4078 ira_assert (! ira_use_lra_p);
4080 /* Set up allocnos can be coalesced. */
4081 coloring_allocno_bitmap = ira_allocate_bitmap ();
4082 for (i = 0; i < n; i++)
4084 regno = pseudo_regnos[i];
4085 allocno = ira_regno_allocno_map[regno];
4086 if (allocno != NULL)
4087 bitmap_set_bit (coloring_allocno_bitmap, ALLOCNO_NUM (allocno));
4089 allocno_coalesced_p = false;
4090 processed_coalesced_allocno_bitmap = ira_allocate_bitmap ();
4091 allocno_coalesce_data
4092 = (coalesce_data_t) ira_allocate (sizeof (struct coalesce_data)
4093 * ira_allocnos_num);
4094 /* Initialize coalesce data for allocnos. */
4095 FOR_EACH_ALLOCNO (a, ai)
4097 ALLOCNO_ADD_DATA (a) = allocno_coalesce_data + ALLOCNO_NUM (a);
4098 ALLOCNO_COALESCE_DATA (a)->first = a;
4099 ALLOCNO_COALESCE_DATA (a)->next = a;
4101 coalesce_allocnos ();
4102 ira_free_bitmap (coloring_allocno_bitmap);
4103 regno_coalesced_allocno_cost
4104 = (int *) ira_allocate (max_regno * sizeof (int));
4105 regno_coalesced_allocno_num
4106 = (int *) ira_allocate (max_regno * sizeof (int));
4107 memset (regno_coalesced_allocno_num, 0, max_regno * sizeof (int));
4108 setup_coalesced_allocno_costs_and_nums (pseudo_regnos, n);
4109 /* Sort regnos according frequencies of the corresponding coalesced
4110 allocno sets. */
4111 qsort (pseudo_regnos, n, sizeof (int), coalesced_pseudo_reg_freq_compare);
4112 spilled_coalesced_allocnos
4113 = (ira_allocno_t *) ira_allocate (ira_allocnos_num
4114 * sizeof (ira_allocno_t));
4115 /* Collect allocnos representing the spilled coalesced allocno
4116 sets. */
4117 num = collect_spilled_coalesced_allocnos (pseudo_regnos, n,
4118 spilled_coalesced_allocnos);
4119 if (flag_ira_share_spill_slots
4120 && coalesce_spill_slots (spilled_coalesced_allocnos, num))
4122 setup_coalesced_allocno_costs_and_nums (pseudo_regnos, n);
4123 qsort (pseudo_regnos, n, sizeof (int),
4124 coalesced_pseudo_reg_freq_compare);
4125 num = collect_spilled_coalesced_allocnos (pseudo_regnos, n,
4126 spilled_coalesced_allocnos);
4128 ira_free_bitmap (processed_coalesced_allocno_bitmap);
4129 allocno_coalesced_p = false;
4130 /* Assign stack slot numbers to spilled allocno sets, use smaller
4131 numbers for most frequently used coalesced allocnos. -1 is
4132 reserved for dynamic search of stack slots for pseudos spilled by
4133 the reload. */
4134 slot_num = 1;
4135 for (i = 0; i < num; i++)
4137 allocno = spilled_coalesced_allocnos[i];
4138 if (ALLOCNO_COALESCE_DATA (allocno)->first != allocno
4139 || ALLOCNO_HARD_REGNO (allocno) >= 0
4140 || ira_equiv_no_lvalue_p (ALLOCNO_REGNO (allocno)))
4141 continue;
4142 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4143 fprintf (ira_dump_file, " Slot %d (freq,size):", slot_num);
4144 slot_num++;
4145 for (a = ALLOCNO_COALESCE_DATA (allocno)->next;;
4146 a = ALLOCNO_COALESCE_DATA (a)->next)
4148 ira_assert (ALLOCNO_HARD_REGNO (a) < 0);
4149 ALLOCNO_HARD_REGNO (a) = -slot_num;
4150 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4151 fprintf (ira_dump_file, " a%dr%d(%d,%d)",
4152 ALLOCNO_NUM (a), ALLOCNO_REGNO (a), ALLOCNO_FREQ (a),
4153 MAX (PSEUDO_REGNO_BYTES (ALLOCNO_REGNO (a)),
4154 reg_max_ref_width[ALLOCNO_REGNO (a)]));
4156 if (a == allocno)
4157 break;
4159 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4160 fprintf (ira_dump_file, "\n");
4162 ira_spilled_reg_stack_slots_num = slot_num - 1;
4163 ira_free (spilled_coalesced_allocnos);
4164 /* Sort regnos according the slot numbers. */
4165 regno_max_ref_width = reg_max_ref_width;
4166 qsort (pseudo_regnos, n, sizeof (int), coalesced_pseudo_reg_slot_compare);
4167 FOR_EACH_ALLOCNO (a, ai)
4168 ALLOCNO_ADD_DATA (a) = NULL;
4169 ira_free (allocno_coalesce_data);
4170 ira_free (regno_coalesced_allocno_num);
4171 ira_free (regno_coalesced_allocno_cost);
4176 /* This page contains code used by the reload pass to improve the
4177 final code. */
4179 /* The function is called from reload to mark changes in the
4180 allocation of REGNO made by the reload. Remember that reg_renumber
4181 reflects the change result. */
4182 void
4183 ira_mark_allocation_change (int regno)
4185 ira_allocno_t a = ira_regno_allocno_map[regno];
4186 int old_hard_regno, hard_regno, cost;
4187 enum reg_class aclass = ALLOCNO_CLASS (a);
4189 ira_assert (a != NULL);
4190 hard_regno = reg_renumber[regno];
4191 if ((old_hard_regno = ALLOCNO_HARD_REGNO (a)) == hard_regno)
4192 return;
4193 if (old_hard_regno < 0)
4194 cost = -ALLOCNO_MEMORY_COST (a);
4195 else
4197 ira_assert (ira_class_hard_reg_index[aclass][old_hard_regno] >= 0);
4198 cost = -(ALLOCNO_HARD_REG_COSTS (a) == NULL
4199 ? ALLOCNO_CLASS_COST (a)
4200 : ALLOCNO_HARD_REG_COSTS (a)
4201 [ira_class_hard_reg_index[aclass][old_hard_regno]]);
4202 update_costs_from_copies (a, false, false);
4204 ira_overall_cost -= cost;
4205 ALLOCNO_HARD_REGNO (a) = hard_regno;
4206 if (hard_regno < 0)
4208 ALLOCNO_HARD_REGNO (a) = -1;
4209 cost += ALLOCNO_MEMORY_COST (a);
4211 else if (ira_class_hard_reg_index[aclass][hard_regno] >= 0)
4213 cost += (ALLOCNO_HARD_REG_COSTS (a) == NULL
4214 ? ALLOCNO_CLASS_COST (a)
4215 : ALLOCNO_HARD_REG_COSTS (a)
4216 [ira_class_hard_reg_index[aclass][hard_regno]]);
4217 update_costs_from_copies (a, true, false);
4219 else
4220 /* Reload changed class of the allocno. */
4221 cost = 0;
4222 ira_overall_cost += cost;
4225 /* This function is called when reload deletes memory-memory move. In
4226 this case we marks that the allocation of the corresponding
4227 allocnos should be not changed in future. Otherwise we risk to get
4228 a wrong code. */
4229 void
4230 ira_mark_memory_move_deletion (int dst_regno, int src_regno)
4232 ira_allocno_t dst = ira_regno_allocno_map[dst_regno];
4233 ira_allocno_t src = ira_regno_allocno_map[src_regno];
4235 ira_assert (dst != NULL && src != NULL
4236 && ALLOCNO_HARD_REGNO (dst) < 0
4237 && ALLOCNO_HARD_REGNO (src) < 0);
4238 ALLOCNO_DONT_REASSIGN_P (dst) = true;
4239 ALLOCNO_DONT_REASSIGN_P (src) = true;
4242 /* Try to assign a hard register (except for FORBIDDEN_REGS) to
4243 allocno A and return TRUE in the case of success. */
4244 static bool
4245 allocno_reload_assign (ira_allocno_t a, HARD_REG_SET forbidden_regs)
4247 int hard_regno;
4248 enum reg_class aclass;
4249 int regno = ALLOCNO_REGNO (a);
4250 HARD_REG_SET saved[2];
4251 int i, n;
4253 n = ALLOCNO_NUM_OBJECTS (a);
4254 for (i = 0; i < n; i++)
4256 ira_object_t obj = ALLOCNO_OBJECT (a, i);
4257 COPY_HARD_REG_SET (saved[i], OBJECT_TOTAL_CONFLICT_HARD_REGS (obj));
4258 IOR_HARD_REG_SET (OBJECT_TOTAL_CONFLICT_HARD_REGS (obj), forbidden_regs);
4259 if (! flag_caller_saves && ALLOCNO_CALLS_CROSSED_NUM (a) != 0)
4260 IOR_HARD_REG_SET (OBJECT_TOTAL_CONFLICT_HARD_REGS (obj),
4261 call_used_reg_set);
4263 ALLOCNO_ASSIGNED_P (a) = false;
4264 aclass = ALLOCNO_CLASS (a);
4265 update_curr_costs (a);
4266 assign_hard_reg (a, true);
4267 hard_regno = ALLOCNO_HARD_REGNO (a);
4268 reg_renumber[regno] = hard_regno;
4269 if (hard_regno < 0)
4270 ALLOCNO_HARD_REGNO (a) = -1;
4271 else
4273 ira_assert (ira_class_hard_reg_index[aclass][hard_regno] >= 0);
4274 ira_overall_cost
4275 -= (ALLOCNO_MEMORY_COST (a)
4276 - (ALLOCNO_HARD_REG_COSTS (a) == NULL
4277 ? ALLOCNO_CLASS_COST (a)
4278 : ALLOCNO_HARD_REG_COSTS (a)[ira_class_hard_reg_index
4279 [aclass][hard_regno]]));
4280 if (ALLOCNO_CALLS_CROSSED_NUM (a) != 0
4281 && ira_hard_reg_set_intersection_p (hard_regno, ALLOCNO_MODE (a),
4282 call_used_reg_set))
4284 ira_assert (flag_caller_saves);
4285 caller_save_needed = 1;
4289 /* If we found a hard register, modify the RTL for the pseudo
4290 register to show the hard register, and mark the pseudo register
4291 live. */
4292 if (reg_renumber[regno] >= 0)
4294 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4295 fprintf (ira_dump_file, ": reassign to %d\n", reg_renumber[regno]);
4296 SET_REGNO (regno_reg_rtx[regno], reg_renumber[regno]);
4297 mark_home_live (regno);
4299 else if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4300 fprintf (ira_dump_file, "\n");
4301 for (i = 0; i < n; i++)
4303 ira_object_t obj = ALLOCNO_OBJECT (a, i);
4304 COPY_HARD_REG_SET (OBJECT_TOTAL_CONFLICT_HARD_REGS (obj), saved[i]);
4306 return reg_renumber[regno] >= 0;
4309 /* Sort pseudos according their usage frequencies (putting most
4310 frequently ones first). */
4311 static int
4312 pseudo_reg_compare (const void *v1p, const void *v2p)
4314 int regno1 = *(const int *) v1p;
4315 int regno2 = *(const int *) v2p;
4316 int diff;
4318 if ((diff = REG_FREQ (regno2) - REG_FREQ (regno1)) != 0)
4319 return diff;
4320 return regno1 - regno2;
4323 /* Try to allocate hard registers to SPILLED_PSEUDO_REGS (there are
4324 NUM of them) or spilled pseudos conflicting with pseudos in
4325 SPILLED_PSEUDO_REGS. Return TRUE and update SPILLED, if the
4326 allocation has been changed. The function doesn't use
4327 BAD_SPILL_REGS and hard registers in PSEUDO_FORBIDDEN_REGS and
4328 PSEUDO_PREVIOUS_REGS for the corresponding pseudos. The function
4329 is called by the reload pass at the end of each reload
4330 iteration. */
4331 bool
4332 ira_reassign_pseudos (int *spilled_pseudo_regs, int num,
4333 HARD_REG_SET bad_spill_regs,
4334 HARD_REG_SET *pseudo_forbidden_regs,
4335 HARD_REG_SET *pseudo_previous_regs,
4336 bitmap spilled)
4338 int i, n, regno;
4339 bool changed_p;
4340 ira_allocno_t a;
4341 HARD_REG_SET forbidden_regs;
4342 bitmap temp = BITMAP_ALLOC (NULL);
4344 /* Add pseudos which conflict with pseudos already in
4345 SPILLED_PSEUDO_REGS to SPILLED_PSEUDO_REGS. This is preferable
4346 to allocating in two steps as some of the conflicts might have
4347 a higher priority than the pseudos passed in SPILLED_PSEUDO_REGS. */
4348 for (i = 0; i < num; i++)
4349 bitmap_set_bit (temp, spilled_pseudo_regs[i]);
4351 for (i = 0, n = num; i < n; i++)
4353 int nr, j;
4354 int regno = spilled_pseudo_regs[i];
4355 bitmap_set_bit (temp, regno);
4357 a = ira_regno_allocno_map[regno];
4358 nr = ALLOCNO_NUM_OBJECTS (a);
4359 for (j = 0; j < nr; j++)
4361 ira_object_t conflict_obj;
4362 ira_object_t obj = ALLOCNO_OBJECT (a, j);
4363 ira_object_conflict_iterator oci;
4365 FOR_EACH_OBJECT_CONFLICT (obj, conflict_obj, oci)
4367 ira_allocno_t conflict_a = OBJECT_ALLOCNO (conflict_obj);
4368 if (ALLOCNO_HARD_REGNO (conflict_a) < 0
4369 && ! ALLOCNO_DONT_REASSIGN_P (conflict_a)
4370 && bitmap_set_bit (temp, ALLOCNO_REGNO (conflict_a)))
4372 spilled_pseudo_regs[num++] = ALLOCNO_REGNO (conflict_a);
4373 /* ?!? This seems wrong. */
4374 bitmap_set_bit (consideration_allocno_bitmap,
4375 ALLOCNO_NUM (conflict_a));
4381 if (num > 1)
4382 qsort (spilled_pseudo_regs, num, sizeof (int), pseudo_reg_compare);
4383 changed_p = false;
4384 /* Try to assign hard registers to pseudos from
4385 SPILLED_PSEUDO_REGS. */
4386 for (i = 0; i < num; i++)
4388 regno = spilled_pseudo_regs[i];
4389 COPY_HARD_REG_SET (forbidden_regs, bad_spill_regs);
4390 IOR_HARD_REG_SET (forbidden_regs, pseudo_forbidden_regs[regno]);
4391 IOR_HARD_REG_SET (forbidden_regs, pseudo_previous_regs[regno]);
4392 gcc_assert (reg_renumber[regno] < 0);
4393 a = ira_regno_allocno_map[regno];
4394 ira_mark_allocation_change (regno);
4395 ira_assert (reg_renumber[regno] < 0);
4396 if (internal_flag_ira_verbose > 3 && ira_dump_file != NULL)
4397 fprintf (ira_dump_file,
4398 " Try Assign %d(a%d), cost=%d", regno, ALLOCNO_NUM (a),
4399 ALLOCNO_MEMORY_COST (a)
4400 - ALLOCNO_CLASS_COST (a));
4401 allocno_reload_assign (a, forbidden_regs);
4402 if (reg_renumber[regno] >= 0)
4404 CLEAR_REGNO_REG_SET (spilled, regno);
4405 changed_p = true;
4408 BITMAP_FREE (temp);
4409 return changed_p;
4412 /* The function is called by reload and returns already allocated
4413 stack slot (if any) for REGNO with given INHERENT_SIZE and
4414 TOTAL_SIZE. In the case of failure to find a slot which can be
4415 used for REGNO, the function returns NULL. */
4417 ira_reuse_stack_slot (int regno, unsigned int inherent_size,
4418 unsigned int total_size)
4420 unsigned int i;
4421 int slot_num, best_slot_num;
4422 int cost, best_cost;
4423 ira_copy_t cp, next_cp;
4424 ira_allocno_t another_allocno, allocno = ira_regno_allocno_map[regno];
4425 rtx x;
4426 bitmap_iterator bi;
4427 struct ira_spilled_reg_stack_slot *slot = NULL;
4429 ira_assert (! ira_use_lra_p);
4431 ira_assert (inherent_size == PSEUDO_REGNO_BYTES (regno)
4432 && inherent_size <= total_size
4433 && ALLOCNO_HARD_REGNO (allocno) < 0);
4434 if (! flag_ira_share_spill_slots)
4435 return NULL_RTX;
4436 slot_num = -ALLOCNO_HARD_REGNO (allocno) - 2;
4437 if (slot_num != -1)
4439 slot = &ira_spilled_reg_stack_slots[slot_num];
4440 x = slot->mem;
4442 else
4444 best_cost = best_slot_num = -1;
4445 x = NULL_RTX;
4446 /* It means that the pseudo was spilled in the reload pass, try
4447 to reuse a slot. */
4448 for (slot_num = 0;
4449 slot_num < ira_spilled_reg_stack_slots_num;
4450 slot_num++)
4452 slot = &ira_spilled_reg_stack_slots[slot_num];
4453 if (slot->mem == NULL_RTX)
4454 continue;
4455 if (slot->width < total_size
4456 || GET_MODE_SIZE (GET_MODE (slot->mem)) < inherent_size)
4457 continue;
4459 EXECUTE_IF_SET_IN_BITMAP (&slot->spilled_regs,
4460 FIRST_PSEUDO_REGISTER, i, bi)
4462 another_allocno = ira_regno_allocno_map[i];
4463 if (allocnos_conflict_by_live_ranges_p (allocno,
4464 another_allocno))
4465 goto cont;
4467 for (cost = 0, cp = ALLOCNO_COPIES (allocno);
4468 cp != NULL;
4469 cp = next_cp)
4471 if (cp->first == allocno)
4473 next_cp = cp->next_first_allocno_copy;
4474 another_allocno = cp->second;
4476 else if (cp->second == allocno)
4478 next_cp = cp->next_second_allocno_copy;
4479 another_allocno = cp->first;
4481 else
4482 gcc_unreachable ();
4483 if (cp->insn == NULL_RTX)
4484 continue;
4485 if (bitmap_bit_p (&slot->spilled_regs,
4486 ALLOCNO_REGNO (another_allocno)))
4487 cost += cp->freq;
4489 if (cost > best_cost)
4491 best_cost = cost;
4492 best_slot_num = slot_num;
4494 cont:
4497 if (best_cost >= 0)
4499 slot_num = best_slot_num;
4500 slot = &ira_spilled_reg_stack_slots[slot_num];
4501 SET_REGNO_REG_SET (&slot->spilled_regs, regno);
4502 x = slot->mem;
4503 ALLOCNO_HARD_REGNO (allocno) = -slot_num - 2;
4506 if (x != NULL_RTX)
4508 ira_assert (slot->width >= total_size);
4509 #ifdef ENABLE_IRA_CHECKING
4510 EXECUTE_IF_SET_IN_BITMAP (&slot->spilled_regs,
4511 FIRST_PSEUDO_REGISTER, i, bi)
4513 ira_assert (! conflict_by_live_ranges_p (regno, i));
4515 #endif
4516 SET_REGNO_REG_SET (&slot->spilled_regs, regno);
4517 if (internal_flag_ira_verbose > 3 && ira_dump_file)
4519 fprintf (ira_dump_file, " Assigning %d(freq=%d) slot %d of",
4520 regno, REG_FREQ (regno), slot_num);
4521 EXECUTE_IF_SET_IN_BITMAP (&slot->spilled_regs,
4522 FIRST_PSEUDO_REGISTER, i, bi)
4524 if ((unsigned) regno != i)
4525 fprintf (ira_dump_file, " %d", i);
4527 fprintf (ira_dump_file, "\n");
4530 return x;
4533 /* This is called by reload every time a new stack slot X with
4534 TOTAL_SIZE was allocated for REGNO. We store this info for
4535 subsequent ira_reuse_stack_slot calls. */
4536 void
4537 ira_mark_new_stack_slot (rtx x, int regno, unsigned int total_size)
4539 struct ira_spilled_reg_stack_slot *slot;
4540 int slot_num;
4541 ira_allocno_t allocno;
4543 ira_assert (! ira_use_lra_p);
4545 ira_assert (PSEUDO_REGNO_BYTES (regno) <= total_size);
4546 allocno = ira_regno_allocno_map[regno];
4547 slot_num = -ALLOCNO_HARD_REGNO (allocno) - 2;
4548 if (slot_num == -1)
4550 slot_num = ira_spilled_reg_stack_slots_num++;
4551 ALLOCNO_HARD_REGNO (allocno) = -slot_num - 2;
4553 slot = &ira_spilled_reg_stack_slots[slot_num];
4554 INIT_REG_SET (&slot->spilled_regs);
4555 SET_REGNO_REG_SET (&slot->spilled_regs, regno);
4556 slot->mem = x;
4557 slot->width = total_size;
4558 if (internal_flag_ira_verbose > 3 && ira_dump_file)
4559 fprintf (ira_dump_file, " Assigning %d(freq=%d) a new slot %d\n",
4560 regno, REG_FREQ (regno), slot_num);
4564 /* Return spill cost for pseudo-registers whose numbers are in array
4565 REGNOS (with a negative number as an end marker) for reload with
4566 given IN and OUT for INSN. Return also number points (through
4567 EXCESS_PRESSURE_LIVE_LENGTH) where the pseudo-register lives and
4568 the register pressure is high, number of references of the
4569 pseudo-registers (through NREFS), number of callee-clobbered
4570 hard-registers occupied by the pseudo-registers (through
4571 CALL_USED_COUNT), and the first hard regno occupied by the
4572 pseudo-registers (through FIRST_HARD_REGNO). */
4573 static int
4574 calculate_spill_cost (int *regnos, rtx in, rtx out, rtx insn,
4575 int *excess_pressure_live_length,
4576 int *nrefs, int *call_used_count, int *first_hard_regno)
4578 int i, cost, regno, hard_regno, j, count, saved_cost, nregs;
4579 bool in_p, out_p;
4580 int length;
4581 ira_allocno_t a;
4583 *nrefs = 0;
4584 for (length = count = cost = i = 0;; i++)
4586 regno = regnos[i];
4587 if (regno < 0)
4588 break;
4589 *nrefs += REG_N_REFS (regno);
4590 hard_regno = reg_renumber[regno];
4591 ira_assert (hard_regno >= 0);
4592 a = ira_regno_allocno_map[regno];
4593 length += ALLOCNO_EXCESS_PRESSURE_POINTS_NUM (a) / ALLOCNO_NUM_OBJECTS (a);
4594 cost += ALLOCNO_MEMORY_COST (a) - ALLOCNO_CLASS_COST (a);
4595 nregs = hard_regno_nregs[hard_regno][ALLOCNO_MODE (a)];
4596 for (j = 0; j < nregs; j++)
4597 if (! TEST_HARD_REG_BIT (call_used_reg_set, hard_regno + j))
4598 break;
4599 if (j == nregs)
4600 count++;
4601 in_p = in && REG_P (in) && (int) REGNO (in) == hard_regno;
4602 out_p = out && REG_P (out) && (int) REGNO (out) == hard_regno;
4603 if ((in_p || out_p)
4604 && find_regno_note (insn, REG_DEAD, hard_regno) != NULL_RTX)
4606 saved_cost = 0;
4607 if (in_p)
4608 saved_cost += ira_memory_move_cost
4609 [ALLOCNO_MODE (a)][ALLOCNO_CLASS (a)][1];
4610 if (out_p)
4611 saved_cost
4612 += ira_memory_move_cost
4613 [ALLOCNO_MODE (a)][ALLOCNO_CLASS (a)][0];
4614 cost -= REG_FREQ_FROM_BB (BLOCK_FOR_INSN (insn)) * saved_cost;
4617 *excess_pressure_live_length = length;
4618 *call_used_count = count;
4619 hard_regno = -1;
4620 if (regnos[0] >= 0)
4622 hard_regno = reg_renumber[regnos[0]];
4624 *first_hard_regno = hard_regno;
4625 return cost;
4628 /* Return TRUE if spilling pseudo-registers whose numbers are in array
4629 REGNOS is better than spilling pseudo-registers with numbers in
4630 OTHER_REGNOS for reload with given IN and OUT for INSN. The
4631 function used by the reload pass to make better register spilling
4632 decisions. */
4633 bool
4634 ira_better_spill_reload_regno_p (int *regnos, int *other_regnos,
4635 rtx in, rtx out, rtx insn)
4637 int cost, other_cost;
4638 int length, other_length;
4639 int nrefs, other_nrefs;
4640 int call_used_count, other_call_used_count;
4641 int hard_regno, other_hard_regno;
4643 cost = calculate_spill_cost (regnos, in, out, insn,
4644 &length, &nrefs, &call_used_count, &hard_regno);
4645 other_cost = calculate_spill_cost (other_regnos, in, out, insn,
4646 &other_length, &other_nrefs,
4647 &other_call_used_count,
4648 &other_hard_regno);
4649 if (nrefs == 0 && other_nrefs != 0)
4650 return true;
4651 if (nrefs != 0 && other_nrefs == 0)
4652 return false;
4653 if (cost != other_cost)
4654 return cost < other_cost;
4655 if (length != other_length)
4656 return length > other_length;
4657 #ifdef REG_ALLOC_ORDER
4658 if (hard_regno >= 0 && other_hard_regno >= 0)
4659 return (inv_reg_alloc_order[hard_regno]
4660 < inv_reg_alloc_order[other_hard_regno]);
4661 #else
4662 if (call_used_count != other_call_used_count)
4663 return call_used_count > other_call_used_count;
4664 #endif
4665 return false;
4670 /* Allocate and initialize data necessary for assign_hard_reg. */
4671 void
4672 ira_initiate_assign (void)
4674 sorted_allocnos
4675 = (ira_allocno_t *) ira_allocate (sizeof (ira_allocno_t)
4676 * ira_allocnos_num);
4677 consideration_allocno_bitmap = ira_allocate_bitmap ();
4678 initiate_cost_update ();
4679 allocno_priorities = (int *) ira_allocate (sizeof (int) * ira_allocnos_num);
4680 sorted_copies = (ira_copy_t *) ira_allocate (ira_copies_num
4681 * sizeof (ira_copy_t));
4684 /* Deallocate data used by assign_hard_reg. */
4685 void
4686 ira_finish_assign (void)
4688 ira_free (sorted_allocnos);
4689 ira_free_bitmap (consideration_allocno_bitmap);
4690 finish_cost_update ();
4691 ira_free (allocno_priorities);
4692 ira_free (sorted_copies);
4697 /* Entry function doing color-based register allocation. */
4698 static void
4699 color (void)
4701 allocno_stack_vec.create (ira_allocnos_num);
4702 memset (allocated_hardreg_p, 0, sizeof (allocated_hardreg_p));
4703 ira_initiate_assign ();
4704 do_coloring ();
4705 ira_finish_assign ();
4706 allocno_stack_vec.release ();
4707 move_spill_restore ();
4712 /* This page contains a simple register allocator without usage of
4713 allocno conflicts. This is used for fast allocation for -O0. */
4715 /* Do register allocation by not using allocno conflicts. It uses
4716 only allocno live ranges. The algorithm is close to Chow's
4717 priority coloring. */
4718 static void
4719 fast_allocation (void)
4721 int i, j, k, num, class_size, hard_regno;
4722 #ifdef STACK_REGS
4723 bool no_stack_reg_p;
4724 #endif
4725 enum reg_class aclass;
4726 enum machine_mode mode;
4727 ira_allocno_t a;
4728 ira_allocno_iterator ai;
4729 live_range_t r;
4730 HARD_REG_SET conflict_hard_regs, *used_hard_regs;
4732 sorted_allocnos = (ira_allocno_t *) ira_allocate (sizeof (ira_allocno_t)
4733 * ira_allocnos_num);
4734 num = 0;
4735 FOR_EACH_ALLOCNO (a, ai)
4736 sorted_allocnos[num++] = a;
4737 allocno_priorities = (int *) ira_allocate (sizeof (int) * ira_allocnos_num);
4738 setup_allocno_priorities (sorted_allocnos, num);
4739 used_hard_regs = (HARD_REG_SET *) ira_allocate (sizeof (HARD_REG_SET)
4740 * ira_max_point);
4741 for (i = 0; i < ira_max_point; i++)
4742 CLEAR_HARD_REG_SET (used_hard_regs[i]);
4743 qsort (sorted_allocnos, num, sizeof (ira_allocno_t),
4744 allocno_priority_compare_func);
4745 for (i = 0; i < num; i++)
4747 int nr, l;
4749 a = sorted_allocnos[i];
4750 nr = ALLOCNO_NUM_OBJECTS (a);
4751 CLEAR_HARD_REG_SET (conflict_hard_regs);
4752 for (l = 0; l < nr; l++)
4754 ira_object_t obj = ALLOCNO_OBJECT (a, l);
4755 IOR_HARD_REG_SET (conflict_hard_regs,
4756 OBJECT_CONFLICT_HARD_REGS (obj));
4757 for (r = OBJECT_LIVE_RANGES (obj); r != NULL; r = r->next)
4758 for (j = r->start; j <= r->finish; j++)
4759 IOR_HARD_REG_SET (conflict_hard_regs, used_hard_regs[j]);
4761 aclass = ALLOCNO_CLASS (a);
4762 ALLOCNO_ASSIGNED_P (a) = true;
4763 ALLOCNO_HARD_REGNO (a) = -1;
4764 if (hard_reg_set_subset_p (reg_class_contents[aclass],
4765 conflict_hard_regs))
4766 continue;
4767 mode = ALLOCNO_MODE (a);
4768 #ifdef STACK_REGS
4769 no_stack_reg_p = ALLOCNO_NO_STACK_REG_P (a);
4770 #endif
4771 class_size = ira_class_hard_regs_num[aclass];
4772 for (j = 0; j < class_size; j++)
4774 hard_regno = ira_class_hard_regs[aclass][j];
4775 #ifdef STACK_REGS
4776 if (no_stack_reg_p && FIRST_STACK_REG <= hard_regno
4777 && hard_regno <= LAST_STACK_REG)
4778 continue;
4779 #endif
4780 if (ira_hard_reg_set_intersection_p (hard_regno, mode, conflict_hard_regs)
4781 || (TEST_HARD_REG_BIT
4782 (ira_prohibited_class_mode_regs[aclass][mode], hard_regno)))
4783 continue;
4784 ALLOCNO_HARD_REGNO (a) = hard_regno;
4785 for (l = 0; l < nr; l++)
4787 ira_object_t obj = ALLOCNO_OBJECT (a, l);
4788 for (r = OBJECT_LIVE_RANGES (obj); r != NULL; r = r->next)
4789 for (k = r->start; k <= r->finish; k++)
4790 IOR_HARD_REG_SET (used_hard_regs[k],
4791 ira_reg_mode_hard_regset[hard_regno][mode]);
4793 break;
4796 ira_free (sorted_allocnos);
4797 ira_free (used_hard_regs);
4798 ira_free (allocno_priorities);
4799 if (internal_flag_ira_verbose > 1 && ira_dump_file != NULL)
4800 ira_print_disposition (ira_dump_file);
4805 /* Entry function doing coloring. */
4806 void
4807 ira_color (void)
4809 ira_allocno_t a;
4810 ira_allocno_iterator ai;
4812 /* Setup updated costs. */
4813 FOR_EACH_ALLOCNO (a, ai)
4815 ALLOCNO_UPDATED_MEMORY_COST (a) = ALLOCNO_MEMORY_COST (a);
4816 ALLOCNO_UPDATED_CLASS_COST (a) = ALLOCNO_CLASS_COST (a);
4818 if (ira_conflicts_p)
4819 color ();
4820 else
4821 fast_allocation ();